Targeted and universal intervention
Leon Feinstein
18-July-2006
www.learningbenefits.net
Five questions





How well can we identify those children at risk of
adult social exclusion?
Which measures are most predictive?
What difference does it make?
What are the implications for the delivery of
services?
What are the implications for the targeting of
interventions?
www.learningbenefits.net
E.g. Relationship of G1 maternal interest at G2 age 7 and G2G3 parenting at G2 age 33
Home_Cog
Low interest
Home_Em
0
0
Medium interest
4.7 ***
2.5 ***
High interest
7.2 ***
2.7 ***
Over-concern
8.7 ***
1.0
3873
3697
N
www.learningbenefits.net
Relative cognitive shifts, 7 -11, by SES,
1958 Cohort
80
Probability of escaping low quartile
74.3
70
65.0
60
57.0
54.3
34.7
34.8
55.7
54.2
Prob. (%)
50
40
30.3
30
28.8
22.6
20
20.4
10
Probability of falling from top quartile
0
1
2
3nm
3m
4
5
SES
Source: Feinstein, L. ‘Schools and Educational Attainment’, Oxford Review of Economic Policy, (in press)
www.learningbenefits.net
Predicted changes in probability of not obtaining Level 4; by age 10
capabilities, men
0.08
0.06
0.04
0.02
0
-0.02
-0.04
-0.06
-0.08
-0.1
Maths
Reading
Self-esteem Loc. of Control Externalising
Internalising
Peer relations
Error bars indicate 95% confidence
www.learningbenefits.net
Skills and capabilities;
Probability of being in workless household with children at age 30
0.040
Marginal effect (change in prob.)
0.030
0.020
0.010
0.000
-0.010
-0.020
-0.030
Maths
Reading
Self Esteem
Locus of
control
Externalising Internalising
behaviour
behaviour
Good peer
relations
Source: Feinstein, L. and Bynner, J. (Dec 2003), ‘The benefits of assets in childhood as protection against adult social exclusion: the relative
effects of financial, human, social and psychological assets’, Note to HM Treasury.
www.learningbenefits.net
Predicted changes in probability of offending; by age 10 capabilities, men
0.05
0.04
0.03
0.02
0.01
0
-0.01
-0.02
-0.03
-0.04
-0.05
Maths
Reading
Self-esteem Loc. of Control Externalising
Internalising
Peer relations
Error bars indicate 95% confidence
www.learningbenefits.net
Conceptual model for family influences on child
development
www.learningbenefits.net
Some cost-benefit notation





C(T+) = The social cost of the outcome (T=truth)
P(T+) = The unconditional probability of the
outcome
P(D+)= The level of targeting of the intervention
C(D+)= The cost of the intervention
f= the effectiveness of the intervention; 0 ≤ f ≤ 1.
www.learningbenefits.net
True and false positives
Truth
Decision
(Predicted
outcome)
Positive
Negative
Positive
True
positive
A
False
positive
B
Negative
False
negative
C
True
negative
D
P(TP) = P(D+ | T+) = A / (A+C)
P(FP) = P(D+ | T-) = B / (B+D)
www.learningbenefits.net
Costs and benefits




