new direct and spillover effects on health of increased income for … · 2019. 11. 26. · 1...
TRANSCRIPT
-
Direct and spillover effects on health of increasedincome for the elderly: evidence from a Chilean pension
program ∗
Enrico Miglino† Nicolas Navarrete H.‡ Gonzalo Navarrete H.∓Pablo Navarrete H.?
Abstract
We estimate the effect of a permanent income increase on health outcomes of the elderly
poor and their relatives. Our regression discontinuity design exploits an eligibility cut-off in a
Chilean basic pension programme that grants monthly payments to pensionless retirees. Using
administrative data we find that, four years after applying, basic pension recipients are 2.5
percentage points less likely to die. We observe spillover effects on working-age relatives,
who are more likely to have a newborn child. Results suggest that increasing the incomes of
the elderly could reduce health inequalities across income groups and benefit younger relatives.
Keywords: pension, elderly, spillover effects, mortality.JEL Classification: I14, I38, J14.
∗We thank the Chilean National Institute of Pensions and the Ministry of Health for the data provided. We alsothank Clément de Chaisemartin, Sascha O. Becker, Dolores de la Mata, Gabriel Facchini and seminar participantsat the University of Warwick, UCL, PSE, UAB, Universidad de Chile, Universidad Diego Portales, Universidad Al-berto Hurtado, and the Royal Economic Society for their useful comments. We would also like to thank ClaudioAllende for his outstanding research assistance. The authors gratefully acknowledge the financial support of the De-velopment Bank of Latin America (CAF). † Enrico Miglino: Department of Economics, University College London(email: [email protected]). ‡Nicolas Navarrete H: Department of Economics, Paris School ofEconomics (email: [email protected]). ∓Gonzalo Navarrete H: Department of Public Health,Universidad de Chile ([email protected]). ?Pablo Navarrete H: Department of Geography and Environ-ment, London School of Economics, and Center for Sutainable Urban Development, Pontificia Universidad Católicade Chile (email: [email protected]).
-
1 Introduction
In light of an ageing population, health inequalities amongst the elderly have received increased
attention over the last two decades. Researchers and policymakers have documented large and
ever-widening life expectancy inequalities across income groups in both developed and develop-
ing countries [Hoffman, 2008; Brønnum-Hansen and Baadsgaard, 2012; Tarkiainen et al., 2012;
OECD, 2017]. For instance, an OECD [2018] report shows that, at retirement age, high-income
earners live longer than low-income earners: 1.6 years longer in the US, 3.6 in Chile, 3.25 in the
UK, and 2.9 in South Korea.1
Despite a large body of literature documenting that, at all ages, wealthier people live longer on
average [Marmot, 2005; Braveman et al., 2010; Waldron, 2013; Chetty et al., 2016], substantial
debate remains on whether an income increase in old age can partially reduce these life expectancy
inequalities. It may be the case that unobserved characteristics, such as genetic factors, are able
to explain both higher life expectancy and better health. Alternatively, better health in general,
the underlying cause of higher life expectancy, could itself be the cause of higher income (reverse
causality). Finally, differences in life expectancy may be the result of cumulative conditions related
to income inequalities at earlier ages, such as different education levels or exposure to air pollution.
The non-contributory pension programme in Chile provides an ideal regression discontinuity
design (RDD) that allows us to identify the causal effect of a permanent income increase on the
mortality rates of the elderly, and also on the number of days they spend hospitalised. Chile is
a high-middle income country with a per-capita GDP (PPP) of US$24,644 in 2017,2 and a life
expectancy at age 65 of 19.45 years (similar to the US).3 Since 2008, Chileans who are aged 65 or
over and who do not have a contributory pension can apply to receive a non-contributory pension,
which provides lifelong monthly payments of approximately 40 percent of the national minimum
wage (basic pension). Upon receiving an application, the government calculates a specifically de-
signed pension score and assigns a basic pension to applicants who fall below the 60th percentile
(cut-off) of the score distribution.
We obtained access to administrative information on the universe of basic pension applicants
and their household members in 2011 and 2012, and matched it with their health history between
1In the report, high-income earners are defined as those who earn more than three times the average wage andlow-income earners are defined as those who earn half of the average wage or less.
2https://data.worldbank.org/indicator3https://data.oecd.org/healthstat/life-expectancy-at-65.htm
2
https://data.worldbank.org/indicator https://data.oecd.org/healthstat/life-expectancy-at-65.htm
-
2011 and 2016.4 We first observed that the pool of applicants consists mostly of women without a
regular employment history. As individuals can apply multiple times, we defined applicants whose
first application score fell below the cut-off, and within a certain bandwidth, as our ‘treatment
group’. Conversely, we defined applicants whose first application score was above the cut-off,
and within a certain bandwidth, as our ‘control group’. We document that nearly all applicants
in the treatment group obtained a basic pension, and that 18 percent of applicants in the control
group submitted another application at a later time that was successful. Therefore, being in the
treatment group increases the probability of receiving a basic pension within four years of the first
application by 82 percent. We also show that density and balance tests cannot reject the hypothesis
that the basic pension is as good as ‘locally’ randomly assigned between treatment and control
group applicants. Following these findings, we implement a fuzzy regression discontinuity design
to explore the causal effect of the pension on mortality and days of hospitalisation.
We find that receiving a basic pension decreases the probability of dying within four years af-
ter applying to the programme by 2.5 percentage points (pp.). This represents 33.8 percent fewer
deaths for those who receive the pension. The decrease is statistically significant and remains
unaffected when using nonparametric estimations, and different sets of controls, bandwidths, or
polynomial orders. As the basic pension increases applicants’ income by 63.1 percent for those in
the bandwidth, we are able to estimate a negative mortality-income elasticity of 0.54.
A survival analysis shows that the effect on mortality becomes visible one year after the first
payment and grows monotonically over time. Improvements in health outcomes appear to be driven
by fewer acute incidents of circulatory and respiratory diseases associated with chronic conditions,
such as heart attacks in patients with diabetes or acute respiratory problems in patients with se-
vere asthma. We provide evidence suggesting that treatment group applicants spend their pension
on goods that improve their overall health (e.g. nutritional food) and thus prevent acute medical
episodes from occurring.5
The heterogeneous analysis shows that applicants living alone or with only elderly household
members are significantly less likely to die and spend fewer days hospitalised, while those living
with working-age household members remain unaffected. An exploratory analysis suggests that
the lack of effects for the latter might be explained by ‘crowding-out’ effects: there exist intra-4The programme did not systematically collect information on applicants and their household members prior to
2011, making it unfeasible to conduct the analysis on earlier years.5In Chile, public health insurance grants free access to medical appointments, medication, and surgery for the
most common diseases in the country. The coverage is almost universal amongst the elderly population without acontributory pension.
3
-
household transfers of income from working-age household members to applicants in the absence
of the pension, which then cease when applicants begin receiving the basic pension.
We also document the presence of spillover effects on working age household members, who
are 3.6 pp. more likely to have a newborn child nine months or later after the application was
made. These results are also in line with the presence of ‘crowding-out’ effects that have been
well-documented in the literature on government transfers [Cox and Jimenez, 1992; Jensen, 2003;
Edmonds, 2006].
Finally, we provide a ‘back-of-the-envelope’ computation showing that the basic pension may
be able to correct around 29 percent of the life expectancy gap between the income group of treated
applicants at the cut-off and the income group that they reach after receiving the basic pension. This
suggests that a relevant part of life expectancy inequalities across income groups in old age are due
to contemporaneous income inequalities.
Our paper makes two main contributions. First, we provide causal evidence that a permanent
income increase for the elderly can reduce their mortality rates at the present time. Salm [2011]
finds that two pension increases that occurred in the early 1900s reduced the mortality rates of
US veterans. This effect was driven by infectious diseases, decades before the medical revolu-
tion of antibiotics [Hare, 1982]. Whether or not these results hold true a century later is unclear.
Cheng et al. [2016] find that a Chinese rural pension scheme, which represents 12-17 percent of
the average income per capita, leads to an insignificant reduction in mortality.6
In contrast with these studies, Snyder and Evans [2006] estimate that a notch in US social secu-
rity payments for the cohorts 1916-1917, which reduced the later cohort’s income by 4 percent on
average, significantly reduced men’s mortality rates with respect to the wealthier cohort. They jus-
tify this result by showing that the poorer cohort undertook more part-time work, suggesting that
a later transition to retirement reduces social isolation and improves health. Feeney [2017, 2018]
exploits the age eligibility cut-off and the staggered introduction of a Mexican non-contributory
pension across small rural towns, finding that this pension increases recipients’ transition to re-
tirement, consumption of ‘obesogenic’ food, and mortality rates; but reduces tobacco and alcohol
consumption. Snyder and Evans’s and Feeney’s results can be reconciled with our findings, consid-
6Lindahl [2005], Cesarini et al. [2016], and Schwandt [2018] show mixed results regarding the impact of increasesin wealth, such as lottery prizes, on mortality rates amongst the elderly. Jensen and Richter [2003] consider the effectof payment interruptions during the Russian pension crisis and found an increase in mortality. Although these studiesdo belong to a related literature, the effects of unexpected wealth increases or temporary income reductions might notmirror the effect of a permanent income increase guaranteed by the government.
