A Review of the Behrens–Fisher Problem and Some of its Analogs

Abstract

The traditional Behrens–Fisher (B-F) problem is to test the equality of the means µ1 and µ2 of two normal populations using two independent samples, when the quotient of the population variances is unknown. Welch [43] developed a frequentist approximate solution using a fractional number of degrees of freedom t-distribution. We make a a comprehensive review of the existing procedures, propose new procedures, evaluate these for size and power, and make recommendation for the B-F and its analogous problems for non-normal populations. On the other hand, we investigate and answer a question: does the same size fit all all, i.e. is the t-test with Welch’s degree of freedom correction robust enough for the B-F problem analogs, and what sample size is appropriate to use a normal approximation to the Welch statistic.

Keywords

• the Behrens–Fisher problem
• the beta-binomial model
• the negative binomial model
• the Weibull model