2.2 Placement Stability for Children in Out-of-Home Care

advertisement
DIFFERENTIAL EFFECTS OF SUBSIDIZED
GUARDIANSHIP ON PLACEMENT STABILITY
FOR CHILDREN IN KINSHIP CARE
by
Ge Hong
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland Baltimore County in partial fulfillment of the
requirements for the degree of Doctor of Philosophy
2006
ABSTRACT
Title of Dissertation: Differential Effects of Subsidized Guardianship on
Placement Stability for Children in Kinship Care
Ge Hong, Doctor of Philosophy, 2006
Dissertation directed by: Dr. Marvin Mandell, Professor, Department of
Public Policy, UMBC
Dr. Dave Marcotte, Associate Professor,
Department of Public Policy, UMBC
Dr. Donna Harrington, Professor, School of Social
Work, University of Maryland, Baltimore
Dr. Claudia Lawrence-Webb
Dr. Willie Tompkins, Washington D.C. Child and
Family Services Agency
For over three decades, placement stability of children in foster care has been a
serious concern for policy makers as well as the general public. Frequent placement
disruption has been found to have negative impacts on children’s psychological
development and educational progress. Yet, a substantial proportion of children in foster
care have experienced three or more placements while they are in foster care. The
objective of this dissertation is to examine the effects of guardianship subsidy availability
on placement stability for a cohort of children in the Maryland’s Guardianship Assistance
Demonstration Project (GAP) who were placed with relative caregivers before May 2000
in Baltimore City. The evaluation of GAP was a true experiment in which study
participants were randomly assigned to either the experimental group or the control
group. Members of the experimental group were eligible for a monthly subsidy for
guardianship participation. Members of the control group, on the other hand, were
ineligible for the subsidy.
This study examines the differential effects of the eligibility of the subsidy on
placement stability between four distinctive groups of children who were placed in
kinship care at the beginning of the project and in the absence of the subsidy:
1. Those who would remain in kinship care in the absence of the subsidy;
2. Those who would move to restricted foster care in the absence of the subsidy;
3. Those who would exit out-of-home care through guardianship in the absence of
the subsidy; and
4. Those who would exit out-of-home care through reunification.
Subgroups are defined using the propensity score matching method and number
of placement changes is examined using a Poisson regression model. The study did not
find any subsidy effects on placement stability regardless of children’s permanency
outcome in the absence of the subsidy. Given that earlier evaluation found the subsidies
to accelerate the pace of children leaving out-of-home care, the implementation of a
guardianship subsidy program should be implemented as a universal policy.
Additionally, this dissertation demonstrates that propensity score matching has the
potential to improve sample balance. The estimates of program effects obtained with the
propensity score methods, however, is sensitive to the selection of predictors.
Table of Contents
Chapter 1. Introduction ................................................................................................... 1
1.1
The Foster Care Drift: Length of Stay and Placement Stability ......................... 2
1.2
Out-of-home Placements with Relatives ............................................................ 3
1.3
Permanency and Placement Stability for Children in Formal Kinship Care ...... 7
1.4
Legal Guardianship ........................................................................................... 11
1.5
Maryland Guardianship Assistance Project ...................................................... 12
Chapter 2. Literature Review ........................................................................................ 17
2.1 Relative Foster Care Vs. Non-relative Foster Care ................................................ 17
2.1.1 Child and Caregiver Characteristics ................................................................ 17
2.1.2 Services Received by Caregivers and Children in Relative Foster Care ......... 21
2.1.3 Child Outcomes ............................................................................................... 22
2.2 Placement Stability for Children in Out-of-Home Care ......................................... 26
2.2.1 Placement Disruption Rates ............................................................................. 26
2.2.2 Predictors of Placement Disruption ................................................................. 28
2.2.3 Consequences of Placement Instability ........................................................... 31
2.3 Outcomes for Children in Legal Guardianship ....................................................... 32
2.4 IV-E Waiver Demonstrations on Subsidized Guardianship ................................... 33
2.5 Propensity Score Matching ..................................................................................... 35
Chapter 3. Methodology ................................................................................................. 38
3.1 Research Question .................................................................................................. 38
3.2 Sample and Data ..................................................................................................... 41
3.3 Measurement of Placement Stability ...................................................................... 43
3.4 Statistical Analysis .................................................................................................. 43
3.4.1 Propensity Score Matching .............................................................................. 44
3.4.2 Poisson Regression .......................................................................................... 48
3.4.3 Sensitivity Analysis ......................................................................................... 50
Chapter 4. Results ........................................................................................................... 52
4.1 Sample Characteristics ............................................................................................ 52
4.2 Bivariate Analysis ................................................................................................... 54
I
4.3 Subgroup Analysis: Children who Would Exit and Children who Would Stay ..... 56
4.3.1 Predicting the Probabilities of Exiting Foster Care in the Absence of the
Subsidy...................................................................................................................... 57
4.3.2 Sample Balance ................................................................................................ 59
4.3.3 Placement Stability .......................................................................................... 64
4.4 Subgroup Analysis: Guardianship, Reunification, Restricted Foster Care, or
Kinship Care ................................................................................................................. 67
4.4.1 Predicting Child Outcomes using a Multinomial Logit Model ....................... 68
4.4.2 Propensity Score Matching .............................................................................. 73
4.4.3 Placement Stability .......................................................................................... 80
4.4.4 Estimates of the Effects of the Eligibility of the Guardianship Subsidy Based
on Multinomial Logit Model 1 ................................................................................. 86
4.5 Sensitivity Analysis – African-American Children Only ....................................... 87
Chapter 5. Discussions and Policy Implications .......................................................... 90
References ........................................................................................................................ 94
Appendix: Additional Statistical Tables ..................................................................... 103
List of Figures
Figure 1- 1 Number of Children in Foster Care (1980 – 2002) .......................................... 1
Figure 1- 2 Types of Kinship Care in Maryland ................................................................. 7
List of Tables
Table 3. 1 Predictors of child permanency and placement outcomes ............................... 47
Table 4. 1 Sample Characteristics ..................................................................................... 52
Table 4. 2 Number of placement changes ......................................................................... 55
Table 4. 3 Permanency Outcomes .................................................................................... 56
Table 4. 4 Coefficients obtained from the logit models predicting the probabilities of
exiting from foster care for the members of the control group ....................... 57
Table 4. 5 Sample Balance without Propensity Score Matching ...................................... 59
Table 4. 6 Sample Balance after Propensity Score Matching based on Logit Model 1 ... 61
Table 4. 7 Sample Balance after Propensity Score Matching Based on Logit Model 2 .. 63
II
Table 4. 8 Exposure Adjusted Mean Number of Placement Change by Permanency Status
......................................................................................................................... 65
Table 4. 9 Coefficients obtained from Poisson and negative binomial models on exposure
adjusted numbers of placement changes for children who would exit from outof-home placements in the absence of the subsidy ....................................... 66
Table 4. 10 Coefficients obtained from Poisson and negative binomial models on
exposure adjusted numbers of placement changes for children who would not
exit from out-of-home placement in the absence of the subsidy ................... 67
Table 4. 11 Coefficients from the multinomial logit models on permanency outcomes 69
Table 4. 12 Sample balance without propensity score matching ...................................... 75
Table 4. 13 Sample Balance after Propensity Score Matching (Based on Multinomial
Logit Model 2) ................................................................................................ 77
Table 4. 14 Exposure-adjusted mean change in placements per year by predicted
outcomes and eligibility for the guardianship subsidy ................................... 81
Table 4. 15 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
exit through guardianship in the absence of the subsidy ............................... 82
Table 4. 16 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placements for children who would exit
through reunification in the absence of the subsidy ...................................... 83
Table 4. 17 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placements for children who would move to
Restricted Foster Care in the absence of the subsidy .................................... 84
Table 4. 18 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placements for children who would remain
in kinship care in the absence of the subsidy ................................................ 85
Tables in Appendix
Table A- 1 Sample Balance after Propensity Score Matching (Not considering
participating in caregiver interview) .......................................................... 104
Table A- 2 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
exit through guardianship in the absence of the subsidy (Not considering
participation in caregiver interview) ........................................................ 106
Table A- 3 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
exit through reunification in the absence of the subsidy (Not considering
participation in caregiver interview) ........................................................ 106
III
Table A- 4 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
move to restricted foster care in the absence of the subsidy (Not considering
participation in caregiver interview) ........................................................ 107
Table A- 5 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
remain in kinship care in the absence of the subsidy (Not considering
participating in caregiver interview) ........................................................ 108
Table A- 6 Coefficients obtained from a multinomial logit model on permanency status
(African-American children only)  ........................................................... 108
Table A- 7 Sample Balance After Propensity Score Matching (African-American children
only, not considering participation in caregiver interviews) ..................... 110
Table A- 8 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placements for children who would exit
through guardianship in the absence of the subsidy – African-American
children only ............................................................................................. 112
Table A- 9 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
exit through reunification in the absence of the subsidy (Not considering
participation in caregiver interview) - African-American children only . 112
Table A- 10 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placements for children who would move
to restricted foster care in the absence of the subsidy – African-American
children only ............................................................................................. 113
Table A- 11 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
remain in kinship care in the absence of the subsidy (Not considering
participation in caregiver interview) - African-American children only .. 114
Table A- 12 Sample Balance after Propensity Score Matching (African-American
children only, considering participation in caregiver interviews) ............. 115
Table A- 13 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
exit through guardianship in the absence of the subsidy (Considering
participation in caregiver interview) - African-American children only  . 117
Table A- 14 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
exit through reunification in the absence of the subsidy (Considering
participation in caregiver interview) - African-American children only  . 118
Table A- 15 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
move to RFC in the absence of the subsidy (Considering participation in
caregiver interview) - African-American children only  .......................... 119
IV
Table A- 16 Coefficients obtained from Poisson and negative binomial regression models
on exposure adjusted numbers of placement changes for children who would
remain in kinship care in the absence of the subsidy (Considering
participation in caregiver interview) - African-American children only  . 120
V
Chapter 1. Introduction
More than four decades have passed since Maas and Engler (1959) drew the nation’s
attention to the ordeals of children in foster care. Although child welfare in the U.S.
underwent great changes during these four decades, many believe the foster care system
has yet to walk out of the quagmire. These beliefs are not unfounded. Figure 1 illustrates
the trend of foster care population since 1980. The number of children in out-of-home
care increased from 302,000 in 1980 to 581,000 in 1999. The size of the foster care
population has been on a slight downward trend since the turn of the 21st century.
Nevertheless, some 532,000 children were in foster care as of September 2002.
Figure 1- 1 Number of Children in Foster Care (1980 – 2002)
700
600
500
400
in thousands
300
200
100
0
80
83
86
89
92
95
98
01
Source: 1980-1997 Ladner (2000); 1998 – 2002 from Adoption and Foster Care Analysis and Reporting System (AFCARS).
Available at http://www.acf.dhhs.gov
While the number of children placed in out of home care has almost doubled from
its 1980 level, the number of non-relative foster parents declined over the 1980s,
dropping from 134,000 in 1984 to 125,000 in 1989 (U.S. Department of Health and
Human Services [USDHHS], 1994). The shortage was most felt in urban areas and for
foster parents who were minorities and who were willing to accept teenagers. This was
1
when relative foster homes became the place where a considerable proportion of children
in out-of-home care lived. This chapter will discuss the cultural and policy contexts of
relative foster care, and the motivation for this study.
1.1 The Foster Care Drift: Length of Stay and Placement Stability
The concerns for children’s experience in foster care center on “foster care drift” – a
phrase used to describe children’s experience of staying in foster care for lengthy periods
of time without permanency plans, while drifting from one placement to another.
Although permanency planning has become a standard practice, permanency remains an
elusive goal for many children in foster care. The dramatic growth of the foster care
population in a large part is due to the fact that a large proportion of children stay in care
for prolonged periods of time (Ladner, 2000).
When children cannot exit from care in a timely manner, their risks for placement
disruptions increase (Olsen, 1982; Proch & Taber, 1985). A substantial proportion of
children in foster care today have experienced three or more placements while they are in
foster care (Chamberlain & Whaley, 2001; Newton, Litrownik & Landsverk, 2000; Smith,
Stormshak, USDHHS, 2003a). This has serious consequences for children’s healthy
development. Research suggests that children who do not receive continuous and
responsive care from seven months of age through toddlerhood are more likely to
develop difficulties in forming attachments to their primary caregivers and healthy
relationships in the future than are children who receive such care (McRoy, 1994). In fact,
placement disruption may be more detrimental to the child’s development than out-ofhome care itself. Frequent changes in placement for children in foster care have been
found to result in increased behavioral problems (Newton et al., 2000). Additionally,
2
multiple placements were shown to disrupt educational progress and lead to increased
criminality, while having been placed in out-of-home care alone was not (Eckenrode,
Rowe, Laird & Braithwaite, 1995; Widom, 1991)
In response to the concerns about the experience of children in foster care, the
congress passed the Adoption Assistance and Child Welfare Act (P.L. 96-272) in 1980 in
the hope of limiting the time it takes for children to be reunified with their parents and to
accelerate adoption efforts when children cannot be returned home. It also encouraged
states to place children in family-like settings. P.L. 96-272, however, fell short of
achieving its goal as the number of children in foster care continued to increase. The
Adoption and Safe Families Act (P.L. 105-89), hereafter referred to as ASFA, of 1997
placed a new emphasis on adoption, as it further shortened the timeframe within which
reunification efforts must be made to 12 months. Although ASFA exempted kinship
placements from the timeframe of reasonable efforts for reunification, and recognized
relative placements could constitute “planned permanent living arrangement” (USDHHS,
2000), states appeared to have increased their efforts to encourage relative caregivers to
pursue adoption and legal guardianship in the wake of ASFA (Chipungu, 2004).
1.2 Out-of-home Placements with Relatives
The rise of kinship care in the child welfare system is closely related to the overrepresentation of African-American children in foster care. The reason for this overrepresentation is manifold. Factors associated with foster care placement such as chronic
poverty, homelessness, joblessness, and teen pregnancy disproportionately affect the
African-American population (Hampton, 1991; Hofferth & Owens, 2001; Leashore,
McMurray & Bailey, 1991). This is reflected by the fact that the majority of African-
3
American children in foster care are placed in out-of-home care because of neglect,
which is closely related to poverty (Smith, 2004; Williams, 1991). The situation is
exacerbated by the drug epidemic, which has led to a drastic increase in the number of
children needing out-of-home placement services (Danzy & Jackson, 1997). For example,
it was estimated that between one third and two thirds of children known to the child
welfare agencies had families beset with substance abuse problems (Courtney, 1999).
Some local programs reported as high as 81% of children in out-of-home placements
with drug addicted mothers (Bonecutter & Gleeson, 1997). Furthermore, differential
treatment by medical professionals and the juvenile justice system as well as the child
welfare system contribute to the rapid increase of African-American children in out-ofhome care (Hill, 2003). For example, some research findings suggested that given the
same situation, African-American children were consistently more likely than white
children to be placed in out-of-home care while white children were more likely to
receive in-home services. Additionally, doctors were more likely to diagnose physical
injury as accidents for affluent families, and abuse for poor families, who were
predominantly African Americans (Hill, 2003).
No matter what the dominant cause for the over-representation of AfricanAmerican children in out-of-home care is, it is an undeniable fact that the influx of
African-American children has greatly changed the landscape of the child welfare system.
However, the current system is not equipped to meet the unique needs of AfricanAmerican children. The current child welfare system was developed based on white
middle-class values such as individualism and the work ethic (Chipungu, 1991). Even in
the 20th century, African-American children were denied child welfare services in the
4
modern child welfare state for a prolonged period of time (Chipungu, 1991). It should not
come as a surprise that the child welfare system does not take into consideration the
unique culture and beliefs of African-American families and the needs of AfricanAmerican children. Many researchers have been calling for an Africentric perspective in
planning and working towards permanency for African-American children (Bonecutter &
Gleeson, 1997; Chipungu, 1991). This perspective would take advantage of the AfricanAmerican value that individuals within extended families view the extended family as
one collective entity. Members of the extended family share resources and care for each
other’s children (Chipungu, 1991).
Kinship care is a natural outcome of the collective identity of African-American
families and the need for out-of-home care for African-American children. Its roots can
be traced back to slavery, when the extended families would take care of a child when the
parents were sold to other slave owners (Danzy & Jackson, 1997). Today, extended
families continue to be an important resource for child rearing when African-American
parents are not able to care for their children, especially after the passage of P.L. 96-272,
which encouraged state child welfare agencies to place children in the least restrictive
placement, which, in many cases, would be with kin.
In the U. S., approximately 2.13 million children were living with relatives
without the presence of a parent in 1998 (USDHHS, 2000). Depending on the degree of
the involvement of the child welfare system, a child may be placed in informal kinship
care or formal kinship care. Children who live with relatives without the involvement of
child welfare agencies are said to be in informal kinship care. The vast majority of
children living with relatives fall into this category. It was estimated that only 200,000
5
children in 1997 were formally placed in relative care by child welfare agencies
(USDHHS, 2000). Children in this type of care are said to be in formal kinship care. In
the past, although the custodies of children in formal kinship care belonged to the state,
many states did not offer foster care board payments to kinship caregivers. Instead,
relative caregivers might receive an AFDC child-only grant. However, in most cases, the
AFDC payment was considerably less than the foster care board payment. Further, AFDC
paid less for each additional child in care while foster care had a flat monthly rate for
each child. The discrepancy in payment rates triggered a class-action lawsuit in Illinois
(Miller vs. Youakim, 1979), for which the Supreme Court ruled that the states must pay
relative caregivers at the same rate as traditional foster parents if the relative caregivers
meet the IV-E requirements.1 Currently a continuum of kinship care policies exists
among the 50 states and the District of Columbia (USDHHS, 2000). All states offer kin
the option of being licensed and receiving the full foster care board payment.2 Many
states waive certain licensing requirements for kin. Relative caregivers who are not
licensed may receive payments at a reduced level, often at the TANF rates.
Figure 2 illustrates the system of relative care in Maryland. There are two types of
relative caregivers in formal kinship care. The first group, referred hereafter as kinship
caregivers, includes relative foster parents who are not licensed. They receive TANF
child-only grants, which averaged $188 per month per child at the time of this study. The
second group, referred to as restricted foster care (RFC) caregivers, are relative foster
parents who are licensed to care for specific children. They receive the foster care board
1
Title IV-E of the Social Security Act provides federal matching funds for foster care payments and
adoption subsidies.
2
California, Oregon, and New Jersey only allow IV-E eligible relative caregivers to be fully licensed.
6
payment, which averaged approximately $600 per month among the participants of this
study.
Figure 1- 2 Types of Kinship Care in Maryland
Informal Kinship Care
No child welfare
involvement
Unlicensed Formal
Kinship Care
Kinship Care
Receiving TANF payments
Formal Kinship Care
With child welfare
involvement
Restricted Foster Care
Receiving Foster Care
payments
1.3 Permanency and Placement Stability for Children in Formal Kinship Care
There is evidence indicating that placements with relative caregivers are associated with
longer stays in care due to lower adoption and reunification rates, the two major channels
of children exiting from foster care (Gleeson, 1999a). African-American children placed
in formal kinship care have the lowest rates of adoption and reunification compared with
other groups of children in out-of-home placements (Barth, Courtney, Berrick, & Alert,
1994),
The unique cultural heritage of African-American families plays an important role
in African-American relative caregivers’ reluctance to pursue adoption. AfricanAmerican families’ sense of collective identity determines that relative care is viewed by
caregivers more as a family responsibility than a legal arrangement. It also determines
that the caregivers’ relationships with the birth parents will play an important role
7
throughout the permanency planning process. Many relative caregivers had been taking
care of the children prior to the involvement of the child welfare agencies. They tend to
feel that adoption is unnecessary because the child is already family (Barth et al., 1994;
Gleeson, O’Donnell, & Bonecutter, 1997; Testa, Shook, Cohen, & Woods, 1996). Some
studies revealed that kinship caregivers, especially grandmothers, were not willing to
adopt children in their care because they did not wish to terminate the rights of the
biological parents (Meyer & Link, 1990; Testa et al., 1996; Thornton, 1991). Some
relative caregivers viewed adoption as a punishment to the birth parents (Geen, 2003).
Many hoped that the birth parents will eventually regain custody of the children (Geen,
2003; Gleeson et al., 1997; Testa et al., 1996). Some caregivers felt that they had a tacit
agreement with the biological parents that they would care for the children until majority
age and the parents would, in some way, remain involved in the children’s lives.
Williams (1991) described a case where a mother placed her child with a friend
voluntarily, believing that her child would be taken good care of in their custody. While
the mother and the kinship caregivers viewed the arrangement as in the child’s best
interest, the child welfare agency considered it abandonment. The kinship caregivers
were reluctant to pursue adoption in fear of undermining their relationship with the
biological mother.
Financial difficulties were also cited in research as one reason why kinship
caregivers did not pursue adoption, although subsidies were available for special needs
adoptions (Barth et al., 1994; Testa et al., 1996). This may be due to the fact that a sizable
proportion of caregivers were never informed of the existence of subsidized adoption
(Geen, 2003; Testa et al., 2001). The fact that special needs adoptions, which qualify for
8
the adoption subsidies, outnumbered regular foster care adoptions between 1983 and
1987 after the passage of P.L. 96-272 suggests that financial assistance attracted at least
some caregivers to pursue adoption (Schwartz & Fishman, 1999).
More recent research in Illinois indicated that a considerable proportion of kin
were in fact willing to adopt (Gleeson et al., 1997). The authors attributed this change to
the pressure on achieving permanency created by litigations in the state, which motivated
caseworkers to discuss adoption more fully with the caregivers. Unfortunately, this has
not been the case in other states. Caseworkers, seeking middle-class two-parent adoptive
families, often deem relative caregivers, predominantly elderly African-American
females with limited income, unsuitable to be adoptive parents (McRoy, 2004).
Apart from lower adoptions rates, slower reunification with birth parents is a
contributing factor to the longer stay in care for children placed with relatives. The
financial difference in foster care board payments and the TANF payments, with the
former considerably higher than the latter, has been identified as an obstacle to
reunification for children cared for by relatives (Takas, 1992). When a child is returned
home, the total subsidies that the extended family receives often fall from the higher
foster care board payment to the lower AFDC/TANF rates. This is evidenced by research
findings in California, which showed that reunification rates were lower for IV-E eligible
children than children who were not IV-E eligible (Berrick & Needell, 1999). In
California, licensed relative caregivers of the former group received the foster care board
payment while those of the latter group were not allowed to become foster parents and
could only receive the AFDC payments if they were eligible. As such, for the former
group, if the child in relative care was returned home, the extended family as a whole
9
would see their total subsidies drop from the foster care rate to the AFDC rate if the
parents qualified for AFDC, and nothing if they did not. The latter group, already
receiving AFDC, would not face this disincentive if the parents qualified for AFDC.
Even if their parents did not qualify for AFDC, the reduction in assistance would be
shallower compared with that for IV-E eligible children. Caseworkers indicated that IV-E
eligible families might avoid reunification to maximize their combined benefits.
Financial disincentive is not the only factor in delaying reunification. Ironically,
the advantages of relative foster care, such as stability and safety, may have stalled
permanency. Research indicated that birth parents were often less motivated to work
towards reunification than their counterparts whose children were placed with non-kin
placements (Geen, 2003). Birth parents felt less propelled to meet case plan requirements,
partly because relative care does not carry as much a stigma as having their children
raised by foster care parents. Caseworkers as well as the parents may deem the
placements to be safe and stable, and therefore, lack the urge and commitment to make
efforts towards reunification (Courtney, 1999).
The close relationship between placement stability and length of stay in care
cannot be stressed enough. Length of stay in care is the single most important predictor of
multiple placements (Proch & Taber, 1985). The number of placements that a child
experiences increases as s/he spends more time in foster care. Multiple placements, in
turn, reduce a child’s probability of being adopted or returned home (Usher, Randolph &
Gogan, 1999). Given the long length of stay in care for children in formal kinship care,
placement stability appears to be of paramount importance for the well being of these
children. Relative placements are found to be more stable than non-relative foster care
10
placements (Courtney & Needell, 1997; Leslie et al., 2000; Scannapieco, 1999; Usher et
al., 1999; Webster, Barth & Needell, 2000). It is nonetheless not uncommon for children
placed with relatives to experience more than one placement during their stay in care
(Terling-Watt, 2001). The data indicated that as high as nearly 30% of children in relative
care experienced more than one placement during the first six months in care. Relative
caregivers, although emotionally committed to the children in their care, receive little
training and meager assistance, and may lack the resources to take care of children for
extended periods of time.
1.4 Legal Guardianship
There is contention about whether reunification and adoption are safer and less
detrimental to the development of children than growing up in kinship care. Research has
repeatedly pointed out the danger of sending a child home when the same risk for which
the child was removed from home still exists (Schwartz & Fishman, 1999). Further, some
children in out-of-home care, especially older children, may be emotionally attached to
their biological parents despite neglect and abuse (Christian & Ekman, 2000). However,
it is undeniable that long-term kinship care lacks the judicial sanction necessary for a
child’s protection and healthy development. In particular, legally sanctioned permanency
options such as adoption provides clarity as to who has the authority to make decisions
related to the child’s life while long-term kinship care usually does not (Williams, 1991).
Combined with the expectation of continued care, the psychological effects may be
significant for the child (Williams, 1991)
For children who cannot be returned home or be adopted safely and in a timely
manner, an alternative permanency option is legal guardianship, which ASFA defines as
11
“a judicially created relationship between child and caregiver which is intended to be
permanent and self-sustaining as evidenced by the transfer to the caretaker of the
following parental rights with respect to the child: protection, education, care and control
of the person, custody of the person, and decision-making” (Christian & Ekman, 2000).
The establishment of a guardianship agreement must be sanctioned by the court, as well
as its dissolvent. Because the legal guardianship arrangement between the child and the
relative caregiver allows children to live in a relatively stable environment and yet stay in
the family, it provides a “middle-ground” between the instability associated with foster
care and the traumatization associated with having to be separated from their families
permanently (Leashore, 1984). For relative caregivers, this may be an attractive option
because it does not require the termination of parental rights, and yet grants the
caregivers considerable decision-making power.
Despite the desirable nature of legal guardianship, only 3% of all foster care cases
in the country have the permanency goal of guardianship (USDHHS, 2003b). This is, in
part, due to the fact that payments to legal guardians are not reimbursable under Title IVE of the Social Security Act. Many caseworkers think that caregivers will not be able to
provide adequate care for the children in their care with only the TANF payments
(Gleeson, 1999b).
1.5 Maryland Guardianship Assistance Project
In order to test the idea of providing caregivers who assume guardianship with a subsidy,
the federal government granted seven states with Title IV-E waivers to conduct assisted
guardianship demonstration projects. The objectives of these projects were to determine
12
whether the subsidy accelerates children’s exit from foster care, and to examine its
impact on key child well-being indicators such as placement stability.
Maryland is one of the seven states that received IV-E waivers to implement a
subsidized guardianship program. The Maryland’s Guardianship Assistance
Demonstration Project (GAP), which started in 1998 and was completed in June 2003,
provided a $300 monthly subsidy to relative caregivers who assumed legal guardianship.
In order to be eligible for the GAP project, participants had to meet the following criteria:

