ECON1980o: Health, Education and Development. Lecture 3 October 2, 2008

Transcription

1 ECON1980o: Health, Education and Development Lecture 3 October 2, 2008

2 Today s Class: Part I: Returns to health, in form of: Labor productivity Cognitive development (investment effects) Schooling attendance (investment effects) Big externalities Part II: Health production How much does health improve with income? Evidence from a natural experiment in South Africa

3 Part II: How do societies produce health? Why we care: Evidence that health status has high economic returns Also care about health as fundamental aspect of ind welfare What needed to increase health status of poorest countries? Most interventions pay for themselves, but what is most effective way, and when are interventions needed at all? Old-fashioned view: Income most direct way to produce health Empirical question: How does health improve with income?

4 Biggest challenge: How can we isolate changes in income that are not potentially driven by/influenced directly by health status? First possibility: Experiment (hand out money) PROGRESA in Mexico Complication: Tied to health components More realistic possibility: Natural experiment What kind of thing?

5 Non-experimental method #1: Natural experiment (NE) Natural experiment: An observational study of the effect of X on Y that relies on an incidence or setting in which exposure to X is quasi-random -- or as if random Idea: An NE is a set of conditions that determines quasirandom exposure to X To implement, need to find setting in which otherwise similar individuals were either exposed and not exposed to an income shock for a clearly identified reason Then, using data collected afterwards, compare average health among exposed (quasi-treatment group) to average health among unexposed ( quasi-control ) group in difference-in-difference framework (is change necessary?)

6 Comments on natural/ quasi experiments: Used to refer to situation that is not experimental but is as if it was Not a precise definition - refers to case where variation in X is good variation A Famous Example: London, 1854

7 John Snow and the case of the Broad Street pump Regular cholera epidemics in 19 th century London, not much known about disease contagion Widely believed to be caused by bad air (miasma) John Snow thought bad water was cause Experimental design would be to randomly give some people good water and some bad water but ethical problems with this (often the reason we need to come up with more creative approaches to identification)!

8 Soho Outbreak August/September 1854 Snow noticed that people closest to a specific source (Broad Street Pump) most likely to die Why doesn t this resolve air vs. water hypotheses? People closest to pump also breath the same air!

9 Snow s idea: Two water companies Lambeth and Southwark/Vauxhall served most of London In 1852 Lambeth moved its intake pipe upstream to a cleaner water source while S/V left its intake pipe downstream This move provided Snow with his natural experiment during the cholera outbreak of How so? He could compare S/V customers who live next door to Lambeth customers but get water from different place Why is next door important?

10 Why is this a natural experiment? Variation in water supply as if it had been randomly assigned, other factors ( air ) held constant How do you use this to estimate treatment effect of water on health? What is Y? What is X? What is Z? Can estimate treatment effect using difference in means Run regression of death on water source. What else? Distance to pump, whether drink water (or beer), etc.

11 Death rate from cholera in London by source of water Water Company Supplying the District Rate per 10,000 houses Southwark & 315 Vauxhall Lambeth 37 Rest of London 59 (adapted from Snow 1855, Table IX, presented in Freedman 2004)

12 Conditions for valid natural experiment To be internally valid, quasi-experimental source of variation (Z) (think of Z as an indicator variable akin to treatment assignment): 1. Must be correlated with X 2. Must be uncorrelated with error: Mechanism of assignment to treatment must be arguably arbitrary ( as if random, no selection) Z must not influence Y through channels other than X Which of these conditions is guaranteed with experimental data (in which Z is treatment assignment)?

13 Favorable conditions for natural experiment To ensure that Z uncorrelated with error, better if: Mechanism of assignment to treatment (source of water supply) arguably arbitrary ( as if random) -Knowledge of the details of the case is crucial Little sample selection bias. What do we mean by this? Exposed and unexposed similar baseline characteristics What does this imply? Relatively blind assignment (ex-post and ex-ante). What do we mean by this and why does this help?

14 How does Snow do on these? Arbitrary source of variation? A few houses are supplied by one Company and a few by the other, according to the decision of the owner or occupier at the time when the Water Companies were in active competition. Sample selection bias? In many cases a single house has a supply different from that on either side. Each company supplies both rich and poor, both large houses and small; there is not any difference either in the condition or occupation of the persons receiving the water of different Companies. Blind experiment? [people] were divided into two groups without their choice, and, in most cases, without their knowledge...

15 What about external validity? Conditions that maximize EV: 1. Large scale: The experiment, too, was on the grandest scale. No fewer than three hundred thousand people of both sexes, of every age and occupation, and of every rank and station, from gentle folk down to the very poor 2. Universal: The pipes of each Company go down all the streets, and into nearly all the courts and alleys. But perhaps most important: 3. Lesson (fact that water carries disease) independent of time or place

16 Pluses and minuses of approach: Merits of natural experiments: Stronger on internal validity than observational studies that just control for observables Finding natural experiments often requires detailed knowledge of cases, which provides additional leverage for inferences Demerits of natural experiments Really clean natural experiments are rare External validity issues

17 Natural experiment: Pensions in South Africa Starting in 1991, South African government expanded coverage of pension system to Coloured & Blacks Large: 520 Rand/month = 2.5 X median income for Blacks Universal: All women above 60 & men above 65 eligible subject to loosely applied means test, irrespective of work history Most Blacks eligible, 85% of eligible people end up getting it Research question: How does family health respond to income? Research challenge: How does Case identify the effect of the pension with cross section data? (what is her NE?)

18 First: Why is this almost as good as an experiment? Unanticipated income shock Exogenous to individual behavior Why is it better than most experiments? Permanent income shock Explain? Why is it worse than an experiment?

19 Problem: Everybody (in same SES cat) reached Why a problem? Since eligible families poorer (extended families, more likely to be rural, poor), no obvious control group -- kids in pension families would have been in worse shape w/o pension Naïve solution: Look at changes over time Treatment group: Post-shock Control group: Pre-shock But: Need panel data, and always need to worry about time effects Cross-section solution: Quasi-control groups Which individuals were affected more than others within same income category?

20 Who can we use as a quasi-control group? Strategy 1: Compare in households with and without an eligible member Treatment group: Adults and children living with pensioner Control group: Adults and children not living with pensioner What is problem with this simplistic an approach? Problem: How do we separate effect of pension from effect of living with a pensioner? How do they solve it? Solution: Condition on having a person over 55 in household Explain what this does?

21 Specification: H + + E + ij = α β NP β 1 j 2 Does this work? What assumptions are necessary for this to be accurate estimate of effect of income on health? Problem: How do we separate effect of pension from effect of living with a young versus old pensioner? What else do they do? Solution A: Look at predicted variation in effect of pension across households (like what?) Solution B: Look at predicted variation in effect of pension within households (like what?) j u ij

22 Specification: Run separate regressions on two sub-samples: H E + ij = α β P β NP β 1 ij 2 j 3 j u ij for all households that pool income AND: H ij = λ P λ NP λ 1 ij 2 j 3 γ E + j ε ij for all households that do not pool income. What is P? What are the hypotheses regarding β 1, β 2, λ 1, λ 2? 1. β 1 ==0; β 2 >0 Explain? 2. λ 1 >0; λ 2 ==0 Explain?

23 Results: Adding an extra pensioner to household improves adult SRH by 0.5 points In households that pool income, this true for everyone (no difference in health effect between pensioners and non-pensioners). For instance, kids grow 5 cm higher (half year s growth) in families with pensioner in which income pooled In households that don t pool, only true for the pensioner s health (and higher 1.03 point increase) Interpretation? Money protects health status, but only if money received by individual or shared within the house Are you convinced?

24 Why did health respond to income? Possible mechanisms: 1. Help for people with disabilities Fact: People with physical limitations have increase in self-reported health if they live with a pensioner 2. Health investments Fact: Households with pensioner more likely to buy toilet 3. Ate better (positive income elasticity of calorie demand) Kids: Height-for-age also suggests nutrition channel 4. Higher self-esteem

25 Conclusions: (1) Getting extra R520 from the government improves health (2) Living with someone who gets R520/month gives you better health but only if income is pooled So (some) households and not just individuals produce health Discussion questions: Can we conclude that income matters for health? With what caveats? What other characteristics of the person who receives income likely to matter for health production?

26 Threats to external validity Characteristics of the population? Setting or context? Exact nature of the intervention? Outcome examined?

27 Backtracking What do we learn from Kremer/Miguel ( Illusion of sustainability ) experiment about type of health investments that are unlikely to improve with income?

28 Impacts of Deworming (Miguel and Kremer 2003) Substantial reduction in infection levels for treated, and very strong externality effects (2/3 total effect through externalities) How do they use social experiment to study willingness to pay for deworming drugs and what do they find?

29 Estimating Deworming Social Effects Experimental design means that exogenous number of social links were in early treatment (Group 1, Group 2) schools, conditional on number of links in all program schools Test whether this changes adoption, knowledge, beliefs 1678 Parent Questionnaires in two survey rounds (Table 1) 10.2 total links, 4.4 to own school, 3.0 nonprogram schools, 2.8 links to program schools (not own), 1.9 to early treatment schools 3164 Pupil Questionnaires in two rounds (Table 1)

30 Econometric identification strategy Expected private benefits from deworming are a function of the number of early treatment links (N E ), total social links (N), cost-sharing (C), and household characteristics (Z): (4) Pr( T ij =1) = Φ E λ N + N λ + γc + Z θ + ε } { 1 ij ij 2 j ij ij Identification issues: Validate randomization (Table 2) Infection social effects likely to be small Pupil transfers, stability of social networks

31 Parent Social Effect Estimates Early treatment social links lead to lower takeup (Table 3) Even weak social links are influential (Table 4) Parents with more education have particularly large effects Parents learn deworming drugs are not effective (Table 6)

32 Conclusions from social learning Social learning may lead to less adoption of a new health technology when treatment externalities are large Self-defined reference groups have major impact on behavior Experimental, non-experimental estimates very different evidence for large positive omitted variable bias Large subsidies may be needed to boost adoption Child-to-child public health programs focusing on teens are likely to be particularly effective

33 Policy implications Increasing popularity of sustainability Local ownership Cost recovery Education Why? Failure of many projects after donors leave ideological commitment to a more participatory and empowering approach to development policy Contrast with standard public finance approach: fund activities with positive externalities indefinitely, in order to overcome market failure and free-riding Leads to changes in what kinds of projects are financed: Microfinance Water Public health

34 Other Factors Affecting Take-up Traditional view is that worms are benign Immediate and salient side effects (e.g., vomiting), but diffuse health benefits over the months following treatment Many other health shocks (e.g., malaria) further complicates the signal-extraction problem

35 Questions What are important policy implications from this study in your views? Do you agree with interpretation that exposure to drugs reduced take-up because people learned about true private return to deworming drug? What evidence convincing/unconvincing? Was take-up possibly low among those with many links just because they need the drug less? (what evidence against this) Examples of communities that have created institutions that encourage sustainable contribution to public goods, but paper suggests that this can t be imposed from outside only happens if it evolves from within community. Do you agree? Is it something particular about worms or would same thing have happened with other contagious diseases such as vaccinations?