Combining Difference-in-difference and Matching for Panel Data Analysis

Transcription

1 Combining Difference-in-difference and Matching for Panel Data Analysis Weihua An Departments of Sociology and Statistics Indiana University July 28, / 15

2 Research Interests Network Analysis Social influence and networks Network and measurement Text networks (social media, citation, biographies, sports records) Causal Inference Matching and propensity score methods Instrumental variable methods Causal inference under interference Applied Research Social policy (e.g., network and neighborhood) Organizations (e.g., network and cognition) 2 / 15

3 Outline Previous Parametric Models for Panel Data Analysis Combining Difference-in-difference and Matching One Example Summary 3 / 15

4 Previous Parametric Models Assume that in the absence of treatment, the potential outcome for subject i at time t is Y 0 it = λ t + X it γ + C i + e it, Assuming an additive constant treatment δ, we can write the potential outcome for subject i at time t under treatment as Y 1 it = Y 0 it + δ. This implies that using the observed outcome, we can write Y it = λ t + δd it + X it γ + C i + e it, (1) The detail of the models and methods can be found in chapter 9 of Morgan and Winship (2007) and Wooldridge (2001). 4 / 15

5 Parametric Models Random effects: C i N(0, σ 2 ) Fixed effects: using differenced outcomes to remove C i. FE model: Y it Ȳi = δ(d it D i ) + (X it X i )γ + (e it ē i ) FD model: Y it = λ t + δ D it + X it γ + e it Random trend and slope: Y it = λ t + g i t + δd it + h i D it + X it γ + C i + e it MA(1): e it = ρe it 1 + v it AR(1): Y it = λ t + θy it 1 + δd it + δ 1D it 1 + X it γ + C i + e it 5 / 15

6 Matching Definitions of treatment effects: ATE: δ = E[Y 1 Y 0 ] ATT: δ 1 = E[Y 1 Y 0 D = 1] The basic idea of matching is to match units with exact covariates in the opposite treatment group to impute the missing potential outcomes. The key of matching is to measure similarity between units. The distance between two units in their covariate values is often measured according to d(x i, X j ) = (X i X j ) T W 1 (X i X j ). There are two popular choices of the weight matrix: (1) the variance-covariance matrix of X ; and (2) the sample variances of X. Estimating ATE: ˆδ = 1 N N (Ŷi 1 Ŷi 0 ) = 1 N N (2D i 1)(1 + K i )Y i σ 2ˆδ = 1 N {1 + K N 2 i } 2 σ 2 (Y i X i, D i ) where K i is the number of times unit i serves as a match to other units. 6 / 15

7 Difference-in-difference (DID) DID assumes that in the absence of treatment the original difference between treated units (denoted by (1)) and control units (denoted by (0)) in the outcome will remain constant over time. E[Y 0 t 1(1) Y 0 t 1(0)] = E[Y 0 t (1) Y 0 t (0)] which implies E[Y 0 t (1) Y 0 t 1(1)] = E[Y 0 t (0) Y 0 t 1(0)], or say, Y 0 t D t. With this assumption, we can estimate the ATT as Y t(1) Y t(0). E[ Y t(1) Y t(0)] = E[Y t(1) Y t 1(1) (Y t(0) Y t 1(0))] = E[Y 1 t (1) Y 0 t 1(1) (Y 0 t (0) Y 0 t 1(0))] = E[Y 1 t (1) Y 0 t 1(1) (Y 0 t (1) Y 0 t 1(1))] = E[Y 1 t (1) Y 0 t (1)] = ATT 7 / 15

8 DID + Matching Assumptions: Strong form of ignorability Weak form of ignorability: Yt 1, Yt 0 D t X t, D t 1. (2) Yt 1, Yt 0 D t Xt,s, D t 1,s. (3) In short, what we propose is to first-difference the outcome and then apply matching to estimate treatment effects at each wave. Using the differenced outcome helps remove the effects of time-invariant confounding factors. Matching is nonparametric and helps balance covariates and create a more focused causal inference. 8 / 15

9 Aggregating Treatment Effects across Waves ˆδ W = ˆσ 2ˆδW = T t=2 T t=2 N t N ˆδ t N 2 t N 2 ˆσ2ˆδt g<h T N g N h N 2 Cov(ˆδ g, ˆδ h ) where Cov(ˆδ g, ˆδ h ) is the covariance of treatment effects across waves g and h. Estimating the covariance terms: ˆδ g = 1 N g (2D ig 1)(1 + K ig ) Y ig N g ˆδ h = 1 N h N h Cov(ˆδ g, ˆδ h ) = Cov( 1 n (2D ih 1)(1 + K ih ) Y ig n J ig Y ig, 1 n n J ih Y ih ) = 1 n 2 where n is the number of common observations in waves g and h, J ig = (2D ig 1)(1 + K ig ), and J ih = (2D ih 1)(1 + K ih ). n J ig J ih Cov( Y ig, Y ih ) 9 / 15

