Abstract There is typically an uneasiness among a large class of experimenters in presenting regression results making full use of the time-series cross-section data from experiments. This uneasiness is expressed as concern about unaccounted dependencies in the error structure of the regression model, and in some extreme cases, they argue that each experimental session provides at most a single observation. The alternative, often recommended in these cases, is to use the Wilcoxon/Mann-Whitney test using session level mean data. Here discuss the use of panel data econometrics when there are correlations across subjects. We show that such techniques can deal with very complicated correlation structures, although we acknowledge that the small sample properties of allowing for such structures are not well known. We discuss two relatively simple estimators for less complex cases. In Section 4 we consider the Wilcoxon/Mann-Whitney test and note that this approach is based on the assumption that session level mean observations are identically and independently distributed. Thus, the test will be invalid if, for example, there are different numbers of subjects in experimental sessions (resulting in different variances for the session means) or there is heteroskedacticity in the underlying data. Finally we compare the approach of estimating the effect of learning, gender, SAT/ACT score and college major on the bids of experienced subjects in a common value auction experiment using the panel data econometrics and the Wilcoxon/Mann-Whitney approaches respectively.
|