1、QuestionspThe relationship of interestpThe ideal experimentpThe identification strategypThe mode of inference随机控制实验了解接受远端祈祷心脏病病人的治疗效果1The relation of interestpMost questions are about casual and effectpExamplenSmaller classes are better for the learning of the studentsnLower copayment encourages the
2、 health utilizations随机控制实验了解接受远端祈祷心脏病病人的治疗效果2Coefficient Equations11112121nniiniiiixyinxxiniiixyx ySSnSSxxnSlopey-intercept01 yxPrediction Equation01yx随机控制实验了解接受远端祈祷心脏病病人的治疗效果3pThe estimates of regression shows the extent of“correlation”,but we are interested to know“causation”pKey issue in empirica
3、l analysis is separating causation from correlation.nCorrelated means that two economic variables move together.nCasual means that one of the variables is causing the movement in the other.随机控制实验了解接受远端祈祷心脏病病人的治疗效果4THE IMPORTANT DISTINCTION BETWEEN CORRELATION AND CAUSATIONpThere are many examples wh
4、ere causation and correlation get confused.pIt is critical for government policy to understand the difference;otherwise policy may not have the intended impact.随机控制实验了解接受远端祈祷心脏病病人的治疗效果5THE IMPORTANT DISTINCTION BETWEEN CORRELATION AND CAUSATIONpOne example concerns SAT preparation courses.nIn 1988,H
5、arvard interviewed its freshmen and found those who took SAT“coaching”courses scored 63 points lower than those who did not.nOne dean concluded that the SAT courses were unhelpful and“the coaching industry is playing on parental anxiety.”随机控制实验了解接受远端祈祷心脏病病人的治疗效果6The ProblempIn both examples,there is
6、 a common problem:an attempt to interpret a correlation as a causal relationship,without sufficient thought to the underlying data generating process.pFor any correlation between two variables A and B,there are three possible explanations for a correlation:nA is causing B.nB is causing A.nSome other
7、 factor is causing both.随机控制实验了解接受远端祈祷心脏病病人的治疗效果7The ProblempIn the Harvard SAT example,the possibilities could be:nSAT prep courses worsen preparation for the SATs.nThose with poorer test taking ability take prep courses to try to catch up.nThose who are generally nervous both like to take prep cou
8、rses and do the worst on standardized exams.pHarvard dean thought the first possibility was correct.随机控制实验了解接受远端祈祷心脏病病人的治疗效果8The ProblempAlthough the peasants or the Harvard dean could actually be correct,odds are they are misinterpreting the underlying process at work.pFor policy purposes,what we c
9、are about is causation.pKnowing that two factors are correlated gives you no predictive power.随机控制实验了解接受远端祈祷心脏病病人的治疗效果9The Problem of BiaspIn this case,the assignment of the intervention was not random.pThis means the treatment and control groups are not identical.pNon-random assignment,in turn,coul
10、d cause bias.随机控制实验了解接受远端祈祷心脏病病人的治疗效果10 The Problem of BiaspBias represents any source of difference between treatment and control groups that is correlated with the treatment,but not due to the treatment.nIn the SAT example,the impact of SAT courses is biased by the fact that those who take the pre
11、p course are likely to do worse on the SAT for other reasons.随机控制实验了解接受远端祈祷心脏病病人的治疗效果11The Problem of BiaspBy definition,such differences do not exist in a randomized trial,since the groups are not different in any consistent fashion.pAs a result,randomized trials have no bias,and it is for this rea
12、son they are the“gold standard”for empirically estimating causal effects.随机控制实验了解接受远端祈祷心脏病病人的治疗效果12The ideal experimentpWhat sort of experiment can ideally be used to capture the casual effect?pGolden standard:controlled trialsnRandomize subjects into the treatment and control group,then compare the
