Health Insurance Reduces Stress: Evidence From A
Randomized Experiment in Kenya⇤
Matthieu Chemin†, Johannes Haushofer‡, and Chaning Jang§
March 10, 2016
Abstract
We show that health insurance reduces levels of self-reported stress and the stress
hormone cortisol using a randomized controlled trial among informal workers in Nairobi,
Kenya. The e↵ects of health insurance on cortisol and stress levels are larger than those
of equally valued unconditional cash transfers, suggesting that reductions in the volatility of consumption through insurance may be more e↵ective in improving psychological
well-being than increases in the level of consumption through cash transfers. The median insurance taker does not use the insurance, further supporting the hypothesis that
the e↵ect of insurance on welfare is primarily a “peace of mind” e↵ect.
⇤ We are grateful to the study participants for generously giving their time. We thank Justin Abraham,
Marie Collins, Faizan Diwan, Bena Mwongeli, Joseph Njoroge, James Vancel, and Matthew White for
excellent research assistance; the Cooperative Insurance Company for fruitful collaboration; and Abhijit
Banerjee, Fenella Carpena, Esther Duflo, Simon Galle, Alexis Grigorie↵, Michael Kremer, Michala Iben
Riis-Vestergaard, Chris Roth, Simone Schaner, Jeremy Shapiro, and Tom Vogl for comments and discussion.
All errors are our own. This research was supported by NIH Grant R01AG039297 and Cogito Foundation
Grant R-116/10 to Johannes Haushofer.
† Department of Economics, McGill University Leacock Building-419, 55 Sherbrooke St. West, Montreal,
QC H3A 2T7, Canada. matthieu.chemin@mcgill.ca
‡ Peretsman Scully Hall 427, Princeton University, Princeton, NJ 08544, and Busara Center for Behavioral
Economics, Nairobi, Kenya. haushofer@princeton.edu
§ Peretsman Scully Hall 407, Princeton University, Princeton, NJ 08544, and Busara Center for Behavioral
Economics, Nairobi, Kenya. cjang@princeton.edu
1
I.
Introduction
Stress has significant welfare costs. First, it is an undesirable outcome in its own right; individuals are generally averse to stress, and in the cross-section, it correlates with undesirable
psychological outcomes such as unhappiness and depression. Second, through its link to
depression, stress may have economic consequences. The relationship between stress and
depression is intimate: roughly 80% of depression diagnoses are preceded by major stressful
life events (Hammen 2005), and one of the most prominent accounts of the neurobiology of
depression is dysfunction of the hormonal system that regulates the release of the stress hormone cortisol (Holsboer 2000). Depression, in turn, has significant economic costs; current
estimates hover around 1-1.5% of GDP per year in terms of lost productivity, not including
treatment costs (Sobocki et al. 2006). Finally, it has emerged recently that stress may have
economic consequences by a↵ecting economic choice; laboratory studies have demonstrated
that stress increases risk aversion and impatience, and reduces the e↵ort people expend
on learning about economic decisions (Porcelli and Delgado 2009; Cornelisse et al. 2013;
Delaney, Fink, and Harmon 2014).
Together, these factors amount to an economic case for reducing stress as a matter of
policy. How can this be achieved? One approach is to reduce stress directly, e.g. through
psychotherapy-type interventions. The efficacy of such approaches have recently begun to
be evaluated in field experiments (Blattman, Jamison, and Sheridan 2015; Bolton et al.
2003) . Here we take a di↵erent approach by asking whether economic interventions can
reduce stress. To choose our interventions, we depart from two observations. First, in the
cross-section, stress is strongly associated with income and consumption volatility; if this
relationship is partly causal, we would expect that interventions that reduce this volatility
should also be e↵ective at reducing stress. Second, programs that increase the level of
consumption, such as cash transfers, have recently been shown to reduce self-reported stress
and depression symptoms, as well as, in some cases, cortisol levels (Haushofer and Shapiro
2013; Fernald and Gunnar 2009). Together, these relationships suggest that both increases
in consumption levels, as well as decreases in consumption volatility, might lead to reductions
in stress.
This study tests these two hypotheses by asking whether health insurance and unconditional cash transfers reduce levels of self-reported stress and the stress hormone cortisol.
We conducted a randomized controlled trial with 789 informal workers in Nairobi, Kenya, in
which one group received free health insurance for themselves and their families for one year,
a second group received an unconditional cash transfer worth retail price of the insurance
by mobile money transfer, and a third group received no intervention. We report the e↵ects
of these treatments on a variety of health and welfare outcomes one year after the beginning
of the intervention. Importantly, the comparison between the insurance group and the unconditional cash transfer group allows us to ask whether an intervention that should mainly
2
reduce consumption volatility, i.e. insurance, a↵ects stress di↵erentially compared to an
intervention which is more likely to raise consumption levels, i.e. cash transfers (Blattman,
Fiala, and Martinez 2013; Haushofer and Shapiro 2013; Baird, McIntosh, and Özler 2011;
Baird et al. 2012; Baird, Hoop, and Özler 2013; Baird et al. 2014). To measure stress, our
endline questionnaire included a battery of measures of psychological well-being such as
self-reported stress, depression, and happiness. In addition, we collected saliva samples to
analyze levels of the stress hormone cortisol. Together, these measures capture the “peace
of mind” e↵ect of health insurance relative to cash transfers and control.
A year after the beginning of the intervention, we find a large reduction in self-reported
stress and cortisol levels in the insurance group relative to control, with no such reduction
in the cash group, suggesting that the health insurance indeed had a “peace of mind” e↵ect
which equivalent cash transfers did not have. Intriguingly, the median insurance beneficiary
does not make use of the health insurance policy a single time during the entire year, and the
reduction in self-reported stress and cortisol persists when we exclude insurance beneficiaries
who made at least one claim during the year. This finding suggests that the reduction in
self-reported stress and cortisol reflects an ex ante “peace of mind” e↵ect, rather than an
ex post e↵ect of consumption smoothing or better healthcare.
This study contributes to several literatures in development economics. First, the study
extends an emerging literature on the psychological consequences of poverty and poverty
alleviation. It has recently been argued that poverty may have adverse psychological consequences, which could have negative e↵ects on economic choice and thereby perpetuate
poverty (Shah, Mullainathan, and Shafir 2012; Mani et al. 2013; Haushofer and Fehr 2014).
One crucial element in providing evidence for this feedback loop is to demonstrate causal
e↵ects of poverty alleviation on psychological well-being. We have previously shown that
unconditional cash transfers improve self-reported psychological well-being and that some
types of transfers reduce cortisol levels (Haushofer and Shapiro 2013)1 . Similar e↵ects have
been shown for other cash transfer programs (Baird et al., 2013, 2014, 2015), and for the
provision of piped water (Devoto et al. 2012), the Moving to Opportunity Program (Kling,
Liebman, and Katz 2007), and lotteries (Gardner and Oswald 2006; Cesarini et al. 2016).
The present study adds to this literature by showing that di↵erent types of poverty alleviation, even when they di↵er neither in price nor in their e↵ects on economic and health
outcomes, may have di↵erential e↵ects on psychological well-being. In the present case,
insurance was superior to cash in improving psychological wellbeing, while the di↵erences
between the two on other dimensions were small. This finding is intriguing because the
cash recipients could have used their transfer to buy insurance; the fact that none of the
cash recipients did this suggests that people either fail to completely anticipate the welfare
1 We
note that the unconditional cash transfers used in the present study were smaller than those in
Haushofer and Shapiro (2013) by a factor of 0.46. This di↵erence may account for the fact that we find no
e↵ects of cash transfers on psychological well-being and cortisol levels in the present study.
3
benefits of insurance, or that they do not value them.
Second, the study broadens the scope of the literature on the welfare e↵ects of health
insurance in developing countries. Specifically, we use outcome variables beyond those
traditionally studied in this literature, and show that even in the absence of improved
health outcomes, insurance confers “peace of mind”, reflected in reduced levels of selfreported stress and the stress hormone cortisol. This finding is in line with the improvement
in mental health documented for insurance beneficiaries in the Oregon Health Insurance
Lottery (Baicker et al. 2013). More broadly, it suggests that the existing literature, by
focusing on healthcare access and health outcomes, might underestimate the e↵ect of health
insurance. According to the model we outline in Section II, the value of health insurance is
consumption smoothing, not improved health. In support of this argument, a study of the
phased rollout in 2002 of Mexico’s Seguro Popular program found that the program did not
a↵ect health outcomes, although it reduced health spending (Barros 2008), possibly because
beneficiaries consumed more health care but had lower catastrophic health expenditures
(Gakidou et al. 2006). Similarly, Wagsta↵ (2007) report that Vietnam’s Health Care Fund
for the Poor, which subsidized healthcare for the poor, reduced catastrophic spending, but
did not decrease in average out-of-pocket spending. These findings support the argument
that the true value of health insurance lies in consumption smoothing and providing “peace
of mind”, a common argument used by insurance companies to sell insurance. Our study
captures and provides support for this phenomenon.
