I have 3 measurement points, N=4000, N=2000, N=3000 (week 0, week 6, week 12).

My goal is to estimate the relative importance of 14 regressors on one dependent variable. Now I can do that separately for each time point, but the N are comparably small to estimate relative regressor importance with such a huge number of regressors.

So I thought about simply "pooling" participants, pretending the data are cross-sectional, and using N=4000+2000+3000. Obviously this violates the assumption of independence of observations, so I should not do that.

Alternatives? I can't use a regression in which I simply add time as regressor because the bottleneck week 6 (only N=2000) will lead to listwise deletion of 50% of the subjects.

  • 1
    $\begingroup$ Have you thought about linear mixed models or marginal models? $\endgroup$
    – Davide
    Commented Dec 18, 2013 at 14:30

1 Answer 1


Fourteen predictors is usually not an unworkable number with 2,000 - 4,000 observations, so a separate regression at each time point does not seem out of the question. Also, it's not a foregone conclusion that if you include time as a 15th predictor you have to use listwise deletion. Have you found that the data are missing in such a way that pairwise deletion would lead to unsound findings?

  • $\begingroup$ 14 predictors for the regression is fine. After that I run relative importance analysis, and bootstrap CIs of the RelaImpo estimates. These CIs are pretty large (due to correlation structure of predictors and the way they are bootstrapped). N=2k is not large enough for this analysis. The missings are weird. There is about 25% dropout from week 0 to week 12, and week 6 seems to be just a smaller sample (reanalysis don't know why that is). I don't know of any other way than listwise deletion in regression analysis with multiple timepoints. What would you suggest? Thank you $\endgroup$
    – Torvon
    Commented Jan 21, 2013 at 12:22
  • $\begingroup$ Sounds as if 3 separate regressions are the way to go, and that you'd want to condense your 14 correlated predictors into a smaller number of more less correlated ones. Whether it's best do that via judgment-based selection, or PCA, or EFA is hard for me to say. $\endgroup$
    – rolando2
    Commented Jan 21, 2013 at 20:44
  • $\begingroup$ My dissertation is about a higher resolution of symptoms instead of simplifying with EFA/PCA/CFA/LCA, so I cannot do that. I need to do this regression longitudinal, minimizing loss of subjects. What happens when I simply go with this: "y ~ x1 + x2 + x3 ... + age + sex + TIME" instead of running "y ~ x1 + x2 + x3 ... + age + sex" per measurement point ? Is that feasible? $\endgroup$
    – Torvon
    Commented Jan 22, 2013 at 13:51

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.