129 – Customers and Clickstreams
The Remarkable Robustness of Ordinary Least Squares in Randomized Clinical Trials
David Ross Judkins
Abt Associates
Kristin Porter
MDRC
There has been a series of occasional papers about robust covariate control in the analysis of clinical trials in Statistics in Medicine and other journals. The robust semiparametric and nonparametric methods for statistical inference of estimated effects are fairly easy to apply with 21st century computers, but many prefer to continue using t-tests and confidence intervals based on ordinary least squares (OLS) for outcomes that clearly do not follow normal distributions. Presumably, issues of tradition and communication make it very hard to deflect this inertia. In addition, recent papers have demonstrated that the tests are asymptotically equivalent, and the more complex but less parametric procedures make little difference in practice. However, in the literature, there is not sufficient examination of whether the tests and confidence intervals based on OLS are robust to substantial excess kurtosis, particularly in small sample sizes. This paper indicates through simulation where the boundaries lie for two types of strongly nonnormal outcomes: binary outcomes and compound binary/gamma outcomes. We found that traditional OLS methods work very well down to very small sample sizes for these outcomes.