Relatedly, the same argument suggests that these “peace of mind” benefits of insurance
should accrue not only to use the insurance product, but also those who do not use it. Even
if a negative health shock does not occur, an insured individual may still benefit from a
reduction in expected variation in income. In line with this view, we find that the reduction
of self-reported stress and cortisol persists in those who do not make insurance claims.
Finally, the study adds to the evidence base that health insurance provision in developing countries may not result in improved health outcomes. Indeed, a number of previous
randomized evaluations of health insurance in developed and developing countries have have
found limited e↵ects of health insurance provision on healthcare utilization and health. Field
et al. (2010) randomized the price of health insurance among informal workers in Nicaragua
(a similar setting to the one we use here), and find no significant impact of health insurance
on visits to health facilities. Out-of-pocket health care expenditures are reduced, but by
an amount less than the cost of the insurance premium, underscoring the argument that
insurance provision has to be distinguished from its income e↵ect. Ansah et al. (2009) find
no e↵ect of insurance provision in rural Ghana on anemia and mortality and King et al.
(2009) find no impact of Mexico’s Seguro Popular program on self-reported health. Gertler
and Solon (2002) show that the introduction of statutory health insurance in the Philippines
did not reduce out-of-pocket health expenditure of insurance beneficiaries, mainly because
4
providers price-discriminated between insured and uninsured patients. Dow and Schmeer
(2003) analyze the e↵ect of the rollout of health insurance for children in Costa Rica in the
1970s, and find little e↵ect on health that cannot be explained by pre-existing trends. The
RAND health insurance experiment also studied free or subsidized health insurance and
therefore contains an income e↵ect (Manning 1987). The experiment varied the copay from
0 to 25%, 50%, and 95%, up to a maximum expenditure of USD 1000 per year, or 5%, 10%,
or 15% of family income, whichever was less. Healthcare was free in all groups above this
$1000 threshold, and the only di↵erence between the groups was for health expenditure below USD 1000. Brook et al. (1983) find no e↵ect of free care versus copay on eight measures
of health status and health habits. Finally, in the Oregon Health Insurance experiment, insurance had a positive impact on self-reported health, but not physical health (Finkelstein
et al. 2012; Baicker et al. 2013). In sum, the e↵ects of health insurance provision on health
outcomes are limited; Acharya et al. (2012) survey 34 experimental and non-experimental
studies on the e↵ects of health insurance and conclude that “there is little evidence on the
impact of social health insurance on changes in health status”.2 Together with the findings
from the present study, this failure of insurance provision to improve health outcomes may
partly explain the low take-up of health insurance in developing countries (Jowett, Contoyannis, and Vinh 2003; Banerjee, Bénabou, and Mookherjee 2006). At the same time, our
finding that health insurance nevertheless improves psychological outcomes suggests that
individual may value these outcomes less than others, or may be less than fully aware of
these benefits of health insurance.
The rest of the paper is structured as follows. Section II presents a theoretical framework
that illustrates the identification strategy that isolates the e↵ects of health insurance in the
present study. Section III describes the interventions, and Section IV the experimental
design. Section V lays out the econometric framework. Section VI presents the main results
and Section VII concludes.
II.
Theoretical Framework
In this section, we present a simple model for the value of insurance that illustrates the
contribution of this study to distinguishing the e↵ect of health insurance from that of free
healthcare (income e↵ect). Let y be income, p the probability of an accident, c the cost
2 A small number of papers report positive health e↵ects as a result of health insurance provision. Alcaraz
et al. (2012) find that public provision of health insurance in Mexico increases standardized test scores
among primary school children. Bloom et al. (2006) find that in Cambodia, areas where government health
services were expanded by contracting with NGOs showed an improvement of 0.5 standard deviations in
targeted service outcomes (provider absence and supervisory visits), and some evidence for improved health
outcomes (rates of immunizations, incidence of diarrhea in children under five). The Accelerated Benefits
(AB) Demonstration funded by the U.S. Social Security Administration in 2006, which provided previously
uninsured individuals with Social Security Disability Insurance (SSDI) benefits, increased health care usage
during the first year (Michalopoulos et al. 2012).
5
of medical treatment, I the insurance premium, and B the benefit paid by the insurance
company.
The expected utility of having no insurance is
EUN oInsurance = (1
p)u(y) + pu(y
c)
(1)
c + B).
(2)
while the expected utility of having insurance is
EUInsurance = (1
p)u(y
I) + pu(y
The value of insurance is thus EUInsurance
I
EUN oInsurance , which can be shown to be
positive with a concave utility function, i.e. risk-averse individuals. In practice, however,
this value is impossible to estimate for two reasons. First, the exact form of the utility
function is unknown. Structural models can be used, but must assume a functional form
for u. Second, the decision to take up insurance is endogenous: one cannot compare people
who choose to purchase insurance or not to estimate the causal impact of health insurance.
However, random assignment to health insurance vs. a cash transfer makes it possible to
estimate this di↵erence, as we illustrate below.
Denote by IN S the treatment group that receives free health insurance for one year, and
by U CT the group that receives an unconditional cash transfer amounting to the market
value I of this insurance product. This experimental design allows us to control for the
income e↵ect of receiving free health insurance, as both groups receive a transfer of the
same amount.3 Thus, this design measures the e↵ect of obtaining health insurance, rather
than free healthcare, since compared to the U CT group, the IN S group has health insurance
and a lower income by an amount I. Formally:
EUU CT = (1
EUIN S = (1
p)u(y + I) + pu(y + I
p)u(y) + pu(y
c)
(3)
c + B)
(4)
Equation 4 can be rewritten as follows:
EUIN S = (1
p) u ((y + I)
I) + pu ((y + I)
I
c + B)
(5)
The expected utility of the insurance group compared to the cash transfer group is
therefore:
3 This analysis assumes that individuals value the health insurance at market rates. If this is not the
case, our estimates would provide either upper or lower bounds on the impact of insurance. Participants
in our study reporterd a willingness-to-pay of USD 12.41 per month for inpatient insurance, USD 6.82 for
outpatient insurance with copay, and USD 8.13 for outpatient insurance without copay compared to the
actual monthly premium of USD 27.35 for the policy we provided. This evidence suggests that our results
provide lower bounds on the e↵ects of insurance.
6
EUIN S
EUU CT = (EUInsurance
EUN oInsurance ) estimated at y + I
(6)
Thus, this design allows us to estimate the expected utility of having health insurance,
controlling for the income e↵ect of having received the insurance free of charge. For riskaverse individuals, if p is large enough, we have EUIN S
EUU CT 0. In contrast, if p is
very low, or perceived to be very low, then EUU CT ⇡u(y + I) u(y)⇡EUIN S , and expected
utility from cash is greater than from free insurance. Thus, theoretical predictions as to the
value of insurance are ambiguous.
One concern with this experimental design is that it evaluates the impact of insurance at
y + I, not y. Because the utility of receiving insurance relative to cash is decreasing in y 4 ,
it is conceivable that I is large enough to make the e↵ect of insurance very small. However,
this is unlikely to be a concern in our case, since the market value of health insurance is at
an average of KES 12,745, which corresponds to only 8 percent of our sample households’
average yearly income.
III.
Interventions
To identify the causal impact of micro-insurance, the project randomly selected a sample
of informal workers into two treatment groups and one control group. The first treatment
group received the CIC Afya Bora (“good health”) Health Insurance Policy free of charge,
while the second treatment group received an unconditional cash transfer equal to the cost
of the insurance. Comparing the two treatment arms controls for any income e↵ect and
allows us to evaluate the impact of providing insurance relative to a cash transfer.
A.
Micro-health insurance
Respondents in the insurance group were enrolled in the CIC Afya Bora plan, a combined
inpatient and outpatient family health insurance policy. These treated households received
inpatient benefits of up to USD 6,437 per family that covered the costs of a broad array of
services, including hospital accommodation, doctor’s fees, routine lab tests, UCI charges,
medications, and maternity services. Chronic and pre-existing conditions were covered up
to USD 1,931. Households also received outpatient benefits of up to USD 1,287 per family
that covered routine outpatient consultations, medication (including ARVs), laboratory services, pre- and post-natal care, oncology, and psychiatry and psychotherapy. Pre-existing
and chronic conditions were covered up to USD 515. Beneficiaries paid around USD 2.60
for each outpatient visit. Both covers included chronic and pre-existing conditions, including HIV/AIDS, but excluded treatment outside Kenya, cosmetic treatment, treatment by
4 Set
@(u(y I
p
=
1 and notice that,
c+B) u(y c))
@y
by Jensen’s inequality,
<0
7
@(EUInsurance
EUN oInsurance )
@y
=
non-qualified persons, infertility, self-inflicted injury, experimental treatment, and dental
treatment unless occasioned by accidental injury. Beneficiaries could access these benefits
through CIC’s network of providers that included 26 mission and faith-based hospitals in
Nairobi. Full details of the insurance cover are given in section B.2 in the appendix.
