4
$\begingroup$

I have a response variable measured at three time points per individual (week 0, 18, and 36). I am interested in differences in the change of the response over the 36 weeks within some categorical variable X.

I see two ways of modeling this.

  1. One way ANOVA with response = week_36_score - week_0_score (this seems like the simplest option)

  2. Repeated Measures ANCOVA with response = week_18_score and week_36_score. Covariate = Week_0_score.

Question

  • Which would you prefer (if any, maybe there is a better choice)?

I know there has been a lot of articles on this and they seem to say each has its own strengths and weaknesses. I believe the second model would have more power. I am not worried about a ceiling effect here.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
2
$\begingroup$

In general, I would go with a repeated measures design.

There is nothing technically wrong with the first option. However, you are essentially throwing away 1/3 of your data (and 1/2 of your non-baseline data!), which may result in a lost of power. Additionally, since you have a measurement in between baseline and 36 weeks, you cannot conclude anything about the shapes of the response profiles (i.e. test for a quadratic/cubic effects over time.)

With that being said, for your repeated measures design, I would define a new variable called response_change. I am assuming that you are interested in testing if X has some effect on changes in the responses. At week 18, this variable would take value week_18_score - week_0_score. At week 36, this would take value week_36_score - week_0_score. Using response_change is more suited to this hypothesis.

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ Great suggestions. I like the response change idea. With this approach would I still need the week 0 score as a covariate? Also what would be different if the week 36 value was week_36_score-week_18_score? $\endgroup$
    – Glen
    Commented May 3, 2011 at 16:42
7
$\begingroup$

There are difficulties in computing change. This doesn't work on ordinal repsonses, and for continuous responses makes a strong assumption of proper choice of transformations for the variables. I recommend adjusting for baseline and modeling the 2nd and 3rd measurements as longitudinal measurements.

Repeated measures ANOVA is becoming obsolete in favor of generalized least squares or mixed effects models.

Improve this answer
$\endgroup$

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.