Merit of t-test vs repeated measures

Question

In a cohort study, patients have been enrolled for a new method of treatment of a condition. The measurement of interest, for example, tumour size is taken at baseline (BL) ie before treatment, the treatment is then given and the tumour size is then measured every 4 weeks for 40 weeks (ie we have a total of 11 measurements with one BL and the rest post treatment). The analysis of interest is change in tumour size from BL to week 40.

This was recommended by someone else to be done using a paired sample t-test between BL and week 40 and supplement with a p-value and confidence interval. However I feel that using a repeated measures mixed effects model with tumour size as dependant variable and time and possibly the BL value as independent variables (ancova), would be more suitable. This way we utilise all of the data rather than just the two time points.

Please can someone explain to me the benefit of using one method over the other? Can the t-test give a realistic view of the treatment impact whilst ignoring all the middle values?

Answer 1

This is not so much a question about what test to use, but what question you're asking.

If this were a randomised controlled trial, you would be forced to define a primary outcome. For example, it might be survival at 40 weeks in those receiving drug A versus those receiving drug B. Given that you have defined the outcome, it is easy enough to choose the test statistic.

In your case, I think you and four friend are both interpreting the situation in different, yet equally valid, ways. That is, you are interested in one research question and he is interested in another.

I suggest that you think carefully about the research question you wish to answer. Then, the analysis plan should flow naturally from that point.

Good luck.

---

ADDED FOLLOWING COMMENT BY OP

It seems to me that your initial response as outlined in your question and the comment below is spot on. In your situation, a paired t-test would compare changes in tumour size in one group, so you would conduct two paired t-tests, each testing the statistical significance of changes in each of the groups. The problem with cohort studies is that the two cohorts may be different in ways that relate to the outcome. That is, there is likely to be some degree of confounding that you will need to consider. If so, then a t-test is a fairly limited test to deploy. In addition, you might wish to explore additional secondary questions, such as the present of effect modification. Again, a t-test won't be able to help.

For the reasons above plus the reasons you yourself suggest, it would be better to propose a comprehensive analytic plan involving regression techniques. Not only do these offer the chance to include the entire experience of your subjects, but it allows you the flexibility to adjust for confounding, to explore effect modifiers and make out-of-sample predictions.

I hesitate recommending a particular type of regression as these will depend on the questions you wish answered. However, this will give you the chance to clarify the intent of your investigators.