The plan provided benefits to principals and spouses under 72 years old and children
dependents younger than 25 years with proof of enrollment in school or college. Respondents
were enrolled in the Afya Bora plan free of charge for one year, a value of USD 328 for the
principal, spouse and up to five dependents. Each additional child dependent increased the
annual premium by USD 52 per child. The project fully covered households for the base
cost and any added premium.
B.
Unconditional cash transfer
Respondents in the second treatment group, “UCT”, received an unconditional cash transfer
equal to the net value of the annual premium they would have had to pay had they enrolled
in the CIC Afya Bora scheme. The magnitude of this transfer was USD 328 for households
with up to five dependents, with an additional USD 52 for each dependent above five. The
transfer was delivered to recipients electronically using the M-Pesa mobile money service.
M-Pesa is a mobile money system o↵ered by Safaricom, the largest Kenyan mobile phone
operator. Using M-Pesa requires a registered SIM card and a valid Kenyan national ID card.
The project transferred the money from Innovations for Poverty Action (IPA) Kenya’s MPesa account to that of the recipient. To facilitate the transfers, we encouraged recipients
to sign up for M-Pesa and helped them obtain, where necessary, the documents required for
registration5 . The money was transferred to the registered SIM card and the recipient could
withdraw the balance at any of the large number of M-Pesa agents in Kenya by putting the
SIM card into the agent’s cell phone or by using their own phone.
IV.
A.
Experimental Design
Setting
The setting for the study was Kenya’s informal sector, commonly known as “Jua Kali”
(“under the hot sun”). Employment in Jua Kali accounts for over 70 percent of non-farm
employment in Kenya (Adams, Silva, and Razmara 2013). The artisans, vendors, and mechanics in the Jua Kali sector face extreme vulnerability to illness, economic dislocation,
and natural disasters. Jua Kali workers supply goods to local markets using predominantly
manual labor, little capital, and often handmade tools. The Jua Kali area in Kamukunji,
Nairobi consist mostly of metalworkers and vendors who work in hazardous conditions with
5 As
a consequence, encouragement to sign up for M-Pesa should be considered part of the UCT treatment.
8
minimal safety equipment. In our sample, 21 percent of the control group was sick or injured in the month prior to being surveyed. Our respondents spent on average 0.4 nights in
the hospital in the past year, and average monthly medical expenses for the main respondent were USD 17. However, very few people had insurance (on average, 0.05 policies per
household, SD = 0.27).
B.
Sampling strategy
We studied a randomly selected sample of metalworkers of the Kamukunji Jua Kali Association (JKA) in Nairobi, an organization of an estimated 4,000 Jua Kali workers. All adult
JKA members working in an area of Kamukunji Jua Kali that makes him or her eligible for
voting rights with the JKA were eligible to participate in the study. We first conducted a
preliminary survey with 1,392 JKA members between August 2008 and December 2010. On
the basis of this survey, we randomly selected 855 participants for the randomized controlled
trial. These respondents were stratified into three groups by weekly household income: 313
respondents with a weekly income greater than USD 103 comprised the high income group;
300 participants with a weekly income between USD 52 and USD 103 comprised the middle
income group; and 242 respondents with a weekly income under USD 52 comprised the low
income group. Within each income stratum, we randomly selected a third of respondents
for one of the two treatment arms and the control group. We exclude from our analysis
respondents who did not have a valid national ID by the time we conducted the baseline
survey for a final sample size of 789 individuals.
C.
Data collection
Data collection for baseline occurred between March 2011 and December 2011. Endline data
collection occurred between January 2013 and April 2013, over a year after the baseline.
Figure 1 is a timeline displaying when respondents completed each phase of the experiment.
Trained interviewers used netbooks to administer the surveys in an office at JKA or at the
respondent’s place of work. Respondents received around USD 5 as payment for participating in each interview, in addition to further payouts determined by responses in the time
and risk preferences section of the survey. To ensure data quality, we performed back-checks
on 10 percent of all interviews, focusing on non-changing information. This procedure was
known to field officers ex ante.
9
0
50
Frequency
100
150
200
250
Figure 1: Project timeline
Jan-11
Jul-11
Jan-12
CIC start date
Baseline survey
Jul-12
Jan-13
Jul-13
Cash transfer
Endline survey
Notes: Histogram of the number of respondents who were surveyed and receive treatment during each phase
of the project. Data collection for baseline occurred between March 2011 and December 2011. endline data
collection occurred between January 2013 and April 2013. The interventions were delivered in early 2012.
10
Respondents were informed of their treatment status after completing the baseline survey. In March 2012, respondents in the cash transfer group who completed the baseline and
had a registered SIM card received an unconditional cash transfer via M-Pesa equal to the
amount of the annual premium they would have had to pay under CIC Afya Bora. Respondents in the insurance group were o↵ered to enrol in CIC’s Afya Bora insurance free of cost
for one year. Project sta↵ assisted this group received with preparing required documents,
and submitted the applications to CIC on their behalf. Beneficiaries then received an ID
card from CIC which they could use to claim benefits in CIC’s network of 26 providers
across Nairobi.
The survey instruments asked respondents about household characteristics, consumption, workplace, insurance usage, health, self-reported well being, and time and risk preferences. An important feature of this study is that, in addition to questionnaire measures of
psychological well-being, we also obtained saliva samples from all respondents, which were
assayed for the stress hormone cortisol. Cortisol has been used extensively in psychological
and medical research (Kirschbaum and Hellhammer 1989), and more recently in randomized
trials in developing countries similar to this one (Fernald and Gunnar 2009; Haushofer and
Shapiro 2013). Cortisol has several advantages over other outcome variables. First, it is an
objective measure and not prone to survey e↵ects such as social desirability bias (Zwane
et al. 2011), and it has several practical advantages which make it attractive to analyze in
field studies. Second, cortisol is a useful indicator of both acute stress (Kirschbaum and
Hellhammer 1989) and more permanent stress-related conditions such as major depressive
disorder (Holsboer 2000; Hammen 2005). Third, cortisol is a good predictor of long-term
health through its e↵ects on the immune system. To measure cortisol levels, we collected
saliva samples using the Salivette (Sarstedt, Germany). Respondents chewed on the cellulose swab for two minutes, and it was then centrifuged, stored at
20 C, and analyzed for
salivary cortisol. Field officers collected a total of two saliva samples from each respondent
at both baseline and endline, one each before and after the survey.6
6 In addition, we collected blood samples from respondents at the end of each survey. Trained phlebotomists took blood draws in the JKA office. These samples have not been fully analyzed and are not
included in this paper.
11
0
Frequency
50
100
150
Figure 2: Raw and residual log average cortisol with boundaries at the 1 and 99 percentiles
0
2
Log avg. cortisol level
4
6
-4
-2
0
Residual log avg. cortisol
2
4
0
Frequency
50
100
150
-2
Notes: This figure displays the distribution of cortisol levels at baseline, averaged across the two samples
taken from each respondent, in log(nmol/L). The lower panel displays the baseline distribution of the
residuals after regressing the raw cortisol values on control variables. Vertical bars denote the 1st and 99th
percentiles.
12
Cortisol levels were analyzed as specified in our pre-analysis plan, and briefly summarized
here: we first obtained the average cortisol level in each participant by averaging the values of
the two samples. Because cortisol levels in population samples are usually heavily skewed,
it is established practice to log-transform them before analysis. We follow this standard
approach here. Salivary cortisol is subject to a number of confounds; it is a↵ected by food
and drink, alcohol and nicotine, medications, and strenuous physical exercise. Cortisol
levels also follow a diurnal pattern: they rise sharply in the morning, and then exhibit
a gradual decline throughout the rest of the day. To control for these confounds, but at
the same time avoid the risk of “cherry-picking” control variables, we present results for
several versions of the cortisol variable in the analysis. First, we use the log-transformed
raw cortisol levels without the inclusion of control variables. Second, we construct a “clean”
version of the raw cortisol variable, which consists of the residuals of an OLS regression
of the log-transformed cortisol levels on dummies for having ingested food, drinks, alcohol,
nicotine, or medications in the two hours preceding the interview, for having performed
vigorous physical activity on the day of the interview, and for the time elapsed since waking
(rounded to the next full hour). Third, we analyze the same variables after trimming at
100 nmol/L and Winsorization at the 99 percent level to address potential outliers. The
distribution of log cortisol for the sample at baseline is displayed in Figure 2.
V.
Identification Strategy7
Our basic specification to capture the impact of the health insurance and cash transfers on
outcomes is:
yi =
0
+
1 IN Si
+
2 U CTi
+ yiB + "i
(7)
where yi is the outcome of interest for individual i. IN Si is a treatment indicator
that takes the value 1 for individuals that received the insurance and 0 otherwise. U CTi
is a second treatment indicator that takes the value 1 for individuals that received the
unconditional cash transfer and 0 otherwise. "i is the idiosyncratic error term. The omitted
category in this specification is the control group.
insurance,
1
2
1
thus captures the e↵ect of free health
the e↵ect of a cash transfer equal to the value of the insurance policy, and
2 captures the value of purchasing health insurance, as outlined in Section II. Following
McKenzie (2012), equation 7 conditions on the baseline level of the individual outcome yiB
to improve statistical power.