4
-
ering that the basic pension in Chile is given only to people without a regular employment history.
Therefore, the Chilean basic pension does not cause a break from working routine that may prove
detrimental to mental and physical health. Changes in consumption patterns when people exit the
labour force could be relevant in explaining the different impacts found in this literature [Browning
and Meghir, 1991; Zantinge et al., 2013].
Our second contribution is to provide causal evidence that a permanent income increase for
the elderly poor can have spillover effects on the fertility of working-age relatives. We are not
aware of previous papers testing this directly using a regression discontinuity design and admin-
istrative data. This result complements previous findings regarding the spillover benefits of non-
contributory pensions on children’s height, weight, school enrolment, and attendance [Duflo, 2000,
2003; Edmonds, 2006]; and on working-age relatives’ self-reported nutrition, sanitation, and em-
ployment [Case, 2004; Case and Menendez, 2007; Ardington et al., 2009]. Finally, our discussion
of ‘crowding-out’ effects suggests that the elderly poor can impose a financial burden on younger
relatives. This could be one of the channels whereby poverty is transmitted across generations, and
thus pension policies may have the potential to alleviate this transmission.
The paper is organised as follows. Section 2 presents the basic pension programme and Sec-
tion 3 describes the data used. Section 4 explains the empirical strategy and analyses the regression
discontinuity validity. Section 5 presents the results, and Section 6 explores potential mechanisms.
Section 7 presents our ‘back-of-the-envelope’ computation and Section 8 concludes.
2 The basic pension
Since 1980, Chile has had a fully funded individual capitalisation system, run by private pension
fund companies, in which workers make monthly mandatory savings of 10 percent of their wage.
Once retired, workers receive a pension that is dependent on the amount saved over the course
of their working life (contributory pension). Until recently, those who had never undertaken paid
work received no pension.
During the 2005 presidential campaign, this pension system was judged to be particularly unfair
to stay-at-home mothers, with their role being recognised as unpaid yet central to a fully function-
ing society in Chile. As a result of this debate, most of the candidates, including future president
Michelle Bachelet, agreed to grant a non-contributory government-funded pension to stay-at-home
mothers in the poorest households.
5
-
On March 11th, 2008, President Bachelet signed ACT 20255, which established that, not only
stay-at-home mothers, but every citizen aged 65 or above with no retirement savings would be
eligible for a pension consisting on lifelong monthly payments provided by the government (basic
pension). The introduction of the basic pension took place across all of Chile simultaneously, and
the first payments were delivered on July 1st, 2008. Between 2011 and 2016, our period of analysis,
basic pension payments have been US$160 per month on average, which is equivalent to approx-
imately 40 percent of the national minimum wage. This corresponds to a 63.1 percent increase in
our treatment group’s income [OECD, 2011; Ministerio Trabajo y Previsión Social, 2011].7
To obtain the basic pension, individuals must first apply to the Pension Institute (PI). The ap-
plication process is free and involves filling in a form in the municipality where the applicant lives.
Following the application, the PI calculates a pension score, specifically designed to assign the ba-
sic pension. The pension score comprises two factors: household income from assets (e.g. contrib-
utory pensions from household relatives) and labour income from all household members. Admin-
istrative records show that these two factors account for 60 and 40 percent of total wealth, respec-
tively. The pension score is then adjusted for household size and the number of disabled household
members (for more details on the basic pension and score calculation, see Appendix Section A).
To identify households, and compute the factors mentioned above, the PI follows the same
definition of a household used by all Chilean government agencies: a group of people, whether
relatives or not, who live in the same house and share income. Households are registered by the
Ministry of Planning (MP). The MP ensures that every person belongs to one household only,
and visits the residence of each household to verify that each registered person lives there. The
household records are then provided to the PI, along with each person’s unique national identifier
number, so that the PI can collect information about each household member and compute the
factors affecting the pension score.
The pension score uses a more comprehensive set of information, and is computed in a dif-
ferent way, than other governmental indices such as the social security score.8 The pension score
complements public agencies (e.g. Tax Service) and self-reported information, with administrative
7To obtain applicants’ income, we divide each applicant’s household income by the number of household mem-bers. Household income per capita is equivalent to an applicant’s income under the assumption of perfect householdincome pooling. Every year, the government increases the basic pension payments to at least match inflation.
8The social security score (“Puntaje de la Ficha de Protección Social”) is a proxy means test based on potentialand self-reported actual income, disability status, and household composition that allows the government to assignsocial benefits. The social security score does not use administrative data on labour income or income from othersources such as contributory pensions.
6
-
data from private companies (e.g. pension fund companies and banks). As constructing the pen-
sion score requires the coordination of several public and private offices, it is calculated only for
people who apply.
Following the assigning of pension scores, the PI uses an arbitrary cut-off to determine ba-
sic pension recipients. This cut-off has changed progressively over time, with changes occurring
simultaneously nationwide: it has increased from covering the poorest 40 percent of the elderly
population in July 2008, to covering the poorest 60 percent since July 2011 (Appendix Table B1
shows the cut-off by date). To determine the 60th percentile for the Chilean population in 2011, the
PI used data from the national household survey and estimated a pension score for each household
in the survey.9 The cut-off then corresponds to the 60th percentile of the estimated pension score
for the sample of households in the survey. There have been no updates to the pension score cut-off
since 2011, when it was estimated to be 1,206 pension score points.
After the decision has been made, applicants know whether or not they will receive the basic
pension, but they do not have access to the score assigned during the application process. The
government considered reassessing recipients’ eligibility every two years. However, this proposal
was never implemented, and virtually all of those who obtained the basic pension have continued
receiving payments every month thereafter.10 Moreover, according to PI records, 96 percent of
recipients in our sample collect their pension in person, by showing their identity card in a bank or
PI office.
Overall, the majority of the elderly population who do not receive a contributory pension apply
to obtain a basic pension. In 2011, 64.5 percent of retirees without contributory pensions received
a basic pension [Ministerio de Desarrollo Social, 2011], and 8 percent have submitted an unsuc-
cessful application according to our records (for more details on the elderly Chilean population
who do not have contributory pensions, see Appendix Table B10).
9The PI could not use the population distribution of pension scores to set the cut-off, because pension scores arecomputed only for the fraction of households that apply for the basic pension. Therefore, data must be drawn fromthe national household survey to estimate the pension score associated with the 60th poorest percentile of the entirepopulation of Chilean households.
10Less than 0.05 percent of recipients who obtained the basic pension between 2008 and 2015 stopped receiving itat some point [Subsecretarı́a de Previsión Social, 2015]. All of these were for reasons unrelated to the pension score(e.g. moving to another country).
7
-
3 Data and descriptive statistics
3.1 Pension and health datasets
Our analysis is based on administrative data provided by the Chilean government. First, we have
access to all applications made in 2011 and 2012. For each applicant and each applicant’s house-
hold member, the PI provided us with information on gender, age, town of residency, household so-
cial security score, national unique identifier number (henceforth ID number), and national unique
household identifier number. The applicant’s ID number allows us to observe who the pension
applicant is in each household, whereas the household identifier number allows us to perfectly
match each applicant with all of their household members. This dataset also includes application
data, such as the pension score, application date, and the outcome of the application.11 All of the
variables provided by the PI were collected at the moment of application. Unfortunately, we do not
have access to applicants’ data from previous years, as it was not systematically recorded before
2011.
The second dataset we use is provided by the Ministry of Health and covers the health history
of every Chilean citizen, for each year from 2011 to 2016. This dataset contains information about
citizens who died, the date of death, and the cause of death (ICD-10 disease code).12 It also pro-
vides information about any type of vaccinations given to each person and the date given, as well
as any hospitalisation that occurred, together with its date, duration, and cause (ICD-10 disease
code). The Ministry of Health also granted us access to the unique national ID number of each
person recorded in this dataset. As all physicians and health centres have the legal obligation to
report this information to the central government on a monthly basis, the dataset covers people
receiving medical attention in both private and public health institutions.
Finally, the Ministry of Health also granted us access to a dataset covering all births that oc-
curred in Chile between 2011 and 2016. This dataset contains the date of birth, along with the
weeks of pregnancy, height, and weight of each newborn, as well as the unique national ID num-
ber of the mother. This data also covers births in public and private health centres.
As all datasets contain the ID number of each person, we were able to merge the PI and the
health data at the level of individuals. We performed several checks to ensure that no observation
11We also have household-level data on the factors that determine the pension score for 2012, the year in which thePI started to systematically record them. Along with these, the PI also provided us with other household characteristicsthat they started to record systematically in 2012.