10 One Example The goal is to study race-of-interviewer effects (ROIE) in the General Social Survey (GSS). Specifically, we are interested in whether an interviewer s race will affect respondents responses. Ten outcome measures: intelligence gap between blacks and whites, respondents views on the government s responsibility and spending and their confidence in political and financial institutions. The major explanatory variable: interviewer s race(1 = black; 0 = otherwise). Other interviewer s characteristics include age, sex, and years of working for the GSS. Control variables: respondent s age, sex, race, employment status, family income, number of children, marital status, years of education, party identification, religious denomination, residential place (e.g., large city vs. small town), and residential region. 10 / 15

11 Table 3. Estimated Race of Interviewer Effects on Ten Selected Outcomes (1) (2) (3) (4) (5) (6) Outcomes RE FE FD RTS MA(1) AR(1) Perceived Intelligence Gap 0.43*** 0.36*** 0.32** 0.43*** 0.43*** ,616 2,617 1,407 2,616 2, Confidence in Executive Branch of Fed. Govt ,695 2,696 1,449 2,695 2, Confidence in Congress 0.07* * 0.07* ,700 2,701 1,451 2,700 2, Confidence in Supreme Court * ,684 2,685 1,432 2,684 2, Confidence in Bank and Financial Institutions ,713 2,714 1,461 2,713 2, Spending on Welfare 0.19*** 0.19** 0.16* 0.19*** 0.19*** 6* ,990 1,991 1,059 1,990 1, Spending on Blacks 0.32*** 9*** 8*** 0.32*** 0.32*** 0.30* ,858 1, ,858 1, Should Help Blacks 0.57*** 0.53*** 0.47*** 0.56*** 0.57*** 0.46*** ,662 2,663 1,405 2,662 2, Should Help Poor ,676 2,677 1,423 2,676 2, Should Help Sick ,683 2,684 1,428 2,683 2, Note: Models 1-6 are random effects, fixed effects, first difference, random trend and slope, dynamic models with 11 / 15

12 Table A4. Matching Estimates of the Race of Interviewer Effects at Waves 2 and 3 ATE ATT Outcomes Est SE N Est SE N Wave 2 Perceived Intelligence Gap ** *** 82 Confidence in Executive Branch of Fed. Govt Confidence in Congress Confidence in Supreme Court Confidence in Bank and Financial Institutions Spending on Welfare Spending on Blacks ** *** 63 Should Help Blacks ** * 89 Should Help Poor Should Help Sick Wave 3 Perceived Intelligence Gap *** Confidence in Executive Branch of Fed. Govt Confidence in Congress Confidence in Supreme Court Confidence in Bank and Financial Institutions *** Spending on Welfare *** * 74 Spending on Blacks *** 65 Should Help Blacks * * 97 Should Help Poor Should Help Sick Note: Panel 1 shows matching estimates of the ROIE for all respondents regardless whether they were interviewed by a black, i.e., the average treatment effect (ATE). Panel 2 shows matching estimates of the ROIE for respondents who were interviewed by a black, i.e., the average treatment effect on the treated 12 / 15

13 Table 6. Combined Matching Estimates of the Race of Interviewer Effects from Waves 2 and 3 ATE ATT Outcomes Est SE N Est SE N Perceived Intelligence Gap *** 1, *** 173 Confidence in Executive Branch of Fed. Govt , Confidence in Congress , Confidence in Supreme Court , Confidence in Bank and Financial Institutions , * 197 Spending on Welfare *** 1, ** 140 Spending on Blacks ** *** 128 Should Help Blacks *** 1, *** 186 Should Help Poor , Should Help Sick , Note: Panel 1 shows matching estimates of the ROIE for all respondents regardless whether they were interviewed by a black, i.e., the average treatment effects (ATE). Panel 2 shows matching estimates of the ROIE for respondents who were indeed interviewed by a black, i.e., the average treatment effects on the treated (ATT). For each measure, the first differenced outcome is used. Robust standard errors are reported. Significance code: *, P < 0.05; **, P < 0.01; ***, P < / 15

14 Comparing the Results Perceived Intelligence Gap Confidence in Presidency Confidence in Congress Confidence in Supreme Court Confidence in Financial Inst Spending on Welfare Spending on Blacks Should Help Blacks Should Help Poor Should Help Sick / 15

15 Summary Race-of-interviewer effects seems to concentrate on racially charged survey items, but not on other items. Developing DIDMatch in Stata to implement the method. The key is to estimate Cov(ˆδ g, ˆδ h ) = 1 n 2 n J ig J ih Cov( Y ig, Y ih ). 15 / 15