13、ir outcome difference between controlled and treatment groupnCommonly used to answer questions in natural science,but difficult to implement to answer questions in social science for various issues随机控制实验了解接受远端祈祷心脏病病人的治疗效果13MEASURING CAUSATION WITH DATA WED LIKE TO HAVE:RANDOMIZED TRIALSpWith random
14、assignment,the assignment of the intervention is not determined by anything about the subjects.pAs a result,the treatment group is identical to the control group in every facet but one:the treatment group gets the intervention.随机控制实验了解接受远端祈祷心脏病病人的治疗效果14Example I:does the payer work?pExample(Harris e
15、t al.):“隨機控制實驗:瞭解接受遠端祈禱心臟病病人的治療效果“随机控制实验了解接受远端祈祷心脏病病人的治疗效果15Harris et el.p實驗設計原則:“隨機,控制,雙盲,事前,同時實驗.“n隨機:病人隨機分配到禱告與否n控制:有些病人沒有禱告n雙盲:病人或醫師不知道為實驗或對照組n事前:在治療前隨機分配n同時:實驗同時進行随机控制实验了解接受远端祈祷心脏病病人的治疗效果16Harris et el.的設計随机控制实验了解接受远端祈祷心脏病病人的治疗效果17随机控制实验了解接受远端祈祷心脏病病人的治疗效果18Harris et el.的結論p“結論:遠端禱告有效”随机控制实验了解接受
16、远端祈祷心脏病病人的治疗效果19TANF and labor supply among single motherspTANF is“Temporary Assistance for Needy Families.”pCash welfare for poor families,mainly single mothers.nFor example,in New Mexico,family of three receives$389 per month.pAssume the two“goods”in utility maximization problem are leisure and fo
17、od consumption.pWhatever time is not devoted to leisure is spent working and earning money.随机控制实验了解接受远端祈祷心脏病病人的治疗效果20Randomized Trials in the TANF ContextpIt is believed that an increases in labor supply when TANF benefits are cut,but the magnitude of the effect is unclear.pOne could design a random
18、ized trial to learn about the elasticity of employment with respect to TANF benefits.随机控制实验了解接受远端祈祷心脏病病人的治疗效果21Randomized Trials in the TANF ContextpImagine a large group(say,2000)of single mothers were randomly assigned to one of two groups with a coin flip:nThe“control”group continues to receive a
19、 guarantee of$5,000.nThe“treatment”group now has their TANF benefit cut to$3,000.pFollow groups for a period of time,and measure the work effort.随机控制实验了解接受远端祈祷心脏病病人的治疗效果22Randomized Trials in the TANF ContextpIn an experiment like this in California in 1992,the elasticity of employment with respect
20、to welfare benefits was estimated to be-0.67.pThus,a 10%decrease in benefits resulted in a 6.7%increase in employment.随机控制实验了解接受远端祈祷心脏病病人的治疗效果23Why We Need to Go Beyond Randomized TrialspRandomized trials present some problems:nThey can be expensive.nThey can take a long time to complete.nThey may r
21、aise ethical issues(especially in the context of medical treatments).nThe inferences from them may not generalize to the population as a whole.nSubjects may drop out of the experiment for non-random reasons,a problem known as attrition.随机控制实验了解接受远端祈祷心脏病病人的治疗效果24Why We Need to Go Beyond Randomized Tr
22、ialspFor these reasons(especially the first one about randomized trials being expensive),economists often take different approaches to try to assess causal relationships in empirical research.随机控制实验了解接受远端祈祷心脏病病人的治疗效果25What is your identification strategypHow can you obtain the casual effect using yo
23、ur observational data?pDrawing inference from the observational data needs to be very carefulpLikely to suffer from the selection bias 随机控制实验了解接受远端祈祷心脏病病人的治疗效果26Example Ip根據9999泛亞人力銀行調查顯示,有六成八的大學應屆畢業生求職處處碰壁,到現在還找不到工作,而這些失業青年目前生活經費的主要來源,有七成以上是在家靠父母親友養,更令人憂心的是,有一成的社會新鮮人是以借貸、舉債過日。9999泛亞人力銀行營運長楊肯誠認為,八月份