Instead of estimating these equations separately for each unique yi , we can estimate the
system of seemingly unrelated regressions (SUR) to improve the precision of the coefficient
estimates (Zellner 1962). SUR estimation is equivalent to multivariate OLS when the error
7 We
registered a pre-analysis plan written and published before the beginning of data analysis: https:
//www.socialscienceregistry.org/trials/647
13
terms are in fact uncorrelated between the regressions or when each equation contains the
same set of regressors. Simultaneous estimation allows us to perform Wald tests of joint
significance on the treatment coefficients. We also test whether the impact of health insurance and cash transfers varies with pre-determined respondent characteristics measured at
baseline and denoted by X‘i in the following equation.
yi =
A.
0
+
1 IN Si
+
2 U CTi
+
3 Xi
+
4 (IN Si
⇥ Xi ) +
5 (U CTi
⇥ Xi ) + yiB + "i (8)
Accounting for multiple inference
Because our interventions are likely to impact a large number of economic behaviors and
dimensions of welfare and given that our survey instrument often included several questions
related to a single outcome, we account for multiple inference in three ways. First, we
pre-defined primary outcome groups in the pre-analysis plan before the beginning of analysis. Second, for each of these outcome groups, we construct an index variable, following
the procedure proposed by Anderson (2008)8 . For each outcome, we invert scores where
necessary so that the positive direction always indicates a “better” outcome. We demean
all outcomes and convert them to e↵ect sizes by dividing each outcome by its control group
standard deviation. Finally, we weight each outcome by the sum of the entries in the row
of the inverted covariance matrix corresponding to that outcome to create a single index.
Third, because combining individual outcome variables in indices as described above
still leaves us with multiple index variables, we additionally control for the family-wise error
rate (FWER) using the free step-down resampling method to compute adjusted p-values.
(Westfall and Young 1993)9 . This approach sets the size of the test to exactly the desired
crticial value. For each index variable, we report both unadjusted p-values as well as p-values
corrected for multiple inference.
B.
Attrition
Three factors made attrition a concern in this study: the high mobility among informal
workers in Kenya, the collection of biomarkers, and the requirement for a national ID to
obtain insurance or an M-Pesa account. In a pilot study which did not involve biomarkers
nor national IDs, we found attrition rates of over 20 percent. To mitigate the attrition, we
therefore compensated each respondent who completed both baseline and endline with USD
26 in addition to the individual fees for each survey. Moreover, we conducted a lottery in
which three respondents among those who had completed all surveys won prizes of USD
515, USD 258, and USD 129, respectively.
8 See
9 See
section C.1 of the appendix for details on the procedure.
section C.2 of the appendix for details on the procedure
14
Table 1: Treatment group by survey participation
Participation
Baseline
Attrited
Endline
Control
282
46
236
Insurance
259
60
199
UCT
248
41
207
Total
789
147
642
Notes: This table displays a cross-tabulation of
treatment assignment and participation status. The
first column includes all respondents surveyed at
baseline. The second column includes respondents
who attrited between baseline and endline surveys.
The third column includes the respondents who successfully completed the endline survey.
Despite these e↵orts, rates of attrition were relatively high, at 17 percent in the control
group, 24 percent in the cash group, and 28 percent in the insurance group. Some factors
that may have contributed to higher dropout rates include tracking issues and unwillingness
to provide saliva and blood samples. In addition, due to a miscommunication with the field
team, respondents were not interviewed in the endline if they were assigned to the insurance
group but did not enroll in the insurance, or assigned to the cash transfer group but did
not receive the cash transfer. This occurred mainly for respondents without a national ID
card; these respondents could not be enrolled in insurance, or receive the cash, because CIC
requires a valid ID to register insurance recipients, and Safaricom requires one to register
an M-Pesa account. To mitigate potential selection bias arising from this issue, we exclude
from our analysis respondents who did not have a valid national ID during baseline. Table
1 reports attrition between baseline and endline surveys for this sample.
To assess di↵erences between attriters and non-attriters, Table 2 presents summary
statistics of selected baseline variables, separately for each of the treatment arms and the
control group with t-tests to compare means. We find that the baseline characteristics are
largely similar across all three groups.
VI.
A.
Results
Stress and cortisol
Tables 3 - 8 present the main results. In each table we report the intent-to-treat e↵ects
with and without control variables. Table 3 shows that the insurance has no e↵ect on any
15
Table 2: Summary statistics - Selected baseline variables by treatment group
Mean (SD, N)
HH Size
Years of education
Total weekly HH inc. last week
Total expenditure last mo.
Savings
Sick/injured last mo.
Nights hospitalized last year
Contribution to hosp. costs
Insurance ownership index
No. of insurance policies owned
Subjective well-being index
Stress (Cohen)
Avg. cortisol level
Log avg. cortisol level
Residual log avg. cortisol
Di↵erence p-value
Control
Insurance
UCT
3.88 (2.03)
282
8.57 (2.50)
282
122.48 (176.33)
282
1293.55 (1167.21)
174
369.97 (937.47)
282
0.21 (0.41)
282
0.44 (4.00)
282
68.27 (283.61)
279
0.02 (1.07)
282
0.05 (0.29)
282
0.04 (0.97)
282
37.73 (6.42)
282
12.39 (19.63)
281
2.18 (0.71)
281
-0.05 (0.68)
277
4.08 (2.00)
259
8.60 (2.50)
258
141.80 (302.01)
259
1519.10 (1905.50)
137
300.05 (661.47)
259
0.20 (0.40)
258
0.35 (2.35)
258
89.58 (510.65)
258
0.03 (1.01)
258
0.07 (0.34)
258
-0.06 (0.97)
259
37.92 (6.20)
259
15.35 (31.21)
255
2.25 (0.83)
255
0.04 (0.80)
250
4.23 (2.20)
248
8.33 (2.74)
248
159.31 (343.72)
248
1294.04 (1219.55)
132
386.61 (971.99)
248
0.21 (0.41)
248
0.33 (3.85)
248
43.91 (142.11)
247
0.04 (1.01)
248
0.06 (0.27)
248
-0.00 (1.02)
248
37.69 (5.97)
248
14.79 (24.32)
245
2.19 (0.88)
245
-0.00 (0.85)
242
Ins. Control
UCT Control
Ins. UCT
0.24
0.06⇤
0.45
0.88
0.29
0.24
0.36
0.12
0.54
0.20
1.00
0.25
0.32
0.84
0.24
0.74
0.90
0.66
0.76
0.76
0.95
0.55
0.22
0.18
0.92
0.86
0.94
0.64
0.77
0.84
0.23
0.60
0.54
0.73
0.94
0.67
0.18
0.21
0.82
0.30
0.84
0.47
0.16
0.45
0.59
Notes: The first three columns report means of each row variable for each treatment group. SD are in parentheses and N is displayed on the
second line. The last three columns report the p-value for a di↵erence of means t-test between each group. * denotes significance at 10 pct., ** at
5 pct., and *** at 1 pct. level.
16
of the pre-specified index variables.1011 The SWB index is 0.18 SD higher in the insurance
compared to the control group. This result is significant at the 10 percent level, although
it is no longer significant when correcting for multiple inference using the family-wise error
rate. However, the fact that we see some evidence for an e↵ect of insurance on subjective
well-being suggests that insurance can have e↵ects on broader dimensions of welfare beyond
health status alone. We observe no significant e↵ects of insurance or cash on any of the
other indexed outcomes.
Table 4 shows detailed results for the individual components of the SWB index. We find
that the driving force of changes in SWB in the insurance group is a decrease in self-reported
stress. Insurance leads to a relatively large (0.29 SD) decrease in scores on the perceived
stress scale in this group, and this decrease is significant compared to both the cash and
control groups.
To corroborate the results obtained from self-reports, Table 5 reports treatment e↵ects
on the stress hormone cortisol. Columns 1 and 2 show that cortisol decreases by between
0.19 - 0.20 SD in the insurance group relative to control. This e↵ect is large, is consistent
across di↵erent transformations of the cortisol variable, and is also robust to the inclusion
of a full set of baseline control variables. Cash transfers do not a↵ect cortisol in any of the
specifications. Moreover, we reject the hypothesis that IN Si = U CTi for four out of the
six specifications that test cortisol and the specification that tests self-reported stress. This
result is evidence that insurance reduces stress above and beyond the income e↵ect from
receiving a free plan.
To add context to these results, we briefly discuss the e↵ect of comparable economic
and psychological interventions on stress. Baicker et al. (2013) conducted one of the few
evaluations to examine the e↵ect of health insurance provision on mental health outcomes.
