If you have to compare a variable among 4 groups (A, B, C and D -progressive disease stages) using a comparison test (student T test) between A and B, B and C, C and D is it necessary to perform a Bonferroni`s correction? or this is required only for ANOVA test?
Thank you!
For the Student T test you should use the Cohen's D to calculate the effect size.
Generally a post hoc correction for multiple testing is sensible. Bonferroni is a poor choice as it has relatively low power. I'd suggest a modified Bonferroni procedure such as Holm or Hochberg corrections. Hochberg's correction has good power and good control of Type I error for pairwise comparisons relative to Bonferroni.
Hello Baisan Andrei. I have some questions.
Thanks for clarifying.
Dear Baisan Andrei,
According to your explanation, it does not seem correct to perform Student's T tests. It would be better to perform an ANOVA (with repeated measures if it is a follow-up design or a simple ANOVA if you have one common measure and four groups). Then, if your ANOVA is significant, you can perform a post-hoc test in order to see which group(s) has(have) significant higher or lower scores in comparison with the other groups.
If you perform your t tests like this, you will increase Type 1 error.
Baisan Andrei sent me a private message with answers to the questions I posed above. He is using SPSS. He has 4 independent groups (A-D), but wishes to make only 3 comparisons: AvB, BvC and CvD (i.e., adjacent or repeated contrasts). The sample sizes range from 20 to 33, and the SDs range from 15.7 to 42. For that reason, I would use robust tests that do not assume equal variances.
In SPSS, the UNIANOVA and GLM commands have an option for REPEATED contrasts. But unfortunately, it appears one can only get contrast results that assume equal variances (even if one uses the ROBUST option to get robust parameter estimates). Fortunately, if one uses the ONEWAY command with CUSTOM contrasts, one can get both sets of contrast results, assuming and not assuming equal variances. Here's an example. Baisan, you'll have to change the variable names, of course. And if you have not installed the SPSSINC GETURI extension command, you'll have to download the data file to your hard disk and open it in the usual manner (GET FILE).
* Use the data file from this UCLA webpage to illustrate:
* https://stats.idre.ucla.edu/spss/faq/how-can-i-do-anova-contrasts-in-spss/
.
SPSSINC GETURI DATA
URI="https://stats.idre.ucla.edu/wp-content/uploads/2016/02/crf24.sav"
FILETYPE=SAV
/OPTIONS
SHEETNUMBER=1 READNAMES=YES ASSUMEDSTRWIDTH=32767.
* Variable b has 4 levels, so use it as the group variable.
* The ONEWAY command is nice because it has two
* options for two "robust" F-tests, Brown-Forsythe
* and Welch. If the sample sizes are not all the
* same (or very similar), the ordinary F-test is
* more sensitive to heterogeneity of variance,
* so one of the robust F-tests should be used in that case.
* ONEWAY does not have a REPEATED contrast option,
* but one can use custom contrasts to get the same results.
ONEWAY Y BY b /
CONTRAST= -1 1 0 0 /
CONTRAST= 0 -1 1 0 /
CONTRAST= 0 0 -1 1 /
STATISTICS DESCRIPTIVES BROWNFORSYTHE WELCH
.
* UNIANOVA (and GLM) do have a REPEATED contrast option.
UNIANOVA y BY b
/CONTRAST(b)=Repeated
/METHOD=SSTYPE(3)
/EMMEANS=TABLES(b)
/ROBUST=HC3
/DESIGN=b.
* Notice that the REPEATED contrast results here
* assume equal variances, and there appears to be
* no way to get contrasts that do not assume
* equal variances. For that reason, I'd stick with
* the CUSTOM contrasts from ONEWAY if I had variation
* in the sample sizes and variance heterogeneity.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA