Two different algorithms solve the same problem 30 times each, producing two sets of 30 results (60 results in total). I want to determine if the null hypothesis is correct (that both algorithms perform equally well), hopefully using a parametric test (unpaired t-test).
Assumptions for t-test include equal variances between the two groups, and that the data results from a normal distribution. Applying the Anderson-Darling test for normality invalidates the normality assumption for a single group, but not the other. I still have to implement the Levene test to check on variances, so for the moment I can't comment on the other assumption.
My question is simple: Is it statistically valid to transform the one non-normal group using a Box-Cox transformation and then performing the relevant t-test? (either normal t-test, or Welch t-test if variances differ).
As a sub-question, how do you go about reporting your analysis in a paper (do you mention transformation, box-cox coefficient etc)?
2007-03-23
03:29:08
·
2 answers
·
asked by
Anonymous
in
Mathematics