the child must be legally committed to the state;

the child must have been with the present caregiver for at least six months;

adoption or reunification is not a viable permanency alternative;

the caregiver is able to provide a stable, healthy, and safe home for the
child;

the child is adjusting well to the caregiver's home;

the caregiver must have a means of financial support other than the
guardianship subsidy; and

the child and caregiver must reside in Maryland (University of Maryland
School of Social Work [UMSSW], 2003)
The GAP project was formally evaluated by the University of Maryland School of
Social Work. The evaluation was a randomized social experiment and adopted an intentto-treat (ITT) approach. The results indicated that children who were initially placed in
kinship care and who were eligible for the guardianship subsidy exited from foster care at
a significantly faster pace than their peers who were not eligible for the subsidy. Among
the children who exited, those who were eligible for the guardianship subsidy were more
13
likely to exit via guardianship. The guardianship subsidy was not found to significantly
affect children in restricted foster care in terms of exit rates or exit types. The effects of
the eligibility of the guardianship subsidy on placement stability while program
participants were in foster care were inconclusive. The literature review section will
discuss the results of this evaluation in further detail.
Although the GAP evaluation was methodologically sound, two factors tend to
limit the usefulness of the evaluation results. The first factor concerns the eligibility of
program participants. To participate in the GAP project, reunification and adoption must
have been ruled out as viable permanency goals. This requirement was intended to
preclude the possibility of legal guardianship substituting for either reunification or
adoption as a permanency outcome for children. However, whether reunification and
adoption can be successfully ruled out as permanency outcomes is questionable. In fact,
among GAP participants who exited from care during the life of the program,
approximately 17% exited via reunification, even though at the time of study recruitment
caseworkers had ruled out reunification as a permanency option. This indicates that the
estimated program impact may not represent the true program effects on the target
population of GAP – those who would not eventually be adopted or be reunified with
their natural parents. As such, the estimated program effects may not be relevant when
GAP is implemented at a time when program eligibility can be determined more
effectively.
The second factor that introduces uncertainty to the external validity of the
evaluation is the distinction between kinship care and restricted foster care. As discussed
previously, Maryland has a two-tired payment system for relative caregivers. Unlicensed
14
formal kinship caregivers receive the TANF payments, averaging $188 per child per
month among program participants. RFC caregivers, on the other hand, received
approximately $600 per child per month. The existence of the dual payment system to
relative caregivers means that a typical RFC caregiver, who received a $600 monthly
subsidy for being a foster parent, stands to lose $300 per month if s/he assumed
guardianship, which paid only $300 per month. It was hypothesized, however, that some
of the RFC caregivers might opt to take up guardianship because the guardianship
subsidy continues until the child is 18 years of age while the foster care board payment
may be temporary (UMSSW, 2003). In addition, there is no stigma associated with
guardianship as that of foster care. It was nonetheless expected that the subsidy would
have a greater impact on children in kinship care rather than those in RFC. The
discrepancy in subsidy rates also entails that for a child who was initially placed in
kinship care, several possible permanency outcomes existed: he/she may have exited
from foster care, have moved to RFC, or remained in kinship care and continued to
receive subsidy at the TANF rate. Children who would exit from care, move to RFC, or
remain in kinship in the absence of the guardianship subsidies may have very different
characteristics. As such, the effects of the availability of the subsidy may be very
different. The GAP estimate of program impact may not be generalizable to a kinship
care population that has a different composition than the GAP participants.
The GAP evaluation, with an experimental design and an ITT approach to
analyses, ensured that causal effects can be convincingly inferred. However, the settings
of this experiment limit the generalizability of the estimates of the program impact. In
particular, what are the true effects of the availability of the subsidy on children who
15
could not exit via adoption or reunification? What will be the effects of the GAP program
if the composition of the children in relative foster care changes in the future, with more
children entering restricted foster care or exiting from care in the absence of the subsidy?
How will this change affect program effects? This dissertation is an attempt to answer
these questions. More specifically, this study will examine the differential effects of the
eligibility of the guardianship subsidy on placement stability between four distinctive
groups of children who were placed in kinship care at the beginning of the project: in the
absence of the subsidy, those who would remain in kinship care, those who would move
to RFC, those who would exit relative foster care via guardianship, and those who would
exit from care via reunification. By differentiating the program effects for these four
groups of children, one is able to generalize the estimated program effects to kinship care
population with different characteristics than the current cohort of children. Innovative
methods of propensity score matching proposed by Frangakis and Rubin (2002) were
used to define the four groups.
16
Chapter 2. Literature Review
2.1 Relative Foster Care Vs. Non-relative Foster Care
Empirical studies on relative foster care started to appear only in the last two decades.
Most of these studies focused on comparison between formal relative foster care and nonrelative foster care (Beeman, Kim & Bullerdick, 2000; Berrick, Barth & Needell, 1994;
Brooks & Barth, 1998; Grogan-Kaylor, 2000; Scannapieco, 1999). Many were studies on
child and caregiver characteristics (Beeman et al., 2000; Grogan-Kaylor, 2000). Other
research focused on services provided for the children and their caregivers (Scannapieco,
Hegar & McAlpine, 1997), and child outcomes including permanency, length of stay, and
long-term outcomes (Buehler, Orme & Post, 2000; Link, 1996; Wells & Guo, 1999;
Courtney & Needle, 1997)
2.1.1 Child and Caregiver Characteristics
The majority of studies on relative foster child characteristics are in comparison to nonrelative foster child characteristics. Most studies find that children in relative foster care
are predominantly African-American (Beeman, Kim & Bullerdick, 2000; Berrick, Barth
& Needell, 1994; Grogan-Kaylor, 2000; Scannapieco, 1999). Race and ethnicity appear
to be a factor in determining the type of placement when a child enters out-of-home care.
Being African American significantly increases a child’s probability of being placed with
relatives (Grogan-Kaylor, 2000).
Very young children appear not to be the focus of relative care. Infants and
toddlers under the age of two are less likely to be placed with kin than non-relative foster
17
parents (Beeman et al., 2000; Grogan-Kaylor, 2000). Placement with relatives has not
been found to be related to the gender of the child (Scannapieco, 1999).
In addition to race and age, the presence of any child problems appear to affect
the caseworker’s decision to place a child with a relative, possibly reflecting
caseworkers’ lack of confidence in kin’s capability to cope with difficult situations.
Placement with relative caregivers as opposed to non-relative foster caregivers is
consistently found to be associated with the absence of observed physical, developmental,
and behavioral problems (Beeman et al., 2000; Berrick, Barth & Needell, 1994; Brooks
& Barth, 1998; Grogan-Kaylor, 2000; Scannapieco, 1999). Overall, children placed in
relative foster care experience lower mortality rates than children cared for by nonrelative caregivers (Barth & Blackwell, 1998). Furthermore, children cared for by
relatives have lower death rates from “congenital abnormalities, SIDS, and ill-defined
conditions” than children in the general population, suggesting that caseworkers may tend
to place healthier children with relatives. Additionally, compared with children in
traditional foster care, children placed in relative care tend to have no prior placements
(Beeman et al., 2000; Brooks & Barth, 1998; Grogan-Kaylor, 2000). In Maryland,
children with mental health problems and disruptive behavior are more likely to be
placed in non-relative foster care than with kin (Chipungu, Everett, & Verduik, 1998).
Similarly, a child is more likely to be placed with non-relative foster parents if he/she is
in special education.
Somewhat conflicting results were reported concerning the health status of
children in relative foster care. Some studies revealed that they had similar health
problems to children in non-relative foster care, though the rates of health problems for
18
both groups were higher than those of children in the U.S. in general (Dubowitz et al.,
1992). Others found children in relative foster care to have better health than their peers
in regular foster care (Berrick et al., 1994; Grogan-Kaylor, 2000). A review of early
empirical works concluded that the lack of consensus was due to the different sources of
assessment. Relative foster caregivers tended to rate the health of children in their care as
good whereas medical evaluations tended to reveal more health problems (Scannapieco,
1999).
Children placed with relatives differ from children placed with non-relatives with
regard to family background and reason for removal. Studies showed that compared with
children in traditional foster care, children placed with kin were less likely to be removed
from families receiving AFDC/TANF than families not receiving AFDC/TANF, although
they were more likely to be removed from single-parent families rather than those with
both parents (Grogan-Kaylor, 2000). A significantly higher fraction of children in relative
foster care than their peers in traditional foster care had been removed from home
because of neglect or maltreatment (Grogan-Kaylor, 2000; Scannapieco, 1999). Children
who had been removed from parents because of parental substance abuse were more
likely to be placed in relative foster care than children who were placed because of other
parent-related issues such as parents’ illness, disability, incarceration, death, inability to
cope, or abandonment (Beeman et al., 2000).
Findings regarding relative caregiver characteristics have been fairly consistent.
Compared with non-relative foster caregivers, relative caregivers appeared to be more
disadvantaged socio-economically as a group. Similar to children in relative care, relative
caregivers were predominantly African-American (Berrick et al., 1994; Gebel, 1996;
19
Link, 1996; Scannapieco, 1999). The majority of them were grandmothers of the children
in their care (Berrick, 1997; McLean & Thomas, 1996; Testa et al., 1996). It is therefore,
not surprising that they were on average at a more advanced age than non-relative foster
parents, with an over-representation of people approaching retirement age (Berrick et al.,
1994; Gebel, 1996; Link, 1996; Scannapieco, 1999). For example, one study reported that
29% of the relative caregivers were over 55 years of age while that age group only
accounted for 19% of the non-relative caregivers (Berrick et al., 1994).
With regard to economic status, relative caregivers, on average, tended to be less
educated and with more meager income than non-relative caregivers (Berrick et al., 1994;
Gebel, 1996; Scannapieco, 1999). One study indicated that over half of the relative
caregivers did not have high school diplomas (Gebel, 1996). The proportion of non-high
school graduates among non-relative caregivers was 19.5%. A significantly higher
fraction of relative caregivers than non-relative caregivers were single parents (Berrick et
al., 1994; Scannapieco, 1999). For many of them, foster care payments accounted for a
large proportion of their income (Pecora, Le Prohn & Nasuti, 1999). They were less
likely than non-relative foster caregivers to own homes, and more likely to be concerned
with drugs in their families and neighborhoods (Berrick, 1997).
In spite of their compromised economic positions, relative caregivers are
committed to caring for the children in their care (McLean & Thomas, 1996). They are
more willing to help maintain children’s relationships with birth parents, and assist with
the social and emotional development of the children in their care (Pecora, Le Prohn, &
Nasuti, 1999). However, relative caregivers were found to be unwilling to adopt children
in their care (Meyer & Link, 1990; Thornton, 1990).
20
2.1.2 Services Received by Caregivers and Children in Relative Foster Care
Although children and caregivers in relative foster care appear to be a group that has
considerable resource needs, there is well-documented evidence that they receive both
less contact with caseworkers and fewer types of services when compared with nonrelative foster caregivers (Barth et al., 1994; Berrick et al., 1994; Scannapieco et al.,
1997). One early study found that as high as 29% of the relative caregivers had no
contact with their caseworkers at all during the past year (Dubowitz & Feigelman, 1993).
Although both relative and non-relative caregivers appear to be in need of services,
relative caregivers are less likely to receive respite care, support group service, and
general and specialized trainings (Barth et al., 1994). Children in non-relative foster care
received more mental health, as well as transportation services, than those in relative
foster care (Scannapieco et al., 1997). The only type of service which was found to be
offered to children in relative care more frequently than children in non-relative care was
substance-abuse treatment (Scannapieco et al., 1997). This may reflect the great
proportion of drug-exposed children in relative foster care. Surprisingly, despite the
inadequacy of services provided to relative caregivers, they are consistently more likely
to have positive views of their social workers and the services they receive (Berrick et al.,
1994; Brooks & Barth, 1998).
The vast majority of studies on relative out-of-home care used child welfare
agencies’ administrative records (Barth et al., 1994; Beeman et al., 2000; Grogan-Kaylor,
2000; McLean & Thomas, 1996; Scannapieco et al., 1997). Some studies obtained
information through caregiver surveys (Berrick et al., 1994; Dubowitz & Feigelman,
1993; Gebel, 1996). A number of studies relied on descriptive or bivariate analysis
techniques (Berrick et al., 1994; Dubowitz & Feigelman, 1993; Gebel, 1996; McLean &
21
Thomas, 1996; Scannapieco et al., 1997). While these techniques are valid when simply
comparing the background characteristics between relative and non-relative care, it may
be misleading in analyzing other outcomes. For example, the fact that children in relative
care receive less mental health services may be due to the lower proportion of children
with mental health problems among those placed with relatives. It may also reflect the
fact that African Americans are less likely to utilize mental health services than their
Caucasian counterparts (Cooper-Patrick et al., 1999; Mcmiller & Weisz, 1996).
2.1.3 Child Outcomes
Child outcomes have been a frequent subject of foster care research. Studies in this area
focus on length of stay, permanency issues, re-entry rates, and long-term outcomes.
Regarding length of stay, a longer stay in care for foster children in general has
been found to be associated with older ages of children, financial hardship of parents,
problems that affect the mother-child relationship, and maternal mental illness (Maluccio,
Fein, & Davis, 1994). African-American boys are at particularly high risk of staying in
care for prolonged periods of time (Kemp & Bodony, 2000; Vogel, 1999). Studies,
mostly based on administrative data, indicated that children in relative foster care
remained in care longer than their peers in non-relative foster care (Link, 1996;
Scannapieco, 1999; Vogel, 1999).
Longer length of stay in care is the result of slower reunification and lower rates
of adoption. Early studies reported that children in relative foster care returned home at a
slower rate (Berrick, Needell, & Barth, 1999; Scannapieco, 1999). This may be due to the
fact that birth parents considered placement with relatives less undesirable than
placements with nonkin, and therefore, lacked the motivation to work hard towards
22
reunification (Berrick et al., 1999). More recent studies found that reunification rates did
not differ by a big margin between relative and non-relative foster care (Wells & Guo,
1999). When there was a difference, it was exclusively a function of a difference in rates
over the first few months that children were in care (Barth et al., 1994; Wells & Guo,
1999). Children placed with relatives were significantly less likely to be reunified with
birth parents than their peers placed with non-relative foster parents during the first few
months in care. However, the first few months appeared to be crucial, because those who
exited via reunification during the first few months accounted for a large proportion of
the children who would be returned home eventually (Barth et al., 1994). As such, the
majority of children in relative out-of-home care may have missed the best opportunities
of reunification. Overall, the effects of race/ethnicity, poverty, and family structure on
family reunification are similar for children in relative and non-relative care. Being
African-American; having health, emotional, and behavioral problems; coming from
families that are AFDC/TANF eligible; and having only one parent present reduces the
likelihood of returning home (Barth et al., 1994; Courtney & Needle, 1997; Landsverk et
al., 1996; Wells & Guo, 1999). Additionally, for children placed with relatives, being
removed from care because of abandonment negatively affects their chance of returning
home (Landsverk et al., 1996).
With regards to adoption, age at entry is the strongest predictor of being adopted
(Barth et al., 1994; Finch, Fanshel & Grundy, 1986; Olsen, 1982). The probability of
being adopted decreases as the age of a child increases. Other factors include race, length
of stay in care, poverty, and placement with relatives (Barth et al., 1994; Finch et al.,
1986; Olsen, 1982). Having stayed in foster care for a long time and being AFDC/TANF
23
eligible reduces the chance of being adopted. Initial placement with relatives was found
to halve the odds of later adoption, even after controlling for child and family
characteristics (Courtney & Needle, 1997). Caucasians, even when placed with relatives,
are more likely to be adopted than African Americans (Barth et al., 1994). AfricanAmerican children placed with relatives are the least likely to be adopted. The effects of
race, however, do not appear to occur prior to termination of parental rights. On the
contrary, African-American children were found to be freed for adoption at a faster pace
than Caucasian children (Barth, Courtney, & Berry, 1994; Olsen, 1982). The difference
appears to be the time of waiting to be placed in adoptive placements and adoption
legalization, which was significantly shorter for Caucasian children than for AfricanAmerican children (Barth et al., 1994; Olsen, 1982).
Although children in relative foster care stay in care longer, research revealed that
once they exited from foster care, they reentered care at a slower rate. A study using
event history analysis reported that a child whose last placement was kinship care
reentered at a rate 226% slower compared to a child in non-relative foster homes and
232% slower than a child in a group home (Wells & Guo, 1999).
In the long term, people who had experience in both relative and non-relative
foster care fare worse than the average young adults in the U.S. Young adults who spent
at least six months in foster care have more problems with alcohol use than young adults
in general. They are less satisfied with marriage and more likely to live with a spouse
before getting married (Buehler et al., 2000). However, former relative foster care
children and non-relative foster care children are similar when they grow up. There is no
difference between the two groups in terms of education, employment, income, housing,
24
social support, and experience of life stress (Scannapieco, 1999). Former members of
relative foster care are more self-sufficient and in somewhat better health, while former
members of non-relative foster care reported less heroin usage and less trading sex for
drugs (Scannapieco, 1999; Zuravin, Benedict, & Stallings, 1999).
While viewing the results concerning child outcomes, it is noteworthy that the
results are potentially subject to selection bias. Because random assignment of placement
into relative or non-relative foster care is generally impossible, the majority of the
comparative studies of relative and non-relative foster care use some type of regression
technique on cross-sectional data. It is not clear whether differences in outcomes are due
to the differences in type of care or some unobserved heterogeneity. The same is true for
studies that compare foster children with children in general.
A second note of caution concerns the external validity of the empirical studies on
children in relative foster care reviewed above, due to the limited geographic areas from
which the study samples were drawn. National samples were rare (Buehler et al., 2000).
Among the 12 studies reviewed by Scannapieco (1999), six used samples from either
Baltimore City or Baltimore County. A large number of more recent studies focused on
children from various areas of California (Berrick et al., 1999; Courtney & Needell, 1997;
Grogan-Kaylor, 2000; Leslie et al., 2000). Given this limitation, results concerning
children in relative foster care may not be nationally representative. However, studies that
concentrate on children in Baltimore City will be particularly relevant to this dissertation,
because all children in the study sample of this dissertation came from Baltimore City.
25
2.2 Placement Stability for Children in Out-of-Home Care
Many efforts have been made to find out the rates and determinants of placement
disruption (Barber, Delfabbro, & Cooper, 2001; James, Landsverk, & Slymen, 2004;
Olsen, 1982; Smith et al., 2001; Taber & Proch, 1987; Webster et al., 2000; Wulczyn,
Kogan, & Harden, 2003). This attention is duly deserved, because placement instability
has been repeatedly found to be detrimental to children’s emotional development and
lead to increased behavioral problems (Kurtz et al., 1993, McRoy, 1994; Newton et al.,
2000). Some recent studies examined the dynamics of children’s movements in out-ofhome care (James et al., 2004; Wulczyn et al., 2003). However, there is insufficient
amount of knowledge on how to prevent placements from disrupting.
2.2.1 Placement Disruption Rates
Various sources have reported on the placement disruption rates of children in foster care.
In a study as early as 1959, it was estimated that one in four children in foster care had
experienced four or more placements (Mass & Engler, 1959). Similarly, a study in 1961
revealed that 28% of all children in foster care changed placements three or more times
while they were in foster care (Proch & Taber, 1985). The disruption rate appeared to
have declined to 17% by 1977 (Proch & Taber, 1985). However, one study in 1986 found
that the percentage for children having three or more placements after two years in care
was an astounding 56% (Newton et al., 2000). Other studies in the ‘80s and early ‘90s
reported percentages of children having three and more placements ranging from 22 to
48% (Newton et al., 2000; Smith et al., 2001). According to the USDHHS’ child welfare
outcome annual report of 2000, only 4.6% of children who had been in foster care for less
than 12 months had three or more placements in Maryland, considerably lower than the
26
50-state median of 15.7% (USDHHS, 2003a). However, the rate increased to 18.1% for
Maryland children who had been in care between 12 to 24 months. The percentage of
Maryland children who had three or more placements was as high as 50.8% for children
who had been in care for over 48 months.
Placements with relatives are generally deemed to be more stable than those with
traditional foster parents (Courtney & Needell, 1997; Leslie et al., 2000; Scannapieco,
1999; Usher et al., 1999; Webster et al., 2000). However, one qualitative study found that
it was not uncommon for relative foster care placements to fall apart. Terling-Watt (2001)
reported that in a sample of children who were placed in relative foster care in Houston,
Texas between 1993 and 1996, 29.2% of children who had been in care less than six
months experienced placement disruption. The highest disruption rates were among
children who had been in care between 15 to 36 months. Almost 50% of this group of
children had multiple placements. This estimate, however, may not apply to all children
in relative care, because the study population included only those whose placements were
meant to be permanent, and only subsequent abuse or neglect, reentry to foster care, or
disappearance of child were defined as placement disruption.
The estimates of placement disruption rates vary widely among studies conducted
in different times. One possible explanation of this variation is the impact of drugs on the
physical and mental health of children. For example, high rates of placement disruption
were reported in 1986, which coincided with the year when crack cocaine was introduced
and popularized in African-American communities. The inconsistency may also be due to
different definitions of placement disruption. Some studies excluded placement changes
in the first few days of placement, arguing that most of the placements were emergency
27
placements that were meant to be temporary. Some studies only counted unplanned
placement changes while others recorded all changes as placement disruptions. As a
result, comparison of number of placements should be viewed with caution.
Furthermore, statistics prior to 1960 may not be comparable to statistics after the
1960s. Historically, African-American children were not admitted to the modern welfare
system due to racial segregation. Instead, a parallel system of child welfare existed for
children of color. In the early 1960s, a large number of African-American children began
to enter the children welfare system in the wake of the Flemming Rule (Lawrence-Webb,
1997)3. The foster care population prior to the Flemming Rule, which had a much lower
percentage of African-American children, is therefore not comparable with that after the
Flemming Rule.
2.2.2 Predictors of Placement Disruption
Much of the foster care placement literature focuses on finding the predictors of
placement disruptions. The only demographic characteristic that has been found to be
associated with placement stability thus far was age of the child at the time of placement
(Barber, Delfabbro, & Cooper, 2001; James, et al., 2004; Olsen, 1982; Smith et al., 2001;
Taber & Proch, 1987; Webster et al., 2000; Wulczyn et al., 2003). Most studies found
that children of older age were at greater risk of multiple placements. The effects of
gender and race are less conclusive than age (Olsen, 1982; Proch & Taber, 1985; Smith et
3
Prior to the 1960s, it was a common practice of some Southern and Mid-Western states to deny AFDC
benefits to African-American families on the grounds of “unsuitability” for a variety of reasons including if
the child was born out-of-wedlock, or a unmarried male adult was living in the household. In the early
1960s, in response to this discriminatory practice, the U.S. Department of Health, Education, and Welfare
issued the “Flemming Rule”, which limited the states’ use of the “suitability” rule to deny AFDC benefits.
However, as an unexpected consequence of the “Fleming Rule”, many cases that would be denied on the
ground of “unsuitability” were then labeled “child neglect”, which result in a dramatic increase of AfricanAmerican children entering foster care. See Lawrence-Webb (1997) for a full discussion of the impact of
