1
$\begingroup$

I am just following up with my previous question here

In my case, only 2.8% of the data are missing.

enter image description here

I can consider Complete Case Analysis but I would like to study time series model and would like to fill the missing values. Both dependent and independent variables have missing data such that if the data is missing for a particular date then it is missing for all variables (monotone). The missing data is MCAR.

The question is: if I get one dataset only using MICE, what could be the potential problem? I did this using maxit=5000.

RainfallData <- mice(rainfall,m=1,maxit=5000)

I got the following convergence plot showing that the convergence was plausible. enter image description here

enter image description here

Please help.

$\endgroup$
3
  • $\begingroup$ The problem why you only get one dataset is because you specified m=1. This seems more like a R problem. $\endgroup$
    – Björn
    Jul 18, 2019 at 5:20
  • $\begingroup$ Yes I specified m=1 intentionally to achive only one dataset. Could this be a problem if I get only one dataset only. $\endgroup$ Jul 18, 2019 at 10:36
  • $\begingroup$ Of course that's a problem - analyses will not reflect the uncertainty about the missing data. $\endgroup$
    – Björn
    Jul 18, 2019 at 10:43

1 Answer 1

2
$\begingroup$

Although there is a lot of R here, it really isn't an R question, it is about multiple imputation.

Yes, it is definitely a problem to impute only one data set. The whole point of MICE (multiple imputation through chained equations) and of multiple imputation generally, is to generate multiple data sets in order to represent the variablility in the missing data - that is, we don't know what the missing data should be and pretending that we do know (but mean imputation, LOCF, single imputation or whatever) is going to distort the results.

There's debate about how many data sets you need but the lowest number I've seen recommended is 5.

$\endgroup$
4
  • $\begingroup$ Thank you Sir for your comment. I did multiple imputation before using m=10 and achieved 10 complete dataset. I was being told that i have to construct model on all the dataset and pool the results. What I do not understand is suppose i am developing ANN to forecast rainfall then I will have to develop the model on all of the dataset and pool the results. It will be a alot of work.. is there any way we can select only one dataset by means of statistical test or something.. please help Sir. $\endgroup$ Jul 18, 2019 at 11:21
  • 1
    $\begingroup$ You have to do what you say will be a lot of work. There are surely tools for making it less work (although that is off topic here) but using only one imputed data set is not right. $\endgroup$
    – Peter Flom
    Jul 18, 2019 at 11:29
  • 1
    $\begingroup$ Thank you Sir for your comment. $\endgroup$ Jul 18, 2019 at 12:15
  • 1
    $\begingroup$ You are welcome. Since you are new, I will let you know that, if an answer answers your question, the usual thing is to accept it using the little check mark. $\endgroup$
    – Peter Flom
    Jul 18, 2019 at 19:38

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.