Scribd
Upload
84 views2 pages

Triple Difference Statistics Tool

This document discusses difference-in-difference and triple difference estimators. It explains single difference, double difference, and triple difference estimators. It also discusses some concerns with difference-in-difference estimators, including time-invariant selection bias, and provides some examples and strategies to address these concerns, such as propensity score matching.

Uploaded by

aakash_kalyani
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
84 views2 pages

Triple Difference Statistics Tool

This document discusses difference-in-difference and triple difference estimators. It explains single difference, double difference, and triple difference estimators. It also discusses some concerns with difference-in-difference estimators, including time-invariant selection bias, and provides some examples and strategies to address these concerns, such as propensity score matching.

Uploaded by

aakash_kalyani
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

A Note on Difference-in-Difference and Triple Difference Estimators

Xi Chen Cornell University http://web.cenet.org.cn/web/ecocxi/ Single-Difference (D) Estimator - Only require a single survey - Selection bias BTT ( X ) and missing data on YitT for Tit = 0 and YitC for Tit = 1

- More insight if outcomes tracked for both participants and nonparticipants over time The Double-Difference estimator - Essential idea: compare samples of participants and non-participants before and after the intervention (initial baseline survey of both non-participants and (subsequent) participants + follow-up survey of both groups after the intervention) o Calculates the difference between the after and before values of the mean outcomes for each of the treatment and comparison groups o The difference between these two mean differences is the impact estimate - How to solve the missing-data problem? The DD estimator assumes that the selection bias BTT ( X ) is time invariant: outcome changes for non-participants reveal the counterfactual outcome changes.

- Weaker assumption than BTT ( X ) = 0 in Single-Difference Estimation - DD estimator gives the mean treatment effect on the treated for period 1:

- Regression based DD estimator: The data over both time periods and across treatment status are pooled and one runs the regression: o The single-difference estimator o The DD estimator - Notice: balanced panel not necessary; can be generalized to multiple periods -1-

Concerns about the DD Estimator - Ex ante prediction of programs participation o Sometimes it is unknown who will participate in the program o Make an informed guess in designing the sampling for the baseline survey. o Types of observation units with characteristics making them more likely to participate will often have to be over-sampled o Example: Ravallion and Chen (2005) JPubEcon Paper - Time-invariant selection bias o o Example 1: returns to infrastructure - Example 2 with illustration: poor-area development programs (Jalan and Ravallion, 1998 JPubEcon Paper) show that poor-area development projects in rural China have been targeted to areas with poor infrastructure and that these same characteristics resulted in lower growth rates.

o The upshot of these observations is that controlling for initial heterogeneity is crucial to the credibility of DD estimates. (1) Using PSM for selecting the initial comparison group will almost certainly reduce the bias in DD estimates (Ravallion and Chen, 2005 JPubEcon Paper). (2) The DD method can also be combined with a discontinuity design (Jacob and Lefgren, 2004 REconStat Paper). Triple Difference (DDD) Estimator - Difference between the double differences for stayers and leavers - If baseline data unavailable, impacts may still be identified by observing participants outcomes in the absence of the program after the program rather than before it - identify the mean gain to participants at date 1 (TT) under two conditions: o there is no selection bias on who leaves the program o there are no current gains to non-participants - A third survey round allows a joint test of these two conditions (logic: estimation of gains to participants in period 1 independent of drop-out in period 2) - Example: Ravallion et al. 2005 JHumanRes Paper - If no comparison group of nonparticipants, DD estimator for stayers vs. leavers -2-

You might also like