Flemming Rule on African-American children and families.
28
al., 2001; Webster et al., 2000). There is evidence showing weak impact of race, with
Caucasian children averaging higher numbers of placements than ethnic minorities
(Olsen, 1982; Webster et al., 2000). It is uncertain, however, how much of this effect is
due to the greater proportion of African-American children placed with relative
caregivers. Somewhat conflicting results were found with respect to the effects of gender.
While one study found male children to be more prone to placement disruption than their
female peers (Webster et al., 2000), another study found teenage girls to be at greater risk
of multiple placements than children from other gender/age groups (Smith et al., 2001).
Compared with race and gender, the presence of behavioral and emotional
problems of the child was more consistently found to be predictive of frequent placement
changes (Barber et al., 2001; Newton et al., 2000; Proch & Taber, 1985). Specifically,
violent or aggressive behavior, withdrawing behavior, and inability to integrate in the
foster family or cope with authority were the main child behavioral problems leading to
placement disruption.
Children’s maltreatment history and experience with out-of-home care have also
been found to be closely related to placement stability. The strongest predictor of
multiple placements is length of stay in care (Proch &Taber, 1985). The longer a child’s
stay in foster care, the more likely it is that s/he will experience multiple placements.
However, number of placements is not merely a function of time spent in foster care.
Most placement changes occur during the first six months in care (Wulczyn et al., 2003).
Unstable placements in the first year of care are predictive of multiple placements in the
future (Barber et al., 2001). It is, therefore, more likely that the same factors preventing
children from exiting from foster care also affect their chance for stable placements.
29
Reason for removal from home and type of placements may be such factors.
Victims of neglect are less likely to experience unstable placements than children
removed from care for other reasons (Barber et al., 2001; James et al., 2004; Webster et
al., 2000). Sexual abuse appears to be most likely to lead to multiple placements, possibly
due to its more detrimental impact on a child’s emotional health (James, 2004). Children
placed in relative foster care experience fewer placement changes than those placed with
non-relative foster caregivers (Courtney & Needle, 1997; Leslie et al., 2000; Scannapieco,
1999; Testa, 2001; Webster et al., 2000). Closer relationship with the child leads to more
stable placements. However, placements with aunts are less stable compared with those
with grandmothers, although aunts are equally close to children as grandmothers (Testa,
2001). Placements with grandparents are the least likely to break. However, the
advantage of relative placement appears to have waned after approximately three years
into the placement, after which the difference in placement stability between relative and
non-relative placement is negligible (Testa, 2001).
Children placed with relative caregivers face unique risks for placement
disruption. Terling-Watt (2001) found that the most frequently cited reason for relative
foster care placement disruption was the continued influence of the child’s parents.
Relative caregivers frequently failed to abide by rules prohibiting unsupervised visitation
by parents, exposing children to risks of further abuse. Moreover, relative caregivers
often lacked the resources and expertise to care for children with special needs. The old
age and relatively poor health of relative caregivers were also barriers to placement
stability.
30
2.2.3 Consequences of Placement Instability
Although frequent change of primary caregiver has been found to affect a child’s ability
to develop attachment to his/her caregivers, which is considered a developmental
milestone (McRoy, 1994), early research did not find overwhelming evidence that
multiple placements were related to child wellbeing and social effectiveness (Proch &
Taber, 1985). More recent research suggests that frequent changes in placement for
children in foster care results in increased behavioral problems such as poor selfadjustment, low self-concept, and internalizing and externalizing behaviors, including
aggressiveness, withdrawal, and inabilities to integrate with the environment (Newton et
al., 2000; Kurtz et al., 1993). Even the perception of impermanency is significantly
associated with behavioral problems (Harden, 2004). Additionally, placement stability is
an important mediating factor for educational progress. The mobility related to foster care
placements were found to account for much of the poor academic performance of
children in out-of-home placements (Eckenrode et al., 1995). Multiple placements are
also predictive of increased juvenile delinquency. For neglected and maltreated children,
frequent placement disruption leads to increased probability of being arrested, while
having been placed in out-of-home care itself does not (Widom, 1991). Those who spent
over ten years in the first placement have lower rates of criminality than those who did
not. However, the differences disappear when factors such as behavioral problems are
controlled.
One prevalent issue concerning the placement stability research is that the
majority of these studies were retrospective and, as such, lack the baseline information
necessary to establish the direction of causality between children’s emotional and
behavioral problems and placement instability (Webster et al., 2000; Barber et al., 2001;
31
Proch & Taber, 1985; Kurtz et al., 1993). One study that controlled baseline behavioral
problems found that behavioral problems were both the cause and consequence of
frequent placement disruptions (Newton et al., 2000). After controlling for baseline
behavioral scores measured on the Child Behavior Checklist, no demographic
characteristic remained significant. Furthermore, children who initially had no behavioral
problems were at the greatest risk of being harmed by multiple placements. However, this
study did not focus on an incoming cohort of children, which inevitably biased the results
towards those who stayed in care for long periods of time.
2.3 Outcomes for Children in Legal Guardianship
There exists a scant amount of research examining the outcomes for children in the care
of legal guardians or whose permanency goal is legal guardianship. One such study found
guardianship placements to be generally stable (Iglehart, 1994). The majority (81%) of
the guardianship adolescents in the study were in their first and only placement, and half
of them had been with their current guardian for five or more years. However, most of the
adolescents were with non-relative guardians and the study was limited to those who
were 16 years or older. Another study based on a group of abused or neglected children
under legal guardianship revealed that the rate for permanency failure in adjudicated
abuse/neglect guardianship was 36%, which was two to three times higher than that in
most adoption studies (Henry, 1999). The disruption rate of non-adjudicated
abused/neglected children, on the other hand, was only 9%, which was similar to
adoption failure rate of 7% to 14% reported in the literature.
32
2.4 IV-E Waiver Demonstrations on Subsidized Guardianship
In order to better understand subsidized guardianship as a permanency option and its
effect on child outcomes, the federal government granted seven states Title IV-E waivers
to conduct demonstration projects that have assisted guardianship components for
children in relative foster care4. Four states - Illinois, Maryland, New Mexico and
Montana used experimental designs with random assignment. Delaware, North Carolina
and Oregon based their evaluations on quasi-experimental designs or descriptive analyses
(USDHHS, 2005).
As of September 2005, quantitative results regarding permanency and placement
stability were available for the states of Illinois, New Mexico, Oregon, and Maryland.
Illinois, which implemented a large scale true experiment on assisted guardianship,
reported that the availability of the guardianship subsidy increased combined permanency
rate by 6 percentage points while the rate of subsequent reports of abuse or neglect
remained unchanged. However, the final results indicated a possible trade-off between
overall legal permanency rate and the rate of adoption. The experimental and control
groups, however, did not differ in terms of placement stability (Testa, 2002; USDHHS,
no date).
The New Mexico assisted guardianship demonstration consisted of two
components: children under state custody and children under tribal custody. In the state
custody component, non-Native American children under state custody were randomly
assigned to either the experiment group, which was eligible for a guardianship subsidy, or
the control group, which was not eligible for the subsidy. Outcomes of the two groups
4
These seven states are Delaware, Illinois, Maryland, Montana, New Mexico, North Carolina and Oregan.
Two more states – Minnesota and Wisconsin – have received waiver approval and were scheduled to
implement demonstrations involving assisted guardianship in late 2005 to early 2006 (USDHHS, 2005).
33
were compared. In the tribal custody component, conversely, all Native American
children under tribal custody were eligible for assisted guardianship. These children
were compared with Native American children under state custody, who were not offered
the subsidy. The state reported that children in the experimental group of the state
custody component spent less time in out-of-home placements compared with their peers
in the control group (TriWest, 2004). The state report, however, did not indicate whether
the difference was statistically significant. No statistics on placement stability were
reported.
The evaluation of Oregon's assisted guardianship project compared children for
whom assisted guardianship was established with children from a comparison county
where subsidized guardianship was not offered. The evaluation revealed that the
eligibility of a guardianship subsidy did not have any statistically significant effect on
placement stability (Lehman, Liang, O’Dell, & Duryea, 2003). Oregon did not report
statistics regarding permanency outcomes.
The evaluation of Maryland’s GAP project was based on an experimental design
(Mandell, 2001; UMSSW, 2003). The evaluation compared outcome measures for
children who were eligible for the subsidy and those who were not eligible for the
subsidy. Program participants were grouped into two cohorts: cohort one, which was
randomly assigned to the experimental or control group between October 1998 and
February 1999; and cohort two, which was assigned on May 1, 2000. Results indicated
that the guardianship subsidy expedited children’s exit from care. An event history
analysis on the time until exit revealed that the kinship care experimental group, which
was eligible for the subsidy, had a hazard ratio that was 43 to 123% higher than that of
34
the control group, which was not eligible for the subsidy. Further, bivariate analysis
revealed that among those who exited from care, kinship children eligible for subsidy
were more likely to exit via legal guardianship than children who were not eligible for the
subsidy. However, the subsidy did not have any significant effects on the speed of
achieving permanency or exit types for children in the RFC (restricted foster care) group.
Regarding placement stability, Poisson regression and negative binomial regression
models were estimated to examine the effects of the availability of the subsidy (UMSSW,
2003). The results were inconclusive. For some unknown reason, the availability of the
subsidy increased placement instability experienced by the group of children who were
initially placed in restricted foster care. For the rest of the study sample, the effects of the
availability of the subsidy were not statistically significant regarding placement stability.
2.5 Propensity Score Matching
The probability of receiving a particular treatment conditional on observed covariates is
commonly known as the propensity score (Rosenbaum & Rubin, 1983). Since its formal
introduction in 1983, propensity score matching (PSM) has been a commonly used
strategy in observational studies to overcome selection bias. Its advantage is to solve the
dimensionality problem in matching by reducing the often numerous background
characteristics into one single variable (Rubin, 1997). Dehejia and Wahba (1999)
assessed the ability of PSM to produce unbiased estimates of program effects by
examining the impact of the National Supported Work Demonstration using the PSM
method. The results were then compared with those generated by Lalonde (1986) in his
ground-breaking study evaluating various nonexperimental methods. The authors
concluded that the PSM method could produce estimates much closer to the experimental
35
benchmark when the range of propensity scores of those receiving intervention
overlapped with that of those who did not receive the intervention. However, the
estimates from PSM were sensitive to the selection of covariates. A similar study using
data from the National Evaluation of Welfare-to-Work Strategies found that selecting
control cases through PSM improved sample balance and reduced bias (Bloom,
Michalopoulos, Hill, & Lei, 2002).
The PSM strategy has recently been used in a number of social experiments to
generate treatment-on-treated (TOT) estimates. One interesting example is a study
conducted by Barnard, Frangakis, Hill, and Rubin (2001), who estimated the effects of
using a school voucher – not the effects of a voucher offer, which generally requires
straightforward intent to treat (ITT) estimates – by stratifying the study sample into
compliers, always takers, never takers, and defiers. Propensity scores for being in each
group were calculated through probit models. Control group members were thus matched
to the experimental group members with the closest propensity score. The impact of the
voucher was derived by comparing standardized test scores of the compliers who were
offered a voucher and the compliers who were not offered the voucher. A similar study
examined the effects of taking advantage of a housing voucher on participants’ wellbeing for people living in high-poverty areas (Katz, Kling, & Liebman, 2001).
The PSM strategy can also be applied when multiple treatments are involved.
Gibson (2001), for example, examined the impact of the New Hope demonstration, a
randomized anti-poverty program using the PSM method. The benefits offered by the
program included child care subsidies, wage subsidies, health insurance, and temporary
community service jobs. The author divided the experimental group into five subgroups
36
according to the services that they took advantage of, and matched control group
members to the experimental group members with the closest propensity scores. Thus,
she was able to show the differential effects of different program components.
Using the PSM strategy to generate TOT estimates in a randomized experiment is
different from standard PSM in that usually only the members of the experimental group
will be used in the estimation of the model that predicts the likelihood to participate. A
similar application of this strategy is the use of PSM in sub-group analysis where only the
members of the control groups are used in the initial model estimation. One example is a
study that examines the differential effect of high-quality center-based child care on
children who would, in the absence of the center-based child care provided, end up
having no non-maternal care, some non-maternal care but no center-based care, and some
center-based care (Hill, Waldfogel, & Brooks-Gunn, 2002). Another example is a study
that aimed to find out whether the impacts of two anti-poverty programs – the New Hope
demonstration and the Minnesota Family Investment Program – differed by the
participants’ risk for unemployment (Yoshikawa, Magnuson, Bos, & Hsueh, 2003). In
both studies, the defining variables for sub-groups were post-randomization variables that
were correlated with whether or not the program participant received intervention.
Furthermore, the defining variables were only observed for the members of the control
groups. Authors of both studies concluded that the PSM strategy was successful in
achieving sample balance, and the authors of the former study reported that results after
PSM were considerably different than if the sample was stratified using observed types of
child care.
37
Chapter 3. Methodology
3.1 Research Question
For children who have been formally placed with relative caregivers, there are many
possible outcomes. Some may be returned to parents. Some may be adopted by their
caregivers. Some may leave the child welfare system by moving to guardianship care,
and yet some may remain. Even those who remain in out-of-home placements may
experience different types of care. Some caregivers may eventually be licensed and
become restricted foster caregivers. Some may remain in kinship care and their
caregivers continue to receive subsidies at the TANF level.
This dissertation focuses on the children who were placed with relative caregivers
at the time of the random assignment of GAP. Although reunification and adoption had,
in principle, been ruled out for program participants as a condition for eligibility in the
study, predicting permanency outcomes appeared to be difficult. As a result, a substantial
number of children actually exited foster care through reunification (UMSSW, 2003). As
such, the majority of children achieved one of the four types of outcomes: 1) exiting
through reunification, 2) exiting through guardianship, 3) remaining in out-of-home care
but moving to restricted foster care, and 4) remaining in kinship care.
Many studies have shown that permanency outcomes are associated with various
background characteristics of children, as the previous chapter has indicated. Similarly,
the effects of the eligibility for the guardianship subsidy on placement stability may vary
among these children. Although the GAP evaluation indicated that the overall effects of
guardianship subsidy on placement stability were negligible, no efforts thus far have been
38
devoted to examining the distribution of the effects. It is not inconceivable that eligibility
for the subsidy may benefit some groups of children by enhancing their placement
stability while hurting other children by undermining it. As such, this dissertation
examined the following research question:
What are the differential effects of the availability of the guardianship subsidy on
placement stability among four distinctive groups of children who were in kinship care
placements at the beginning of the project:
1. Those who would remain in kinship care in the absence of the subsidy;
2. Those who would move to restricted foster care in the absence of the subsidy;
3. Those who would exit relative foster care via guardianship in the absence of the
subsidy; and
4. Those who would be reunified with their parents in the absence of the subsidy.
There is reason to believe that the availability of a guardianship subsidy may
reduce the number of placements a child experiences in foster care. It is conceivable that
the prospect of receiving a guardianship subsidy will encourage relative caregivers to stay
with the children in their care rather than withdraw from providing care in case of any
problems. In addition, the knowledge that his/her placement with the relative caregiver is
more likely to become permanent may help the child develop attachment to the caregiver
and vice versa.
Regarding the differential effects of the subsidy among the four sub-groups of
children, it is expected that the subsidy will not have a great impact on those children
who would exit from care even if a guardianship subsidy were not available. The effects
on those who would exit from care via guardianship in the absence of the subsidy are
39
expected to be especially small, if any. Regarding those who would be reunified with
their natural families, the availability of the subsidy may result in guardianship,
substituting reunification as a permanency outcome for children. However, there is no
reason to believe that the subsidy may affect placement stability.
Although the subsidy offer may not have much impact on children who would
exit from care in the absence of the subsidy, it may benefit children who would otherwise
remain in kinship care. It has been reported that the availability of the subsidy increased
the overall exit rates from foster care (Mandell, 2001; UMSSW, 2003). It is reasonable to
assume that the eligibility of the subsidy induced a fraction of the group members which
would remain in kinship care in the absence of the subsidy to exit from care and take up
guardianship. This suggests that the subsidy offer’s effects on placement stability may be
the greatest among those who would remain in care in the absence of the subsidy.
Compared with the other three sub-groups, the magnitude of the effects of the
subsidy offer for those who would otherwise move to restricted foster care is more
difficult to predict. Whether the child is eligible for guardianship subsidy may not have
much bearing on the caregiver’s motivation to continue providing care, because RFC
provides a subsidy twice as high as the guardianship subsidy. However, subsidized
guardianship as an additional permanency option may still have psychological effects on
the caregiver as well as the child.
As indicated by the research question above, subgroup status is defined by the
permanency outcome of the child in the absence of the subsidy. Permanency outcomes,
however, were not realized until after randomization. Further, post-randomization
permanency outcomes are expected to be correlated with the eligibility for the subsidy.
40
As such, post-randomization and observed permanency outcomes are not a good
approximation of the defining variable of the subgroups. This issue will be discussed in
more detail when the methodology of the statistical analysis is presented.
3.2 Sample and Data
Two cohorts of children in Baltimore City and one group of children from other
jurisdictions in Maryland participated in GAP. They were randomly assigned to either the
experimental group or the control group. Members of the experimental group were
eligible to receive a $300 monthly subsidy until the child reaches maturity at 18 years of
age if the caregiver assumed guardianship5. Members of the control group, on the other
hand, were not eligible for this subsidy. Random assignment of the first cohort of 772
children, which consisted of 40% of Baltimore City’s eligible relative foster care
population, took place between October 1998 and February 1999. However, a process
study revealed that the offers of guardianship subsidy were never made to many RFC
caregivers even if they were eligible, perhaps because their caseworkers believed that
they would not be interested in participating (UMSSW, 2003). The second cohort of
children, which included the remaining 986 eligible children in relative foster care in
Baltimore City, was randomly assigned to the control or experimental group on May 1,
2000. The offer of subsidy was implemented more successfully with this cohort of
children due to more aggressive training of the caseworkers (UMSSW, 2003). The group
of 107 children from other jurisdictions in Maryland was also randomly assigned.
However, with such a small sample size, the subsidy effects would have to be quite
substantial to be detected for children coming from other jurisdictions of Maryland.
5
In the event that the child is in an approved education setting, the age of maturity is 21.
41
This dissertation focuses on the second cohort of children from Baltimore City,
with whom the subsidy was implemented as originally designed. Among these children,
511 were in kinship care placements at the time of random assignment. Further
examination indicated that 74 children exited from foster care before random assignment,
and never reentered the system. These 74 cases were excluded from any analysis. Further,
13 cases exited from foster care after random assignment and reentered the system at
some point during the study. Placement dynamics were expected to be different for these
cases. These cases were also excluded from the study. As a result, the effective sample
included 424 children who were randomly assigned to the experimental or the control
group on May 1, 2000, and were in kinship care placements at the time of random
assignment.
Data on case history and some child characteristics were downloaded monthly
from the state administrative database Foster Care and Adoption Child Tracking System
(FACTS). During the period between December 2000 and April 2003, 20 monthly
downloads were provided by the Maryland Department of Human Resources.6
In addition to the administrative data, information regarding caregiver
characteristics was available for some caregivers. All caregivers were invited to
participate in a self-administered and computerized interview, which asked for extensive
information about caregivers, including caregiver demographics, economic and health
status, spirituality, and social support, as well as some information about each child in
6
Downloads were not available in 9 of the 29 months between December 2000 and April 2003: March,
April, June, and October of 2001; April, October, and November of 2002; and February and March of 2003.
Because none of the relevant variables involve exact time, it is not expected that the missing downloads
will have a significant impact on the results. Further, even if some information regarding case history was
made unavailable by the missing downloads, the effects of such missing information should have affected
the control and experimental groups equally.
42
their care. However, only 159 caregivers of 258 children from cohort two agreed to and
completed the interview by May 2003. Among the 258 children whose caregivers have
been interviewed, 146 were initially placed in kinship care.
3.3 Measurement of Placement Stability
Placement stability is defined as the change in number of placements between December
2000 and April 2003. Ideally a measure of placement stability should exclude moves
from a more restrictive setting to a less restrictive setting (Smith et al., 2001).
Additionally, one may also argue that placement changes should exclude stays at
emergency shelters, which serve as temporary placements before a more permanent
placement can be arranged (James et al., 2004). However, because distinguishing “good”
placement movements versus “bad” placement movements requires more information
than is currently available, this study only examines the aggregate count of placement
changes during the follow-up period.
3.4 Statistical Analysis
This study examined the differential effects of the availability of a guardianship subsidy
among four groups of children. Group status was defined as the permanency outcome of
the child in the absence of the subsidy, a variable that is only observed after
randomization. The statistical analysis, therefore, adopted a two-stage procedure. In the
first stage, the propensities to achieve the four different types of permanency outcome for
the control group were analyzed with a multinomial logit model. Members of the
experimental group were subsequently matched with members of the control group based
on the propensity scores. In the second stage, separate Poisson regression models were
estimated for each group of children to gauge the effects of the availability of the subsidy.
43
3.4.1 Propensity Score Matching
The research question of this dissertation distinguished between the effects of eligibility
for the subsidy among four groups of children. The first step in the data analysis,
therefore, was to identify the children in each of the four groups. Group identification
was complicated in this dissertation study as the defining variable of groups was a postrandomization variable that was only observed for the control group. Specifically, let Z
denote the variable representing whether the child was in the experimental group and, as
such, was eligible for the subsidy:
0
Z 
1
If the child was not eligible for the guardianship subsidy
If the child was eligible for the guardianship subsidy
Let G denote the variable that represents the permanency outcome of a child:
 g If the child exited from foster care via guardianship
r

