1
$\begingroup$

I have missing values for some of the variables in my data. I am using pooled OLS and have 144 observations. I have missing values for three of the variables. Less than 10% of the data for each variable is missing.

What should I do with these missing values? Should I use multiple imputation or should I just omit the observations with missing values? Can I bias thw results if I use multiple imputation?


Thanks for your answer.

I am using country data for the years 2010-2012, and the missing data is on trade openness and GDP per capita. Data on GDP per capita is missing for Syria and Cuba. Data on trade openness measured as (export+import)/GDP is also missing for Syria along with other countries, like Myanmar, Iran, Libya and for one year for Hungary, Cambodia, and Malawi. I can see that the missingness of data for Syria is not random, because of the war etc.The same holds for Libya and Iran. However, for trade openness there is also missing values for Hungary and Cambodia which I think is random (cant see other reasons why is should not be random).
These variables are not of interest for my regression, they are used as control variables.

$\endgroup$
  • 2
    $\begingroup$ The answers depend on why the values are missing and on the pattern of missingness. (Are all three variables entirely missing from the dataset? Are they missing only from some cases? If they are, are all three always missing in those cases or are only some of them missing?) Answers also depend on the values of non-missing variables in the cases with missing values. They also depend on your method of multiple imputation. Finally, whether any of this matters depends on what you will be using the OLS results for. Please edit your question to clarify these points. $\endgroup$ – whuber Apr 18 '14 at 15:27
  • $\begingroup$ Daja, sometimes in a regression certain cases can have unusually high influence on the results. If such a case has missing values, then the entire analysis could be sensitive to how those values are imputed (or whether they are imputed at all). That is one of several ways in which the recommended action would depend on values of the non-missing variables. $\endgroup$ – whuber Apr 18 '14 at 17:41
2
$\begingroup$

Honaker, James, and Gary King. "What to do about missing values in time‐series cross‐section data." American Journal of Political Science 54.2 (2010): 561-581, (ungated version) should answer all of your questions. It is a paper written exactly for this kind of data.

Honaker and King strongly advocate multiple imputation in general, and the article then gives tailored advice for improving MI for TSCS data (with common shocks, trends, and the possibility to incorporate prior information, e.g. for missingness due to war). For more info and software, see King's research page on missing values.

$\endgroup$

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.