24、二技、四技二專考試放榜後,部分為被錄取的高職、專科生也將投入就業市場,預估九月份的應屆畢業生失業率更為嚴重,有可能超越去年九月的七成二。p該人力銀行昨天在台北的喜來登酒店舉行記者會,公布八月三日至十四日所進行今年找全職工作的應屆畢業生就業狀況調查,針對跟人力銀行資料庫符合條件的九萬二千一百五十二名應屆畢業生發出問卷,有效樣本共一萬九千三百七十八份,回收率百分之廿一,當信心水準為百分之九十五時,誤差值為正負百分之零點八。受訪樣本集中在大專以上學歷為主,占全體百分之八十七點五。試就上述論點一一評論下列敘述:随机控制实验了解接受远端祈祷心脏病病人的治疗效果27pDo you believe the
25、number?pBefore you believe in the observational data,you should nExamine the numbers with your personal experiencesnExamine the numbers with the government datanExamine the numbers with other studiespMore than half times these numbers are incorrect,so betting on disbelieving these numbers is usually
26、 smarter随机控制实验了解接受远端祈祷心脏病病人的治疗效果28pWhat is the problem?nProvide basic survey statistics:sample,response rate,sample size,confidence intervalnThe majority of the studies have the response rate is lower than 70%nFor confidential questions,the response rate could be even lower than 30%随机控制实验了解接受远端祈祷心脏病
27、病人的治疗效果29pWhere does the bias come from?nThose who have found the job is unlikely to respond to the emailnTherefore,the majority of them should be the ones who cannot find the jobsnIf we do the reweighing this factor,then the unemployment rate would drop to less than 15%随机控制实验了解接受远端祈祷心脏病病人的治疗效果30pWh
28、at is the correct estimates随机控制实验了解接受远端祈祷心脏病病人的治疗效果31Example IIpNHIS asks about the health status(1 best and 5 worst)of individuals admitted and not admitted for hospitals pHospital makes you sicker!随机控制实验了解接受远端祈祷心脏病病人的治疗效果32How to solve the problem?pQuasi-random assignmentpInstrumental variable met
29、hodpPropensity score matching methodpRegression随机控制实验了解接受远端祈祷心脏病病人的治疗效果33Quasi-ExperimentspEconomists typically cannot set up randomized trials for many public policy discussions.Yet,the time-series and cross-sectional approaches are often unsatisfactory.pQuasi-experiments are changes in the economi
30、c environment that create roughly identical treatment and control groups for studying the effect of that environmental change.nThis allows researchers to take advantage of randomization created by external forces.随机控制实验了解接受远端祈祷心脏病病人的治疗效果34Quasi-ExperimentspBasic approach is to let outside forces do
31、the randomization for us.In some cases,the situation happens naturally.nSuppose,for example,that Arkansas cut its TANF benefit by 20%in 1997,and that we had a large sample of single mothers in Arkansas in 1996 and 1998.nAt the same time,imagine that Louisianas benefits remained unchanged.随机控制实验了解接受远
32、端祈祷心脏病病人的治疗效果35Quasi-ExperimentspIn principle,the alteration in the states policies has essentially performed our randomization for us.nThe women in Arkansas who experienced the decrease in benefits are the treatment group.nThe women in Louisiana whose benefits were unchanged are the control.nBy com
33、puting the change in labor supply across these groups,and then examining the difference between treatment(Arkansas)and control(Louisiana),we can obtain an estimate of the impact of benefits on labor supply that is free from bias.随机控制实验了解接受远端祈祷心脏病病人的治疗效果36Quasi-ExperimentspImagine we simply studied s
34、ingle mothers in Arkansas alone.pArkansas has essentially performed an“experiment”where single mothers in 1996 are the control group,and those in 1998 are the treatment group.pIn practice,this comparison runs into the criticisms that confront us with time series analysis.nFor example,the national ec
35、onomy was growing exceptionally fast during this period.随机控制实验了解接受远端祈祷心脏病病人的治疗效果37Quasi-ExperimentspBecause of these concerns about national trends,the quasi-experimental approach includes the extra step of comparing the treatment group for whom the policy changed to a control group for whom it did