If the child exited from foster care via reunification
G  
 k If the child remained in kinship care

 f If the child moved to RFC
The defining variable of subgroup status, namely permanency outcomes of the child in
the absence of the subsidy, can be written as [ G Z  0 ]. This however, was only observed
for the members of the control group. What we observed for the members of the
experimental group, on the other hand, was [ G Z  1 ]. For these children, [ G Z  0 ] was
a potential outcome which was not observed. Further, as the previous section points out,
stratifying the sample using the observed permanency does not produce unbiased
estimates of the subsidy effects within strata because permanency outcome G was likely
to be affected by the eligibility of the subsidy, represented by Z. Children from the
44
control group and children from the experimental group who had achieved the same
outcome may not be equivalent.
This issue was addressed by the propensity score matching method described by
Frangakis and Rubin (2002), who showed that treatment effects derived by comparing
cases of different treatment status within principal strata are causal effects. Principal
stratification with respect to a post-treatment variable, such as permanency outcomes,
refers to the process of classifying cases according to the probabilities of potentially
obtaining the values of the post-treatment variable, predicted by pre-treatment covariates.
In this dissertation, a multinomial logit model was estimated with data related to
the control group, for whom permanency outcomes in the absence of the subsidy were
observed. Using the estimated parameters, propensity scores, namely predicted
probabilities, were computed for members of both the control and experimental groups.
For each individual, four propensity scores, representing the probabilities of achieving
each of the four permanency outcomes, are computed. The variable for group definition
([ G Z  0 ]) was observed for members of the control group. Experimental cases, for
whom [ G Z  0 ] was not observed, were matched to control cases with the closest
propensity score. For example, suppose child A from the control group exited from care
via guardianship during the study period. Four propensity scores, namely Pr(GA  g x A ) ,
Pr(G A  r x A ) , Pr(G A  k x A ) , and Pr(G A  f x A ) would be computed from the
multinomial logit model. Four propensity scores would also be computed for each
member of the experimental group, one for each permanency outcome. In reality,
suppose it was observed that the child A exited from care via guardianship. To find the
experimental group member that should be matched to child A, one would
45
compare Pr(GA  g x A ) with Pr(G  g x ) for each member of the experimental group.
Assume that child B is the member of the experimental group that has the closest
Pr(G  g x ) to child A. Child B, then, would be matched to Child A. Both child A and
child B are said to belong to the group of children who would exit through guardianship
in the absence of the subsidy ([ G  g Z  0 ]), regardless of child B’s actual permanency
outcome.
Matching was conducted with replacement. As such, it is possible that one
experimental case was matched to more than one control group case in the same
subgroup. It is also possible for one experimental case to be matched to control cases
belonging to different permanency outcome groups. Because some experimental cases
may be matched to more than one control case, matching with replacement results in a
smaller sample size. However, it produces more balanced treatment and comparison
groups (Bloom, Michalopoulos, Hill, & Lei, 2002).
Table 3.1 presents the variables that were considered as for being included in the
predictive models. The main source of data is FACTS, the state administrative database,
which contains variables regarding demographic characteristics, special needs of the
child, and case history. The caregiver interview provided much needed information
regarding caregiver characteristics, which were considered in the sensitivity analysis
discussed in the end of this chapter.
46
Table 3. 1 Predictors of child permanency and placement outcomes
Variable Type
Variables
Gender and race
Age at the time of entry into out-of-home care
Child characteristics
Whether the child has a sibling group in out-of-home care
Presence of special needs
Length of stay in care prior to random assignment
Case history
Reason for entering into out-of-home placement
Number of placements before the current episode
Child Characteristics
Child race, gender and age at the time of entry into out-of-home care were
included as predictors in the models predicting permanency status. A substantial number
of children in the sample have sibling groups in out-of-home care. An indicator of
whether the child has siblings was included in the model.
Whether the child was in special education was also considered a predictor.
However, because the proportion of child in special education was small, it was not able
to be included in the predictive models of permanency and placement outcomes.
A small proportion of children (less than 10%) have any special needs. These
special needs include a variety of physical, mental, and behavioral problems that would
warrant special attention in placement.7 Because the types of special needs vary widely,
and accounting for each type of these special needs would dramatically reduce the cell
sizes, a dummy variable representing the presence of any special needs, therefore, was
considered in the predictive models.
Case History
Length of stay in care prior to random assignment was computed for all children. Also
included in the model was the number of placements during the previous episode of out7
The percentages of children with different types of special needs are presented in Table 4.1 in Chapter 4.
47
of-home care. Regarding the child’s entry reason, whether the child was placed in out-ofhome care because of abandonment was included in the model. Abandonment was the
second most common reason of entry, accounting for over 10% of children. Almost 80%
of all children in the sample were placed in out-of-care solely or partially due to parental
neglect. A small fraction of children were removed from homes because of abuse and a
number of other reasons.8 Dummy variables representing whether the child was removed
from home because of parental neglect and abandonment were considered as predictored
in the models. Other entry reasons were not included due to the lack of variations.
Another variable not included in the model was primary caregiver at removal.
Although children removed from parents and other caregivers might differ in a way that
affects the permanency outcomes, information on primary caregiver at removal was
missing in a substantial number of the cases (19%, n=80). Including this variable as a
predictor would reduce the effective sample size significantly.
3.4.2 Poisson Regression
The outcome variable of this study, namely changes in the number of placements during
the follow-up period, is a count variable. As common to variables representing patterns of
events occurring independently over a period of time, it was assumed that placement
changes follow a Poisson probability distribution. Specifically,
Exp(-µ)µy
Pr(y|µ) =
y!
8
See Table 4.1 in Chapter 4 for a tabulation of entry reasons by the eligibility of the subsidy.
48
where y was a Poisson random variable representing the number of changes a child
experiences during a certain time period and µ was the mean number of changes in
placements. Poisson regressions, where the conditional mean of a Poisson random
variable is explained by individual characteristics, was employed to estimate the effects
of being in the experimental group on placement stability. Specifically, the structural
model estimated was the following:
i  E( yi xi )  exp(  Zi  xi β) , where yi ~ Poisson (  i ) .
Variable Z represents whether the child is eligible for the guardianship subsidy, and x is
a vector of covariates that have been identified in the literature as predictive of placement
stability, including child race, gender, age at entry, length of stay in care, child special
needs, and entry reason. Due to the fact that some children exited from foster care at
different stages of the study, the time of exposure to the risk of placement disruption
varied among cases. The model, therefore, was also adjusted for the amount of exposure
of the cases. Regressions models are estimated separately for each of the four groups:
children who would exit from care via guardianship, children who would exit care via
reunification, children who would move to RFC, and children who would remain in
kinship care in the absence of the guardianship subsidy.
One critical assumption of the Poisson probability distribution is that the variance
of the Poisson random variable is equal to its expected value. When this assumption is
not satisfied, Poisson regression is inefficient, and the standard errors of the coefficient
estimates are smaller than what they should be (Long, 1997). If this is the case, negative
binomial regression, which allows for overdispersion, should be used. The resulting
model will be the following:
49
~i  E( yi x i )  exp(  Z i  x i β   i )  exp(  Z i  x i β) i ,
where δi represents unobserved heterogeneity and is assumed to have an expected value
of 1 and follow a Gamma distribution with variance 1/vi. It is assumed that vi is constant
across individuals and
vi   1 where   0 .
The equal-dispersion assumption was tested using a log-likelihood ratio test, where the
null hypothesis is   0 . When the null hypothesis was not rejected, the negative
binomial model reduced to a Poisson regression model. Otherwise the negative binomial
regression model was used.
3.4.3 Sensitivity Analysis
Sensitivity Analysis 1: Caregiver Characteristics
As discussed previously, the number of children in the sample whose caregivers
completed the caregiver interview accounted for approximately 25% of all children in the
sample. Although caregiver characteristics were expected to be predictive of permanency
outcomes, we faced a trade off in that including variables retrieved from the caregiver
interviews would reduce the sample size dramatically.
To determine whether the omission of caregiver characteristics has any impact on
the matching results, a multinomial logit model was estimated with child characteristics
and a dummy variable representing whether or not the caregiver of the child participated
in the caregiver interview. While the individual impact of caregiver characteristics, such
as caregiver demographics, household income, and caregiver health status, could not be
determined, the dummy variable served as a proxy for the set of characteristics that
distinguish interview participants from non-participants.
50
Sensitivity Analysis 2: Excluding non-African-American Children
Over 95% of the sample of this study was African-American. A sensitivity analysis was
conducted with the American-American sub-sample. Non-African-American children
were excluded from both the propensity score matching process and the second-stage
Poisson regression or negative binomial models. Results are discussed in the next chapter.
51
Chapter 4. Results
4.1 Sample Characteristics
A total of 424 children who were initially placed in kinship care and did not reenter outof-home care during the study period were included in the analysis. Among these
children, 223 were assigned to the control group, and 201 were assigned to the
experimental group. Table 4.1 presents the characteristics of these children. Both the
control and experimental groups were approximately evenly split between males and
females. African-American children accounted for the overwhelming majority of the
experimental and control groups, while children from other ethnic groups accounted for
less than 5% of either group. There are slightly more African-American children in the
experimental group than the control group. At the time that a typical child entered kinship
care, s/he was approximately 6 years old. Children in the experimental group entered
kinship care at a slightly older age than those in the control group, but not significantly so.
Table 4. 1 Sample Characteristics
Gender
Female
Male
Race
African American
Other
Education
Regular School
Special Education
Not in School
Entry reason
Neglect
Abandonment
Abuse
Control Group
Experimental
Group
49.3%
50.7%
51.2%
48.8%
χ2= 0.16, p=0.69
95.5%
4.5%
98.5%
1.5%
χ2= 3.15, p=0.07
70.9%
2.7%
26.5%
64.8%
3.9%
31.3%
χ2= 1.68, p=0.43
70.7%
11.6%
5.6%
72.5%
11.8%
5.6%
52
Control Group
Other
More than one of the above
Special needs
None
Sibling groups
Other special needs
Both sibling group and other
special needs
Caregiver at removal
Mother only
Father only
Both mother and father
Mother/father and other
adult
Relatives
Non-relatives
Age at entry
Length of stay in care (in months)
prior to random assignment
N
1.5%
10.6%
Experimental
Group
0.6%
9.6%
χ2= 0.95, p=0.92
35.5%
52.7%
7.5%
4.3%
45.2%
45.8%
7.1%
1.8%
χ2= 4.75, p=0.19
79.6%
1.1%
2.8%
4.0%
76.2%
1.2%
1.8%
3.6%
6.3%
6.3%
5.9
43.4
11.9%
5.4%
6.4
41.6
223
201
χ2= 3.74, p=0.59
t=-1.23, p=0.22
t = 0.66, p=0.51

African American, Non-Hispanic

Other entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s
special needs.

Other special needs include child physical, mental and behavioral problems.
Sample mean
Children in the sample were relatively free of problems. Only approximately 10%
of the children have any identified special needs involving physical, mental, and
behavioral problems. About half of the children had sibling groups. Less than 4% of
children were in special education. For sample members five years and older, 4.67% were
in special education. This was considerably lower than the proportion of children in
special education in Maryland during the same time period. During the 2000/2001 school
year, as high as 8.55% of the children from 6 to 21 years of age in Maryland were in
special education (U.S. Department of Education, 2002). However, 10.75% of the sample
members 5 years and older were not in school at the beginning of the study. All of these
children were under 15 years old. It is unclear why these children were not in school. No
53
statistically significant differences were detected between the experimental and control
groups regarding child special needs or education status.
The members of the experimental and the control groups were also similar in
terms of family background and maltreatment history. Most children were removed from
single-parent families. Only about 3% of the control group members and 2% of the
experimental group members were removed from families with both parents present. The
most common reason for out-of-home placement was parental neglect. In over 70% of the
cases in the sample, parental neglect was identified as the only reason why the child was
removed from home. The second most commonly identified reason for removal was
abandonment, followed by abuse. At the time of random assignment, the members of the
experimental group and the control group had spent an average of 41.64 and 43.40
months (or about three and a half years) in out-of-home care, respectively. No significant
difference was found between members of the experimental and control groups in terms
of family background or reasons for entry into out-of-home care.
4.2 Bivariate Analysis
While the background characteristics of the members of the experimental group
and those of the control group were similar at the baseline, their experiences in out-ofhome care started to diverge after the guardianship subsidies were made available to the
experimental group members. Table 4.2 presents the number of placement changes
during the study period between December 2000 and April 2003. The majority of both
the control and experimental groups experienced no placement disruption during this
period. However, a significantly higher percentage of the members of the experimental
54
group (73%) had no placement changes compared to the members of the control group
(57%).
Table 4. 2 Number of placement changes
Control Group
Experimental Group
None
128
146
57%
73%
One
64
33
28%
16%
Two
20
11
9%
5%
Three or more
9
8
4%
4%
N = 424
χ2 = 12.54
P=0.006
The difference in the number of placement changes, however, was not simply a
function of placement stability when the sample members were in out-of-placement care.
It also reflected the higher percentage of children who exited from out-of-home care
before the end of the study. As previous evaluation results indicate, children in the
experimental group exited from care faster than children in the control group (Mandell,
2001; UMSSW, 2003). As children exited from care, they were no longer exposed to the
risks of placement disruption. As such, everything else equal, it was expected that the
group of children who exited from care faster would experience fewer placement changes.
Table 4.3 shows the number of children who exited from kinship care between
December 2000 and April 2003 and the permanency status of these children as of April
2003. Approximately 65% of the members of the experimental group exited from out-ofhome care during this period while only 48 % of the members of the control group exited
from care. Most of the increase in the number of exits appeared to be attributable to
guardianship. A significantly higher percentage of the experimental group members (40%)
exited via guardianship than those in the control group (27%). On the other hand, a
55
higher age of the children in the control group (27%) than children in the experimental
group (15%) remained in kinship care. Additionally, a lower percentage of children in the
experimental group (7%) than those in the control group (13%) were returned to their
birth parents, indicating that guardianship may have substituted for reunification as a
permanency outcome for a small proportion of children.
Table 4. 3 Permanency Outcomes
Control Group
61
27%
28
13%
10
4%
7
3%
56
25%
61
27%
Exited to guardianship
Exited to reunification
Exited to adoption
Exited for other reasons9
Moved to RFC
Remained in KC
N = 424
χ2 = 31.57
Experimental Group
107
40%
14
7%
5
3%
5
2%
41
20%
30
15%
P<0.001
4.3 Subgroup Analysis: Children who Would Exit and Children who Would Stay
The objective of providing a guardianship subsidy is to expedite the pace of children
exiting from out-of-home care. As such, a policy relevant question is whether the
subsidy has a bigger impact on children who would remain in kinship care than children
who would exit out-of-home care in the absence of the subsidy. To address this question,
based on the probabilities of exiting, children in kinship care placements at the beginning
of the project were divided into two groups: 1) those who would exit from foster care in
the absence of the subsidy; and 2) those who would not exit from foster care in the
9
Other exit reasons include independent living, running away, agency service not appropriate, and other.
56
absence of the subsidy. The impact of the availability of the guardianship subsidy was
examined separately for each group.
4.3.1 Predicting the Probabilities of Exiting Foster Care in the Absence of the
Subsidy
Logit models were estimated to determine the impact of demographic characteristics and
case history on the probabilities of exiting for children in the control group. Table 4.4
presents the results obtained from the logit models. In addition to demographic and case
history variables, Logit model 2 included whether the caregiver participated in the
caregiver interview as a predictor while Logit model 1 did not. Among the 223 children
in the control group who were in kinship care placements at the time of random
assignment, 44 children were not included in the logit models due to missing data,
resulting in a sample size of 179. Because more than one child might come from the same
sibling group and be cared for by the same caregiver, the independence assumption of the
maximum likelihood procedure may be violated. The clustering effects were adjusted for
by using the Huber/White robust estimator of variance to compute the standard errors of
the coefficient estimates.
Table 4. 4 Coefficients obtained from the logit models predicting the probabilities of
exiting from foster care for the members of the control group
Logit Model 1
Logit Model 2
Male
-0.05
-0.05
(0.33)
(0.33)
African American
-0.96
-0.91
(0.88)
(0.92)
Age at entry (in months)
-0.00
-0.01
(0.00)
(0.00)
Length of stay in care prior to random -0.02***
-0.02***
assignment (in months)
(0.01)
(0.01)
Entered foster care due to neglect
-0.21
-0.04
(0.55)
(0.56)
Entered foster care due to abandonment 0.85*
0.91**
57
Logit Model 1
(0.47)
Number of placements in the previous 0.11*
episode
(0.06)
Having sibling groups
-0.30
(0.34)
Having any special needs
-0.30
(0.51)
Participated in caregiver interview
---Constant
N=
Model Significance p<=
Pseudo R2 =
Log Likelihood =
2.18**
(1.06)
179
0.01
0.10
-111.53
Logit Model 2
(0.46)
0.12**
(0.06)
-0.25
(0.34)
-0.27
(0.50)
-0.62
(0.40)
2.13*
(1.10)
179
0.01
0.11
-110.25
Significant at .1 level.; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level

The reference group is non-African-American children
As table 4.4 indicates, three variables were predictive of the probability of exiting from
out-of-home care. Having spent more time in out-of-home placement significantly
reduces the likelihood of exiting foster care. This is consistent with the exiting body of
literature, which has provided voluminous evidence on the negative correlation between
length of stay in care and the chances of achieving permanency (Barth et al. 1994;
Courtney & Needle, 1997; Finch et al., 1986; Olsen, 1982). Children who were
abandoned by their parents had a higher probability of exiting from care, probably
because relative caregivers were more motivated to pursue permanency options other
than reunification when the parents of the children were not involved. A greater number
of placements in the previous episode led to a higher probability of exiting care. This
result is somewhat unexpected because the number of placements is likely to be related to
physical, behavioral, and emotional problems, which normally reduces the likelihood of
achieving permanency.
58
The results indicate that both logit model 1 and logit model 2 are statistically
significant. A log-likelihood ratio test showed that adding the variable representing
whether the caregiver completed the caregiver interview did not significantly add to the
precision of the model (p=0.11).
4.3.2 Sample Balance
Table 4.5 presents the baseline characteristics of children in kinship care by observed
outcomes and whether the child was in the control or the experimental group. Only one
child in each sibling group was randomly chosen for this analysis, regardless of whether
the child was eligible for the guardianship subsidy.
As table 4.5 indicates, the control and experimental groups were similar in most
background characteristics. The only statistical difference concerns the race of the
children who did not exit during the life of the GAP study. In the control group, about
5% of the children who remained in out-of-home care were not of African-American
decent. None of the non-African-American children in the experimental group remained
in foster care.
Table 4. 5 Sample Balance without Propensity Score Matching
Exited Foster Care
Did Not Exit Foster Care
Control
Experimental Control Experimental
Group
Group
Group
Group
Gender
Female
48.7%
44.4%
56.6%
56.0%
Male
51.3%
55.6%
43.4%
44.0%
χ2
0.28
0.00
Race
African American
93.4%
97.5%
94.7%*
100.0%*
Other
6.6%
2.5%
5.3%*
0.0%*
2
χ
1.47
2.72
Education
Regular School
63.5%
55.6%
72.1%
78.7%
Special Education
0.0%
4.2%
4.4%
6.4%
59
Not in School
χ2
Entry reason
Neglect
Abandonment
Abuse
Other
More than one of the
above
2
χ
Sibling Groups
Yes
No
χ2
Special needs
Yes
No
2
χ
Caregiver at removal
At least one parent
present
No parent present
χ2
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the previous
episode
t
N
Exited Foster Care
Control
Experimental
Group
Group
36.5%
40.3%
3.11
67.7%
10.8%
4.6%
1.5%
15.4%
67.6%
13.5%
8.1%
1.4%
9.5%
Did Not Exit Foster Care
Control Experimental
Group
Group
23.5%
14.9%
1.41
70.0%
15.7%
7.1%
1.4%
5.7%
1.87
76.2%
9.5%
4.8%
0.0%
9.5%
2.26
45.9%
54.1%
47.0%
53.0%
0.01
58.2%
41.8%
43.2%
56.8%
2.40
13.1%
86.9%
7.6%
92.4%
1.06
14.9%
85.1%
13.6%
86.4%
0.04
90.2%
80.9%
83.6%
82.1%
9.8%
16.4%
36.26
19.1%
2.21
5.82
-0.06
37.78
49.0
18.0%
0.04
6.60
-0.65
47.9
0.58
-0.40
1.73
0.40
0.19
0.59
76
-0.92
81
76
-0.99
50
5.79
6.11
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level
African American, Non-Hispanic
Other entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s
special needs.
Other special needs include child physical, mental and behavioral problems.
Sample mean

Tables 4.6 and 4.7 present sample characteristics by predicted outcome and
whether the child was eligible for the guardianship subsidy. The matched sample for
60
Table 4.6 was based on the propensity scores computed from logit model 1 in Table 4.4,
where the variable representing participation in the caregiver interview was not included,
whereas the matched sample for Table 4.7 was based on the propensity scored computed
from logit model 2, where the variable representing participation in caregiver interview
was included as a predictor.
As Table 4.6 shows, the propensity matching procedure produced a more
balanced sample than the unmatched sample. No statistically significant difference with
regard to background characteristics was detected between the control and experimental
groups, regardless of whether the child would exit out-of-home care in the absence of the
subsidy.
Table 4. 6 Sample Balance after Propensity Score Matching based on Logit Model 1
Would Exit
Would Not Exit
Control
Experimental Control Experimental
Group
Group
Group
Group
Gender
Female
47.5%
47.2%
51.6%
42.0%
Male
52.5%
52.8%
48.4%
58.0%
χ2
0.00
1.03
Race
African American
91.5%
97.2%
96.9%
100.0%
Other
8.5%
2.8%
3.1%
0.0%
χ2
1.23
1.59
Education
Regular School
64.3%
55.6%
70.5%
60.0%
Special Education
0.0%
2.8%
4.9%
6.0%
Not in School
35.7%
41.7%
24.6%
34.0%
χ2
2.03
1.36
Entry reason
Neglect
64.4%
61.1%
68.8%
72.0%
Abandonment
11.9%
16.7%
17.2%
10.0%
Abuse
5.1%
8.3%
6.3%
8.0%
Other
1.7%
0.0%
1.6%
0.0%
More than one of the
17.0%
13.9%
6.3%
10.0%
above
χ2
1.53
2.48
Sibling Groups
61
Yes
No
χ2
Special needs
Yes
No
χ2
Caregiver at removal
At least one parent
present
No parent present
χ2
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the previous
episode
t
Number of placement
changes in GAP
t
N
Would Exit
Control
Experimental
Group
Group
47.5%
52.8%
52.5%
47.2%
0.00
Would Not Exit
Control Experimental
Group
Group
60.9%
50.0%
39.1%
50.0%
1.36
13.6%
86.4%
8.3%
91.7%
0.60
14.1%
85.9%
12.0%
88.0%
0.10
89.3%
85.7%
86.5%
86.8%
10.7%
13.5%
37.89
14.3%
0.26
5.47
-0.22
38.05
53.94
13.2%
0.00
6.04
-0.04
54.22
0.46
-0.03
0.17
0.31
-0.04
0.30
0.46
1.53
0.47
1.02**
0.08
0.38**
5.28
6.02
-0.06
59
36
2.04
64
50
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level
 African American, Non-Hispanic
Other entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s
special needs.

Other special needs include child physical, mental and behavioral problems.

Sample mean
Matching based on the propensity scores produced by logit model 2 was less
successful than that of logit model 1. As Table 4.7 indicates, the experimental group
members and the control group members were different in two background
characteristics. Among those children who would exit out-of-home care in the absence of
the subsidy, the control group had a significantly higher average number of placement
changes in the previous episode than those in the experimental group. Additionally,
62
among children who would not exit out-of-home care in the absence of the subsidy, a
higher percentage of children in the control group belonged to sibling groups than that of
children in the experimental group. This result shows that the propensity score matching
method is highly sensitive to unobserved heterogeneity. Models with different sets of
predictors may produce vastly different matching results.
Table 4. 7 Sample Balance after Propensity Score Matching Based on Logit Model
2
Would Exit
Would Not Exit
Control
Experimental Control Experimental
Group
Group
Group
Group
Gender
Female
47.5%
59.5%
51.6%
44.0%
Male
52.5%
40.5%
48.4%
56.0%
2
χ
1.43
0.64
Race
African American
91.5%
97.6%
96.9%
100.0%
Other
8.5%
2.4%
3.1%
0.0%
χ2
1.63
1.59
Education
Regular School
64.3%
52.4%
70.5%
64.0%
Special Education
0.0%
4.8%
4.9%
6.0%
Not in School
35.7%
42.7%
24.6%
30.0%
χ2
3.56
0.53
Entry reason
Neglect
64.4%
69.1%
68.8%
74.0%
Abandonment
11.9%
16.7%
17.2%
8.0%
Abuse
5.1%
2.4%%
6.3%
6.0%
Other
1.7%
0.0%%
1.6%
0.0%
More than one of the
17.0%
11.9%
6.3%
12.0%
above
χ2
2.07
3.75
Sibling Groups
Yes
47.5%
54.8%
60.9%**
36.0%**
No
52.5%
45.2%
39.1%**
64.0%**
2
χ
0.52
6.98

Special needs
Yes
13.6%
7.1%
14.1%
16.0%
No
86.4%
92.9%
85.9%
84.0%
χ2
1.04
0.08
Caregiver at removal
63
At least one parent
present
No parent present
Would Exit
Control
Experimental
Group
Group
89.3%
81.6%
10.7%
37.89
18.4%
1.13
5.66
-0.47
40.04
0.46**
59
χ2
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the previous
episode
t
N
5.28
Would Not Exit
Control Experimental
Group
Group
86.5%
85.7%
13.5%
53.94
14.3%
0.01
6.18
-0.21
48.4
-0.42
0.12**
0.31
0.92
0.32
2.01
42
64
-0.05
50
6.02
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level
African American, Non-Hispanic

Other entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s
special needs.
Other special needs include child physical, mental and behavioral problems.

Sample mean

4.3.3 Placement Stability
Compared with logit model one, logit model two, which included the variable
representing participation status in the caregiver interviews, did not significantly improve
model fit. Additionally, logit model two produced a less balanced sample. As such,
matched sample produced by logit model one was used in estimating the impact of the
subsidy on placement stability.
Number of placement changes between December 2000 (the time of the first
FACTS download) and April 2003 (the time of the last available FACTS download) was
computed for each child in the sample. Because some children exited from out-of-home
care faster than others, those who exited earlier were exposed to lower risks of placement
64
disruption. Exposure-adjusted mean number in placement change, therefore, was
computed separately for the experimental group and the control group on a yearly basis 10.
Table 4.8 presents the exposure adjusted mean number of placement changes by
permanency status. For the group of children who would exit out-of-home care in the
absence of the subsidy, members of the control group experienced slightly higher rates of
placement change than members of the experimental group. However this difference was
not statistically significant. Similar results were found for the group of children who
would remain in out-of-home care in the absence of the subsidy.
Table 4. 8 Exposure Adjusted Mean Number of Placement Change by Permanency
Status
Control Group
Experimental Group
t-statistic
Exit out-of-home care
0.70
0.51
0.45
Remain in out-of-home care
0.47
0.34
0.62
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
Table 4.9 presents the estimates from Poisson and negative binomial (NB) models on the
number of placement changes for children who would exit out-of-home care in the
absence of the subsidy. The first two columns of the tables present results from two
Poisson models while the last two columns present results of Models 3 and 4 are negative
binomial models. A log-likelihood ratio test indicates that the set of covariates in Poisson
model two significantly increases model precision over model one (p=0.001). Further,
Poisson model two satisfies the equi-dispersion assumption of the Poisson distribution.
Poisson model 2, therefore, is the most appropriate model. The availability of the
10
To adjust for exposure, number of change in placement was divided by the number of days that the child
was under observation after random assignment. The result was then scaled up to a yearly basis. Because
exposure was calculated by subtracting the last day of available information – the exit date if the child
exited from care during the study period, and the date of the last download if the child did not exit –
exposure would be negative if the child exited after random assignment (May 2000) and the first FACTS
download (December 2000). These children (n=42) were excluded from the analyses.
65
guardianship subsidy was not found to have any impact on placement stability for this
group of children.
Table 4. 9 Coefficients obtained from Poisson and negative binomial models on
exposure adjusted numbers of placement changes for children who would exit from
out-of-home placements in the absence of the subsidy
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental group
0.03
-0.21
-0.02
-0.27
(0.31)
(0.36)
(0.42)
(0.40)
Male
----0.29
----0.27
(0.34)
(0.37)
Age at entry (in months)
---0.00
---0.00
(0.00)
(0.00)
Sibling groups
---0.44
---0.45
(0.39)
(0.41)
Neglected
---0.40
---0.47
(0.66)
(0.72)
Abandoned
---0.66
---0.68
(0.48)
(0.51)
Number of placements in
---0.38**
---0.34*
previous episode
(0.16)
(0.19)
Length of stay in care prior ---0.02***
---0.02***
to random assignment
(0.01)
(0.01)
African-American
----0.69
----0.74
(0.62)
(0.66)
Special needs
----0.03
---0.00
(0.67)
(0.70)
Constant
-6.49****
-7.77****
-6.47****
-7.80****
(0.19)
(1.01)
(0.25)
(1.06)
Log likelihood
-74.68
-60.82
-68.64
-60.70
N
63
63
63
63
Standard
errors are in parentheses.
Models one and two are Poisson models. Models three and four are negative binomial models.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level

*
Table 4.10 presents the coefficients obtained from Poisson and negative binomial
models on exposure adjusted number of placement changes for children who would not
exit from out-of-home placements in the absence of the subsidy. For both Poisson models
one and two, the equi-dispersion assumption was not satisfied. A negative binomial
model, therefore, was appropriate. Further, as a log likelihood ratio test indicated, the set
of covariates in NB model 2 did not add to model precision (p=0.28). As such, NB
66
model one was the most appropriate model. As the results of NB model one revealed, the
availability of subsidy did not have any statistically significant effects on placement
stability for this group of children.
Table 4. 10 Coefficients obtained from Poisson and negative binomial models on
exposure adjusted numbers of placement changes for children who would not exit
from out-of-home placement in the absence of the subsidy
Model 1
Model 2
Model 3
Model 4
Experimental group
-0.19
-0.25
-0.17
-0.15
(0.27)
(0.28)
(0.36)
(0.36)
Male
---0.22
---0.23
(0.23)
(0.32)
Age at entry (in months)
---0.01***
---0.01**
(0.00)
(0.00)
Sibling groups
---0.48*
---0.59
(0.26)
(0.37)
Neglected
---0.47
---0.60
(0.45)
(0.65)
Abandoned
---0.30
---0.46
(0.66)
(0.47)
Number of placement in
----0.27
----0.21
previous episode
(0.19)
(0.25)
Length of stay in care prior ---0.00
---0.00
to random assignment
(0.00)
(0.01)
African-American
---0.14
----0.32
(1.05)
(1.28)
Special needs
---0.01
---0.24
(0.33)
(0.48)
Constant
-6.70****
-8.71****
-6.70****
-9.24****
(0.12)
(1.28)
(0.18)
(1.68)
Log likelihood
-139.42
-128.76
-119.66
-114.23
N
96
96
96
96
Standard
errors are in parentheses.
Models one and two are Poisson models. Models three and four are negative binomial models.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level

4.4 Subgroup Analysis: Guardianship, Reunification, Restricted Foster Care, or
Kinship Care
The previous section analyzed separately the impact of the availability of the subsidy on
placement stability for children who would exit out-of-home care and those who would
stay in the absence of the subsidy. The sample was divided into two sub-groups: those
67
who exit care in the absence of the subsidy and those who would remain. However,
intra-group heterogeneity might still exit. Specifically, among children who exited outof-home care, those who were returned home might have a different profile from those
who exited via guardianship. Among children who remained in care, those who
remained in kinship care for which the caregiver continued to receive subsidy at the
TANF rate might be different from children whose caregivers eventually became licensed
and received the higher foster care board payment. The following section, therefore,
discusses the subsidy effects on the following four groups:
1. Those who would remain in kinship care in the absence of the subsidy;
2. Those who would move to restricted foster care in the absence of the subsidy;
3. Those who would exit relative foster care via guardianship in the absence of the
subsidy; and
4. Those who would be reunified with their parents in the absence of the subsidy.
4.4.1 Predicting Child Outcomes using a Multinomial Logit Model
A multinomial logit model was estimated using data available for the control group
members in this sample who were placed in unlicensed formal kinship care at the time of
random assignment. The outcome variable was a categorical variable representing the
permanency outcomes of these children. Four types of permanency outcomes were
examined: exiting via guardianship, exiting via reunification, moving to RFC, and
remaining in kinship care. Control group members who exited from out-of-home care
through adoption or other reasons, accounting for approximately 7% of the control group,
were not included in the model.
68
Table 4.11 presents the coefficients obtained from two multinomial logit models
with children who exited via guardianship as the reference group. Model one included
such explanatory variables as child gender, age at entry, length of stay, abandoned by
birth parents, number of placements in previous episode, and whether the child belonged
to a sibling group. Child race was excluded because over 95% of the children in the
sample were African-American. The lack of variation does not allow the effects of race to
be estimated. Further, preliminary testing indicated that the lack of variation regarding
whether the child was placed in out-of-home care due to neglect makes it impossible to
compute standard errors of the estimates of the parameter. This variable, therefore, was
not included in the model. The presence of any special needs was excluded from the
models for the same reason. Model 2 added a variable indicating whether the caregiver
completed a caregiver interview. Coefficients were estimated using the statistical
software STATA versions 8 and 9. The Huber/White robust estimator of variance was
used to compute the standard errors of the coefficient estimates to adjust for the
clustering of children cared for by the same caregiver.
Table 4. 11 Coefficients from the multinomial logit models on permanency
outcomes
Multinomial
Multinomial Logit
Logit Model 1
Model 2
Reunification
Male
-0.42
-0.61
Age at entry (in months)
0.01*
0.01*
Length of stay in care prior to -0.02
-0.01
random assignment (in months)
Abandoned by parents
0.34
0.63
Number of placements in the -0.05
0.34
previous episode
Having sibling groups
0.48
0.82
Participated in caregiver interview ----43.11***
Constant
-1.07
-1.58
Moved to RFC
69
Male
Age at entry (in months)
Length of time in care prior to
random assignment
Abandoned by parents
Number of placements in the
previous episode
Having sibling groups
Participated in caregiver interview
Constant
Remained in Kinship Care
Male
Age at entry (in months)
Length of time in care prior to
random assignment
Abandoned by parents
Number of placements in the
previous episode
Having sibling groups
Participated in caregiver interview
Constant
N=
Model Significance p<=
Pseudo R2 =
Log Likelihood =

*
-0.25
0.01*
0.02**
-0.25
0.01*
0.02**
-0.44
-0.13
-0.44
-0.11*
0.09
----1.44*
0.10
0.04
-1.49*
0.14
0.01*
0.02*
0.16
0.01
0.02
-1.42**
-0.25
-1.45**
-0.25
0.59
----1.52*
164
0.02
0.08
-202.85
0.50*
0.71
-1.73**
164
0.001
0.12
-193.77
The reference group is “exited to guardianship”
Significant at .01 level.; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level
As shown in Table 4.11, among the child demographic characteristics, gender did
not have any effect on permanency outcomes. Age at entry, on the other hand, had a
marginal effect on permanency outcome. Children who exited from out-of-home care via
guardianship tended to be younger when they entered into care than children who were
returned home, moved to RFC, or remained in kinship care. For example, everything else
equal, one month older in the age at entry decreased the odds of exiting through
guardianship rather than remaining in kinship care by approximately 1 percentage point 11.
The effects of age at entry on exiting through guardianship versus reunification or
moving to RFC were similar. One month older in age at entry decreased the odds of
11
e0.01=1.01. Unit change in odds in the following paragraphs were computed in similar ways.
70
exiting through guardianship versus reunification and moving to RFC by 1 percentage
point.
Among the case history variables, length of stay in care prior to random
assignment was the strongest predictor of permanency outcomes. Longer length of stay
was associated with decreased chance of achieving permanency. In model one, when a
child stayed in care for one more month prior to random assignment, his/her odds of
exiting through guardianship versus moving to RFC and remaining in kinship care
reduced by approximately 1 percentage point and 2 percentage points, respectively.
Furthermore, the odds of reunification versus moving to RFC and remaining in kinship
care reduced by 2 percentage points and 3 percentage points, respectively, with every one
month increase in the length of stay in care.12 With the exception of the odds of achieving
guardianship versus remaining in kinship care, the same results were found in model 2
after controlling for whether the caregiver completed the caregiver interview, although
the magnitude of coefficients were slightly smaller.
Having entered out-of-home care because of abandonment was found to be
predictive of exiting through guardianship or reunification versus remaining in kinship
care. Children who entered out-of-home care because of abandonment were almost five
times less likely than other children to be reunified versus remaining in kinship care
according to model one and seven times less likely according to model two. Additionally,
abandoned children’s odds of exiting through guardianship versus remaining in kinship
care were more than four times as likely as other children in both models one and two.
12
The effects of any variable on the logit of any two outcomes can be calculated by comparing the
coefficients in Table 4.4. The standard errors can be calculated through the variance-covariance matrix.
See Long (1997) for more details. The variance-covariance matrices of the two models are included in the
appendix.
71
This may be explained by the possibility that the caregiver is more willing to pursue
guardianship when they believe that the child’s chance of returning home is slim, as with
children who were abandoned by their parents.
There was some evidence showing that the number of placements in the previous
episode was related to reunification. When participation in caregiver interview was
controlled, one more placement during the previous episode led to an increase in the odds
of reunification versus moving to RFC or remaining in kinship care by 79 percentage
points and 56 percentage points, respectively. Having a sibling group was not found to be
predictive of permanency outcomes in either multinomial logit model one or model two.
Participation in the caregiver interview was predictive of reunification versus any other
permanency outcome. Children in the care of the caregivers who completed the surveys
were substantially less likely to be returned home. This may be due to the fact that
caregiver interview was conducted well after the inception of the project, and the research
team generally did not interview caregivers who did not care for the participating
children anymore. A proportion of the children may have already returned home when
the caregiver interviews started. Additionally, caregivers who expected that the child
would return home might be less willing to participate in the interviews then caregivers
who expected that the child would stay in their care for the foreseeable future.
Overall, model two with the variable representing participation in the caregiver
interview significantly improved model fit. A log likelihood ratio test indicated that the
null-hypothesis that the two models were equivalent can be rejected (χ2 = 18.16, p<0.001).
As such, multinomial logit model two is the more appropriate model predicting child
outcomes.
72
4.4.2 Propensity Score Matching
Children in the experimental group were matched to the children in the control group
based on predicted probabilities obtained through the multinomial logit models discussed
in the previous section. Because the variable representing caregiver participation in the
caregiver interviews significantly improved model fit, this section will present the results
based on multinomial logit model two. Results based on model one are briefly discussed
in section 4.5.
To address the interdependency issue of children from the same sibling group,
one member of each sibling group in the control group was randomly selected. Only
selected control group members were entered into the propensity score matching process.
This led to significant attrition of the sample, with approximately 32 percent of the
control group members (n=71) dropping out of the sample. Another 33 children in the
control group were not included in propensity score matching due to missing data.
Further, a total of 14 children in the control group exited from care for reasons other than
guardianship or reunification13. Among these 14 children, two were siblings or had
missing data and therefore were already excluded. The remaining 12 children were
further excluded from the analyses. As such, a total of 107 control group members
remained in the sample. Among these children, 33 exited through guardianship, 14 exited
through reunification, 27 moved to RFC, and 33 remained in kinship care at the end of
the study.
For the experimental group, propensity scores were not able to be computed for a
total of 43 children due to missing data. The pool of experimental children eligible for
13
These reasons included adoption, independence, ran away, agency custody not appropriate, and other
reasons.
73
matching contained 158 children, of which 78 children were matched to 107 children in
the control group based on their propensity scores. Among these 78 children, 11 children
were matched to control cases in two subgroups, and 2 were matched to control cases in
three subgroups.
Tables 4.12 and 4.13 compare the balance of matched and unmatched samples.
Table 4.12 presents the comparison of the control group and the experimental group by
actual permanency status. For both the control and experimental groups, only one
member of a sibling group was randomly selected and included in the table. Table 4.13
presents the comparison of background characteristics between the control group
members and the matched experimental group.
74
Table 4. 12 Sample balance without propensity score matching
Exited through
Exited through
Guardianship
Reunification
Control
Experimental Control Experimental
Group
Group
Group
Group
Gender
Female
44.4%
47.8%
44.4%
62.5%
Male
55.6%
52.2%
55.6%
37.5%
2
χ
0.11
0.07
Race
African American
91.1%
98.5%
100.0%
87.5%
Other
8.9%
1.5%
0.0%
12.5%
*
2
3.31
χ
2.34
Education
Regular School
62.2%
56.7%
71.4%
80.0%
Special Education
0.0%
5.0%
0.0%
0.0%
Not in School
37.8%
38.3%
28.6%
20.0%
χ2
1.97
0.14
Entry reason
Neglect
68.4%
65.6%
60.0%
66.7%
Abandonment
15.8%
14.8%
0.0%
0.0%
Abuse
7.9%
8.2%
6.7%
16.7%
Other
0.0%
1.6%
0.0%
0.0%
More than one of the
7.9%
9.8%
33.3%
16.7%
above
χ2
0.77
0.90
Sibling Group
Yes
48.6%
46.4%
57.1%
60.0%
No
51.4%
53.6%
42.9%
40.0%
χ2
0.04
0.01
Special needs
Moved to Restricted
Remained in Kinship
Foster Care
Care
Control Experimental Control Experimental
Group
Group
Group
Group
59.4%
40.6%
48.2%
51.9%
0.74
51.2%
48.8%
47.8%
52.2%
0.07
96.8%
3.2%
100.0%
0.0%
0.88
95.4%
4.7%
100.0%
0.0%
1.10
65.5%
0.0%
34.5%
88.0%
12.0%
0.0%
12.99***
72.2%
8.3%
19.4%
63.6%
0.0%
36.4%
3.49
58.6%
27.6%
6.9%
3.5%
3.5%
59.1%
18.2%
9.1%
0.0%
13.6%
79.0%
5.3%
7.9%
0.0%
7.9%
95.2%
0.0%
0.0%
0.0%
4.8%
2.96
50.0%
50.0%
31.8%
68.2%
1.67
3.38
63.9%
36.1%
59.1%
40.9%
0.13
75
Yes
No
χ2
Caregiver at removal
At least one parent
Relatives or friends
2
χ
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the previous
episode
t
N
Exited through
Guardianship
Control
Experimental
Group
Group
11.4%
7.14%
88.6%
92.9%
0.43
94.3%
5.7%
37.1
81.0%
19.0%
3.18*
6.3
-1.45
36.8
1.3
0.07
2.0
45
-0.37
67
5.1
Exited through
Reunification
Control Experimental
Group
Group
7.1%
40.0%
92.9%
60.0%
2.99*
93.3%
6.7%
31.9
83.3%
16.7%
0.50
7.3
-0.36
41.8
0.7
-1.08
0.4
18
0.71
8
6.7
Moved to Restricted
Remained in Kinship
Foster Care
Care
Control Experimental Control Experimental
Group
Group
Group
Group
7.1%
13.6%
19.4%
13.6%
92.9%
86.4%
80.6%
86.4%
0.57
0.32
84.0%
16.0%
45.7
73.9%
26.1%
0.74
7.6
-1.51
40.9
0.5
0.64
0.6
32
0.31
27
6.0
77.1%
22.9%
48.5
100.0%
0.0%
4.59**
5.8
-0.01
50.6
0.5
-0.32
0.6
43
-0.55
23
5.8

African American, Non-Hispanic.

