The publicly available data I used


I addressed the unbalanced followup visits by transforming the visit time to the nearest multiple of 8 then de-duplicated visit times.

I then used last observation carried forward to impute missing log CD4 values. Then ran two linear mixed intercept models (log CD4 ~ time) on the raw data and on the imputed data. I checked the residual plots and I got these.

Raw model residuals Raw data model residuals Imputed model residuals Imputed data model residuals

So I see that my imputation method has made some of the residuals more normal, but why is there that huge group of fitted values around 3? How would I be able to address that?

The models I fit:

raw_3_model <- lme(log_CD4_1 ~ Time, random = ~Time|ID, 
                   data = raw_data, method = "REML",
                   na.action = na.exclude)

treatment_3_model <- lme(measurements ~ time, random = ~time|ID, 
                         data = final_3, method = "REML",
                         na.action = na.exclude)
  • $\begingroup$ Could you post the models you fit? $\endgroup$ – user158565 Dec 7 '18 at 19:30
  • $\begingroup$ @user158565 I just added them in! $\endgroup$ – j681 Dec 7 '18 at 19:33
  • $\begingroup$ I did not get the results like you did. $\endgroup$ – user158565 Dec 7 '18 at 23:25
  • $\begingroup$ @user158565 Would there be a way I could see what your residuals look like? Was it of the raw data? $\endgroup$ – j681 Dec 8 '18 at 22:03
  • $\begingroup$ "transforming the visit time to the nearest multiple of 8 then de-duplicated visit times." In fact, I think I misunderstood this sentence. What exactly did you do on time variable? $\endgroup$ – user158565 Dec 8 '18 at 22:05

A couple of points:

  • Mixed models work with unbalanced data. Hence, you do not need to transform the visit times to be balanced nor to delete any measurements. Actually, if you are interested in the longitudinal evolution it is better not to have balanced data.
  • Mixed models will provide you with valid inferences under the missing at random missing data mechanism. Hence, you do not need to impute any missing data. Moreover, the last observation carried forward is a terrible method of imputation that does not even provide correct inferences under the missing completely at random mechanism.
  • In the residuals plots you see two things: (1) the vertical lines around 3 come from the baseline measurement because all subjects were measured at 0; (2) the diagonal lines on the bottom left come from the bounded nature of the CD4 cell count outcome (i.e., it is greater or equal than zero and you have measurements at and close to this boundary).
  • $\begingroup$ That's really helpful! Thank you Dimitris! Should I impute missing baseline information then? And keep the followup data missing as is? $\endgroup$ – j681 Dec 9 '18 at 21:21
  • $\begingroup$ You mean imputing the baseline value of the outcome CD4 cell count or baseline covariates? Regarding the former, if you do not include the baseline CD4 cell count measurement as a baseline covariate in the model, you do not need to impute it. Regarding the latter, you could consider imputing covariates with missing data. $\endgroup$ – Dimitris Rizopoulos Dec 10 '18 at 8:44

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.