1
$\begingroup$

I understand that people use Games Howell, Tukey or sometimes pair-wise t-test (with correction) test as post hoc analysis procedures following one-way ANOVA.

My question is following. For example, I have four groups, A, B, C and D. I did a ANOVA and found a significant p-value. If I only care about A vs B, A vs C and C vs D, what post hoc analysis procedure do I need to use?

Thanks in advance.

$\endgroup$

1 Answer 1

0
$\begingroup$

Whichever you like. Anything advertised as a post-hoc test for a one-way ANOVA will tell you about such pairwise relationships. The fact that you don't care about some of the pairs won't invalidate any of these methods.

$\endgroup$
7
  • $\begingroup$ Thanks. How about repeated measured ANOVA? The only post-hoc I know is the pair-wise paired t-test with Bonferroni (or other similar) correction. Suppose I have to use repeated ANOVA and I have, say 20 repeated measured groups but I only care the 1st group against other 19 groups. Would you recommend me to use pair-wise paired t-test with Bonferroni correction on all-comparison or just on 1st-rest compairson? $\endgroup$
    – WCMC
    Jul 3, 2017 at 3:10
  • $\begingroup$ @WCMC If you're using a Bonferroni (or, better, Holm-Bonferroni) correction, then you do need to use it for every comparison you're looking at. This said, if you're using (Holm-)Bonferroni-corrected $t$-tests on pairs, and all you care about is the pairwise differences, there's no need to do the ANOVA itself. $\endgroup$ Jul 3, 2017 at 3:20
  • 1
    $\begingroup$ Hi @Kodiologist, maybe I was unclear. For example I have 20 repeated measured groups. If I only care the 1st group vs 2nd, 1st group vs 3rd, 1st group vs 4th, ..., 1st group vs 20th, should I use pair-wise paired t-test with Holm correction on these 20 comparisons or on all-compairsons (1 vs 2, ..., 1vs20, 2vs3,...,2vs20, ..., 19vs20)? From your answer, I think on 20 comparisons was better? $\endgroup$
    – WCMC
    Jul 3, 2017 at 3:28
  • 1
    $\begingroup$ Surely the choice has a power implication? I.e. even if any valid test will be valud, some may be better choices? $\endgroup$
    – Björn
    Jul 3, 2017 at 5:50
  • 1
    $\begingroup$ You wouldn't want to make any adjustment for comparisons you didn't make. Some canned tests have been developed for common situations - e.g. Dunnett's control vs treatments test & Hsu's best vs rest test - or you can use Bonferroni adjustment for just the comparisons you make. $\endgroup$ Jul 3, 2017 at 11:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.