Other entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s special needs.

Other special needs include child physical, mental, and behavioral problems.
Sample mean.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level.
76
Table 4. 13 Sample Balance after Propensity Score Matching (Based on Multinomial Logit Model 2)
Exit through Guardianship
Exit through
Move to Restricted
Remain in Kinship
Reunification
Foster Care
Care
Control Experimental
Control Experimental Control Experimental Control Experimental
Group
Group
Group
Group
Group
Group Group
Group
Gender
Female
39.4%
55.6%
42.9%
45.5%
51.9%
68.2%
45.5%
38.7%
Male
60.6%
44.4%
30.8%
27.3%
48.2%
31.8%
54.6%
61.3%
2
χ
1.56
0.02
1.34
0.30
Race
African American
87.9%
100.0%
100.0%
100.0%
92.6%
100.0% 100.0%
100.0%
Other
12.1%
0.0%
0.0%
0.0%
7.4%
0.0%
0.0%
0.0%
2
χ
3.51*
N/A
1.70
N/A
Education
Regular School
64.5%
51.9%
69.2%
72.7%
61.5%
63.6%
71.0%
71.0%
Special Education
0.0%
0.0%
0.0%
0.0%
0.0%
4.6%
9.7%
0.0%
Not in School
35.5%
48.2%
30.8%
27.3%
38.5%
31.8%
19.4%
29.0%
χ2
0.95
0.04
1.34
3.60
Entry reason
Neglect
63.6%
85.2%
57.1%
45.5%
55.6%
63.6%
78.8%
80.7%
Abandonment
18.2%
11.1%
0.0%
27.3%
29.6%
22.7%
6.1%
12.9%
Abuse
9.1%
0.0%
7.1%
18.2%
7.4%
4.6%
6.1%
3.2%
Other
0.0%
3.7%
0.0%
0.0%
3.7%
0.0%
0.0%
0.0%
More than one of the
9.1%
0.0%
35.7%
9.1%
3.7%
9.1%
9.1%
3.2%
above
χ2
7.57
6.42*
1.90
1.96
Sibling Groups
Yes
51.5%
44.4%
57.1%
45.5%
51.6%
63.6%
66.7%
51.6%
No
48.5%
55.6%
42.9%
54.6%
48.2%
36.4%
33.3%
48.4%
χ2
0.30
0.34
0.69
1.50

Special needs
77
Exit through Guardianship
Yes
No
χ2
Caregiver at removal
At least one parent
Relatives or friends
2
χ
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the previous
episode
t
N
Control Experimental
Group
Group
12.1%
3.7%
87.9%
96.3%
1.38
93.6%
6.5%
38.3
84.0%
16.0%
1.32
4.8
0.05
38.7
1.4
-0.06
0.4
33
0.90
27
4.8
Exit through
Reunification
Control Experimental
Group
Group
7.1%
0.0%
92.9%
100.0%
0.82
93.3%
6.7%
33.7
60.0%
40.0%
4.17**
8.0
-1.12
39.7
0.8
-0.79
0.6
14
0.29
11
6.1
Move to Restricted
Remain in Kinship
Foster Care
Care
Control Experimental Control Experimental
Group
Group Group
Group
7.41%
4.6%
18.2%
6.5%
92.6%
96.5%
81.8%
93.6%
0.17
2.01
86.4%
13.6%
50.3
79.0%
21.1%
0.40
7.2
-1.07
48.9
0.4
0.17
0.3
27
0.45
22
5.9
82.1%
17.9%
51.9
84.6%
15.4%
0.06
6.0
-0.25
50.5
0.2
0.21
0.3
33
-0.25
31
5.8

African American, Non-Hispanic

Other entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s special needs.

Other special needs include child physical, mental, and behavioral problems.

Sample mean.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
78
As Table 4.12 indicates, experimental group members who exited through
guardianship were slightly more likely to be African American than their peers in the
control group. Additionally, children in the control and experimental groups also differed
with regard to their caregivers at the time of removal. Members of the experimental
group who exited to gurdianship were less likely to be removed from their parents when
compared with children in the control group. Members of the experimental group who
remained in kinship care, conversely, were more likely to be removed from relatives and
friends than their peers in the control groups. Table 4.12 also shows that the
experimental group members who moved to RFC were significantly more likely to be in
special education and less likely to be out of school than members of the control group in
the same permanency category. This is a significant difference which may potentially
bias the results of comparative analyses if comparison is made between the experimental
and the control groups based on observed outcome for this group of children.
Propensity score marginally improved sample balance. As Table 4.13 shows, the
difference between the experimental and control groups with regard to education status
was eliminated through propensity score matching. However, the racial difference
remained for those who exited through guardianship remained. Additionally, the control
and experimental groups who exited through guardianship and remained in kinship care
no longer differed with respect to caregiver at removal. However, such difference was
found with children who would exit through reunification, with the control group
members more likely to be removed from the parents than similar experimental group
members. Furthermore, when identifying children in the experimental group who would
be reunified with their parents in the absence of the guardianship subsidy, the model
79
failed to match children with respect to entry reasons. Children in the control group were
more likely to have entered out-of-home care because of neglect, and less likely to have
entered because of abandonment or abuse. This difference was marginally significant. As
such, it may be necessary to control for entry reasons when placement stability was
examined for this group of children.
The drawback of the propensity score matching procedure is that it has resulted in
considerable sample attrition. For example, the guardianship subgroup dropped from 112
without matching to 60 with matching. This is because the matching procedure with
replacement entails that one child in the experimental group may be matched to more
than one control group member. Furthermore, propensity scores have to be computed
based on background characteristics, some of which may not be available for all children.
4.4.3 Placement Stability
Table 4.14 presents the mean yearly change in placement for the control group members
randomly selected from their sibling groups and the matched experimental group
members14. Children who would be reunified with their biological parents on average had
the highest rate of change in placement. Those who would exit through guardianship or
remain in kinship care, on the other hand, had relatively low numbers of change in
placement. For none of the four groups was there any statistically significant difference
between the control and the experimental group.
14
As discussed previously, because exposure was calculated by subtracting the last day of available
information – the exit date if the child exited from care during the study period, and the date of the last
download if the child did not exit – exposure would be negative if the child exited after random assignment
(May 2000) and the first FACTS download (December 2000). These children (n=36) were excluded from
the analyses.
80
Table 4. 14 Exposure-adjusted mean change in placements per year by predicted
outcomes and eligibility for the guardianship subsidy
Control Group
Experimental Group
t-statistic
Exit via guardianship
0.22
0.30
-0.50
Exit via reunification
1.72
0.60
1.06
Move to RFC
0.70
0.38
0.91
Remain in kinship care
0.24
0.29
-.038
Exposure-adjusted numbers of change in placement were subsequently examined in the
framework of Poisson models or, in the case that the equidispersion assumption of the
Poisson distribution was not satisfied, negative binomial models. Separate models with
different sets of explanatory variables were estimated. The first model was estimated with
eligibility for the subsidy as the only independent variable. A set of covariates were
included sequentially in the second model. These covariates were the following: child
race, age at entry, entry reasons, length of stay in care, number of placements in the
previous episode, and whether had a sibling group, and the presence of any special needs.
Because the propensity score matching procedure used all of these variables as predictors,
adding the covariate should not significantly change the direction or the size of the
coefficient for the main dependent variable: the eligibility for the subsidy. It may,
however, reduce the variance of the estimate if the covariates improve the precision of
the model significantly.
Table 4.15 presents the coefficients obtained from Poisson and negative binomial
(NB) models on exposure adjusted number of placement changes for children who would
exit through guardianship in the absence of the subsidy. The first two columns present
results from the two Poission models and the last two columns present the results from
the NB models. None of the explanatory variables in Poisson model two were
statistically significant. As expected, a log likelihood ratio test indicate that the set of
81
covariates did not contribute significantly to the model (χ2 = 6.88, p = 0.44). Further, a
comparison of Poisson model one and NB model one indicated that the equidispersion
assumption was not satisfied, and that NB model one was more appropriate. NB model
one was also found to be more appropriate than NB model two (χ2 = 3.71, p = 0.81). In
no model was eligibility for the subsidy or any of the covariates found to have any
significant effect on placement stability. This is consistent with our hypothesis that the
subsidy would have little effect on those children who exit via guardianship even in the
absence of the subsidy.
Table 4. 15 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
exit through guardianship in the absence of the subsidy
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
0.10
-0.07
0.14
-0.06
(0.45)
(0.54)
(0.55)
(0.56)
Male
---0.05
---0.05
(0.56)
(0.58)
Age at entry (in months)
---0.01
---0.01
(0.01)
(0.01)
Sibling Groups
---0.35
---0.34
(0.59)
(0.61)
Neglect
---1.06
---1.07
(1.35)
(1.37)
Abandonment
----0.09
----0.03
(1.03)
(1.07)
Number of placement in
----0.32
----0.30
previous episode
(0.39)
(0.41)
Length of stay in care prior ----0.00
----0.00
to random assignment
(0.01)
(0.01)
Participated in caregiver ---0.05
---0.63
interview
(0.56)
(0.65)
Constant
-6.98****
-8.35****
-7.03****
-8.33****
(0.33)
(1.52)
(0.40)
(1.55)
Log likelihood
-35.25
-31.08
-33.67
-31.07
N
43
43
43
43
Standard
*
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
82
Table 4.16 presents the coefficients obtained from the Poisson and negative
binomial regression models on number of placement changes for children who would exit
through reunification in the absence of the subsidy. Results from the Poisson regression
models and the negative binomial regression models were almost identical. The only
covariate that appeared to have a significant effect was sibling group. Those having
sibling groups had an average number of placement changes more than six times higher
than those without a sibling group. This difference was marginally significant (p=0.08). It
may also reflect the difficulties faced by the child welfare agency in its efforts to place
sibling groups together.
Although sibling group affected a child’s chance of having a stable placement, the
set of covariates as a whole did not. After the specification tests, it was determined that
Poisson model one, with only the variable representing the eligibility of the guardianship
subsidy as the independent variable, was the most appropriate model. Being eligible was
not found to be related to the number of placement changes that children may have, if
they would be reunified with their parents even in the absence of the subsidy.
Table 4. 16 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placements for children who would exit
through reunification in the absence of the subsidy
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-0.51
0.54
-0.51
0.54
(0.47)
(1.23)
(0.47)
(1.23)
Male
---0.45
---0.45
(0.85)
(0.85)
Age at entry (in months)
----0.00
----0.00
(0.01)
(0.01)
Sibling Groups
---2.03*
---2.03*
(1.17)
(1.17)
Neglect
----0.80
----0.80
(1.42)
(1.42)
Abandonment
----0.80
----0.80
83
Poisson
Model 1
Number of placement in
previous episode
Length of stay in care prior
to random assignment
Participating in caregiver
interview
Constant
Log likelihood
N
NB Model
1
----
Poisson
Model 2
(1.06)
0.12
(0.27)
-0.01
(0.03)
----
----
NB Model
2
(1.06)
0.12
(0.27)
-0.01
(0.03)
----
-5.98****
(0.35)
-22.67
19
-6.51****
(1.40)
-19.74
19
-5.98****
(0.35)
-22.67
19
-6.51****
(1.40)
-31.51
19
-------
-------
Standard
errors are in parentheses.
“Participation in caregiver interview” dropped out of the model because none of the caregivers of children in this group participated
in caregiver interview.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.

Table 4.17 presents the results obtained through Poisson and negative binomial
regressions for children who would move to RFC in the absence of the guardianship
subsidy. Comparison of the Poisson models with the NB models indicated that the
equidispersion assumption was not satisfied in Poisson, and NB models should be used
instead. Further, the set of covariates in NB model two did not contribute to model
precision. As such, NB model one with the eligibility of the subsidy as the sole
explanatory variable (model three) was the most appropriate model. As NB model one
indicates, being eligible for the guardianship subsidy did not affect placement stability for
children who would moved to RFC in the absence of the subsidy. None of the covariates
had any effect on placement stability for this group of children.
Table 4. 17 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placements for children who would move
to Restricted Foster Care in the absence of the subsidy
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-0.46
-0.83**
-0.46
-0.56
(0.34)
(0.37)
(0.50)
(0.49)
Male
---0.08
----0.11
(0.29)
(0.44)
Age at entry (in months)
---0.01***
---0.01*
84
Poisson
Model 1
Sibling Groups
----
Neglect
----
Abandonment
----
Number of placement in
previous episode
Length of stay in care prior
to random assignment
Participating in caregiver
interview
Constant
----
Log likelihood
N
-------6.26****
(0.15)
-79.22
42
Poisson
Model 2
(0.00)
0.51*
(0.32)
0.30
(0.51)
-0.27
(0.43)
-0.38
(0.24)
-0.00
(0.01)
0.09
(0.36)
-7.51****
(0.99)
-67.35
42
NB Model
1
-------------------6.26****
(0.26)
-62.86
42
NB Model
2
(0.01)
0.20
(0.47)
0.22
(0.75)
-0.02
(0.61)
-0.31
(0.32)
-0.00
(0.01)
0.30
(0.52)
-7.30****
(1.35)
-59.13
42

*
Standard errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
Table 4.18 presents the results of various models for children who would remain
in kinship care in the absence of the subsidy. Using the specification tests discussed
previously, model three, the NB model one with the eligibility as the sole independent
variable, was found to be the most appropriate model. No statistically significant effect of
the eligibility for the subsidy was found, nor was any covariate found to have any effect
on placement stability.
Table 4. 18 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placements for children who would remain
in kinship care in the absence of the subsidy
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
0.32**
0.32**
0.31
0.35
(0.39)
(0.43)
(0.49)
(0.52)
Male
---0.47
---0.44
(0.46)
(0.56)
Age at entry (in months)
----0.00
----0.00
(0.01)
(0.01)
85
Sibling Groups
Poisson
Model 1
----
Neglect
----
Abandonment
----
Number of placement in
previous episode
Length of stay in care prior
to random assignment
Participating in caregiver
interview
Constant
----
Log likelihood
N
-------7.32**
(0.24)
-52.87
52
Poisson
Model 2
0.40
(0.44)
-0.69
(0.66)
-0.58
(0.82)
-0.44
(0.48)
-0.01
(0.01)
0.42
(0.46)
-6.72****
(1.12)
-49.54
52
NB Model
1
-------------------7.32****
(0.29)
-50.16
52
NB Model
2
0.42
(0.54)
-0.61
(0.82)
-0.60
(0.96)
-0.41
(0.51)
-0.01
(0.01)
0.44
(0.56)
-6.85***
(1.34)
-47.98
52

*
Standard errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
4.4.4 Estimates of the Effects of the Eligibility of the Guardianship Subsidy Based
on Multinomial Logit Model 1
Estimates of the effects of guardianship subsidy on placement stability were also obtained
with a experimental group selected based on the multinomial logit model where
participation in caregiver interview was not considered (multinomial logit model one).
Results were similar to those based on multinomial logit model two with one exception.
For the subgroup of children who would move to restricted foster care in the absence of
the subsidy, members of the experimental group experienced placement disruption 75%
less frequently than their peers in the control group15. However, because multinomial
logit model one was considered less accurate than multinomial logit model two, the
results based on the former model may be less reliable than those discussed in the
previous section.
15
Coefficient=-1.38; Standard error=0.54. Percent increase is computed as follows: e -1..38 *100%=25%.
See the appendix for table of coefficients and standard errors.
86
4.5 Sensitivity Analysis – African-American Children Only
Because the overwhelming majority of the sample was African-American
children, the effect of race was not able to be assessed and controlled for due to the lack
of variation. However, existing literature has repeatedly pointed out that AfricanAmerican children are less likely than Caucasian children to achieve permanency in a
timely manner (Barth et al., 1994; Courtney & Needell, 1997; Landsverk et al., 1996).
The sensitivity analysis attempted to address the effects of race by excluding nonAfrican-American children from the analyses. Only African-American children were
included in the propensity score matching process and the estimation of the effects of the
subsidy. Propensity scores were computed separately based on multinomial logit models
without and with the variable representing whether the caregiver participated in the
caregiver interview (multinomial logit model one and multinomial logit model two
respectively) . A log likelihood ratio test indicated that the multinomial logit model with
the caregiver interview participation was superior in predicting permanency status.
Moreover, the experimental group members selected based on the results of multinomial
logit two were more similar to those selected based on multinomial logit model one.
Tables of coefficients obtained from the multinomial logit models and sample balance are
included in the Appendix.
Poisson models and negative binomial models were estimated to examine the
effects of the eligibility of the subsidy on placement stability. Results are presented in
the appendix. When propensity score matching were based on the multinomial logit
model with the variable representing caregiver participation in the interview as a
predictor, the results of the analysis with African-American children did not differ from
that with the overall sample. The availability for the subsidy did not have any significant
87
effects on placement stability for any of the four subgroups. When members of the
experimental group were matched to the control group based on the multinomial logit
model without the caregiver interview information, however, one major difference
emerged. The results indicated that among children who would exit through guardianship
in the absence of the subsidy, those who were eligible for the guardianship subsidy
experienced placement disruption 316% as frequently as those who were not eligible for
the subsidy (p=0.04)16.
This result was somewhat unexpected. Although one would not anticipate the
availability of the subsidy to increase placement stability for this group of children, it was
not expected to hurt them. One explanation is that the availability of the subsidy may
have selectively expedited some caregiver’s decision to assume guardianship. These
caregivers were more likely than not to have been in stable relationship with the children
in their care. As a result, more children in the experimental group who had been in stable
placements may have exited before the first FACTS download were retrieved in
December 2000 and, therefore, were excluded from the analyses. As the data indicated,
36% (n=8) of the experimental group members exited before December 2000, compared
to 26% (n=7) of the control group members. However, the difference in early exiting
appears to be marginal, and not likely to explain all the differences in terms of placement
changes between members of the control and experimental groups given the size of the
coefficient.
Overall, as previously discussed, the multinomial logit model which included
caregiver interview participation as a predictor explained more variation in permanency
16
Coefficient=1.15; Standard error=0.56. Percentage increase is computed as follows: e 1.15 *100%=316%.
See the appendix for table of coefficients and standard errors.
88
status than the model without caregiver interview information, and, therefore, rendered
more balanced samples of the control and experimental groups. The relationship between
the availability of the subsidy and placement stability estimated based on the latter model,
as such, may very well be spurious. The inconsistency in the results based on the two
slightly different models, however, indicate that the propensity score matching method
may be very sensitive to model specification and the existence of unobserved
heterogeneity.
89
Chapter 5. Discussions and Policy Implications
This dissertation examined whether the availability of a guardianship subsidy has
differential effects on different groups of children initially placed with unlicensed kinship
caregivers. The sample was divided in two groups: children who would exit out-of-home
care and children who would remain in care in the absence of the subsidy. Further, a
four-group analysis was conducted, which examined the differential effects of the
availability of the subsidy on children who would:
1. remain in kinship care in the absence of the subsidy;
2. move to restricted foster care in the absence of the subsidy;
3. exit out-of-home care via guardianship in the absence of the subsidy; and
4. be reunified with their parents in the absence of the subsidy.
Propensity score matching methods were used to identify the subgroups, and
Poisson and negative binomial models were estimated to determine the differential
effects of the availability of the guardianship subsidy on placement stability for the four
subgroups of children.
Subgroups identified through the propensity score matching procedure marginally
improved sample balance when compared with subgroups identified based on observed
permanency status at the end of the GAP project. The matched sample in the two group
analysis eliminated differences between the experimental and control groups in race. In
the four-group analysis, propensity score matching eliminated differences in the
percentage of children in special education, and reduced the differences between the
experimental and control groups with respect to the caregiver at removal. These two
factors were likely to be correlated with the probability of achieving permanency within a
90
short timeframe. However, racial differences between the control and experimental
groups remained, and propensity scores failed to match the experimental group members
and the control groups with respect to entry reason.
Sensitivity analyses were conducted with different sets of predictors and with the
African Americans only sub-samples. When analyses of placement stability were based
on the predictive models that produced the more balanced samples, no subsidy effects on
placement stability was found for any subgroup. However, alternative predictive models
produced considerably different samples and results. More specifically, when the
variable presenting whether the caregiver participated in the caregiver interview was
omitted from the multinomial logit models, members of the experimental group who
would move to restricted foster care were found to have experienced placement
disruptions 75% less frequently than their peers in the control group. Further, for AfricanAmerican children who would exit through guardianship in the absence of the subsidy,
members of the experimental group experienced placement disruption 316% as
frequently as their peers in the control group. This indicates that although propensity
score matching has the potential to improve sample balance and reduce bias, estimates
obtained through propensity score matching is sensitive to the selection of covariates in
the predictive model, and the existence of unobserved heterogeneity. This is consistent
with previous studies using the propensity score methods, which found that the accuracy
of propensity score match varies with regard to the observables selected (Dehejia &
Wahba, 1999).
This dissertation revealed that, while the availability of the subsidy did not
enhance placement stability for children in kinship care, it did not harm any subgroup of
91
children, regardless of their potential permanency outcomes in the absence of the subsidy.
Additionally, previous evaluations indicated that the subsidy accelerated children’s pace
leaving out-of-home care (Mandell, 2001; UMSSW, 2003). As such, one may conclude
that both the kinship care population and the society as a whole will benefit from a
universal subsidized guardianship for children in kinship care.
The current study did not find any subsidy effects on placement stability.
However, as discussed previously, propensity score matching is susceptible to sample
selection bias if insufficient amount of information exists to build the prediction model.
As demonstrated by this study, propensity score matching may be especially fragile given
a small sample size. To ascertain the findings of this study, it is desirable that the
evaluations of similar subsidized guardianship programs in other states include analyses
examining the subsidy effects on placements for similar subgroups, especially in places
where the out-of-care population is over-represented by African-American children, and
where the percentage of kinship caregivers who become licensed foster caregivers is
historically high.
As one considers the results of this study, several limitations must be noted. First,
an ideal study would follow an incoming cohort of children from entry into foster care to
exit. Due to the design of the GAP project, however, this study included a cohort of
children who had been in foster care for varied periods of time. As such, children who
spent long periods of time in foster care were over-represented in the sample. Second,
the children in this study were randomly assigned to the experimental and control groups
in May 2000. However, no information regarding their placement changes was available
until December 2000. Their placement experience during these seven months is unknown.
92
Finally, regarding the measure of placement stability, the data did not distinguish
desirable placement changes from undesirable placement changes. More nuanced data is
needed to assess the extent to which the guardianship subsidy affected children’s abilities
to stay in stable and family-like environments during their stay in out-of-home
placements.
93
References
Barber, J. G., Delfabbro, P. H. & Cooper, L.L. (2001). The predictors of unsuccessful
transition to foster care. Journal of Child Psychology & Psychiatry & Allied
Disciplines, 42(6), 785-791.
Barnard, J., Frangakis, C., Hill, J. & Rubin, D. B. (2001). School choice in NY City: A
Bayesian Analysis of an imperfect randomized experiment. In C. Gatsonis, R. E.
Kass, B. Carlin, A. Carriquiry, A. Gelman et al. (Eds.) Case Studies in Bayesian
Statistics Vol. V. New York: Springer-Verlag New York, Inc.
Barth, R. P., Courtney, M., Berrick, J. D. & Altert, V. (1994). From child abuse to
permanency planning. New York: Aldine De Bruyter.
Barth, R. P., Courtney, M. & Berry, M. (1994). Timing is everything: An analysis of the
time to adoption and legalization. Social Work Research, 18(3), 139-148.
Barth, R. P. & Blackwell, D. L. (1998). Death rates among California’s foster care and
former foster care populations. Children and Youth Services Review, 20(7), 577604.
Beeman, S., Kim, H., & Bullerdick, S. (2000). Factors affecting placement
of children in kinship and nonkinship foster care. Children and Youth Services
Review, 22(1), 36-54.
Berrick, J.D. (1997). Assessing quality of care in kinship and foster family care. Family
Relations, 46(3), 273-280.
Berrick, J. D. & Needell, B. (1999). Recent trends in kinship care: Public policy,
payments, and outcomes for children. In P. A. Curtis, G. Dale, Jr. & J. C. Kendall
(eds.), The foster care crisis: Translating research into policy and practice.
Lincoln: Nebraska Press.
Berrick, J. D., Barth, R. P. & Needell, B. (1994). A comparison of kinship foster homes
and foster family homes: Implications for kinship foster care as family
preservation. Children and Youth Services Review, 16(1/2), 33-63.
Berrick, J. D., Needell, B. & Barth, R. P. (1999). Kin as a family and child welfare
resource: The child welfare worker’s perspective. In R. L. Hegar & M.
Scannapieco (Eds.), Kinship foster care: Policy, practice and research. New
York: Oxford University Press.
Bloom, H. S., Michalopoulos, C. M., Hill C. J. & Lei, Y. (2002). Can nonexperimental
comparison group methods match the findings from a random assignment
94
evaluation of mandatory welfare-to-work programs? [Online]. Available:
www.mdrc.org [Retrieved: 2003, November, 26].
Bonecutter, F. J. & Gleeson, J. P. (1997). Broadening our view: Lessons from kinship
foster care. In G. Anderson, A. S. Ryan & B. R. Leashore (Eds.), The challenge
of permanency planning in a multicultural society. New York: Haworth Press.
Brooks, D., & Barth, R. P. (1998). Characteristics and outcomes of drug-exposed
and non drug-exposed children in kinship and non-relative foster care. Children
and Youth Services, 20(6), 475-501.
Buehler, C., Orme, J. G. & Post, J. (2000). The long-term correlates of family foster care.
Children and Youth Services Review, 22(8), 595-625.
Chipungu, S. P. (1991). A value-based policy framework. In J. E. Everett, S. P. Chipungu
& B. R. Leashore (Eds.), Child Welfare: An Afrocentric Perspective. New
Brunswick: Rutgers University Press.
Chipungu, S. P. (2004). The impact of child welfare policies on African American
families: A decade later. In J. Everett, S. Chipungu & B. Leashore (Eds.), Child
welfare revisited: An Africentric perspective. New Brunswick: Rutgers University
Press.
Chipungu, S. P., Everett, J. E. & Verduik, M. J. (1998). Children placed in foster care
with relatives: A multi-state study. Washington D. C.: Department of Health and
Human Services.
Christian, S & Ekman [Ekerman in text], L. (2000). A place to call home: Adoption and
guardianship for children in foster care. [Online]. Available:
http://www.ncsl.org/programs/pubs/bkfstrs.htm [Retrieved: 2003, August 5]
Cooper-Patrick, L, Gallo, J. J., Powe, N. R., Steinwachs, D. M., Eaton, W. W. & Ford, D.
E. (1999). Mental health service utilization by African Americans and whites:
The Baltimore epidemiologic Catchment area follow-up. Medical Care, 37(10),
1034-1045.
Courtney, M. E. (1999). Foster care and the costs of welfare reform. In P. A. Curtis, G.
Dale, Jr. & J. C. Kendall (Eds.), The foster care crisis: Translating research into
policy and practice. Lincoln: Nebraska Press.
Courtney, M. & Needell, B. (1997). Outcome of kinship care: Lessons from California.
In R. Barth, J. D. Berrick & N. Gilbert (Eds.), Child Welfare Research Review:
Vol 2. New York: Columbia University Press.
95
Danzy, J. & Jackson, S. M. (1997). Family preservation and support services: A missed
opportunity for kinship care. S. Jackson & S. Brissett-Chapman (Eds.), Serving
African American children. New Brunswick: Transaction Publishers.
Dehejia, R. H. & Wahba, S. (1999). Causal effects in nonexperimental studies:
Reevaluating the evaluation of training programs. Journal of the American
Statistical Association, 94(448), 1053-1062.
Dubowitz, H. & Feigelman, S. (1993). A profile of kinship care. Child Welfare, 72(2),
153-170.
Dubowitz, H., Feigelman, S & Zuravin, S. (1992). The physical health of children in
kinship care. American Journal of Diseases of Children, 146, 603-610
Eckenrode, J., Rowe, E., Laird, M. & Braithwaite, J. (1995). Mobility as a mediator of
the effects of child maltreatment on academic performance. Child Development,
66, 1130-1142.
Finch, S. J., Fanshel, D. & Grundy, J. F. (1986). Factors associated with discharge from
foster care. Social Work Research & Abstracts, 22(1), 10-18
Frangakis, C. & Rubin, D. B. (2002). Principal stratification in causal inference.
Biometrics, 58(1), 20-29.
Gebel, T. J. (1996). Kinship care and non-relative family foster care: A comparison of
caregiver attributes and attitudes. Child Welfare, 75(1):5-18.
Geen, R. (2003). Finding Permanent Homes for Foster Children: Issues Raised by
Kinship Care. [Online]. Available: www.urban.org. [Retrieved 2004, November
5].
Gibson, C. (2001). Privileging the participant: The importance of take-up rates in social
welfare evaluations. Chicago: Joint Center for Poverty Research.
Gleeson, J. P., O’Donnell, J. & Bonecutter, F. J. (1997). Understanding the complexity of
practice in kinship foster care. Child Welfare, 76(6), 801-826.
Gleeson, J. P. (1999a). Kinship care as a child welfare services: What do we really know.
In J. P. Gleeson & C. F. Hairston (Eds.), Kinship care: Improving practice
through research. Washington D.C.: Child Welfare League of America.
Gleeson, J. P. (1999b). Who decides? Predicting caseworkers’ adoption and guardianship
discussion with kinship caregiver. In J. P. Gleeson & C. F. Hairston (Eds.),
Kinship care: Improving practice through research. Washington D.C.: Child
Welfare League of America.
96
Grogan-Kaylor, A. (2000). Who goes into kinship care? The relationship of child and
family characteristics to placement into kinship foster care. Social Work
Research, 24(3), 143-155.
Hampton, R. L. (1991). Child abuse in the African American community. In J. E. Everett,
S. P. Chipungu & B. R. Leashore (Eds.), Child Welfare: An Afrocentric
Perspective. New Brunswick: Rutgers University Press.
Harden, B. J. (2004). Safety and stability for foster children: A developmental
perspective. Children, Families, and Foster Care, 14(1): 31-48.
Henry, J. (1999) Permanency outcomes in legal guardianship of abused/neglected
children. Families in Society. 80, 561-568.
Hill, R. (2003). Disproportionality of minorities in Child Welfare: Synthesis of research
findings. [Online]. Available: http://www.chs-wa.org/2_cfkreportsData.htm.
[Retrieved: 2004, November 5].
Hill, J. Waldfogel, J. & Brooks-Gunn, J. (2002). Differential effects of high-quality child
care. Journal of Policy Analysis and Management, 21, 601-627.
Hofferth, S. L. & Owens, T. J. (2001). Children at the Millennium: Where have we come
from, where are we going? Amsterdam: JAI.
Inglehart, A. (1994). Kinship foster care: Placement, services, and outcome issues.
Children and Youth Services Review, 16, 107-122.
James, S., Landsverk, J. & Slymen, D. J. (2004). Placement movement in out of home
care: Patterns and Predictors. Children and Youth Service Review, 26, 185-206.
Katz, L, Kling, J & Liebman, J. (2000). Moving to opportunity in Boston: Early results of
a randomized mobility experiment. Quarterly Journal of Economics, 116(2), 607654.
Kemp, S. P. & Bodony, J. M. (2000). Infants who stay in foster care: Child characteristics
and permanency outcomes of legally free children first placed as infants. Child
and Family Social Work, 5(2), 95-106.
Kurtz, P. D., Gaudin, P. Howing, P. & Wodarski, J. (1993). The consequences of
physical abuse and neglect on the school age child: Mediating factors. Children
and Youth Services Review, 15(2), 85-104.
Ladner, A. J. (2000). Children in Out-of-Home Placements. [Online]. Available:
http://www.urban.org [Retrieved: 2004, November 5].
97
LaLonde, R. (1986). Evaluating the econometric evaluations of training programs.
American Economic Review, 76, 604-620.
Landsverk, J., Davis, I., Ganger, W., Newton, R. & Johnson, I. (1996). Impact of child
psychosocial functioning on reunification from out-of-home placement. Children
and Youth Services Review, 18(4/5), 447-462.
Lawrence[Laurence in text]-Webb, C. (1997). African American children in the modern
child welfare system: A legacy of the Flemming Rule. In S. Jackson & S.
Brissett-Champman (Eds). Serving African American children. New Brunswick:
Transaction Publishers.
Leashore, B.R. (1984). Demystifying legal: An unexplored option for dependent children.
Journal of Family Law, 23, 391-400.
Leashore, B. R., McMurray, H. L. & Bailey, B. C. (1991). Reuniting and preserving
African American families. In J. E. Everett, S. P. Chipungu & B. R. Leashore
(eds), Child welfare: An Africentric perspective. New Brunswick: Rutgers
University Press.
Lehman, C., Liang, S., O’Dell, K., & Duryea, M. (2003). Evaluation of Oregon’s Title
IV-E waiver demonstration project: Final report. [Online]. Available:
http://www.cwp.pdx.edu/pdfs/Waiver%20Final%20Report%203-27-03.pdf
[Retrieved: 2003, August, 5].
Leslie, L. K., Landsverk, J., Ezzet-Lofstrom, R., Tschann, J. M., Slymen, D.J., Garland,
A. F. (2000). Children in foster care: Factors influencing outpatient mental
health service use. Child Abuse and Neglect, 24, 465-476.
Link, M. K. (1996). Permanency outcomes in kinship care: A study of children placed in
kinship care in Erie County, NY. Child Welfare, 75(5): 509-528.
Long, S. L. (1997). Regression models for categorical and limited dependent variables.
Thousand Oaks: Sage Publications.
Maluccio, A., Fein. E., & Davis, I. (1994). Family reunification: Research findings,
issues, and directions. Child Welfare, 73(5), 489-504
Mandell, M. B. (2001). The Effect of Subsidized Guardianship on Exits from Kinship
Care: Results from Maryland's Guardianship Assistance Demonstration Project.
The 23rd Annual Association of Public Policy Administration and Management (APPAM)
Conference, November 1-3, 2001, Washington D.C.
Mass, H. S., & Engler, R. E. (1959). Children in need of parents. New York: Columbia
University Press
98
McLean, B. & Thomas, R. (1996). Informal and formal kinship care populations: A study
in contrasts. Child Welfare, 75(5), 489-508.
Mcmiller, W. P. & Weisz, J. R. (1996). Help-seeking preceding mental health clinic
intake among African-American, Latino, and Caucasian youths. Journal of the
American Academy of Child & Adolescent Psychiatry, 35(8), 1086-1094.
McRoy, R. (1994). Attachment and racial identity issues: Implication for child placement
decision making. Journal of Multicultural Social Work, 3(3), 59-74.
McRoy, R. (2004). African American Adoptions. In J. Everett, S. Chipungu & B.
Leashore (Eds.), Child Welfare Revisited: An Africentric Perspective. New
Brunswick: Rutgers University Press.
Meyer, B. S. & Link, M. K. (1990). Kinship foster care: The double edge dilemma. New
York: Task Force on Permanency Planning for Foster Children.
Newton, R. R., Litrownik, A. J. & Landsverk, J. A. (2000). Children and youth in foster
care: Disentangling the relationship between problem behaviors and number of
placements. Child Abuse & Neglect, 24(10), 1363-1374.
Olsen, L. J. (1982). Predicting the permanency status of children in foster care. Social
Work Research and Abstracts, 18(1), 9-20.
Pecora, P.J., Le Prohn, N.S. & Nasuti, J.J. (1999). Role perceptions of kinship and other
foster parents in family foster care. In R. L. Hegar & M. Scannapieco (Eds.),
Kinship foster care: Policy, practice and research. New York: Oxford
University Press.
Proch, K. & Taber, M.A. (1985). Placement disruption: A review of research. Children
and Youth Services Review, 7, 309-320.
Rosenbaum, P. R. & Rubin, D. (1983). The central role of the propensity score in
observational studies for causal effects. Biometrika, 70(1), 41-55.
Rubin, D. B. (1997). Estimating causal effects from large data sets using propensity
scores. Annals of Internal Medicine, 127(8), 757-763.
Scannapieco, M (1999). Kinship care in the public child welfare system. In R. L. Hegar
& M. Scannapieco (Eds.), Kinship foster care: Policy, practice and research.
New York: Oxford University Press.
Scannapieco, M., Hegar, R. L. & McAlpine, C. (1997). Kinship care and foster care: A
comparison of characteristics and outcomes. Families in Society, 78(5), 480-488.
99
Smith, D. K., Stormshak, E., Chamberlain[Camberlain in text], P. & Whaley, R.B. (2001).
Placement disruption in treatment foster care. Journal of Emotional &
Behavioral Disorder, 9(3), 200-212.
Smith, J. (2004). Demography of African American families. In J. Everett, S. Chipungu
& B. Leashore (Eds.), Child Welfare Revisited: An Africentric Perspective. New
Brunswick: Rutgers University Press.
Schwartz, I. M. & Fishman, G. (1999). Kids raised by the government. Westport: Praeger.
Taber, M. A., & Proch, K. (1987). Placement stability for adolescents in foster care:
findings from a program experiment. Child Welfare, 66, 433-445.
Takas, M (1992). Kinship care: Developing a safe and effective framework for protective
placement of children with relatives. Zero to Three, 13(3), 12-17.
Terling-Watt, T. (2001). Permanency in kinship care: An exploration of disruption rates
and factors associated placement disruption. Children and Youth Services
Review, 23(20), 111-126.
Testa, M. F., Shook, K. L., Cohen, L. & Woods, M.G. (1996). Permanency planning
options for child in formal kinship care. Child Welfare, 75(5), 451-470.
Testa, M. F. (2001). Kinship care and permanency. Journal of Social Service Research,
28(1): 25-43.
Testa, M. F. (2002). Subsidized guardianship: Testing an idea whose time has finally
come. Social Work Research, 26(3), 145-158.
Thornton, J. L. (1991). Permanency planning for children in kinship care. Child Welfare,
70, 593-601.
TriWest Group (2004). New Mexico Title IV-E waiver evaluations: Evaluation Update.
[Online]. Available: http://www.cyfd.org. [Retrieved: 2006, April 22].
University of Maryland, School of Social Work (2003). Maryland subsidized
guardianship demonstration project evaluation final report. Unpublished.
U.S. Department of Education. (2002). Twenty-Fourth Annual Report to Congress on the
Implementation of the Individuals with Disabilities Education Act. [Online].
Available: http://www.ed.gov/about/reports/annual/osep/2002/index.html.
[Retrieved: 2005, October 3].
U.S. Department of Health and Human services, Administration for Children and
Families, Children’s Bureau. (2000). Report to the Congress on Kinship Foster
100
Care. [Online]. Available: http://www.acf.dhhs.gov. [Retrieved: 2000,
December 20].
U.S. Department of Health and Human Services. (2003a). Child welfare Outcomes 2000:
Annual Report. [Online]. Available:
http://www.acf.hhs.gov/programs/cb/publications/cwo.htm [Retrieved: 2004,
April 10].
U.S. Department of Health and Human Services. (2003b). AFCARS report (Preliminary
FY2001 estimates as of March 2003). [Online]. Available:
www.acf.dhhs.gov/programs/cb/publications/ afcars/report8.pdf [Retrieved:
2003, September 19].
U.S. Department of Health and Human Services. (1994). The national survey of current
and former foster parents. Washington D.C.: Author.
U.S. Department of Health and Human services, Administration for Children and
Families, Children’s Bureau (no date). Assisted guardianship/kinship
permanence. [Online]. Available:
http://www.acf.dhhs.gov/programs/cb/initiatives/cwwaiver/assisted.htm
[Retrieved: 2003, July 23].
U.S. Department of Health and Human Services. (2005). Synthesis of findings: Assisted
guardianship child welfare waiver demonstrations. [Online]. Available:
http://www.acf.hhs.gov/programs/cb/programs_fund/cwwaiver/agissue/index.ht
m. [Retrieved: 2006, April 21].
Usher, C., Randolph, K. & Gogan, H. (1999). Placement patterns in foster care. Social
Service Review, 73(1), 22-36.
Vogel, C.A. (1999). Using administrative database to examine factors affecting length of
stay in substitute care. Children and Youth Services Review, 21(8), 677-690.
Webster, D., Barth, R. & Needell, B. (2000). Placement stability for children in out-ofhome care: A longitudinal analysis. Child Welfare, 79(5), 614-632.
Wells, K. & Guo, S. (1999). Reunification and Reentry of Foster Children. Children and
Youth Services Review, 21 (4), 273-294.
Widom, C. (1991). The role of placement experiences in mediating the criminal
consequences of early childhood victimization. American Journal of
Orthopsychiatry, 61(2), 195-209.
Williams, C. C. (1991). Expanding the options in the quest for permanence. In J. E.
Everett, S. P. Chipungu & B. R. Leashore (Eds.), Child Welfare: An Africentric
Perspective. New Brunswick: Rutgers University Press.
101
Wulczyn, F., Kogan, J & Harden, B. J. (2003). Placement stability and movement
trajectories. Social Service Review, 77(2), 212-237.
Yoshikawa, H., Magnuson, K, Bos, J. M. & Hsueh, J. (2003). Effects of earningssupplement policies on adult economic and middle-childhood outcomes differ
for the “hardest to employ”. Child Development, 74(5), 1500-1521.
Zuravin, S., Benedict, M & Stallings, R. (1999). The adult functioning of former kinship
and non-relative foster care children. In R. L. Hegar & M. Scannapieco (Eds.),
Kinship foster care: Policy, practice and research. New York: Oxford
University Press.
102
Appendix: Additional Statistical Tables
103
Table A- 1 Sample Balance after Propensity Score Matching (Not considering participating in caregiver interview)
Exit through Guardianship
Exit through
Move to Restricted
Remain in Kinship
Reunification
Foster Care
Care
Control Experimental
Control Experimental Control Experimental Control Experimental
Group
Group
Group
Group
Group
Group
Group
Group
Gender
Female
39.4%
42.9%
50.0%
69.2
51.9%
60.9%
54.3%
48.4%
Male
60.6%
57.1%
50.0%
30.8%
48.2%
39.1%
45.7%
51.6%
2
χ
0.08
1.03
0.41
0.23
Race
African American
87.9%
100.0%
100.0%
100.0%
92.6%
100.0% 100.0%
100.0%
Other
12.1%
0.0%
0.0%
0.0%
7.4%
0.0%
0.0%
0.0%
χ2
3.63*
N/A
1.70
N/A
Education
Regular School
62.5%
50.0%
71.4%
76.9%
61.5%
78.3%
72.7%
61.3%
Special Education
3.1%
7.1%
0.0%
7.7%
3.9%
17.4%
9.1%
6.5%
Not in School
34.4%
42.9%
28.6%
15.4%
34.6%
4.4%
18.2%
32.3%
χ2
2.39
1.63
1.86
1.72
Entry reason
Neglect
60.6%
64.3%
57.1%
69.2%
55.6%
82.6%
80.0%
77.4%
Abandonment
18.2%
21.4%
0.0%
7.7%
29.6%
8.7%
8.6%
12.9%
Abuse
9.1%
3.6%
7.1%
7.7%
7.4%
4.4%
2.9%
3.2%
Other
0.0%
3.6%
0.0%
0.0%
3.7%
0.0%
0.0%
0.0%
More than one of
12.1%
7.1%
35.7%
15.4%
3.7%
4.4%
8.6%
6.5%
the above
χ2
2.38
2.31
5.11
0.41
Sibling Group
Yes
48.5%
60.7%
57.1%
38.5%
55.6%
56.5%
68.6%
67.8%
No
51.5%
39.3%
42.9%
61.5%
44.4%
43.5%
31.4%
32.3%
2
χ
0.52
0.94
0.00
0.01
104
Exit through Guardianship
Control Experimental
Group
Group
Special needs
Yes
No
2
χ
Caregiver at removal
At least one parent
Relatives or friends
2
χ
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the
previous episode
t
N
Exit through
Reunification
Control Experimental
Group
Group
Move to Restricted
Foster Care
Control Experimental
Group
Group
Remain in Kinship
Care
Control Experimental
Group
Group
19.2%
81.8%
7.1%
82.9%
1.62
0.0%
100.0%
15.4%
84.6%
2.33
7.4%
92.6%
4.4%
95.7%
0.21
17.1%
82.9%
9.7%
90.3%
0.78
92.6%
6.5%
92.9%
7.1%
50.9
94.1%
5.9%
0.14
5.9
0.11
50.3
82.8%
17.2%
33.7
83.4%
16.7%
0.57
7.9
-0.91
33.9
90.9%
9.1%
38.1
91.7%
8.3%
0.07
4.9
0.48
40.0
53.0
92.0%
8.0%
0.41
5.8
0.04
58.3
0.48
-0.33
0.57
0.8
-0.03
0.2
0.3
0.06
0.2
0.3
-0.74
0.3
33
-0.34
28
14
1.55
13
27
0.48
23
35
0.12
31
4.4
6.4
6.1
5.9

African American, Non-Hispanic
Other entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s special needs.

