I have longitudinal data from N = 80 people who participated in 12 short monthly assessments. Around 40 participated in 10-12 of the interviews, the rest dropped out due to different reasons. Aim of the study is to assess the pattern of the course of symptoms during the 12 months. I checked whether the missings are MCAR, MAR, or MNAR - as far as I can see from the dataset they are MAR. To be able to calculate with the data, I would like to do multiple imputation with the missings (as far as I read this would be the optimal procedure). So far, MICE was the package I was looking at. Doing this several questions arose:

  1. Would multiple imputation be the right way to impute the missings? The persons have very different symptom courses (some decrease, others increase, others don't change over the 12 months) and I don't want the pattern of other persons to influence the pattern of a specific person (e.g. one with 2 time points missing). Is there a way that the pattern of other persons does not influence the imputation of a specific person?

  2. So far I used the following formula

    dataset_withoutmissings <- mice(dataset_withmissings) data_nomissings <- complete(dataset_withoutmissings, 1)

to impute the missings - I assume this is only the basic function and I am not quite sure whether I should add something important?

  • $\begingroup$ You have lots of questions here and it would be better to ask one (or perhaps two) at a time with links back to the previous one. At the moment people who could help with one question are not going to be able to answer. $\endgroup$ – mdewey Aug 2 '17 at 13:57
  • $\begingroup$ What kind of test are you wanting to do? Are you even sure that you want to impute missing data? Is the loss to follow-up random? You might consider doing a linear mixed model, which allows missing data without imputation. $\endgroup$ – Tami Aug 2 '17 at 13:57
  • $\begingroup$ Thank you for the comments! I want to do plots on the structure of the data (spaghetti plots), calculations on the changes between specific time points and its influencing variables and as you already pointed out mixed models. I do know that I don't need it for the mixed models but I do need it for the other calculations. $\endgroup$ – ElI Aug 2 '17 at 14:59
  • $\begingroup$ It all depends on the type of missingness, Missing Completely at Random (MCAR) , Missing at Random (MAR) or Missing Not at Random (MNAR). Mice works fine for MCAR and MAR but not for MNAR. MNAR occurs when the missingness depends on the value of the variable that you measure, e.g. you want to compute the average the length of people and the longer people tend not to answer the survey. see e.g. statisticalassociates.com/missingvaluesanalysis_p.pdf or google for MCAR, MAR, MNAR $\endgroup$ – user83346 Aug 2 '17 at 15:37
  • $\begingroup$ I checked for MCAR, MAR, MNAR - they are MAR. $\endgroup$ – ElI Aug 3 '17 at 7:29

Multiple imputation is an appropriate approach for your situation but you need to account for the multilevel nature of your data. The observations are nested within participants and this fact needs to be considered when making the imputations. So you will need to select a multilevel imputation method. MICE offers several such methods and they all begin with 2L dot (e.g., 2l.norm). You can read more about the approach in this article: "Multilevel multiple imputation: A review and evaluation of joint modeling and chained equations imputation" http://doi.org/10.1037/met0000063 and you can read more about implementing it in MICE in this article: "Multivariate imputation by chained equations" http://doc.utwente.nl/78938/1/Buuren11mice.pdf


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.