12https://www.who.int/classifications/icd/icdonlineversions/en/
8
https://www.who.int/classifications/icd/icdonlineversions/en/
-
was lost in this process.
We restrict our attention to applications submitted between July 1, 2011 and December 31,
2012. We do not use applications from earlier in 2011 because the 60 percent cut-off point for
eligibility was introduced by the government in July 2011 (Section 2). Given the fact that the most
recent health data to which we have access is from December 2016, this allows us to measure
health outcomes for up to four years from the date of application.
We focus our main analysis on basic pension applicants and then conduct a separate analysis
for their household members. In Chile, the minimum legal age to claim private pension benefits
is 65 for men and 60 for women, and the minimum legal working age is 15. We thus define three
exclusive groups of household members, based on household members’ age: 1) men above 64 and
women above 59 years of age (elderly); 2) men aged 16-64 and women aged 16-59 years (working
age); and, 3) individuals below 16 years of age (school-age children). Given the small number of
observations in this last group of household members (931), we focus the analysis on the first two
groups.
Finally, we observe that 21.2 percent of those who did not receive the basic pension in their first
application applied more than one time. To avoid multiple-counting that could potentially bias our
estimates, we keep in the sample only the first application made by each applicant in our period
of interest.13 Since the focus of our paper is to use the basic pension to analyse the effect of an
income increase on mortality rates and days of hospitalisation of the elderly, we accommodate later
changes in pension holder status using the ‘fuzzy’ regression discontinuity framework presented
in the next section. In the final sample we have 49,552 applicants, 44,053 working-age household
members, and 27,057 elderly household members.
3.2 Descriptive statistics
Table I reports descriptive statistics for applicants, working-age household members, and elderly
household members at the moment of their first application. Columns (1), (2), and (3) report
the mean of the covariates for all applicants, for those who received the basic pension in their
13As those submitting more than one application are self-selected, counting each repeated application separatelycould bias our estimates of the treatment effect. First, it would give an excessive weight to applicants that appliedmore than one time. Second, if the type of applicants that submit more than one application are somehow differentfrom the rest (for example, more motivated), we would give an excessive weight to this particular type of applicant.Also, there could be multiple applicants within a household. We observe that less than 1 percent of applicants in ourworking sample share a household with another applicant.
9
-
Table I: Characteristics of applicants, and their household members, at the moment of application
All applications Bandwidthsample
All Below cut-off Above cut-off P-value(1) (2) (3) (4) (5)
Panel A: applicants
Female 0.758 0.738 0.889 0.000 0.871Age (years) 66.19 65.91 68.02 0.000 66.85Household size 2.482 2.460 2.629 0.000 2.643Elderly cohabitant 0.533 0.503 0.738 0.000 0.661Working-age cohabitant 0.532 0.534 0.515 0.005 0.571Social security score 6969.9 6475.5 10289.4 0.000 9385.7Days hospitalised 0.579 0.604 0.410 0.106 0.504Influenza vaccination 0.273 0.262 0.342 0.000 0.320Pneumonia vaccination 0.0635 0.0657 0.0494 0.000 0.0609Urban town 0.686 0.672 0.785 0.000 0.762Submitted >1 application 0.028 0.000 0.213 0.000 0.090Received a basic pension 0.882 0.999 0.0953 0.000 0.699Observations 49,552 43,129 6,423 8,499
Panel B: working-age household members
Female 0.410 0.414 0.385 0.000 0.363Age (years) 38.80 38.59 40.29 0.000 40.36Household size 3.726 3.720 3.768 0.015 3.685Elderly cohabitant 0.434 0.417 0.553 0.000 0.470Days hospitalised 0.261 0.256 0.292 0.458 0.249Observations 44,053 38,589 5,464 8,047
Panel C: elderly household members
Female 0.242 0.270 0.112 0.000 0.120Age (years) 71.22 71.11 71.73 0.000 71.07Household size 2.883 2.917 2.725 0.000 2.749Working-age cohabitant 0.459 0.469 0.409 0.000 0.434Days hospitalised 0.577 0.601 0.465 0.079 0.490Observations 27,057 22,234 4,823 5,722
Notes: This table reports the mean of several covariates for applicants and their household members. Column (1)reports means for the whole sample, Column (2) reports means for applicants who received the basic pension intheir first application and Column (3) reports means for applicants who did not receive the basic pension in theirfirst application. Column (4) reports the p-value of the mean difference between Columns (2) and (3). Column(5) reports means for the working sample in the bandwidth. Health covariates are computed 6 months beforeapplicants submit their first application.
10
-
first application (below cut-off), and for those who did not receive the basic pension in their first
application (above cut-off), respectively. Column (4) reports the p-value of the mean difference
between applicants below and above the cut-off, and Column (5) displays means for the applicants
in a 500-score-point bandwidth from the cut-off.
This table shows that 88.2 percent of applicants have their first application approved and 75.8
percent of all applicants are female. This large share of female applicants could be the result of
women being less likely to have private pension savings and suggests that the basic pension pro-
gramme was successful in targeting women as its main beneficiaries. We also observe that the
share of women above the cut-off is higher. A potential explanation for this pattern is that women
with no retirement savings may be more likely to have a partner who has some source of income
(such as a contributory pension) than men with no retirement savings, which would make women
more likely to have a pension score above the cut-off.
The average applicant’s age is around 66, which suggests that applications are submitted shortly
after reaching the minimum application age (83 percent are 65 years old). Applicants above the
cut-off are around 2 years older than those below. This could be explained by older people with
unsuccessful applications deciding to reapply after the change in the threshold, only to receive the
same result.
The average pension applicant is below the 30th percentile of the social security score distri-
bution, an indication that applicants are poorer than the median Chilean.14 Applicants above the
cut-off are on average around the 40th percentile, and those below the cut-off are far lower than
them in the distribution. Even though the pension score cut-off is set at the 60th percentile, the
average social security score for applicants close to the cut-off (bandwidth sample) is well below
the 60th percentile. This is likely to be caused by the pension score incorporating administrative
information and factors that do not form the social security score, such as income from household
relatives’ contributory pensions, which may constitute a significant portion of household income.
Looking at the health characteristics of applicants we observe that, over the six months prior to
their application, they spend approximately half a day hospitalised, with no significant difference
between applicants below and above the cut-off.
Applicants within the bandwidth have a higher social security score and are more likely to be
female (87 percent), and to have an elderly and working-age household member (66 and 57 per-
14According to the Ministry of Social Development [Ministerio de Desarrollo Social, 2010], the 20th percentileequates to 6,036 points, the 30th percentile to 8,500, and the 40th percentile to 10,320 points of the social security score.
11
-
cent respectively) than the average applicant. In Section 4.4 we show that, within the bandwidth,
applicants below and above the cut-off are comparable.
There are slightly more male than female working-age household members and, on average,
they are 39 years old, which could indicate that they are the applicants’ children. On the other
hand, elderly household members are five years older than applicants, mostly male (75.8 percent)
and, on average, spend half a day hospitalised. These characteristics seem to indicate that most
elderly household members are the husbands of female applicants.
3.3 Serial applicants
Table I shows that a small fraction of applicants below the cut-off did not receive the basic pen-
sion. This is explained by reasons unrelated to the pension score (e.g. not redeeming the pension
in time). This table also shows that 21.3 percent of applicants above the cut-off submitted a further
application (henceforth referred to as serial applicants), and 9.5 percent of applicants above the
cut-off obtained a basic pension at a later time.
To analyse the characteristics of serial applicants, we regress an indicator for whether the per-
son is a serial applicant against baseline covariates. Column 1 of Table II presents a series of
bivariate regressions in which each baseline characteristic is entered separately, while columns 2,
3, and 4 show estimations that regress on multiple covariates simultaneously. This table shows that
applicants above the cut-off who are older and have a higher social security score are less likely
to be serial applicants, while those in a larger household are more likely. This could be because:
1) older applicants might perceive a lower present value of basic pension income (they expect to
live for a shorter time); and, 2) wealthier people see themselves as less likely to obtain the pen-
sion. In contrast, people in larger families might be more likely to see changes in their household
composition or income, which could affect their pension score and encourage them to reapply.15
15In Appendix Table B2, we analyse the characteristics of serial applicants who eventually obtain the basic pension.We show that serial applicants with a higher social security score are less likely to eventually obtain the basic pensionfor the reasons mentioned above. Living with an elderly relative also reduces the probability of obtaining the basicpension later, likely due to elderly relatives having the least volatile source of income: a contributory pension.
12
-
Table II: The effect of baseline covariates on the probability of applying multiple times
(1) (2) (3) (4)
Female -0.034 0.031 0.031 0.026(0.016) (0.016) (0.017) (0.015)
Age (years) -0.012 -0.009 -0.009 -0.006(0.001) (0.001) (0.001) (0.001)
Social security score -0.023 -0.021 -0.022 -0.019(0.001) (0.001) (0.001) (0.001)
Days hospitalised -0.001 0.001 0.001 0.001(0.002) (0.002) (0.002) (0.001)
Received influenza vaccination 0.050 0.069 0.070 0.018(0.011) (0.011) (0.011) (0.011)
Received pneumonia vaccination -0.034 -0.097 -0.098 -0.005(0.024) (0.024) (0.024) (0.021)
Household size 0.014 0.021 0.023(0.005) (0.007) (0.006)
Elderly cohabitant -0.055 -0.017 -0.012(0.012) (0.014) (0.012)
Working-age cohabitant 0.037 -0.010 -0.021(0.010) (0.016) (0.014)
Live with child under 16 0.074 0.004 -0.045(0.052) (0.051) (0.048)
FIXED EFFECTS NO NO NO YESN 6,423 6,423 6,423 6,423
Notes: Using the sample of all applicants above the cut-off, this table reports results fromOLS regressions of a binary indicator equal to 1 if the individual applied for the basicpension more than once (and 0 otherwise) on several covariates. Column (1) reports coef-ficients of bivariate regressions. Columns (2), (3) and (4) report coefficients of multivariateregressions on the specified variables. Fixed effects are at the month-of-application andthe health-district level. Standard errors are clustered at the province level. For ease ofinterpretation, the social security score is rescaled (divided by 1,000).
13
-
4 Empirical strategy and RD validity
Since most applicants below the cut-off obtained their basic pension, and some applicants above
the cut-off managed to obtain a basic pension at a later time, we can predict that the cut-off discon-
tinuously changes the probability of receiving a basic pension, but not from zero to one hundred
percent. Following this, we implement a fuzzy regression discontinuity (FRD) design to estimate
the causal effect of a basic pension on health outcomes.
4.1 Fuzzy regression discontinuity design
To estimate the causal effect of the basic pension on each health outcome, we perform a two-stage
least square regression. The set of equations is as follows:16
Health Outcomei,h = f 0(Scoreh) + βLATEPensionh + Pensionh × f 1(Scoreh) + ui,h, (1)
where Pensionh is instrumented using:
Pensionh = g0(Scoreh) + βDh + Dh × g1(Scoreh) + vh, (2)
and Dh is a dummy variable defined as follows:
Dh =
1 if Scoreh ≤ 00 if Scoreh > 0
In this set of equations, Health Outcomei,h is one of the health outcomes described in Section
3.1, four years after applying, for person i in household h. Pensionh is a dummy indicator equal to
1 if the applicant of household h has received a basic pension within four years of the first applica-
tion, and 0 otherwise. Dh is a dummy indicator equal to 1 if the applicant of household h obtained
16Note that in these equations the score is centred at the cut-off to ensure that the treatment effect is the coefficientβLATE .
14
-
a pension score below the cut-off in their first application, and 0 otherwise. Scoreh is the distance
of the first application score from the cut-off point, for the pension applicant of household h. f j and
gj , with j = 0, 1, are different functional forms of Scoreh. In our preferred specification, f j and
gj , are polynomials of order 1 in Scoreh. We check the robustness of our results to different spec-
ifications by estimating equations 1 and 2 using polynomials of order 2 in Scoreh, nonparametric
estimations, logistic regression, and different sets of controls.
In each regression, we use triangular kernels such that the weight of each observation decreases
with the distance from the cut-off. The sample is restricted to a bandwidth of 500 points on either
side of the threshold. As a robustness check, we also repeat our analysis using the mean-squared
error optimal bandwidth approach as proposed by Calonico et al. [2014]. Standard errors are clus-
tered at the province level.17
In equation 1, βLATE captures the Local Average Treatment Effect (LATE) of a permanent in-
crease in monthly income on the health outcomes of applicants and household members close to
the cut-off.
4.2 The effect of the pension score on receiving the basic pension
Figure I displays the probability of receiving a basic pension in the four years following the first
application, for various pension score distances from the cut-off (the ‘First Stage’). This figure
shows that applicants in the bandwidth with a score below the cut-off in their first application
(treatment group) are more likely to receive a basic pension in the following four years than those
whose score is just above the cut-off (control group). Column (1) of Table III confirms this result,
and shows that being in the treatment group increases the likelihood of receiving a basic pension
in the next four years by 82 pp. (p-value< 0.001). For reasons explained in Section 3.3, Figure I
also shows that some applicants did not obtain the basic pension in their first application, but still
went on to receive it at a later time.17There are 33 health districts and 54 provinces in Chile. The standard errors are clustered at the province level in
our preferred specification, because provinces serve as a good proxy of health districts, but their number is sufficientlyhigh to employ the law of large numbers and make correct use of clustered standard errors. Clustering at the health-district level does not change the results of our estimates.
15
-
Figure I: First-stage graph
0.2
.4.6
.81
With
in-b
in a
vera
ge
-500 0 500Distance of the pension score from threshold
Mean bin Quadratic fit Confidence interval
Notes: Each circle represents the average probability of obtaining the basic pension, four years after the firstapplication, for the applicants in each 50 score-point bin.
4.3 Continuity of applicants’ frequency around the cut-off
Identification of the treatment effect requires that applicants cannot manipulate their pension score
in order to receive the basic pension. This assumption would fail if, for instance, more motivated
applicants, who happened to be healthier, were able to adjust their pension score to fall below the
cut-off.
To formally confirm the absence of manipulation, we perform the test developed by McCrary
[2008]. Here, we use the density of applicants in 10 score-point bins as the dependent variable in
equation 2. As shown in Figure II, the test does not reject the null hypothesis of no discontinuity
in the density of applicants (t-statistic of -1.019 and p-value of 0.309). Appendix Figure C1 also
shows no discontinuity in the density of applicants’ working-age household members (t-statistic of
-0.013 and p-value of 0.999) and elderly household members (t-statistic of -1.576 and p-value of
0.115) at the cut-off.
This result is not surprising given that the pension score is not computed until an individual
16
-
Table III: First stage
Variables RD Coef. S.E. t-stat P-value BW N Control(1) (2) (3) (4) (5) (6) (7)
Received a basic pension 0.820 (0.011) 72.946 0.000 500 8,499 0.181
Notes: This table reports results from an OLS regression of a dummy indicator equal to 1 if the applicant receivesthe basic pension four years after their first application (and 0 otherwise) on a treatment dummy indicator anddeviation of the pension score from the cut-off. Column (1) reports the treatment indicator coefficient and Column(2) reports the standard error of this coefficient, clustered at the province level. Column (3) reports the t-statisticof the treatment dummy indicator coefficient, and Column (4) reports the p-value of this coefficient. Column(5) indicates the range of pension score points from the cut-off in which regressions are performed. Column (6)reports the number of observations used in the regression. Column (7) reports the fraction of control applicantsat the cut-off who received the pension.
applies, so the applicant does not know the score beforehand. This piece of evidence suggests that
applicants do not manipulate their first application score in order to receive the pension.
4.4 Continuity of predetermined covariates around the cut-off
In order to have comparable treatment and control groups in the RD design, a series of pre-
determined characteristics that could affect applicants’ health should change smoothly at the cut-
off [Lee and Lemieux, 2010]. Appendix Figures C2 and C3 graphically demonstrate that pre-
determined covariates do indeed vary smoothly at the cut-off for applicants. Column (3) of Table
IV reports the results of the t-test performed on the coefficient β, in equation 2, using as a de-
pendent variable one of the 11 individual and household characteristics at the time of application.
Panel A of this table confirms the results and shows that only 1 out of the 11 estimations (pneu-
monia vaccinations) is significant at conventional levels for applicants. We do not believe that this
represents a systematic difference between treatment and control groups around the cut-off: given
the large number of estimations, it is common to have one significant estimation at this significance
level. For the covariates used to compute the pension score, we only have data for applicants in
2012. Appendix Table B3 reports balancing checks using these covariates, and shows that only
1 out of 11 estimations is significant at the 10 percent significance level. The evidence presented
above suggests that the basic pension is as good as (locally) randomly assigned, after conditioning
on pension score deviations from the cut-off.
As we also explore the effect of the pension on household members living with applicants at
the moment of application, we examine whether the same covariates change smoothly at the cut-
off for working-age and elderly relatives. Panel B of Table IV shows that 1 out of the 11 available
17
-
Table IV: Balancing tests
Variables RD Coef. S.E. t-stat P-value BW N Control(1) (2) (3) (4) (5) (6) (7)
Panel A: applicants
Female -0.016 (0.015) -1.016 0.314 500 8,499 0.890Age (years) -0.372 (0.236) -1.578 0.121 500 8,499 67.57Days hospitalised -0.131 (0.154) -0.851 0.399 500 8,499 0.458Influenza vaccination -0.025 (0.020) -1.281 0.206 500 8,499 0.357Pneumonia vaccination 0.017 (0.008) 2.019 0.049 500 8,499 0.043Household size -0.008 (0.040) -0.192 0.849 500 8,499 2.634Social security score 64.687 (181.386) 0.357 0.723 500 8,499 9737.Elderly cohabitant 0.016 (0.018) 0.872 0.387 500 8,499 0.693Working-age cohabitant -0.004 (0.018) -0.214 0.832 500 8,499 0.548Live with child under 16 0.002 (0.004) 0.396 0.694 500 8,499 0.006Municipal income -2.465 (4.250) -0.580 0.564 500 8,483 146.7
Panel B: working-age household members
Female 0.030 (0.024) 1.240 0.221 500 8,047 0.358Age (years) -1.090 (0.656) -1.661 0.103 500 8,047 40.96Days hospitalised -0.046 (0.062) -0.737 0.465 500 8,047 0.172Influenza vaccination -0.015 (0.012) -1.204 0.235 500 8,047 0.094Pneumonia vaccination -0.001 (0.003) -0.271 0.788 500 8,047 0.004Number of children born 0.006 (0.005) 1.062 0.293 500 8,047 0.008Household size 0.007 (0.060) 0.121 0.904 500 4,836 3.228Social security score 147.319 (261.230) 0.564 0.575 500 4,836 9857.Elderly cohabitant 0.047 (0.026) 1.767 0.084 500 4,836 0.525Live with child under 16 0.007 (0.006) 1.054 0.297 500 4,836 0.007Municipal income -8.321 (5.181) -1.606 0.115 500 4,828 150.0
Panel C: elderly household members
Female 0.032 (0.016) 2.016 0.049 500 5,722 0.097Age (years) -0.608 (0.358) -1.702 0.095 500 5,722 71.82Days hospitalised -0.038 (0.093) -0.404 0.688 500 5,722 0.312Influenza vaccination -0.026 (0.029) -0.899 0.373 500 5,722 0.364Pneumonia vaccination 0.001 (0.006) 0.083 0.934 500 5,722 0.019Household size 0.050 (0.050) 1.003 0.321 500 5,566 2.679Social security score 96.419 (199.801) 0.483 0.632 500 5,566 1.0e+Working-age cohabitant 0.027 (0.024) 1.147 0.257 500 5,566 0.412Live with child under 16 -0.000 (0.006) -0.044 0.965 500 5,566 0.009Municipal income -2.603 (5.244) -0.496 0.622 500 5,558 147.4
Notes: This table reports results from OLS regressions of pre-determined variables on a treatment dummy indicatorand deviation of the pension score from the cut-off. Columns (1), (2), (3), and (4) report the treatment indicatorcoefficient, its standard error clustered at the province level, the t-statistic, and the p-value, respectively. Columns(5) and (6) report the range of pension score points from the cut-off and the number of observations in the regression,respectively. Column (7) reports the variable mean for control applicants at the cut-off. Health covariates are defined6 months before applying.
18
-
Figure II: McCrary test of applicants
0.0
005
.001
.001
5.0
02D
ensi
ty
-1000 -500 0 500 1000Distance of the pension score from threshold
Notes: This figure shows the density of applicants in 10 score-point bins. The solid line plots fitted values from alocal linear regression of density on the pension score, separately estimated on both sides of the cut-off. The thin linesrepresent the 95% confidence interval.
covariates (elderly cohabitant) is significant for working-age household members, which is in the
expected range of significant coefficients at this significance level. However, Panel C of Table IV
shows that 2 out of the 10 available covariates (female and age) are statistically significant among
elderly household members. This is higher than the expected level. We address the implications of
these imbalances in Section 5 by adding these variables as controls, and show that our results on
elderly household members do not change.
5 Results
5.1 The effect of receiving a pension on applicants’ health
The top left panel of Figure III shows the causal effect of receiving a basic pension from the first
application on the probability of dying within four years after applying (henceforth referred to as
19
-
Figure III: Effect of the basic pension on mortality, cumulative days of hospitalisation and medicalepisodes of applicants
-.10
.1.2
With
in-b
in a
vera
ge
-500 0 500Distance of the pension score from threshold
Mean bin Quadratic fit Confidence interval
Mortality rate
.72.
74.
76.
7W
ithin
-bin
ave
rage
-500 0 500Distance of the pension score from threshold
Mean bin Quadratic fit Confidence interval
Days hospitalised
.2.3
.4.5
With
in-b
in a
vera
ge
-500 0 500Distance of the pension score from threshold
Mean bin Quadratic fit Confidence interval
Medical episode
Notes: Each graph shows the average value of the corresponding variable conditional on the distance of the pensionscore from the cut-off. The circles represent averages across 50-point bins on either side of the threshold, while thesolid and dashed lines represent the predicted values and confidence interval, respectively.
mortality). This panel indicates that applicants in the treatment group are less likely to die than
applicants in the control group. Column (1) of Table V confirms this result and shows that receiv-
ing a basic pension significantly decreases the probability of dying by 2.5 pp. (p-value=0.034).
As the basic pension increases applicants’ income by 63.1 percent in our control group, and the
average probability of dying for control group applicants at the cut-off is 7.4 percent (Column (8)),
we obtain a negative mortality-income elasticity estimate of 0.54. Our elasticity measure is similar
to those found in studies of US veterans in the early 1900s [Salm, 2011], which lie between -0.54
and –0.32.
Since the basic pension itself has an effect on the probability of dying, we cannot estimate its
20
-
Table V: Applicants’ health outcomes over four years from application
Variables 2sls S.E. 2sls RD Coef. S.E. P-value BW N Control(1) (2) (3) (4) (5) (6) (7) (8)
Mortality rate -0.025 (0.012) -0.021 (0.010) 0.034 500 8,499 0.074Days hospitalised -0.762 (0.770) -0.657 (0.640) 0.309 500 8,499 4.811Medical episode -0.050 (0.021) -0.042 (0.018) 0.024 500 8,499 0.342
Notes: This table reports results, within four years from the date of the first application, from regressions of severaldependent variables on a treatment dummy indicator and deviation of the pension score from the cut-off. Column(1) reports the treatment coefficient adjusted by the first stage and Column (2) reports its standard error clusteredat the province level. Column (3) reports the treatment indicator coefficient and Column (4) reports its standarderror clustered at the province level. Column (5) reports the p-value of the adjusted treatment coefficient reportedin Column (1). Column (6) reports the range of pension score points from the cut-off and Column (7) reports thenumber of observations in the regression. Column (8) reports the mean of the outcome variable for control applicantsat the cut-off.
causal effect on the days of hospitalisation: applicants on each side of the cut-off are not compa-
rable as those above the cut-off have fewer days available to be hospitalised, due to their higher
mortality rate. With this caveat, we still show the effect on the number of days that applicants spent
in hospital, from the date of application to the last day that we are able to observe them (four years
after applying, or the moment of death). Even though treated applicants have more potential days
for hospitalisation, the point estimates in Column (1) of Table V suggest that, if anything, they
spend fewer days in hospital, although the effect is not statistically significant.
As a way to summarise treatment effects on health outcomes and circumvent the survival bias,
we follow the medical literature and use as an outcome variable a dummy indicator equal to 1 if the
applicant had either been hospitalised or died (medical episode) in the four years after applying.18
Column (1) of Table V shows that treated applicants are 5.0 pp. (p-value=0.024) less likely to
experience a medical episode in the four years after applying. Since more than one-third of con-
trol group applicants at the cut-off experienced a medical episode in this period, the basic pension
reduces the probability of medical episodes by around 16 percent.
Column (2) of Table VI shows that the causal effect of the basic pension on mortality and med-
ical episodes remains, whether we use logistic regressions, non-parametric estimations, different
sets of controls, or polynomials of order two in Scoreh. This Table and Figure IV also show that
the results do not change if we use different bandwidths around the cut-off.
Figure V displays the share of surviving control and treatment group applicants, at each point
18What we call a ‘medical episode’ is commonly referred to as a ‘primary outcome’ in the medical literature [Pittet al., 2014; Eikelboom et al., 2017].
