7

This might be a silly question, but I could not find an equivalent one on CrossValidated.

I was wondering: if I transform my data in order to meet the assumption of homoscedasticity before running a two-way ANOVA, should I also use the transformed data to perform Tukey's post-hoc test and find out which treatments are significantly different from each other?

I can sense I will get an answer that tells me it depends on my data, the kind of transformation and the resulting data, but there might an obvious explanation or a rule of thumb somewhere.

stragu
  • 429
  • 1
  • 4
  • 12

1 Answers1

9

Yes, you should use the same, transformed data throughout the analysis. Tukey's test makes the same assumptions as the ANOVA.

Placidia
  • 13,501
  • 6
  • 33
  • 62
  • 2
    +1, I think Placidia is right here; if you need to do a transformation for your ANOVA to be valid, I don't see how whether you should use the transformed data for the Tukey test can "depend" on anything. – gung - Reinstate Monica Jun 03 '13 at 01:41
  • 2
    +1 While I agree that Tukey's test makes the same assumptions as the ANOVA, I am concerned more generally about the ANOVA itself - if the data is transformed, the hypothesis on the original, untransformed scale is no longer about means, and estimated effects won't relate to mean differences. Often that won't matter, but sometimes it might. – Glen_b Jun 03 '13 at 02:52
  • 1
    That's true. You don't preserve unbiasedness by transforming the data -- but you get the same rejection region for the hypothesis test. – Placidia Jun 03 '13 at 15:07