2
$\begingroup$

I have data from two groups - let's call them Group A and Group B. - from three different time points. Data are collected on day 1, day 50, and day 100. I collected writing errors at each of those time points. Both groups were reacting to an intervention and both produced errors which I categorized. The problem is that the number of participants varied across time points. So I collected the total numbers of errors so that I can compare general trends.

What I am trying to understand is how to analyze this data. I'm not sure how to interpret the changes beyond a very simple descriptive analysis along the lines of "Group B showed a larger decrease than Group A for Error Type 2".

enter image description here

These numbers refer to the % of errors by category. For example, on Day 1 the Type 2 Errors made by Group A were 22.45% of the total errors they made. Whereas on Day 100, the Type 2 Errors made by Group A had fallen to 6.45% of the total errors they made on Day 100.

Any advice on ways to analyze this data would be greatly appreciated. Even advice on simple qualitative analysis would be great.

$\endgroup$
1
$\begingroup$

If day 1 is a baseline, an ANCOVA model is well suited to characterize change between groups. Here, for each individual, you adjust for the day 1 writing errors, for the group assignment, and for the time point at which data was collected, and the interaction between group assignment and time point. The statistical test of group assignment and the interaction term suffices to test the null hypothesis that change does not differ between the two groups. Robust (sandwich) standard errors contributes greatly to the validity of the inference.

$\endgroup$
  • $\begingroup$ (+!) Although that would require obtaining dis-aggregated data. $\endgroup$ – Robert Long May 24 '19 at 5:59
  • $\begingroup$ How could I do this? Do I need a large sample size and equal samples from the to groups? $\endgroup$ – kandyman May 24 '19 at 9:38
  • $\begingroup$ @kandyman You ought to have a sufficiently powered sample. Do a sample size calculation. You don't need balance between groups, but imbalance will reduce the power. $\endgroup$ – AdamO May 24 '19 at 16:02
  • $\begingroup$ @RobertLong the data seem to be count in nature. A quasipoisson model might be an appropriate method for analyzing results in the aggregate. But I agree, lots more detail is needed... $\endgroup$ – AdamO May 24 '19 at 16:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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