21
-
Table VI: Applicants’ health outcomes in four years from the first application using logit,non-parametric estimations, optimal bandwidth, controls, and quadratic functional form in Scoreh
Variables Regression RD Coef. S.E. P-value BW N(1) (2) (3) (4) (5) (6)
Mortality rate Logit -0.020 (0.009) 0.029 500 8,499Mortality rate Non-parametric -0.021 (0.010) 0.045 500 8,499Mortality rate Optimal bandwidth -0.028 (0.013) 0.027 306 5,048Mortality rate Controls -0.019 (0.010) 0.059 500 8,499Mortality rate Quadratic -0.033 (0.015) 0.024 500 8,499Medical episode Logit -0.041 (0.018) 0.018 500 8,499Medical episode Non-parametric -0.042 (0.023) 0.071 500 8,499Medical episode Optimal bandwidth -0.048 (0.021) 0.021 398 6,605Medical episode Controls -0.038 (0.017) 0.025 500 8,499Medical episode Quadratic -0.057 (0.028) 0.043 500 8,499
Notes: This table reports results, within four years from the date of the first application, from regressionsof several dependent variables on a treatment dummy indicator and deviation of the pension score fromthe cut-off. Column (1) indicates the specification used. Logit reports estimations using a logistic regres-sion. Non-parametric reports non-parametric estimations using kernel local linear regressions. Optimalbandwidth estimates treatment effects using the optimal bandwidth proposed by Calonico et al. [2014].Controls employs our preferred specification, polynomial of order 1 in Scoreh, as well as 17 other controlslisted in Appendix Section B. Quadratic uses polynomial of order 2 in Scoreh. Column (2) reports thetreatment indicator coefficient and Column (3) reports the standard error clustered at the province level.Column (4) reports the p-value of the treatment coefficient. Column (5) indicates the range of pensionscore points from the cut-off and Column (6) reports the number of observations in the regression.
in time after their first application, adjusted by the deviation of their score from the cut-off using a
Cox proportional hazard model. This Figure suggests that the mortality effect manifests itself ap-
proximately one year after the first payment and grows almost monotonically over time, reaching
a maximum at the end of the studied period. This piece of evidence suggests that the beneficial
effects of the basic pension tend to accumulate over time and cannot be ascribed to ‘instantaneous’
mechanisms.
Table VII shows the effect of the basic pension on different sub-groups of applicants. We start
by following common practice in medical literature on ageing and mortality (e.g. Garre-Olmo
et al. [2013]; Hawton et al. [2011]), exploring the heterogeneous effects depending on the house-
hold structure of the applicants. Panel A of this Table shows that treated applicants living alone or
with elderly household members (i.e. without a working-age household member) seem strongly
affected, with a significant reduction in their mortality rate of 5.2 pp. (p-value=0.008). On the other
hand, Panel B suggests that those living with at least one working-age household member remain
22
-
Figure IV: Robustness of results for mortality and medical episodes using different bandwidths-.1
-.05
0.0
5.1
Coe
ffici
ent
250 400 500 600CCTBandwidth
point estimate using different bandwidthsMortality rate
-.1-.0
50
.05
.1C
oeffi
cien
t
250 300 500 600CCTBandwidth
point estimate using different bandwidthsMedical episode
Notes: Each graph shows the point estimate and the standard error of non-parametric estimation for the correspondingbandwidth. CCT is the optimal bandwidth using the approach proposed by Calonico et al. [2014]
unaffected, with an insignificant reduction in their mortality rate of 0.02 pp. (p-value=0.954).19
The same pattern arises when looking at the number of days of hospitalisation and the probability
of having a medical episode. Here, treatment group applicants living without a working-age house-
hold member spend 2.6 fewer days in hospital and are 11 pp. less likely to experience a medical
episode (p-values of 0.001 and 0.012, respectively). These results suggest that the group of treated
applicants most strongly affected by the basic pension is not only living longer, but also enjoying
better health, at least in terms of fewer days spent hospitalised. These findings are also robust to
the use of different specifications (see Appendix Tables B5 and B6).
Not surprisingly, an analysis of survival rates shows that applicants living without working-age
relatives are the main drivers of this dynamic effect across the whole sample, whereas applicants
living with working-age relatives see virtually no improvement over time (Appendix Figure C11).
5.2 The heterogeneous effects of receiving a pension on applicants
We do not find evidence of heterogeneous treatment effects based on the gender of applicants.
However, given that males constitute a small fraction of our sample, we cannot rule out the pres-
19There is no evidence of score manipulation within any of these subsamples of applicants: test statistics for theMcCrary test are -0.486 and -0.976 for applicants living with and without working-age relatives, respectively (FigureC9). Moreover, the subsamples appear to be locally comparable at the baseline: none of the 10 available covariatesis significant for applicants living with working-age relatives, and only 1 out of 10 is significant for applicants livingwithout working-age relatives (Table B4).
23
-
Figure V: Share of surviving applicants over 4 years from date of application, adjusted by thedeviation of pension score from the cut-off.
1st payment
.95
.975
1Sh
are
of s
urvi
vors
0 1 2 3 4Years from application submission
Control group Treatment group
Notes: This figure presents the share of survivors in the treatment and control groups (500 score-point bandwidth),and at each point in time following the first application. Survival rates are computed using a Cox proportional hazardsmodel, adjusted for the deviation of the score from the cut-off and using triangular weights to give more weight to theapplicants closer to the cut-off.
24
-
Table VII: Applicant’s health outcomes over four years from application: household structure andgender
Variables 2sls S.E. 2sls RD Coef. S.E. P-value BW N Control(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: applicants not living with a working-age household member
Mortality rate -0.052 (0.019) -0.045 (0.016) 0.008 500 3,647 0.102Days hospitalised -2.632 (0.769) -2.250 (0.658) 0.001 500 3,647 5.773Medical episode -0.111 (0.042) -0.093 (0.036) 0.012 500 3,647 0.369
Panel B: applicants living with working-age household members
Mortality rate -0.002 (0.018) -0.001 (0.014) 0.954 500 4,852 0.050Days hospitalised 0.846 (1.313) 0.635 (1.061) 0.552 500 4,852 3.918Medical episode 0.001 (0.039) -0.000 (0.032) 0.998 500 4,852 0.317
Panel C: female applicants
Mortality rate -0.024 (0.010) -0.020 (0.008) 0.021 500 7,403 0.067Days hospitalised -0.808 (0.750) -0.703 (0.629) 0.269 500 7,403 4.791Medical episode -0.053 (0.024) -0.044 (0.020) 0.035 500 7,403 0.336
Panel D: male applicants
Mortality rate -0.051 (0.051) -0.039 (0.039) 0.318 500 1,096 0.141Days hospitalised -0.431 (2.275) -0.354 (1.726) 0.838 500 1,096 4.959Medical episode -0.042 (0.110) -0.033 (0.083) 0.694 500 1,096 0.391
Notes: This table reports results, within four years from the date of the first application, from regressions of severaldependent variables on a treatment dummy indicator and deviation of the pension score from the cut-off. Column(1) reports the treatment coefficient adjusted by the first stage and Column (2) reports its standard error clusteredat the province level. Column (3) reports the treatment indicator coefficient and Column (4) reports its standarderror clustered at the province level. Column (5) reports the p-value of the adjusted treatment coefficient reportedin Column (1). Column (6) reports the range of pension score points from the cut-off and Column (7) reports thenumber of observations in the regression. Column (8) reports the mean of the outcome variable for control applicantsat the cut-off.
25
-
ence of differential effects between genders.
5.3 The effect of having a relative who receives a basic pension
Table VIII: Health outcomes over four years from application: household members, by age
Variables 2sls S.E. 2sls RD Coef. S.E. P-value BW N Control(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: elderly household members
Mortality rate 0.001 (0.016) 0.000 (0.013) 0.979 500 5,722 0.125Days hospitalised 0.511 (1.221) 0.443 (1.045) 0.673 500 5,722 5.236Medical episode 0.041 (0.035) 0.034 (0.030) 0.256 500 5,722 0.369
Panel B: working-age household members
Days hospitalised -0.345 (0.611) -0.242 (0.493) 0.625 500 8,047 2.132Newborn child 0.036 (0.009) 0.028 (0.007) 0.000 500 8,047 0.025
Panel C: female household members in fertility age (16-40)
Days hospitalised -0.188 (0.735) -0.137 (0.547) 0.804 500 2,058 1.764Newborn child 0.124 (0.039) 0.091 (0.028) 0.001 500 2,058 0.097
Notes: This table reports results, within four years from the date of the first application, from regressions of severaldependent variables on a treatment dummy indicator and deviation of the pension score from the cut-off. Column(1) reports the treatment coefficient adjusted by the first stage and Column (2) reports its standard error clusteredat the province level. Column (3) reports the treatment indicator coefficient and Column (4) reports its standarderror clustered at the province level. Column (5) reports the p-value of the adjusted treatment coefficient reportedin Column (1). Column (6) reports the range of pension score points from the cut-off and Column (7) reports thenumber of observations in the regression. Column (8) reports the mean of the outcome variable for control applicantsat the cut-off.
Panel A of Table VIII shows that elderly household members are more likely to die than ap-
plicants (the average mortality rate of elderly relatives living with a non-recipient applicant at the
cut-off is 12.5 percent), but this seems to be unaffected by having a relative who receives the basic
pension.
On the other hand, Panel B of Table VIII reveals that working-age relatives are 3.6 pp. (p-
value
-
Mumford, 2013].20 Although the basic pension leads to significantly more children being born to
fertility-age female household members, we do not find evidence that it affects the health of the
newborn child (Appendix Table B7).
Appendix Table B8 shows that the results for elderly and working-age household members do
not change when we use logistic regressions, non-parametric estimations, the optimal bandwidth
approach proposed by Calonico et al. [2014], all available controls, or when we control for a poly-
nomial of order 2 in Scoreh. This also ensures that the null effect on elderly household members is
not driven by the slight imbalance in this group shown in Section 4.4.
