0

I have a dataset that has 6 metabolites that were measured over time in two groups and using a mixed-effects linear model I would like to investigate the group differences for each metabolite. Since the dataset contained missing variables I did imputation using mice (resulted in 50 generated datasets).

As proposed by the mice package and several online resources this is how I analysed the data:

  1. imputed the data
  2. applied mixed-effects model (for all 6 metabolites)
  • metabolie1 = group + time + sex + (1|subjectID))
  • metabolie2 = group + time + sex + (1|subjectID))
  • metabolie3 = group + time + sex + (1|subjectID))
  • metabolie4 = group + time + sex + (1|subjectID))
  • metabolie5 = group + time + sex + (1|subjectID))
  • metabolie6 = group + time + sex + (1|subjectID))
  1. pooled the data (Rubin's rules)
  2. looked at confidence intervals of the beta for the groups
  3. (? What should I do with the p-values ?)
  • p.adjust( model_metabolite1_beta_for_group, model_metabolite2_beta_for_group, ..., model_metabolite6_beta_for_group,) ?

I then looked at my confidence intervals for the studygroups and concluded, that if the CI is not crossing zero, that there is a difference. However, the model also provides me p-values and I should report those. I was now wondering how I should correct for multiple testing? I thought of using False discovery rate correction and then just taking all my p-values for the group generated by each model, but I am not sure if this is correct.

CST
  • 159
  • 1
  • 6

1 Answers1

0

You could use standard approaches for correcting for multiple testing by adjusting the p-values, e.g., Bonferroni, Holm's method, false discovery rate. In R most of these methods are implemented in the p.adjust() function.

Dimitris Rizopoulos
  • 17,519
  • 2
  • 16
  • 37
  • Thank you for your comment, I will try to use p.adjust. Do you have any thoughts on whether my approach of just adding the p values of the studygroup beta estimate of each model would be correct? – CST Mar 19 '21 at 09:01
  • 1
    Just to be clear, I meant you can adjust for multiple comparisons the p-values you obtain based on the pooled results. The p-values from the individual models fitted to each imputed dataset cannot be simply added up. – Dimitris Rizopoulos Mar 19 '21 at 09:08
  • I have one dataset, that I imputed (generated 50 sets). I applied the 6 models on the same dataset (different metabolite variable as dependent variable), pooled the results from the 50 datasets and obtained estimates, CI and p-values for each of the 6 models. Can I now just simply take the 6 p-values of the pooled result of each model and adjust for multiple testing? – CST Mar 19 '21 at 12:40