14.74 Lecture 5: Health (2) Esther Duflo February 17, 2004 1 Possble Interventons Last tme we dscussed possble nterventons. Let s take one: provdng ron supplements to people, for example. From the data, what effect do we expect from ths nterventon? But what doubts can we have about ths nterpretaton of the data? 2 Fndng out what works: the value of experments Ths conversaton should have convnced you that, usng only data from observatons, we can form ntellgent hypotheses, but not resolve them. Before spendng all of our money on somethng, how do we fnd out whether or not t wll work? Why do we have problems teasng out causal relatonshp n reallfe data? To address these, we need to compare comparable people, some of whom were exposed to a partcular polcy and some of whom were not. How would we do t to test a new drug? 1
Why not do t to test an nterventon n ths context? To test the effect of a polcy, we can use randomzed evaluaton, where a randomly selected treatment group receves a treatment, whle the other group does not (ths s the comparson group). We wll collect data on both the treatment and the comparson group, and compare the result. Because the treatment and the comparson group have been randomly selected, we can conclude that any statstcally sgnfcant dfference we observe between the treatment and the comparson group s due to the nterventons. Now, a more formal ntroducton to ths problem: Let us call Y T the earnngs level of an ndvdual wth a det rch n ron and Y NT earnngs of the same ndvdual f hs det s poor n ron. Can we observe Y T same tme? Y T and Y NT are called potental outcomes. We are nterested n the dfference: and Y NT the at the Y T Y NT The effect of havng textbooks for school. The problem: we don t observe ndvdual bothwthandwthoutthedetrchnronatthe same tme. What can we do? We wll never know the effect of havng ron on a partcular ndvdual. We may hope to learn the average effect of a det rch n ron. Y NT ] Imagne we have access to data on lots of ndvduals n the regons. Some ndvduals have a det rch n ron and others do not. We may thnk of takng the average n both groups, and the dfference between the two. Why does t make sense? /hgh ron det] [Y NT Subtract and add E[Y NT /T ] /T ] E[Y NT /T ] E[Y NT The frst term /N T ]+E[Y NT /low ron det] =E[Y T /T ] E[Y NT /N T ] /T ]= Y NT /T ]+E[Y NT /T ] E[Y NT /N T ] Y NT /T ] s the treatment effect that we try to solate: on average, n the treatment schools, what dfference wll the books make? 2
What s: E[Y NT /T ]? E[Y NT /N T ]? The dfference E[Y NT /T ] E[Y NT /N T ]? Whch s lkely to be bgger? Why? The dfference s the selecton bas. It tells me that besde the effect of the textbooks, there maybesystematcdfferences between those who have ron and those who do not. 2.1 What happens when we randomly allocate the treatment? Suppose that we select the ndvdual to whom we gve the ron supplement randomly wthn a populaton of ndvduals. We observe the test scores n both the treatment schools, and the other schools, whch wll form our control (or comparson) group. On average, what do we expect to fnd f we compare the treated schools and the comparson schools before the nterventon? If we compare other characterstcs of these schools? Compare E[Y NT /N T ] and E[Y NT /T ] What s /T ] E[Y NT /N T ] equal to? Example: Iron supplementaton n Indonesa. Baselevelofanema:fgure 1 STEP ONE: desgn. About 3,000 households. Households are randomly selected to be n the placebo or treatment group. Iron s dstrbuted at home n blster packs. STEP TWO: Baselne comparson: table 3. In whch column do we see the baselne comparson? What do we expect for the baselne comparson? Why s t mportant? What s the mean dfference at baselne for men? for women? What s the T statstc? 3
Are these dfferences sgnfcant? STEP THREE: Protectng the desgn. Complance s strctly enforced (over 90%). What s the rght comparson? Why? Those who took the plls versus all of those who dd not? Those who took the plls versus the comparson group? All of those ntally n the treatment group versus (supposed to take the plls) allof those ntally n the comparson group (not supposed to take the plls)? Ths comparson s called the INTENTION TO TREAT estmate. How do we obtan the AVERAGE EFFECT on those who took the plls? Remark: Is t a program that could be scaled up? Why or why not? Why do we care about the results then? STEP FOUR: Attrton What could happen to the sample f the treatment people were much healther because of the experment and the comparson people saw no mprovement? How could that affect the results? What do we need to do to avod that? In ths experment: Attrton was 3% Attrton was no lower n treatment group Attrton s not related to baselne hb levels. STEP FIVE: Results Effect on hb level: Results: fgure 2, table 3: effect on hb level n blood. What s column 3? 4
What s column 5? What s the dfference wth column 3, and whch s best to use? What s column 6? How does t dffer? Do we expect t to be dfferent from 5? What s best to use? What s column 7? How does t relate to fgure 2? Why do we see the pattern we see n fgure 2? What s column 9? Do we observe what we expected? Tables 4 to 7: eesults on work, health, happness. How do we read these results? What are the man conclusons we can draw? 3 Experments n Udapur What experments do you propose to do n Udapur? The experments we propose: Publc Health: chlornaton of wells Nutrton: ron fortfcaton of flour Health care: second ANM n publc centers Health nvestment: ncentves for vaccnatons. Desgn: we wll work n 120 vllages, randomly dvded nto enough groups to test nterventons aganst each other, and n combnaton. What are the advantages of dong several experments n one place? What wll the results from these experments tell us about polcy n Rajasthan? About the broader polcy debate about whether one should nvest n health care? 5