6 Mechanisms
In this section we provide evidence suggesting that the decrease in mortality and medical episodes
reflect fewer cases of respiratory and circulatory diseases. This seems to be explained by the re-
cipients of the basic pension having fewer acute episodes of chronic diseases. Finally, we provide
evidence suggesting that the heterogeneous effects on applicants, as well as the spillover effects on
working-age relatives, may be due to an intra-household transfer of income that ceases when basic
pension payments begin.
6.1 Why do applicants’ mortality and medical episodes decrease?
To gain an insight into which diseases drive the observed effects, Table IX disaggregates medical
episodes by cause. We look at the sub-group of applicants whose health outcomes are most strongly
affected - applicants living without working-age relatives - and find that the effects are driven by
a reduction in the probability of suffering from a circulatory or respiratory disease. Deaths or
hospitalisations for these reasons often result from decompensations of a chronic condition. For
example, medical episodes of circulatory diseases are often heart attacks due to uncontrolled di-
abetes or hypertension, and medical episodes of respiratory diseases are often acute respiratory
problems in patients with severe asthma. Circulatory an respiratory diseases are also the first and
the third leading cause of deaths around the world [World Health Organization, 2011; European
20Since goods with no substitutes are generally considered normal, children should be ‘normal goods’ and their‘consumption’ should increase with income. Our results, along with other recent empirical studies [Lindo, 2010;Lovenheim and Mumford, 2013], shed light on the long-term puzzle of the negative cross-sectional correlationbetween income and fertility that is present in many parts of the world (e.g. Jones and Tertilt [2008]).
27
-
Respiratory Society, 2017], and 47 percent of elderly pension recipients in 2011 had at least one
of the underling chronic conditions leading to them [Ministerio de Desarrollo Social, 2011].
Table IX: Medical episodes by cause over four years from application
Variables 2sls S.E. 2sls RD Coef. S.E. P-value BW N Control(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: applicants
Circulatory -0.015 (0.010) -0.012 (0.008) 0.142 500 8,499 0.057Respiratory -0.008 (0.007) -0.006 (0.006) 0.245 500 8,499 0.016Tumour -0.008 (0.011) -0.006 (0.009) 0.497 500 8,499 0.035Digestive or nutrional -0.014 (0.013) -0.012 (0.011) 0.254 500 8,499 0.060
Panel B: applicants not living with a working-age household member
Circulatory -0.047 (0.015) -0.040 (0.013) 0.002 500 3,647 0.082Respiratory -0.022 (0.008) -0.018 (0.007) 0.006 500 3,647 0.024Tumour -0.016 (0.011) -0.014 (0.009) 0.126 500 3,647 0.037Digestive or nutrional -0.009 (0.023) -0.008 (0.020) 0.685 500 3,647 0.056
Notes: This table reports results, within four years from the date of the first application, from regressions of severaldependent variables on a treatment dummy indicator and deviation of the pension score from the cut-off. Column (1)reports the treatment coefficient adjusted by the first stage and Column (2) reports its standard error clustered at theprovince level. Column (3) reports the treatment indicator coefficient and Column (4) reports its standard error clusteredat the province level. Column (5) reports the p-value of the adjusted treatment coefficient reported in Column (1). Column(6) reports the range of pension score points from the cut-off and Column (7) reports the number of observations in theregression. Column (8) reports the mean of the outcome variable for control applicants at the cut-off.
The decrease in respiratory episodes amongst applicants does not appear to be driven by an
increase in the use of vaccinations. As we can see in Appendix Table B9, we do not find any
significant effect of the basic pension on the number of vaccinations received for influenza or
pneumonia in the four years after applying, for any of the sub-groups.21
Two main mechanisms may explain the results so far. The first proposed mechanism is that re-
cipients are putting part of their the income towards goods or actions that improve their health, such
as vitamin-rich food [Bartali et al., 2006; Cederholm et al., 1995; Ortega et al., 1997; Fogel, 2004;
Salm, 2011] or transport costs that allow them to attend medical check-ups regularly [Dahlgren
and Whitehead, 1991; Jensen and Richter, 2003; Aguila et al., 2015]. This may help recipients to
improve their ability to control chronic conditions and prevent acute episodes from occurring. This
mechanism seems plausible, considering that basic pension recipients spend around 77 percent of
their income on goods conducive to health, such as food, and only 13 percent on goods that impede21Since we observe a slight imbalance in pneumonia vaccination levels at baseline, we repeat the analysis using all
available controls and observe that results remain unchanged. Therefore, this seems unlikely to account for our results.
28
-
Table X: Health outcomes over four years from application: applicants by vaccination statusbefore applying
Variables 2sls S.E. 2sls RD Coef. S.E. P-value BW N Control(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: Applicants vaccinated before applying
Mortality rate -0.046 (0.021) -0.037 (0.017) 0.033 500 2,819 0.101Days hospitalised -2.489 (1.287) -2.035 (1.002) 0.048 500 2,819 6.385Medical episode -0.131 (0.037) -0.104 (0.030) 0.001 500 2,819 0.428
Panel B: Applicants not vaccinated before applying
Mortality rate -0.014 (0.015) -0.011 (0.013) 0.376 500 5,680 0.060Days hospitalised 0.125 (1.147) 0.096 (0.978) 0.922 500 5,680 4.012Medical episode -0.008 (0.033) -0.006 (0.028) 0.823 500 5,680 0.297
Notes: This table reports results, within four years from the date of the first application, from regressions of severaldependent variables on a treatment dummy indicator and deviation of the pension score from the cut-off. Column(1) reports the treatment coefficient adjusted by the first stage and Column (2) reports its standard error clusteredat the province level. Column (3) reports the treatment indicator coefficient and Column (4) reports its standarderror clustered at the province level. Column (5) reports the p-value of the adjusted treatment coefficient reportedin Column (1). Column (6) reports the range of pension score points from the cut-off and Column (7) reports thenumber of observations in the regression. Column (8) reports the mean of the outcome variable for control applicantsat the cut-off.
health, such as alcohol [Ministerio Trabajo y Previsión Social, 2017]. This hypothesis would imply
that the benefits of the income increase should be more pronounced for people who care more for
their health. Unfortunately, we do not have a means of directly observing the health-consciousness
of the applicants. However, we use as a proxy for health-consciousness whether the applicant had
an influenza or pneumonia vaccination in the six months before applying. Table X shows that
applicants with these vaccinations are more positively affected by the pension, showing a signifi-
cantly larger drop in mortality. These pieces of evidence suggest that recipients using part of their
income to buy health-conducive goods that prevent acute episodes from occurring is a mechanism
that is likely to be at play.
This explanation is also in line with what Salm [2011] finds for US veterans. However, Feeney
[2017, 2018] finds that rural Mexicans receiving a non-contributory pension increase their tran-
sitions into retirement, unhealthy food consumption, and mortality from cardiovascular diseases.
One reason that might explain these contrasting results is differences in the characteristics of pen-
sion recipients. In the case of the US veterans, as in our study, recipients had already moved out
of the labour force, while in the Mexican study more than 40 percent of recipients were employed.
Thus, changes in consumption patterns that occur when exiting the labour force may provide a rele-
29
-
vant explanation of these seemingly contradictory results in food consumption patterns [Browning
and Meghir, 1991; Zantinge et al., 2013].
The second proposed mechanism is that the income increase allows pensioners to purchase
health insurance or medical treatment that they could not have afforded in the absence of the ba-
sic pension. Previous work has not found a clear relationship between health insurance coverage
and mortality [Finkelstein and McKnight, 2008; Chetty et al., 2016]. Purchasing a more generous
private health insurance does not seem to be the most relevant mechanism in our case, considering
that 95 percent of basic pension holders in the Chilean population were and remained enrolled
in public health insurance during the time of our study [Ministerio de Desarrollo Social, 2011,
2015]. Furthermore, medicines and medical treatments in Chile are free of charge for all non-rare
diseases, including the underlying chronic conditions that produce the circulatory and respiratory
diseases driving the effects observed here.22
6.2 Why is there no effect on pensioners living with a working-age relative,but there are spillover effects on these relatives?
The null effect of the basic pension on treated applicants living with a working-age household
member could be the result of some ex-ante transfer of income from working-age household mem-
bers to applicants, which then stops when applicants start to receive pension payments. In this
sense, the net income increase would be lower for this particular group of treated applicants, but it
would also imply a positive income shock for their working-age household members. The further
positive spillover effects on the fertility levels of these household members, presented in Table
VIII, are in line with this hypothesis.23 Also consistent with this view is the fact that, according to
a recent government survey, 65 percent of basic pension recipients consider that they could borrow
money from a working-age relative, while only 17 percent could do so from an elderly relative
[Ministerio Trabajo y Previsión Social, 2017]. These pieces of evidence do suggest that ex-ante
intra-household transfers of income are a likely mechanism operating here.
