Currently I am working on a large data set with well over 200 variables (238 to be exact) and 290 observations for each variable (in theory). This data set is missing quite a lot of values, with variables ranging from 0-100% 'missingness'. I will eventually be performing logistical regression on this data, so of my 238 columns I will at most only be using ten or so.

However as almost all of my columns are missing some data, I am turning to multiple imputation to fill in the blanks (using the MICE package).

My question is; given that I have a large amount of variation in the missing data, at what percentage missing should I start to exclude variables from the mice() function?

Can mice function well with variables that are missing 50% of their values? What about 60%, 70%, 80%, 90%?


4 Answers 4


In principle, MICE should be able to handle large amounts of missing data. Variables with lots of missing data points would be expected to end up with larger error terms than those with fewer missing data points, so your ability to detect significant relations to those variables would be limited accordingly. That's an advantage of having multiple imputations and analyzing results from all of the imputations.

The greater the fraction of (a) cases with missing data or (b) predictors that are frequently missing, the more imputed data sets you will need. See Section 3.10 of Frank Harrell's Regression Modeling Strategies for some "rough guidelines" about how to proceed. If the missing values are for variables that you consider important based on your understanding of the subject matter, he says: "Extreme amount of missing data does not prevent one from using multiple imputation, because alternatives are worse."

More important than a "cutoff" for missing data is to consider carefully (1) the intended use of your model and (2) whether the "missing-at-random" assumptions needed for multiple imputation holds in your case.

In terms of (1) if you, say, intend to use the model for prediction but some variables are inherently hard to get, then there's no sense including them in the model. Also, you should use your knowledge of the subject matter to consider variables for inclusion. If you suspect that only 10 or so will be important based on such knowledge, maybe you should just use those 10.

In terms of (2), "missing at random" means that the probability of missingness doesn't depend on unobserved data. As Stef van Buuren says in Section 1.2 of Flexible Imputation of Missing Data (FIMD):

If the probability of being missing is the same only within groups defined by the observed data, then the data are missing at random.

That's a typical starting point for analysis. van Buuren discusses ways to evaluate the sensitivity of the results to violations of the assumption in Section 9.2 of FIMD.


Mice can handle a large amount of missing data. Especially if there are a lot of columns with few missing data, one with 80% is no problem. You can also expect that in most of the times adding this variable leads to better imputation results than leaving it out. ( because more information / correlations available that help estimating the other variables)

But: The hard truth is, you will never know for sure, how good the imputation is anyway. Because the true values are well ... "missing"

If I'd have several imputation options to choose from, I'd take the one which leads to the best results for the prediction model.


This is not a coding question but if you want an answer here it is...

Missing data are very complicated. There is not a percentage value to accept of discard your variables. The variance of your variable is what is important to watch before imputation of data.

If you do not want to take some time to review all the statistic behind missing values, just take your variables with less missing value.

If you take the time read the MICE manual there is some basic information that will help you to impute correctly.

Missing data are not simple task, you have know what you do. Otherwise you will introduce bias!

  • $\begingroup$ I'm not sure that this really answers the question. $\endgroup$
    – Hong Ooi
    Commented Apr 30, 2015 at 16:05
  • 1
    $\begingroup$ I think the advice to consider the variance is good. $\endgroup$ Commented Nov 18, 2017 at 10:18
  • $\begingroup$ I would like to see an extensive elaboration of this point "The variance of your variable is what is important to watch before imputation of data." $\endgroup$
    – TrungDung
    Commented Nov 18, 2022 at 8:57

(Unable to comment yet - sorry! I would have like to have commented on Joel's response.)

I want to point out that, I believe, the quality of the imputation algorithm has bearing on the amount of data that may be validly imputed.

If the imputation method is poor (i.e., it predicts missing values in a biased manner), then it doesn't matter if only 5% or 10% of your data are missing - it will still yield biased results (though, perhaps tolerably so). The more missing data you have, the more you are relying on your imputation algorithm to be valid. E.g., if you are imputing 80% of your data, I believe you would want to be very confident that you are imputing it well; otherwise, you could introduce considerable bias.


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.