They find that insurance coverage reduced rates of depression as measured by the PHQ-8
(-30.5 percent, p < 0.05) and increased overall self-reported mental health. Haushofer and
Shapiro (2013) analyze the e↵ects of a large unconditional cash transfer on cortisol and
self-reported stress and find a 0.14 SD (p < 0.05) reduction in self -reported stress using
the PSS. We find twice the e↵ect size in our insurance arm while the UCT arm showed no
significant e↵ect.
12
10 See appendix for exact definition and composition of the indices as well as analysis on the constituent
outcomes.
11 This index is a weighted standardized average of scores on the Life Orientation Test (Scheier), Rosenberg’s Self-Esteem Scale, the CES-D depression scale, Cohen’s Perceived Stress Scale, the World Value
Survey happiness question (“Taking all things together, would you say you are very happy (1), quite happy
(2), not very happy (3), or not at all happy (4)?”), and the World Value Survey life satisfaction question (“All things considered, how satisfied are you with your life as a whole these days? (1=dissatisfied,
10=satisfied)”).
12 The contrasting results in these two studies may be related to study setting, transfer amounts, or the
delay between intervention and endline surveys. Haushofer and Shapiro (2013) conduct endline surveys on
average four months after the end of the cash transfer, while we conduct endline surveys approximately one
year after the baseline survey. Recent work by Galiani, Gertler, and Undurraga (2015) suggests that hedonic
adaptation can a↵ect the size of self-reported measures of well-being, so the timing of measurement relative
17
to the intervention is crucial for comparing e↵ects. Haushofer and Shapiro (2013) only find decreases in
cortisol for individuals in households where the woman received the transfer relative to the man (-0.22 log
nmol/L, p < 0.05), where households received lump-sum transfers relative to monthly transfers (-0.26 log
nmol/L, p < 0.05), and where households received large (USD 1,520) transfers (-0.16 log nmol/L, p < 0.05).
In comparison, our JKA sample consisted of mostly male workers in an urban setting who received cash
transfers amounting to only 8 percent of average household yearly income.
18
Table 3: Treatment e↵ects – Index variables
No Controls
With Controls
Subjective well-being index
Insurance ownership index
Insurance WTP index
19
Asset ownership index
Labor mobility index
Labor productivity index
Job risk index
Joint p-value
(1)
(2)
(3)
Di↵erence
p-value
Insurance
UCT
0.18⇤
(0.10)
[0.38]
0.01
(0.09)
[1.00]
-0.07
(0.10)
[0.98]
0.02
(0.07)
[0.99]
0.02
(0.12)
[1.00]
-0.03
(0.11)
[0.99]
0.01
(0.09)
[1.00]
0.07
(0.09)
[0.94]
0.02
(0.09)
[0.99]
-0.11
(0.08)
[0.84]
0.03
(0.07)
[0.99]
0.01
(0.11)
[0.99]
-0.14
(0.09)
[0.84]
-0.12
(0.09)
[0.84]
0.25
[0.92]
0.80
0.43
0.25
0.95
[1.00]
0.58
[0.96]
0.84
[1.00]
0.96
[1.00]
0.36
[0.96]
0.19
[0.92]
(4)
(5)
Insurance
UCT
0.18⇤
(0.10)
[0.34]
0.01
(0.09)
[1.00]
-0.06
(0.09)
[0.97]
-0.01
(0.07)
[1.00]
0.02
(0.12)
[1.00]
-0.05
(0.11)
[0.99]
0.00
(0.09)
[1.00]
0.80
Sample
(6)
Di↵erence
p-value
(7)
Control Mean
(SD)
(8)
0.09
(0.09)
[0.91]
0.04
(0.09)
[0.95]
-0.10
(0.08)
[0.88]
0.01
(0.08)
[1.00]
0.01
(0.11)
[1.00]
-0.15
(0.10)
[0.76]
-0.11
(0.09)
[0.88]
0.35
[0.91]
0
(1.01)
642
0.78
[0.95]
.03
(1.06)
640
0.65
[0.88]
.02
(1.03)
640
0.80
[1.00]
.06
(0.96)
640
0.96
[1.00]
.01
(1.06)
626
0.38
[0.76]
.05
(1.00)
638
0.25
[0.88]
.02
(1.00)
640
0.45
0.35
N
Notes: This table reports the estimated treatment e↵ect of insurance and UCT on each row variable. Columns 1 - 2 report estimates from
an intent-to-treat analysis without correcting for selection. Columns 4 - 5 report OLS estimates controlling for baseline covariates. Columns
3 and 6 report the p-values for tests of the equality of the UCT and insurance coefficients. The bottom row reports the p-value for a test of
the treatment e↵ect across models using SUR. Standard errors are in parentheses and FWER adjusted p-values are in brackets. * denotes
significance at 10 pct., ** at 5 pct., and *** at 1 pct. level.
Table 4: Treatment e↵ects – Subjective well-being
No Controls
With Controls
(1)
(2)
Insurance
UCT
(4)
(5)
(6)
Di↵erence
p-value
(7)
Control Mean
(SD)
(8)
Insurance
UCT
-0.29⇤⇤⇤
(0.10)
-0.01
(0.10)
0.01⇤⇤
-0.29⇤⇤⇤
(0.11)
-0.04
(0.10)
0.03⇤⇤
-.02
(1.01)
640
Optimism (Scheier)
0.04
(0.10)
0.16⇤
(0.09)
0.24
0.04
(0.10)
0.18⇤⇤
(0.09)
0.17
.03
(0.97)
640
Self-esteem (Rosenberg)
0.01
(0.10)
0.05
(0.09)
0.67
0.01
(0.10)
0.04
(0.09)
0.77
-.05
(0.99)
640
Depression (CESD)
-0.09
(0.09)
-0.08
(0.09)
0.90
-0.11
(0.09)
-0.12
(0.09)
0.96
-.02
(0.98)
642
Locus of control
0.13
(0.10)
0.22⇤⇤
(0.10)
0.39
0.12
(0.10)
0.19⇤
(0.10)
0.51
-.01
(1.00)
637
WVS Happiness
0.04
(0.09)
0.01
(0.09)
0.79
0.02
(0.09)
0.01
(0.09)
0.94
-.01
(0.96)
640
WVS Satisfaction
0.04
(0.10)
0.02
(0.10)
0.88
0.03
(0.10)
0.01
(0.10)
0.79
.02
(0.99)
640
Joint p-value
0.05⇤
0.14
0.01⇤⇤
0.06⇤
0.12
0.03⇤⇤
Stress (Cohen)
(3)
Di↵erence
p-value
Sample
N
20
Notes: This table reports the estimated treatment e↵ect of insurance and UCT on each row variable. Columns 1 - 2 report estimates
from an intent-to-treat analysis without correcting for selection. Columns 4 - 5 report OLS estimates controlling for baseline covariates.
Columns 3 and 6 report the p-values for tests of the equality of the UCT and insurance coefficients. The bottom row reports the p-value
for a test of the treatment e↵ect across models using SUR. Standard errors are in parentheses. * denotes significance at 10 pct., ** at 5
pct., and *** at 1 pct. level.
Table 5: Treatment e↵ects – Cortisol
No Controls
With Controls
(1)
(2)
Insurance
UCT
Log avg. cortisol level
-0.13⇤⇤
(0.06)
-0.01
(0.06)
Residual log avg. cortisol
-0.13⇤⇤
(0.06)
Log avg. cortisol less 100
(5)
Insurance
UCT
0.04⇤⇤
-0.13⇤⇤
(0.06)
-0.02
(0.06)
0.06⇤
-0.14⇤⇤
(0.06)
-0.06
(0.06)
Residual log avg. cortisol less 100
-0.13⇤⇤
(0.06)
Log avg. cortisol (.99 Wins.)
Residual log avg. cortisol (.99 Wins.)
21
(4)
Joint p-value
(3)
Di↵erence
p-value
Sample
(6)
Di↵erence
p-value
(7)
Control Mean
(SD)
(8)
-0.01
(0.06)
0.05⇤⇤
2.48
(0.66)
579
-0.13⇤⇤
(0.06)
-0.02
(0.06)
0.08⇤
.03
(0.64)
578
0.17
-0.13⇤⇤
(0.06)
-0.06
(0.06)
0.19
2.48
(0.66)
576
-0.06
(0.06)
0.17
-0.13⇤⇤
(0.06)
-0.06
(0.06)
0.21
.03
(0.64)
576
-0.13⇤⇤
(0.06)
-0.02
(0.06)
0.05⇤⇤
-0.13⇤⇤
(0.06)
-0.02
(0.06)
0.06⇤
2.48
(0.66)
579
-0.12⇤⇤
(0.06)
-0.02
(0.06)
0.08⇤
-0.12⇤⇤
(0.06)
-0.03
(0.06)
0.10⇤
.03
(0.64)
578
0.27
0.49
0.04⇤⇤
0.24
0.34
0.05⇤⇤
N
Notes: This table reports the estimated treatment e↵ect of insurance and UCT on each row variable. Columns 1 - 2 report estimates from an intentto-treat analysis without correcting for selection. Columns 4 - 5 report OLS estimates controlling for baseline covariates. Columns 3 and 6 report the
p-values for tests of the equality of the UCT and insurance coefficients. The bottom row reports the p-value for a test of the treatment e↵ect across models
using SUR. Standard errors are in parentheses. * denotes significance at 10 pct., ** at 5 pct., and *** at 1 pct. level.
