What Did (and Do We Still) Learn from the La Londe Dataset (Part I)? 

 In a pioneering paper,  Bob La Londe  (1986) used experimental data from the National Supported Work Demonstration Program (NSW) as well as observational data from the Current Population Survey (CPS) and the Panel Study of Income Dynamics (PSID) to evaluate the reliability with which conventional estimators recover the experimental target estimate. He utilized the experimental data to establish a target estimate of the average treatment effect, then replaced  the experimental controls with several control groups built from the general population surveys. Then he re-estimated the effects using conventional estimators. His crucial finding was that conventional regression as well as tweaks such as instrumental variables etc. get it wrong, i.e. they do not reliably recover the causal effects estimated in the experimental data. This is troubling, of course, because usually we do not know what the correct answer is, so we simply accept the estimates that our conventional estimators spit out, not knowing how wrong we may be. 

 This finding (and others) sparked a fierce debate in both econometrics and applied statistics. Several authors have used the same data to evaluate other estimators, such as several matching estimators and related techniques. In fact, today the La Londe  data  is THE canonical dataset in the causal inference literature. It has not only been used for many articles, it has also been widely distributed as a teaching tool. I think it’s about time we stand back for a second and ask two essential questions: (1) What have we learned from the La Londe debate? (2) Does it makes sense to beat this dataset any further or have we essentially exhausted the information that can be extracted from this data and need to move one to new datasets? I wholeheartedly invite everybody to join the discussion.  I will provide some suggestions in a subsequent post tomorrow.