# Comparing means for 2 sets of (paired?) data

N patients with N samples from site A and N from site B (before and after treatment). In both sites we find 1-4 microbes, which are tested for antimicrobial susceptibility to obtain MIC values.

I wish to determine if the mean-MIC in site-A samples are actually lower than the mean-MIC in site-B.

Which model to choose, when I need to include at least 3 covariates age, (sex), drug and duration and uncertain if the data is actually paired??

I would go for a 2-sample paired t-test - but how will can I integrate 3 covariates?

Please consider also, that in most cases the microbe in site A is the same as in site B - but in 10-15 % of the cases we have different microbes (and also different numbers). So, would it be inappropriate to pool all the data and compare the overall means? Or should I divide the data so I separately consider the groups with identical microbes and the groups with various microbes?

Given that you have missing data and that you have repeated measures (and that the assumption of sphericity is unlikely to be reasonable). I suggest a multi-level model. In R see lme4 or nlme packages. In SAS look at PROC MIXED.

• Now I understand the assumption of sphericity and yes, that would be a mistake to do. But what is my missing data? And the repeated measure is not quite right, as but group A is before and group B is after treatment. So A and B differs in both time AND condition, which is the reason I doubted that the data is paired in the first place. Sep 10, 2013 at 13:04
• Your missing data is the lack of repetition. Sep 10, 2013 at 21:55

So you have group A and group B, and you have a before and after measurement. This means that you have both time and condition, so a two-way repeated measure ANOVA is in place.

Any covariates can be included in this kind of analysis.

Hope that helps!

• Repeated measures ANOVA does not deal well with missing data and assumes sphericity. It is not a good choice here. Sep 10, 2013 at 10:47
• Thank you for both responses!! @Peter Flom However, I do not understand the assumption of sphericity nor the missing data (which is?). I expect normally distributed values for the MIC-measurements. Moreover, sorry for the confusion - but group A is before, while group B is after treatment. So A and B differs in both time AND condition, which is the reason I doubt paired data. Sep 10, 2013 at 12:39