C(T-, D+) = The cost of a positive decision and hence
intervention for those who would not experience the
outcome (deadweight +stigmatisation costs)
C(T+, D-) = The cost of a negative decision and hence
failure to intervene for those who would experience the
outcome
B(T+, D+) = The net benefit of a positive decision and
intervention for those who would experience the outcome
(net of intervention and stigmatisation costs)
B(T-, D-) = The net benefit of a negative decision and
absence of intervention for those who would not
experience the outcome.
www.learningbenefits.net
B(T+, D+) =
P(D+, T+) * f * C(T+) – P(D+, T+) * C(D+)
=
P(TP) * P(T+) {f * C(T+) –C(D+) }
www.learningbenefits.net
C(T-, D+) = P(D-, T+) * C(D+)
= P(FP) * P(T-) * C(D+)
Hence cost effectiveness requires:
P(TP)
P(FP)
>
P(T-) *
P(T+)
C(D+)
{ f * C(T+) –C(D+) }
www.learningbenefits.net
0.00
0.25
0.50
0.75
1.00
ROC Curves: Allowing the allocation threshold to
vary
0.00
0.25
0.50
1 - Specificity
0.75
1.00
Area under ROC curve = 0.9521
www.learningbenefits.net
Optimal point on the ROC curve
S* = P(T-) *
P(T+)
B(T-, D-) + C(T-, D+)
B(T+, D+) + C(T+, D-)
www.learningbenefits.net
2. Estimating P(TP) and P(FP)
www.learningbenefits.net
Predicting teenage parenthood, by age
1958 Cohort
1970 Cohort
P(TP)
P(FP)
P(TP)
P(FP)
22.0
5.7
23.8
4.0
22 months
26.8
3.9
42 months
29.2
3.7
Birth
5/7 years
28.5
4.2
29.2
3.8
10/11 years
30.7
4.3
60.0
1.4
16 years
45.2
3.6
32.4
3.1
www.learningbenefits.net
Predicting adult depression, by age
1958 Cohort
1970 Cohort
P(TP)
P(FP)
P(TP)
P(FP)
Birth
34.2
17.3
32.3
17.4
5/7 years
38.3
15.8
34.4
17.1
10/11 years
34.6
16.8
46.7
11.0
16 years
45.8
12.4
34.5
16.7
www.learningbenefits.net
Predicting teenage parenthood, by domain
1958 Cohort 1970 Cohort
P(TP) P(FP) P(TP)
P(FP)
Distal family background
31.1
2.1
52.6
1.3
Proximal family process
21.3
2.1
19.6
4.2
Neighbourhood
16.0
5.4
11.2
3.0
School
25.6
2.7
23.8
4.3
Child cognitive achievement 25.5
2.4
20.9
4.2
Behavioural and affective
34.0
2.0
35.5
2.8
Health and growth
19.0
2.8
15.9
4.4
www.learningbenefits.net
Predicting adult depression, by domain
1958 Cohort 1970 Cohort
P(TP) P(FP) P(TP)
P(FP)
Distal family background
35.7
15.8
40.1
14.6
Proximal family process
35.0
15.5
32.5
17.1
Neighbourhood
29.9
19.3
6.0
2.8
School
36.4
16.5
33.3
17.2
Child cognitive achievement 32.9
16.5
32.1
17.5
Behavioural and affective
37.7
16.5
34.3
15.9
Health and growth
31.8
18.1
30.0
17.0
www.learningbenefits.net
Predicting other outcomes from age 10
measures, 1970 Cohort
Workless household with children
Mental health problem
Serious addiction to drugs or alcohol
Long periods of NEET (16-30)
Long periods of NEET (16-18)
On benefits at 30
Multiple offending
P(TP)
50.0
45.3
52.2
51.6
26.5
41.6
51.6
P(FP)
1.3
16.3
1.1
2.5
2.9
3.6
1.3
www.learningbenefits.net
Predicting other outcomes, P(TP), 1970
Workless household with children
Mental health problem
Serious addiction to drugs or alcohol
Long periods of NEET (16-30)
Long periods of NEET (16-18)
On benefits at 30
Multiple offending
Distal
44.9
42.0
42.5
43.0
40.6
33.9
53.5
Behav.
26.3
37.5
29.4
39.3
24.8
29.2
39.6
www.learningbenefits.net
3. Preliminary discussion



Targeting depends on identification
Identification capabilities are very strong
So targeted intervention is possible, but:







Ethical issues
Stigmatisation
Need effective and appropriate intervention
Facilitated community engagement crucial
People move in and out of risk
Progressive universalism is essential
Role of schools and GPs?
www.learningbenefits.net