22Since 2005, the Chilean plan for universal access to treatment and drugs (‘Plan de Acceso Universal de GarantiasExplicitas’ or AUGE) granted free access to drugs and procedures for the most common diseases in the country. Thesystem covers virtually the entire population: only 1 percent of patients report that they suffer from a disease thatis not covered by AUGE and more than 90 percent report having access to a doctor when required [Ministerio deDesarrollo Social, 2011].
23We are not able to disentangle the different alternative uses of this income shock. For instance, fertility-agewomen in recipient households may be continuing work after having a newborn child, or they may be exiting thelabour force. We see these alternative paths as variations of the same mechanism: a positive income shock forrecipients’ working-age household members.
30
-
Findings in previous papers also support this ‘crowding-out’ hypothesis. In Peru and South
Africa, Cox and Jimenez [1992] and Jensen [2003] find evidence that social security benefits and
pensions ‘crowd out’ 20-30 percent of private transfers from younger generations to the elderly.
Furthermore, Edmonds [2006] shows that working-age women tend to depart when their elderly
household members start to receive a pension, whereas children and women of child-bearing age
tend to move to live with the elderly. These changes in living arrangements could also potentially
lead to a change in net transfers from the young to the elderly.
A second possible mechanism for the null effect on this group is that, upon receiving the pen-
sion, the elderly start transferring a substantial share of their basic pension to younger relatives.
This would also diminish the positive impact of the pension on the income and health of the elderly,
but it would increase working-age relatives’ income sufficiently to produce the observed spillover
effects. Some evidence from previous research is line with this, as Duflo [2000, 2003] and Ard-
ington et al. [2009] find that grandmothers in rural South Africa transfer some of their pension to
their granddaughters and sons, respectively. However, this hypothesis does not seem very likely for
Chilean grandparents: 82 percent of pension recipients do not share any money with their relatives
or friends, and only 4 percent share more than one-fifth of their pension with others [Ministerio
Trabajo y Previsión Social, 2017].
Finally, a third explanation could be that living with working-age household members is a
proxy for some other household characteristic. For example, families with working-age members
may be larger, which may dilute the effect of the basic pension when household income is pooled;
or the effect of the basic pension might be attenuated if applicants are already receiving proper care
from younger relatives. This hypothesis is not supported by the data. In particular, when we repli-
cate our results, controlling for a rich set of individual, household and residence characteristics, the
heterogeneous results by household structure remain unchanged (Appendix Tables B5 and B6).
7 Back-of-the-envelope computation: to what extent could thebasic pension close the gap in life expectancy?
How far can the basic pension go in correcting health disparities across income levels in old age?
To shed light on this question, we first calculate that treated applicants have an average income of
US$221.04 at the time of application, which then increases to US$386.56 after receiving the basic
pension. We denote the group of elderly people who do not receive a basic or contributory pension,
31
-
Figure VI: Back-of-the-envelope computation: the effect of the basic pension on closing thesurvival gap over a four-year window
1st payment
.95
.975
1Sh
are
of s
urvi
vors
0 1 2 3 4Years from application submission
Control group Treatment groupCounterfactual control group
Notes: This figure presents the share of survivors in the control group, treatment group and counterfactual controlgroup at each point in time following the first application and in a 500 score-point deviation bandwidth. Survival ratesare computed using a Cox proportional hazards model, adjusted for the deviation of the score from the cut-off.
but who have the same income as pension recipients (US$386.56), as the counterfactual control
group.
Ideally, we would like to observe the income and mortality rates of the whole population of
elderly people who are not receiving a contributory pension. We would then be able to calculate
the following figures for the four years after applying for a basic pension: 1) the difference in
life expectancy between the treatment and control groups at the cut-off; and, 2) the difference in
life expectancy between the counterfactual control group and the control group at the cut-off. By
dividing the first difference by the second, we could derive the ratio by which the basic pension
income increase reduces the life expectancy gap between the people with an income equal to the
control group at the cut-off (US$221.04) and the counterfactual control group (US$386.56). This
would tell us to what extent the basic pension income increase can reduce health disparities across
income levels, and to what extent these health disparities are permanent.
32
-
Unfortunately, there is no data to compute the survival rates of elderly people by income level
using the whole population. We circumvent this problem by first finding the pension score corre-
sponding to the counterfactual control group income level in the pool of applicants. According to
our records, applicants in the counterfactual control group received 585 points more than treated
applicants at the cut-off. We then compute the share of survivors at the cut-off for the treated and
control groups as in Figure V. As the characteristics of applicants and their households tend to
change across the pension score distribution, we estimate the share of survivors for an applicant
with the same characteristics as treated applicants at the cut-off, but with a pension score equal to
the counterfactual control group.24 This allows us to compute the survival rates of the treatment
and control groups at the cut-off and of the counterfactual control group.
Figure VI shows the survival functions of each of these three groups at each observable point
in time. The share of survivors drops by 7 percent for the control group, 5 percent for the treat-
ment group and 2 percent for the counterfactual control group over the four-year window. The life
expectancy of each group over the four years after applying is represented by the integral of each
survival function. Taking the ratio of the differences in life expectancies noted above, we estimate
that the basic pension closes 29 percent of the life expectancy gap between the control group at the
cut-off and the counterfactual control group.
8 Concluding remarks
In this paper, we have explored the impact of a permanent increase of income for the elderly poor
on their mortality and days of hospitalisations, as well as the spillover effects on their household
members. By comparing applicants to a Chilean basic pension scheme, and their household mem-
bers, in a narrow window around the eligibility threshold, the effects of this income increase have
been disentangled in a regression discontinuity framework.
This permanent increase in income sees a reduction in the mortality rates of basic pension re-
cipients, at the point in time four years after they apply for the pension. This effect seems to be
driven by higher consumption of goods that improve their health, such as vitamin-rich food, rather
than by increased consumption of medicine or health insurance. This is in line with a wide range
of literature that shows the impact of health-conducive goods on individuals’ health status [Bartali
24We follow this approach, as we are interested in understanding by how much the basic pension income increasecan close the life expectancy gap between two persons with the same observable characteristics at two given incomelevels, rather than the ‘unconditional’ life expectancy gap between two persons at two given income levels.
33
-
et al., 2006; Cederholm et al., 1995; Ortega et al., 1997; Fogel, 2004; Salm, 2011]. These results
suggest that health inequalities in the elderly population are driven, at least in part, by contempo-
raneous income inequalities.
We also observe that the effect of the basic pension is concentrated on recipients who live alone
or with another elderly person, while those who live with a working-age relative are not affected.
An exploratory analysis shows that this last null effect is likely to be due to crowding-out effects:
working-age relatives stop transferring a portion of their income to their elderly cohabitants once
they start to receive the basic pension. In line with this hypothesis, we find that working-age rel-
atives are more likely to have a newborn child nine months or later after the application is made.
These results suggest that the financial burden imposed by the elderly poor on their younger rela-
tives may be one of the channels through which poverty is transmitted, calling for further research
in this area. Moreover, our results show that the effects of a non-contributory pension reach beyond
the recipients to other household members. This is a relevant point for the optimal design of social
pension policies, as the spillover effect can account for a significant share of the overall benefits of
a pension.
References
Aguila, Emma, Arie Kapteyn, and James P Smith. 2015. “Effects of income supplementation
on health of the poor elderly: The case of mexico”. Proceedings of the National Academy of
Sciences 112(1), 70–75.
Ardington, Cally, Anne Case, and Victoria Hosegood. 2009. “Labor supply responses to large
social transfers: Longitudinal evidence from south africa”. American economic journal:
Applied economics 1(1), 22–48.
Bartali, Benedetta, Edward A Frongillo, Stefania Bandinelli, Fulvio Lauretani, Richard D Semba,
Linda P Fried, and Luigi Ferrucci. 2006. “Low nutrient intake is an essential component
of frailty in older persons”. The Journals of Gerontology Series A: Biological Sciences and
Medical Sciences 61(6), 589–593.
Becker, Gary S.. 1960. Demographic and Economic Change in Developed Countries, Chapter An
Economic Analysis of Fertility, pp. 209–240. Princeton University Press.
34
-
Braveman, Paula A., Catherine Cubbin, Susan Egerter, David R. Williams, and Elsie Pamuk.
2010. “Socioeconomic disparities in health in the united states: what the patterns tell us”.
American journal of public health 100(S1), S186–S196.
Brønnum-Hansen, Henrik and Mikkel Baadsgaard. 2012. “Widening social inequality in life
expectancy in denmark. a register-based study on social composition and mortality trends for
the danish population”. BMC Public Health 12(1), 994.
Browning, Martin and Costas Meghir. 1991. “The effects of male and female labor supply on
commodity demands”. Econometrica 59(4), 925–951.
Calonico, Sebastian, Matia D. Cattaneo, and Rocio Titiunik. 2014. “Robust Nonparametric
Confidence Intervals for Regression-Discontinuity Designs”. Econometrica 82(6), 2295–2326.
Case, Anne. 2004. “Does Money Protect Health Status? Ev