What might be the channel through which insurance reduces cortisol levels? There is
cross-sectional evidence that income and consumption volatility is strongly associated with
stress. We therefore argue that insurance reduces stress primarily through its ability to
smooth income. Recall that the conceptual framework we outlined in Section II suggests
that the value of insurance does not depend on the realization of an adverse health shock.
Thus, stress reduction associated with insurance should be observable regardless of whether
insurance was in fact utilized or not. Table 6 asks whether the reduction in cortisol induced
by insurance is robust to exclusion of participants who made insurance claims. Indeed,
we find that that cortisol decreases significantly even for respondents who did not utilize
insurance13 . The magnitude of this e↵ect is between 0.21-0.23 SD, similar to the magnitude
observed when including all respondents. This suggests that insurance may have a “peace
of mind” e↵ect, stemming from income smoothing, that extends beyond direct pecuniary
or health benefits. For those who received cash transfers, financial risk is reduced only
by the amount of the transfer plus any returns. Furthermore, cash transfers are at risk
of appropriation by family members and friends, further reducing any risk mitigation that
these windfalls may provide. In fact, Haushofer and Shapiro (2013) suggest that the reason
lump-sum transfers led to significant reductions in cortisol relative to monthly transfers was
the inability for households in the latter group to protect their newfound wealth.
13 Table 46 of the appendix shows that 88 households enrolled in the insurance made a claim, while 152
did not. Appendix Figures 7 to 12 show the usage data.
22
Table 6: Treatment e↵ects on cortisol excluding subjects who made insurance claims
No Controls
With Controls
(1)
(2)
Insurance
UCT
Log avg. cortisol level
-0.15⇤⇤
(0.06)
-0.01
(0.06)
Residual log avg. cortisol
-0.14⇤⇤
(0.06)
Log avg. cortisol less 100
23
(4)
(5)
Insurance
UCT
0.03⇤⇤
-0.16⇤⇤⇤
(0.06)
-0.02
(0.06)
0.05⇤⇤
-0.15⇤⇤
(0.06)
-0.06
(0.06)
Residual log avg. cortisol less 100
-0.14⇤⇤
(0.06)
Log avg. cortisol (.99 Wins.)
Residual log avg. cortisol (.99 Wins.)
Joint p-value
(3)
Di↵erence
p-value
Sample
(6)
Di↵erence
p-value
(7)
Control Mean
(SD)
(8)
-0.01
(0.06)
0.02⇤⇤
2.48
(0.66)
510
-0.16⇤⇤
(0.06)
-0.02
(0.06)
0.04⇤⇤
.03
(0.64)
509
0.16
-0.15⇤⇤
(0.06)
-0.06
(0.06)
0.11
2.48
(0.66)
507
-0.06
(0.06)
0.18
-0.15⇤⇤
(0.06)
-0.06
(0.06)
0.15
.03
(0.64)
507
-0.15⇤⇤
(0.06)
-0.02
(0.06)
0.03⇤⇤
-0.16⇤⇤⇤
(0.06)
-0.02
(0.06)
0.02⇤⇤
2.48
(0.66)
510
-0.14⇤⇤
(0.06)
-0.02
(0.06)
0.06⇤
-0.16⇤⇤
(0.06)
-0.03
(0.06)
0.04⇤⇤
.03
(0.64)
509
0.25
0.43
0.03⇤⇤
0.20
0.31
0.02⇤⇤
N
Notes: This table reports OLS and Heckman two-step estimates of the treatment e↵ect restricted to the segment of the insurance group that made at
least one claim in the study period. Cortisol levels are coded as log(nmol/L). We include in the analysis raw log cortisol and the cleaned and truncated
measures. Columns 1 - 2 report estimates from an intent-to-treat analysis without control variables. Columns 4 - 5 report estimates from an intent-to-treat
analysis with control variables. Column 7 reports the coefficient for the inverse Mills’ ratio in the second stage. Columns 3 and 6 report the p-values for
Wald tests of the equality of the UCT and insurance e↵ects after estimation. Standard errors are in parentheses. * denotes significance at 10 pct., ** at
5 pct., and *** at 1 pct. level.
B.
Heterogeneous treatment e↵ects
Table 7 reports the results of analysis for heterogeneous treatment e↵ects. Column 1 shows
that the poor (those with weekly log income below the median) show larger reductions in
cortisol levels when they receive insurance than the less poor. This result is not statistically
significant when analyzing the standard cortisol variable in Table 7, but is significant at
the 10 percent level for all log cortisol variables that control for covariates. This finding is
intuitive given that the receipt of insurance constitutes a relatively larger transfer for the
poor compared to the less poor. Another possible explanation is that poorer households
experience relatively more economic uncertainty and would therefore receive greater “peaceof-mind” from health insurance.
Column 2 shows that the cortisol-reducing e↵ect of insurance is larger for those who
have children. This finding can be understood in light of the fact that the health insurance
product covered all household members including children. It is plausible that the di↵erential cortisol reduction reflects an additional the “peace-of-mind” e↵ect from knowing that
children’s health is protected.
Column 3 show that the value of insurance is higher for those with above median depression (as measured by the CES-D), a finding that can possibly be explained by the fact
that depression is associated with a high degree of rumination and worry, which may be
partially alleviated by insurance. While we do not find an overall e↵ect of the unconditional
cash transfer, the transfer reduces cortisol (-0.28 log nmol/L, p > 0.01) for respondents in
our sample with above median CES-D scores. Similarly, Fernald and Gunnar (2009) found
that the Oportunidades conditional cash transfer program reduced salivary cortisol by 0.19
(p < 0.05) log units in children living with high levels of maternal depression.
In sum, we find that insurance reduces cortisol levels compared to a control group or an
equivalent cash transfer, and this e↵ect is particularly strong for those that have children
and score highly on a depression scale.
24
Table 7: Heterogeneous e↵ects – Log avg. cortisol level
(1)
(2)
(3)
Log avg. cortisol level Log avg. cortisol level Log avg. cortisol level
Ins. ⇥ Interactant
Insurance
UCT ⇥ Interactant
UCT
Interactant
Constant
Interactant
Adjusted R2
Ins. p-value
UCT p-value
Observations
-0.196
(0.121)
-0.0218
(0.0912)
-0.115
(0.128)
0.0317
(0.0894)
0.0274
(0.165)
2.313⇤⇤⇤
(0.186)
-0.437⇤⇤⇤
(0.144)
0.211
(0.129)
-0.273⇤⇤
(0.135)
0.194⇤
(0.109)
0.130
(0.0974)
2.203⇤⇤⇤
(0.121)
-0.390⇤⇤⇤
(0.122)
0.0183
(0.0703)
-0.440⇤⇤⇤
(0.134)
0.157⇤
(0.0852)
0.301⇤⇤⇤
(0.0994)
2.191⇤⇤⇤
(0.102)
Below median income
0.0300
0.0100
0.370
546
Have children
0.0400
0
0.310
576
Above median CESD
0.0500
0
0.0100
576
Notes: This table reports the coefficient estimates of the interaction between the treatment and a baseline variable.
Each column corresponds to a model with a unique interacting variable. Standard errors are in parentheses. *
denotes significance at 10 pct., ** at 5 pct., and *** at 1 pct. level. We report p-values for joint tests of the
treatment and interaction coefficients.
25
C.
Health and healthcare use
Table 8 shows that free insurance has little to no e↵ect on health. In addition, when
considering the e↵ect of health insurance, rather than free insurance, Columns 3 and 6
show no e↵ect of health insurance on any of the measured health outcomes. In the absence
of liquidity constraints or with the availability of informal insurance, insurance provision
should not change health status. In our sample, households had median savings of USD
77 (compared to a median weekly income of USD 72), and the majority (65 percent) of
respondents reported feeling secure from their savings, suggesting that liquidity constraints
may not have been binding. In line with this view, we find that the provision of health
insurance does not a↵ect health status, a result which is consistent with our conceptual
model as well as previous experimental literature (Ansah et al. 2009; King et al. 2009; Field
et al. 2010). Similar to these studies, we measure endline outcomes one year after the
provision of health insurance, which may be too soon to observe changes in health status.
However, Baicker et al. (2013) find no significant e↵ects of health insurance on objective
health outcomes (hypertension, high cholesterol levels, or diabetes) even after a two year
evaluation period.