Other special needs include child physical, mental, and behavioral problems.
Sample mean.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
105
Table A- 2 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
exit through guardianship in the absence of the subsidy (Not considering
participation in caregiver interview)
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-0.12
0.01
-0.12
0.01
(0.50)
(0.54)
(0.50)
(0.54)
Male
---0.38
---0.38
(0.58)
(0.58)
Age at entry (in months)
----0.00
----0.00
(0.01)
(0.01)
Sibling Groups
----0.46
----0.46
(0.56)
(0.56)
Neglect
---0.73
---0.73
(0.71)
(0.71)
Abandonment
-------
-------
Number of placement in
---previous episode
Length of stay in care prior ---to random assignment
Constant
-6.90****
(0.33)
Log likelihood
-25.83
N
38
-0.41
(0.37)
-0.01
(0.01)
-7.09****
(0.86)
-24.38
38
-------6.90****
(0.33)
-25.83
38
-0.41
(0.37)
-0.01
(0.01)
-7.09****
(0.86)
-24.38
38
Standard
errors are in parentheses.
due to lack of variation.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
 Dropped
*
Table A- 3 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
exit through reunification in the absence of the subsidy (Not considering
participation in caregiver interview)
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-0.79
-0.46
-0.79
-0.46
(0.57)
(1.04)
(0.57)
(1.04)
Male
---0.08
---0.08
(1.08)
(1.08)
Age at entry (in months)
----0.00
----0.00
(0.01)
(0.01)
Sibling Groups
---1.30
---1.30
(1.22)
(1.22)
Neglect
----0.93
----0.93
(1.31)
(1.31)
Abandonment
----0.67
----0.67
106
Poisson
Model 1
Number of placement in
---previous episode
Length of stay in care prior ---to random assignment
Constant
-5.98****
(0.35)
Log likelihood
-17.28
N
17
Poisson
Model 2
(1.30)
-0.05
(0.39)
-0.01
(0.02)
-5.37****
(2.09)
-16.02
17
NB Model
1
-------5.98****
(0.35)
-17.28
17
NB Model
2
(1.30)
-0.05
(0.39)
-0.01
(0.02)
-5.37****
(2.09)
-16.02
17
Standard
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
*
Table A- 4 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
move to restricted foster care in the absence of the subsidy (Not considering
participation in caregiver interview)
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-1.18***
-1.64***
-1.20**
-1.38**
(0.44)
(0.47)
(0.54)
(0.54)
Male
---0.23
---0.36
(0.31)
(0.42)
Age at entry (in months)
---0.02****
---0.02***
(0.00)
(0.01)
Sibling Groups
---0.94***
---0.77*
(0.33)
(0.43)
Neglect
---0.15
----0.02
(0.55)
(0.69)
Abandonment
----0.58
----0.62
(0.49)
(0.62)
Number of placement in
----0.33
----0.24
previous episode
(0.30)
(0.35)
Length of stay in care prior ---0.01
---0.01
to random assignment
(0.01)
(0.01)
Constant
-6.23****
-9.08****
-6.23****
-8.42****
(0.15)
(1.09)
(0.24)
(1.29)
Log likelihood
-71.70
-55.55
-58.65
-52.52
N
42
42
42
42

*
Standard errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
107
Table A- 5 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
remain in kinship care in the absence of the subsidy (Not considering participating
in caregiver interview)
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
0.24
0.14
0.25
0.12
(0.41)
(0.44)
(0.52)
(0.56)
Male
---0.78
---0.77
(0.42)
(0.51)
Age at entry (in months)
---0.00
---0.00
(0.01)
(0.01)
Sibling Groups
---0.16
---0.22
(0.45)
(0.55)
Neglect
---0.84
---0.72
(1.19)
(1.30)
Abandonment
---0.10
---0.03
(0.73)
(0.85)
Number of placement in
----0.27
----0.27
previous episode
(0.31)
(0.37)
Length of stay in care prior ---0.00
---0.01
to random assignment
(0.01)
(0.01)
Constant
-7.43****
-9.11****
-7.43****
-9.06****
(0.24)
(1.66)
(0.30)
(1.89)
Log likelihood
-53.25
-50.44
-49.77
-47.88
N
56
56
56
56
Standard
*
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
Table A- 6 Coefficients obtained from a multinomial logit model on permanency
status (African-American children only) 
Multinomial
Multinomial
Logit Model 1
Logit Model 2
Reunification
Male
-0.37
-0.51
Age at entry (in months)
0.01*
0.01**
Length of stay in care prior to random -0.02
-0.01
assignment (in months)
Abandoned by parents
0.17
0.52
Number of placements in the previous -0.07
0.31
episode
Having sibling groups
0.41
0.71
Participating in caregiver interview
----41.66****
Constant
-0.79
-1.39
Moved to RFC
Male
-0.30
-0.30
108
Multinomial
Logit Model 1
Age at entry (in months)
0.01**
Length of time in care prior to random 0.02*
assignment
Abandoned by parents
-0.72
Number of placements in the previous -0.18
episode
Having sibling groups
0.16
Participation in caregiver interview
---Constant
-1.26
Remained in Kinship Care
Male
0.17
Age at entry (in months)
0.01**
Length of time in care prior to random 0.01
assignment
Abandoned by parents
-1.61**
Number of placements in the previous -0.28
episode
Having sibling groups
0.57
Participation in caregiver interview
---Constant
-1.24
157
N=
0.01
Model Significance p<=
2
0.09
Pseudo R =
-193.49
Log Likelihood =

*
Multinomial
Logit Model 2
0.01**
0.02*
-0.72
-0.14*
0.17
0.13
-1.34
0.16
0.01**
0.01
-1.65**
-0.28
0.47
0.74
-1.47
157
0.001
0.13
-184.47
The reference group is “exited to guardianship”
Significant at .01 level.; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level
109
Table A- 7 Sample Balance After Propensity Score Matching (African-American children only, not considering participation
in caregiver interviews)
Exit through Guardianship
Exit through
Move to Restricted
Remain in Kinship
Reunification
Foster Care
Care
Control Experimental
Control Experimental Control Experimental Control Experimental
Group
Group
Group
Group
Group
Group
Group
Group
Gender
Female
33.3%
45.5%
46.2%
36.4%
61.5%
66.7%
50.0%
36.7%
Male
66.7%
54.6%
53.9%
63.6%
38.5%
33.3%
50.0%
63.3%
2
χ
0.75
0.24
0.14
1.15
Education
Regular School
56.0%
36.4%
75.0%
54.6%
64.0%
83.3%
68.8%
70.0%
Special Education
0.0%
9.1%
0.0%
0.0%
0.0%
4.7%
9.4%
3.3%
Not in School
44.0%
54.6%
25.0%
45.5%
36.0%
12.5%
21.9%
26.7%
χ2
3.50
1.06
4.43
1.03
Entry reason
Neglect
63.0%
50.0%
61.5%
45.6%
61.5%
79.2%
76.5%
86.7%
Abandonment
14.8%
18.2%
0.0%
36.4%
26.9%
12.5%
8.8%
3.3%
Abuse
11.1%
9.1%
7.7%
9.1%
7.7%
0.0%
5.9%
0.0%
Other
0.0%
0.0%
0.0%
9.1%
3.9%
0.0%
0.0%
0.0%
More than one of
11.1%
22.7%
30.8%
0.0%
0.0%
8.3%
8.8%
10.0%
the above
χ2
1.49
9.59**
6.79
2.76
Sibling Group
Yes
48.2%
50.0%
53.9%
27.3%
57.7%
54.2%
67.7%
46.7%
No
51.9%
50.0%
46.2%
72.7%
42.3%
45.8%
32.4%
53.3%
χ2
0.02
1.73
0.06
2.88*

Special needs
Yes
14.8%
9.1%
7.7%
9.1%
7.7%
16.7%
17.7%
13.3%
No
85.2%
90.9%
92.3%
90.9%
92.3%
83.3%
82.4%
86.7%
110
Exit through Guardianship
χ2
Caregiver at removal
At least one parent
Relatives or friends
χ2
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the
previous episode
t
N
Control Experimental
Group
Group
0.37
96.0%
4.0%
39.9
76.2%
23.8%
3.95**
5.1
-1.02
41.2
1.7
-0.21
4.9
27
-0.77
22
4.1
Exit through
Reunification
Control Experimental
Group
Group
0.02
92.3%
7.7%
34.9
81.8%
18.2%
0.60
7.1
-0.42
31.5
0.9
0.46
0.5
13
0.79
11
6.4
Move to Restricted
Foster Care
Control Experimental
Group
Group
0.95
95.5%
4.6%
50.2
84.2%
15.8%
1.46
6.7
-0.26
50.8
0.4
-0.06
0.2
26
0.90
24
6.3
Remain in Kinship
Care
Control Experimental
Group
Group
0.23
82.8%
17.2%
51.7
92.6%
7.4%
1.24
6.3
-0.68
47.1
0.2
0.68
0.3
34
-0.32
30
5.6
Other
entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s special needs.
Other special needs include child physical, mental, and behavioral problems.
Sample mean.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
111
Table A- 8 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placements for children who would exit
through guardianship in the absence of the subsidy – African-American children
only
Model 1
Model 2
Model 3
Model 4
Experimental or Control
1.15**
0.95
1.18*
1.01
(0.56)
(0.61)
(0.61)
(0.68)
Male
----0.15
----0.21
(0.60)
(0.64)
Age at entry (in months)
---0.01
---0.01
(0.01)
(0.01)
Sibling Groups
----0.09
----0.09
(0.64)
(0.69)
Neglect
----0.05
----0.01
(0.71)
(0.78)
Abandonment
----0.44
----0.34
(0.61)
(0.68)
Number of placement in ----0.09
----0.09
previous episode
(0.13)
(0.15)
Length of stay in care prior ---0.00
---0.00
to random assignment
(0.01)
(0.02)
Constant
-7.54****
-7.53****
-7.56****
-7.44****
(0.50)
(1.17)
(0.52)
(1.21)
Log likelihood
-27.83
-26.86
-27.39
-26.71
N
34
34
34
34
Standard
errors are in parentheses.
Models one and two are Poisson models. Models three and four are negative binomial models.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.

Table A- 9 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
exit through reunification in the absence of the subsidy (Not considering
participation in caregiver interview) - African American children only
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-1.13
-1.65
-1.13
-1.65
(0.80)
(2.22)
(0.80)
(2.22)
Male
----0.24
----0.24
(1.56)
(1.56)
Age at entry (in months)
---0.01
---0.01
(0.02)
(0.02)
Sibling Groups
---1.61
---1.61
(1.58)
(1.58)
Neglect
----3.28
----3.28
(3.39)
(3.39)
Abandonment
---0.19
---0.19
(2.30)
(2.30)
112
Poisson
Model 1
----
Number of placement in
previous episode
Length of stay in care prior ---to random assignment
Constant
-6.02****
(0.38)
Log likelihood
-13.86
N
15
Poisson
Model 2
0.25
(0.55)
0.01
(0.04)
-3.96
(4.25)
-11.73
15
NB Model
1
-------6.01****
(0.38)
-13.86
15
NB Model
2
0.25
(0.55)
0.01
(0.04)
-3.96
(4.25)
-11.73
15
Standard
*
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
Table A- 10 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placements for children who would move
to restricted foster care in the absence of the subsidy – African-American children
only
Model 1
Model 2
Model 3
Model 4
Experimental or Control
-0.38
-0.64*
-0.45
-0.73
(0.31)
(0.36)
(0.50)
(0.53)
Male
---0.16
----0.07
(0.31)
(0.49)
Age at entry (in months)
---0.02****
---0.02***
(0.00)
(0.01)
Sibling Groups
----0.08
----0.11
(0.30)
(0.48)
Neglect
----0.46
----0.57
(0.48)
(0.75)
Abandonment
---0.56
---0.77
(0.44)
(0.65)
Number of placement in
---0.10
---0.01
previous episode
(0.15)
(0.28)
Length of stay in care prior
---0.02***
---0.02**
to random assignment
(0.01)
(0.01)
Constant
-6.24****
-8.62****
-6.24****
-8.96****
(0.15)
(0.88)
(0.29)
(1.45)
Log likelihood
-85.63
-70.39
-63.41
-56.90
N
43
43
43
43
Standard
errors are in parentheses.
Models one and two are Poisson models. Models three and four are negative binomial models.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.

113
Table A- 11 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
remain in kinship care in the absence of the subsidy (Not considering participation
in caregiver interview) - African-American children only
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
0.56
0.54
0.56
0.57
(0.36)
(0.41)
(0.41)
(0.45)
Male
---0.54
---0.54
(0.42)
(0.44)
Age at entry (in months)
---0.00
---0.00
(0.00)
(0.01)
Sibling Groups
---0.65
---0.65
(0.42)
(0.46)
Neglect
---0.73
---0.73
(1.13)
(1.17)
Abandonment
----0.02
----0.01
(0.67)
(0.71)
Number of placement in
----0.05
----0.04
previous episode
(0.27)
(0.29)
Length of stay in care prior ----0.00
----0.00
to random assignment
(0.01)
(0.01)
Constant
-7.35****
-8.94****
-7.35****
-8.96****
(0.24)
(1.47)
(0.41)
(1.54)
Log likelihood
-55.34
-52.30
-54.16
-46.78
N
56
56
56
56
Standard
*
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
114
Table A- 12 Sample Balance after Propensity Score Matching (African-American children only, considering participation in
caregiver interviews)
Exit through Guardianship
Exit through
Move to Restricted
Remain in Kinship
Reunification
Foster Care
Care
Control Experimental
Control Experimental Control Experimental Control Experimental
Group
Group
Group
Group
Group
Group
Group
Group
Gender
Female
33.3%
39.1%
46.2%
76.9%
61.5%
57.7%
50.0%
62.1%
Male
66.7%
60.8%
53.9%
23.1%
38.5%
42.3%
50.0%
37.9%
2
χ
0.18
2.60
0.08
0.92
Education
Regular School
56.0%
52.2%
75.0%
61.5%
64.0%
42.3%
68.8%
51.7%
Special Education
0.0%
8.7%
0.0%
0.0%
0.0%
11.5%
9.4%
3.5%
Not in School
44.0%
39.1%
25.0%
38.5%
36.0%
46.2%
21.9%
44.8%
χ2
2.27
0.52
4.33
3.99
Entry reason
Neglect
63.0%
69.6%
61.5%
38.5%
61.5%
69.2%
76.5%
82.8%
Abandonment
14.8%
21.7%
0.0%
7.7%
26.9%
19.2%
8.8%
3.5%
Abuse
11.1%
8.7%
7.7%
15.4%
7.7%
3.9%
5.9%
10.3%
Other
0.0%
0.0%
0.0%
0.0%
3.9%
0.0%
0.0%
0.0%
More than one of
11.1%
0.0%
30.8%
38.5%
0.0%
7.7%
8.8%
3.5%
the above
χ2
3.04
2.13
3.78
1.90
Sibling Group
Yes
48.2%
34.8%
53.9%
61.5%
57.7%
65.4%
67.7%
55.2%
No
51.9%
65.2%
46.2%
38.5%
42.3%
34.6%
32.4%
44.8%
χ2
0.91
0.16
0.33
1.03

Special needs
Yes
14.8%
17.4%
7.7%
7.7%
7.7%
7.7%
17.7%
10.3%
No
85.2%
82.6%
92.3%
92.3%
92.3%
92.3%
82.4%
89.9%
115
Exit through Guardianship
χ2
Caregiver at removal
At least one parent
Relatives or friends
χ2
Age at entry
t
Length of stay in care
(in months) prior to
random assignment
t
Number of placement
changes in the
previous episode
t
N
Control Experimental
Group
Group
0.06
96.0%
4.0%
39.9
90.5%
9.5%
0.57
4.5
-0.37
41.2
1.7
-0.84
0.3
27
0.28
23
4.1
Exit through
Reunification
Control Experimental
Group
Group
0.00
92.3%
7.7%
34.9
45.5%
54.6%
6.33**
6.0
0.24
43.2
0.9
-1.14
0.5
13
0.67
13
6.4
Move to Restricted
Foster Care
Control Experimental
Group
Group
0.00
95.5%
4.6%
50.2
85.7%
14.3%
1.21
5.3
0.87
54.7
0.4
-0.53
0.3
26
0.48
26
6.3
Remain in Kinship
Care
Control Experimental
Group
Group
0.68
82.8%
17.2%
51.7
95.0%
5.0%
1.65
5.2
0.42
56.8
0.2
-0.65
0.4
34
-0.88
29
5.6
entry reasons include parental incapacity, mental injury due to abuse or neglect, child’s disruptive behavior, and child’s special needs.
Other special needs include child physical, mental, and behavioral problems.
Sample mean.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
Other
116
Table A- 13 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
exit through guardianship in the absence of the subsidy (Considering participation
in caregiver interview) - African-American children only 
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
0.95*
1.18
1.06
1.18
(0.56)
(0.82)
(0.76)
(0.82)
Male
----0.42
----0.42
(0.66)
(0.66)
Age at entry (in months)
---0.00
---0.00
(0.01)
(0.01)
Sibling Groups
----0.40
----0.40
(0.75)
(0.75)
Neglect
----1.08
----1.08
(0.70)
(0.70)
Abandonment
----0.10
---0.10
(0.75)
(0.75)
Number of placement in
----0.05
----0.05
previous episode
(0.11)
(0.11)
Length of stay in care prior ---0.00
----0.00
to random assignment
(0.02)
(0.02)
Participating in caregiver ---1.73**
1.73**
interview
(0.86)
(0.86)
Constant
-7.54****
-7.08****
-7.63****
-7.08
(0.50)
(0.86)
(0.62)
(0.82)
Log likelihood
-39.33
-24.76
-31.34
-24.76
N
37
37
37
37
Standard
*
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
117
Table A- 14 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
exit through reunification in the absence of the subsidy (Considering participation
in caregiver interview) - African-American children only 
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-0.54
-2.44**
-0.54
-2.44**
(0.49)
(1.21)
(0.49)
(1.21)
Male
----0.77
----0.77
(0.87)
(0.87)
Age at entry (in months)
---0.01
---0.01
(0.01)
(0.01)
Sibling Groups
---1.27
---1.27
(1.36)
(1.36)
Neglect
----4.12**
----4.12**
(2.06)
(2.06)
Abandonment
---1.66*
---1.66*
(0.91)
(0.91)
Number of placement in
---0.32
---0.32
previous episode
(0.43)
(0.43)
Length of stay in care prior ---0.02
---0.02
to random assignment
(0.03)
(0.03)
Participating in caregiver ------------interview
Constant
-6.01****
-3.31*
-6.01****
-3.31*
(0.38)
(1.83)
(0.38)
(1.83)
Log likelihood
-23.21
-18.89
-23.21
-18.89
N
20
20
20
20
Standard
errors are in parentheses.
Variable “whether participating in caregiver interview” dropped out of the model because none of the caregivers of children in this
group participated in caregiver interviews.
*
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.

118
Table A- 15 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
move to RFC in the absence of the subsidy (Considering participation in caregiver
interview) - African-American children only 
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
-0.26
-0.59
-0.28
-0.56
(0.30)
(0.37)
(0.50)
(0.57)
Male
----0.23
----0.11
(0.32)
(0.49)
Age at entry (in months)
---0.01***
---0.01**
(0.00)
(0.01)
Sibling Groups
---0.09
---0.22
(0.31)
(0.52)
Neglect
----0.37
----0.37
(0.40)
(0.73)
Abandonment
---0.76*
---0.85
(0.91)
(0.68)
Number of placement in
---0.13
---0.02
previous episode
(0.16)
(0.33)
Length of stay in care prior ---0.02***
---0.02*
to random assignment
(0.01)
(0.01)
Participating in caregiver ----0.13
----0.10
interview
(0.39)
(0.70)
Constant
-6.24****
-8.25****
-6.24****
-8.59****
(0.15)
(1.83)
(0.29)
(1.38)
Log likelihood
-87.36
-75.44
-64.65
-60.47
N
43
43
43
43
Standard
*
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
119
Table A- 16 Coefficients obtained from Poisson and negative binomial regression
models on exposure adjusted numbers of placement changes for children who would
remain in kinship care in the absence of the subsidy (Considering participation in
caregiver interview) - African-American children only 
Poisson
Poisson
NB Model
NB Model
Model 1
Model 2
1
2
Experimental Group
0.37
0.95*
0.38
0.84
(0.41)
(0.50)
(0.54)
(0.59)
Male
---1.22***
---1.16**
(0.46)
(0.52)
Age at entry (in months)
---0.00
---0.01
(0.01)
(0.01)
Sibling Groups
---0.63
---0.49
(0.51)
(0.57)
Neglect
---0.24
---0.26
(0.69)
(0.79)
Abandonment
---0.64
---0.62
(0.65)
(0.73)
Number of placement in
----0.22
----0.18
previous episode
(0.37)
(0.41)
Length of stay in care prior ----0.01
----0.01
to random assignment
(0.01)
(0.01)
Participating in caregiver ----0.38
----0.46
interview
(0.49)
(0.59)
Constant
-7.35****
-8.24****
-7.35****
-8.16****
(0.24)
(1.09)
(0.54)
(1.22)
Log likelihood
-52.37
-46.25
-48.82
-45.14
N
52
52
52
52
Standard
*
errors are in parentheses.
Significant at .1 level; ** Significant at .05 level; *** Significant at .01 level; **** Significant at .001 level.
120
Download