:::

研討會總覽

Comparing the Conley-Taber and the Standard Approaches to Inference in Difference-in-Difference Models Based on Small Policy Variation: The Case of TennCare

2018/10/16
  • 研討會日期 : 2018-10-16
  • 時間 : 14:30
  • 主講人 : Professor John Ham
  • 主持人 : Professor Kamhon Kan
  • 地點 : Conference Room B110
  • 演講者簡介 : Professor Ham received his Ph.D. from Princeton University. He is currently a Professor of Economics at New York University in Abu Dhabi and Global Professor of Economics at the Wagner School of Public Policy at NYU New York. His research is in the general area of applied microeconomics, and specifically in development economics, health economics, labor economics, health economics, and experimental economics.
  • 演講摘要 : A 'weak instruments' type of problem arises in panel data models with little within variation. If there is insufficient policy variation (Conley and Taber 2011, hereafter CT), the differencein-difference (or triple difference) estimator for the impact of the policy change will not have the standard normal asymptotic distribution and hence cannot be used for inference. CT propose an approach that produces consistent estimation of confidence intervals for the impact of the policy change. We investigate whether this is an issue in the recent influential paper by Garthwaite, Gross and Notowidigdo (2014, hereafter GGN); while their analysis is based on data over eight years in 17 states, the only policy change occurs in Tennessee in 2005. To do this, we first investigate the sensitivity of the estimates across three different data sets for all of their models and for the models passing test of the çommon trends’ assumption. We argue that their resutls are decidely not robust across data sets. We take the above non-robustness as evidence that the GGN results do suffer from the identification problem raised by CT; note that many other difference-in-difference estimates also use small policy changes and are likely to have the same problem. We then turn to the CT approach for obtaining consistent confidence intervals for the Tenn Care effect using our three different data sets. We find that the CT approach produces the confidence intervals that are much more robust across data sets and more intuitively plausible.