26
Table 8: Treatment e↵ects – Health
No Controls
With Controls
(1)
(2)
Insurance
UCT
Sick/injured last mo.
-0.04
(0.04)
0.00
(0.04)
Days missed due to sickness
0.04
(0.19)
Prop. of household sick in past mo.
27
(4)
(5)
(6)
Di↵erence
p-value
(7)
Control Mean
(SD)
(8)
Insurance
UCT
0.33
-0.05
(0.04)
-0.02
(0.04)
0.47
.28
(0.45)
640
-0.11
(0.16)
0.43
-0.00
(0.19)
-0.15
(0.15)
0.42
.46
(1.58)
567
-0.01
(0.04)
-0.03
(0.03)
0.55
-0.01
(0.03)
-0.02
(0.03)
0.64
.26
(0.37)
642
Prop. children in household sick
-0.05
(0.04)
-0.09⇤⇤
(0.04)
0.18
-0.04
(0.04)
-0.08⇤⇤
(0.04)
0.28
.23
(0.35)
526
Hospitalization
-0.04
(0.04)
-0.07⇤
(0.04)
0.36
-0.03
(0.04)
-0.06
(0.04)
0.46
.3
(0.46)
640
Children vaccinated
-0.01
(0.03)
0.01
(0.03)
0.41
-0.01
(0.03)
0.02
(0.03)
0.32
.93
(0.26)
517
Child check-up in 6 mo.
-0.04
(0.06)
-0.10⇤⇤
(0.05)
0.22
-0.03
(0.06)
-0.10⇤
(0.05)
0.22
.39
(0.49)
517
Contribution to hosp. costs
47.34
(71.96)
-6.38
(14.63)
0.46
52.21
(74.42)
7.22
(18.34)
0.49
54.89
(146.17)
637
Nights hospitalized last year
-0.01
(0.27)
-0.29⇤
(0.16)
0.20
-0.01
(0.27)
-0.28⇤
(0.16)
0.21
.4
(2.39)
640
Nights should have been hospitalized
-0.68⇤
(0.38)
-0.70⇤
(0.39)
0.71
-0.67⇤
(0.37)
-0.65⇤
(0.35)
0.72
.75
(6.15)
640
0.55
0.04⇤⇤
0.33
0.55
0.05⇤
0.47
Joint p-value
(3)
Di↵erence
p-value
Sample
N
Notes: This table reports the estimated treatment e↵ect of insurance and UCT on each row variable. Columns 1 - 2 report estimates from an intent-to-treat
analysis without correcting for selection. Columns 4 - 5 report OLS estimates controlling for baseline covariates. Columns 3 and 6 report the p-values for
tests of the equality of the UCT and insurance coefficients. The bottom row reports the p-value for a test of the treatment e↵ect across models using SUR.
Standard errors are in parentheses. * denotes significance at 10 pct., ** at 5 pct., and *** at 1 pct. level.
The limited e↵ects of insurance provision on health status may be partly explained by the
limited e↵ect on healthcare utilization. Almost half of those who enrolled in the insurance
plan made no claims during the study period. The median number of claims made among
those who made at least one claim is 5 (SD = 6.33). None of our measures of healthcare
use, however, are significantly di↵erent between the insurance and control groups. This
finding on healthcare use is consistent with results from previous experiments (King et al.
2009; Field et al. 2010). Furthermore, we observe no di↵erence in health expenditures and
out-of-pocket payments for the insurance group relative to the other groups. These two
facts suggest that there may be barriers to insurance usage that were not addressed by
our design. We provided the insurance coverage free of charge to our respondents, but
healthcare utilization may be further hindered by lack of trust in the insurance company,
incorrect beliefs about the product, or time costs. Future work might address these factors
and assess their e↵ect on healthcare access.
Other recent studies report increases in primary care visits for children (Ansah et al.
2009), preventative care, and hospitalizations (Finkelstein et al. 2012), and number of prescription drugs received (Baicker et al. 2013) as a consequence of insurance coverage. One
crucial di↵erence between these and the present study is that they estimate e↵ects using
populations which have selected into enrollment. In the Oregon Medicaid experiment, the
analytic sample consisted of people who signed up for the lottery for the chance of receiving
coverage. This process may induce significant adverse selection, which is likely much reduced in our design because a random sample of households were o↵ered insurance for free.
As a consequence, we may observe limited e↵ects in comparison to these previous studies,
in which people who select into the insurance plan could be those who stand to benefit the
most from coverage. Indeed, Ansah et al. (2009) find that only households who self-enrolled
into the insurance plan benefitted from increased access to care.
VII.
Conclusion
The existing experimental literature on the provision of free health insurance finds limited
impact on health outcomes, a finding that one would expect for populations without binding
liquidity constraints or functioning informal insurance networks. However, for risk averse
individuals, health insurance should nevertheless confer welfare benefits. What are they and
how do we measure them? In this study we randomly provided free health insurance to a
group of informal metal-workers in Nairobi, Kenya. In doing so, we add to the literature
on this subject in two ways. First, in addition to a standard control group who receive no
intervention, a second treatment group received the market value of the health insurance
premium as a cash transfer. This approach controls for income e↵ects of insurance provision.
Second, we collected a host of outcomes including health, psychological well-being, and
salivary cortisol, a biomarker of stress.
28
As predicted by our conceptual framework, we find that those assigned to receive insurance show large reductions in self-reported stress and cortisol compared to a control group,
while those in the cash group do not. The fact that this finding holds even for individuals
who never use the insurance provides evidence for a “peace of mind” e↵ect of health insurance that extends beyond health outcomes alone. This e↵ect is stronger for individuals who
are poorer, have children, and score highly on the CES-D. Thus, even when health insurance
does not improve health outcomes, it may increase overall welfare through “peace of mind”,
particularly among vulnerable populations.
References
Acharya, Arnab, Sukumar Vellakkal, Fiona Taylor, Edoardo Masset, Ambika Satija, Margaret Burke, and Shah Ebrahim. 2012. “The Impact of Health Insurance Schemes for
the Informal Sector in Low- and Middle-Income Countries: A Systematic Review.” The
World Bank Research Observer, November, lks009.
Adams, Arvil V., Sara Johansson de Silva, and Setareh Razmara. 2013, July. Improving
Skills Development in the Informal Sector: Strategies for Sub-Saharan Africa. World
Bank Publications.
Alcaraz, Carlo, Daniel Chiquiar, Marı́a José Orraca, and Alejandrina Salcedo. 2012. “The
E↵ect of Publicly Provided Health Insurance on Academic Performance in Mexico.”
Working Paper 2012-10, Banco de México.
Anderson, Michael L. 2008. “Multiple Inference and Gender Di↵erences in the E↵ects
of Early Intervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early
Training Projects.” Journal of the American Statistical Association 103 (484): 1481–
1495.
Ansah, Evelyn Korkor, Solomon Narh-Bana, Sabina Asiamah, Vivian Dzordzordzi, Kingsley Biantey, Kakra Dickson, John Owusu Gyapong, Kwadwo Ansah Koram, Brian M
Greenwood, Anne Mills, and Christopher J. M Whitty. 2009. “E↵ect of Removing Direct
Payment for Health Care on Utilisation and Health Outcomes in Ghanaian Children:
A Randomised Controlled Trial.” PLoS Med 6 (1): e1000007 (January).
Baicker, Katherine, Sarah L. Taubman, Heidi L. Allen, Mira Bernstein, Jonathan H. Gruber, Joseph P. Newhouse, Eric C. Schneider, Bill J. Wright, Alan M. Zaslavsky, and
Amy N. Finkelstein. 2013. “The Oregon Experiment - E↵ects of Medicaid on Clinical
Outcomes.” New England Journal of Medicine 368 (18): 1713–1722 (May).
Baird, Sarah, Ephraim Chirwa, Jacobus De Hoop, and Berk Özler. 2014. “Girl Power: Cash
Transfers and Adolescent Welfare. Evidence from Cluster-Randomized Experiment in
Malawi.” In African Successes: Human Capital. University of Chicago Press.
29
Baird, Sarah, Jacobus de Hoop, and Berk Özler. 2013. “Income Shocks and Adolescent
Mental Health.” Journal of Human Resources 48 (2): 370–403.
Baird, Sarah, Craig McIntosh, and Berk Özler. 2011. “Cash or condition? Evidence from
a cash transfer experiment.” The Quarterly Journal of Economics 126 (4): 1709–1753.
Baird, Sarah J., Richard S. Garfein, Craig T. McIntosh, and Berk Özler. 2012. “E↵ect of
a cash transfer programme for schooling on prevalence of HIV and herpes simplex type
2 in Malawi: a cluster randomised trial.” The Lancet 379 (9823): 1320–1329.
Banerjee, Abhijit Vinayak, Roland Bénabou, and Dilip Mookherjee. 2006, May. Understanding Poverty. Oxford University Press.
Barros, Rodrigo. 2008. “Wealthier But Not Much Healthier: E↵ects of a Health Insurance
Program for the Poor in Mexico.” Discussion Paper 09-002, Stanford Institute for
Economic Policy Research.
