I have to analyze a study where 2 treatments (t1 and t2) were used to control blood sugar levels. The 't2' is a new treatment which is being compared with an old established treatment 't1'. I have a number of parameters of each patient: ttgroup, pre_bslevel, post_bslevel (main outcome variable), age and weight. There are 200 subjects in each group.

I have to find out if t2 is better than t1 (or not) in having a better (lower) post_bslevel. While analyzing, I have to correct for other variables listed above (pre_bslevel, age, gender and weight). In addition, I am told there may be interaction between these baseline variables with treatment effect and with each other.

To analyze, I think linear modelling with interactions may be appropriate (code in R):

lm(post_bslevel~ (ttgroup+age+gender+weight+pre_bslevel)*(ttgroup+age+gender+weight+pre_bslevel , data=mydata))

However, on reading on the net, I find mention other techniques also: ANCOVA is mentioned as the main method. Also there is mention of repeated measures ANOVA for this purpose using code as:


There is also option of t-test comparing change in blood sugar level (post_bslevel- pre_bslevel).

What will be the best method to do this analysis? Thanks for your help.


1 Answer 1


Most pressingly you need a repeated-measures design to account for the variable you measured twice (blood sugar). You could do this with a repeated-measures ANOVA (or more specifically a mixed-measures ANOVA since you would be including other non-repeated variables in your model).

This is a simple guide to choosing the right test:


And you can use any of Andy Field's texts for a readable introductory reference to the theory and practical points of running these tests in SAS, R, or SPSS.

As an aside, depending on the experimental design, you may not need to control for the other variables in the model. If this is a well-executed RCT, the differences observed in various confounders between groups arise from chance and don't need to be modelled.

  • $\begingroup$ Thanks for your response. Can you point to some reference for this statement that "in a well-executed RCT, the differences observed in various confounders between groups arise from chance and don't need to be modelled." Also if this is a well-executed RCT, what will be the R code to find which treatment results in a lower post-treatment blood sugar level. $\endgroup$
    – rnso
    Commented Nov 20, 2014 at 6:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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