36、not.pSingle mothers in Louisiana did not experience the TANF cut,yet benefit from the growth in the economy.随机控制实验了解接受远端祈祷心脏病病人的治疗效果38Quasi-ExperimentspThat is,by examining hours of work in Arkansas,we obtain:nHOURSAR,1998-HOURSAR,1996nThis contains both the treatment effect and the bias from the ec
37、onomic boom.pIn contrast,by examining hours of work in Louisiana,we obtain:nHOURSLA,1998-HOURSLA,1996nThis contains only the effect of the economic boom.随机控制实验了解接受远端祈祷心脏病病人的治疗效果39Quasi-ExperimentspBy subtracting the change in hours of work in Louisiana from that in Arkansas,we control for the bias c
38、aused by the economic boom.pWe obtain a causal estimate of the effect of TANF benefits on hours of work.pAn example is given in,first focusing on Arkansas alone.随机控制实验了解接受远端祈祷心脏病病人的治疗效果40Table 1Using Quasi-Experimental VariationArkansas19961998DifferenceBenefit Guarantee$5,000$4,000-$1,000Hours of W
39、ork Per Year1,0001,200200CPS data shows actual hours of work for single mothers in Arkansas.At the same time benefits were being cut,hours of work increased for single mothers in Arkansas.Benefits fell by$1,000 in Arkansas during 1997.随机控制实验了解接受远端祈祷心脏病病人的治疗效果41Quasi-ExperimentspWhile benefits fell b
40、y 20%,hours of work increased by 20%;the implied elasticity of labor supply with respect to benefits levels is-1.pThis is larger than the-0.67 elasticity estimate found in the randomized trial in California.随机控制实验了解接受远端祈祷心脏病病人的治疗效果42Quasi-ExperimentspThere is likely to be bias in this“first-differen
41、ce,”because there was major economic growth during this period.nThus,single mothers in Arkansas may have increased their work effort even if TANF benefits had not fallen.pWe examine single mothers in the neighboring state of Louisiana,in the bottom panel of.随机控制实验了解接受远端祈祷心脏病病人的治疗效果43Table 1Using Qua
42、si-Experimental VariationArkansas19961998DifferenceBenefit Guarantee$5,000$4,000-$1,000Hours of Work Per Year1,0001,200200Louisiana19961998DifferenceBenefit Guarantee$5,000$5,000$0Hours of Work Per Year1,0501,10050We can gather the same kind of data for Louisiana.Benefit levels did not fall in Louis
43、iana.But labor supply still increased,perhaps due to the growing economy.It appears that 50 hours of the 200 hour increase was due to economic conditions,not TANF.随机控制实验了解接受远端祈祷心脏病病人的治疗效果44Quasi-ExperimentspThis approach yields the difference-in-difference estimator the difference between the change
44、s in outcomes for the treatment group that experiences an intervention and a control group that does not.pWe are taking the difference in labor supply changes in these states in an attempt to purge the estimate of bias(due to the growing economy).nWhile cross-sectional analysis would suggest that th
45、e reduction in welfare benefits leads to a 100-hour increase in work,the difference-in-difference analysis suggests a 150-hour increase.随机控制实验了解接受远端祈祷心脏病病人的治疗效果45Quasi-ExperimentspThe difference-in-difference estimator is:pThe second term,for Louisiana,nets out the bias from the growing economy.pThu
46、s,the causal effect of TANF benefit cuts would be a 150-hour increase in labor supply.HOURSHOURSHOURSHOURSAKAKLALA,1998199619981996随机控制实验了解接受远端祈祷心脏病病人的治疗效果46Quasi-Experiments:Problems with quasi-experimental analysispThis approach also has problems,however.nIt is possible that the economic boom affected Arkansas differently than it did Louisiana.nMore generally,single mothers may be different across states.pWe can never be completely certain that we have purged the treatment-control comparisons of bias.随机控制实验了解接受远端祈祷心脏病病人的治疗效果47