Blattman, Christopher, Nathan Fiala, and Sebastian Martinez. 2013. “Generating Skilled
Self-Employment in Developing Countries: Experimental Evidence from Uganda*.”
The Quarterly Journal of Economics, December, qjt057.
Blattman, Christopher, Julian Jamison, and Margaret Sheridan. 2015, May. “Reducing Crime and Violence: Experimental Evidence on Adult Noncognitive Investments
in Liberia.” SSRN Scholarly Paper ID 2594868, Social Science Research Network,
Rochester, NY.
Bloom, Erik, Indu Bhushan, David Clingingsmith, Rathavuth Hong, Elizabeth King,
Michael Kremer, Benjamin Loevinsohn, and J. Brad Schwartz. 2006. Contracting
for health: evidence from Cambodia. Brookings Institution.
Bolton, Paul, Judith Bass, Richard Neugebauer, Helen Verdeli, Kathleen F Clougherty,
Priya Wickramaratne, Liesbeth Speelman, Lincoln Ndogoni, and Myrna Weissman.
2003. “Group interpersonal psychotherapy for depression in rural Uganda: a randomized controlled trial.” JAMA: the journal of the American Medical Association 289
(23): 3117–3124 (June).
Brook, Robert H., John E. Ware, William H. Rogers, Emmett B. Keeler, Allyson R. Davies,
Cathy A. Donald, George A. Goldberg, Kathleen N. Lohr, Patricia C. Masthay, and
Joseph P. Newhouse. 1983. “Does Free Care Improve Adults’ Health?” New England
Journal of Medicine 309 (23): 1426–1434 (December).
Cesarini, David, Erik Lindqvist, Robert Ostling, and Bjorn Wallace. 2016. “Estimating
the causal impact of wealth on health: Evidence from the Swedish lottery players.” The
Quarterly Journal of Economics.
Cornelisse, Sandra, Van Ast, Vanessa, Johannes Haushofer, Maayke Seinstra, and Marian Joels. 2013, July. “Time-Dependent E↵ect of Hydrocortisone Administration on
30
Intertemporal Choice.” SSRN Scholarly Paper ID 2294189, Social Science Research
Network, Rochester, NY.
Delaney, Liam, Günther Fink, and Colm P. Harmon. 2014, April. “E↵ects of Stress on
Economic Decision-Making: Evidence from Laboratory Experiments.” SSRN Scholarly
Paper ID 2420705, Social Science Research Network, Rochester, NY.
Devoto, Florencia, Esther Duflo, Pascaline Dupas, William Parient?, and Vincent Pons.
2012. “Happiness on Tap: Piped Water Adoption in Urban Morocco.” American
Economic Journal: Economic Policy 4 (4): 68–99.
Dow, William H., and Kammi K. Schmeer. 2003. “Health insurance and child mortality
in Costa Rica.” Social science & medicine 57 (6): 975–986.
Fernald, L. C. H., and M. R. Gunnar. 2009. “Poverty-alleviation program participation
and salivary cortisol in very low-income children.” Social Science & Medicine 68 (12):
2180–2189.
Field, Erica, Rebecca Thornton, Laurel Hyatt, Mursaleena Islam, and Freddy Solis. 2010.
“Social Security Health Insurance for the Informal Sector in Nicaragua: A Randomized
Evaluation.” Health Economics.
Finkelstein, Amy, Sarah Taubman, Bill Wright, Mira Bernstein, Jonathan Gruber,
Joseph P. Newhouse, Heidi Allen, Katherine Baicker, and the Oregon Health Study
Group. 2012. “The Oregon Health Insurance Experiment: Evidence from the First
Year +.” The Quarterly Journal of Economics, May, qjs020.
Gakidou, Emmanuela, Rafael Lozano, Eduardo Gonzalez-Pier, Jesse Abbott-Klafter,
Jeremy T. Barofsky, Chloe Bryson-Cahn, Dennis M. Feehan, Diana K. Lee, Hector
Hernandez-Llamas, and Christopher JL Murray. 2006. “Assessing the e↵ect of the
2001-06 Mexican health reform: an interim report card.” The Lancet 368 (9550):
1920–1935.
Galiani, Sebastian, Paul J. Gertler, and Raimundo Undurraga. 2015. “The Half-Life of
Happiness: Hedonic Adaptation in the Subjective Well-Being of Poor Slum Dwellers
to a Large Improvement in Housing.” Technical Report, National Bureau of Economic
Research.
Gardner, Jonathan, and Andrew J. Oswald. 2006. “Money and Mental Wellbeing: A Longitudinal Study of Medium-Sized Lottery Wins.” IZA Discussion Paper 2233, Institute
for the Study of Labor (IZA).
Gertler, Paul, and Orville Solon. 2002. “Who benefits from social health insurance?
Evidence from the Philippines.” Unpublished Manuscript, University of California,
Berkeley and the University of the Philippines.
Hammen, C. 2005. “Stress and depression.” Annu. Rev. Clin. Psychol. 1:293–319.
31
Haushofer, Johannes, and Ernst Fehr. 2014. “On the psychology of poverty.” Science 344
(6186): 862–867 (May).
Haushofer, Johannes, and Jeremy Shapiro. 2013. “Household response to income changes:
Evidence from an unconditional cash transfer program in Kenya.” Massachusetts Institute of Technology.
Holsboer, F. 2000. “The corticosteroid receptor hypothesis of depression.” Neuropsychopharmacology 23 (5): 477–501.
Jowett, M, P Contoyannis, and N. D Vinh. 2003. “The impact of public voluntary health
insurance on private health expenditures in Vietnam.” Social Science & Medicine 56
(2): 333–342 (January).
King, Gary, Emmanuela Gakidou, Kosuke Imai, Jason Lakin, Ryan T Moore, Clayton
Nall, Nirmala Ravishankar, Manett Vargas, Martha Marı́a Téllez-Rojo, Juan Eugenio Hernández Ávila, and Maurici Ávila. 2009. “Public policy for the poor? A randomised assessment of the Mexican universal health insurance programme.” The Lancet
373 (9673): 1447–1454 (April).
Kirschbaum, C, and D H Hellhammer. 1989. “Salivary cortisol in psychobiological research:
an overview.” Neuropsychobiology 22 (3): 150–169.
Kling, Je↵rey R., Je↵rey B. Liebman, and Lawrence F. Katz. 2007. “Experimental Analysis
of Neighborhood E↵ects.” Econometrica 75 (1): 83–119.
Mani, Anandi, Sendhil Mullainathan, Eldar Shafir, and Jiaying Zhao. 2013. “Poverty
Impedes Cognitive Function.” Science 341 (6149): 976–980 (August).
Manning, Willard G. 1987. “Health Insurance and the Demand for Medical Care: Evidence
from a Randomized Experiment.” American Economic Review 77 (3): 251–77.
McKenzie, David. 2012. “Beyond baseline and follow-up: The case for more T in experiments.” Journal of Development Economics 99 (2): 210–221.
Michalopoulos, Charles, David Wittenburg, Dina A. R. Israel, and Anne Warren. 2012.
“The E↵ects of Health Care Benefits on Health Care Use and Health: A Randomized
Trial for Disability Insurance Beneficiaries.” Medical Care 50 (9): 764–771 (September).
Porcelli, Anthony J, and Mauricio R Delgado. 2009. “Acute stress modulates risk taking
in financial decision making.” Psychological Science 20 (3): 278–283 (March).
Shah, Anuj K., Sendhil Mullainathan, and Eldar Shafir. 2012. “Some Consequences of
Having Too Little.” Science 338 (6107): 682–685 (November).
Sobocki, P., B. Jönsson, J. Angst, C. Rehnberg, et al. 2006. “Cost of depression in Europe.”
The journal of mental health policy and economics 9 (2): 87.
32
Wagsta↵, Adam. 2007, February. “Health Insurance for the Poor: Initial Impacts of
Vietnam’s Health Care Fund for the Poor.” SSRN Scholarly Paper ID 961760, Social
Science Research Network, Rochester, NY.
Westfall, Peter H., and S. Stanley Young. 1993, January. Resampling-Based Multiple
Testing: Examples and Methods for P-Value Adjustment. John Wiley & Sons.
Zellner, Arnold. 1962. “An Efficient Method of Estimating Seemingly Unrelated Regressions and Tests for Aggregation Bias.” Journal of the American Statistical Association
57 (298): 348–368 (June).
Zwane, Alix, Jonathan Zinman, E. VanDusen, William Pariente, Clair Null, Edward
Miguel, M. Kremer, Dean Karlan, Richard Hornbeck, Xavier Gine, Esther Duflo, Florencia DeVoto, Bruno Crepon, and Banerjee. 2011. “Being surveyed can change later
behavior and related parameter estimates.” Proceedings of the National Academy of
Sciences 108 (5): 1821–1826.
33