2
$\begingroup$

When I construct panel data from large household surveys (say, World Bank's Living Standard Measurement Surveys), I try to construct data set with as many potentially usable variables as possible. This creates a large unbalanced dataset and I don't know how many balanced data left until I run a regression.

I personally have not encountered any problems with this, but I wonder what other people think when they have to use dataset I construct. I have not worked with others so far, but I wonder how I should construct a good dataset to work on if I need to work with professors. Should I construct balanced dataset? Or unbalanced dataset with many variables with many missing values?

$\endgroup$
  • $\begingroup$ Why don't you create both? $\endgroup$ – london Sep 22 '17 at 20:31
0
$\begingroup$

The key question to ask here is this: is the data missing a random selection of the population/sample? If this is the case (a situation called Missing Completely At Random, MCAR, you can simply discard the observations with missing values and run your model on the balanced panel. Conversely, if that is not the case (Missing Not At Random, MNAR), an estimation from the balanced panel will be biased. There is also an intermediate case, Missing At Random (MAR), which depends on distinguishing between missing based on observables (in which case you can do simple imputation), or missing on unobservables, in which case you need to perform a two-step Heckman estimation. Read more about missing data here, and in this very approachable book.

There are plenty of tests that you can do to evaluate the nature of missing data. See here. These slides show a hands-on example of how to deal with unbalanced panel datasets.

A very useful introduction to Unbalanced Panal Data is here, based on the textbook on Panel Data by B. Baltagi.

$\endgroup$
  • $\begingroup$ Thanks for the detailed answers! Yes, I check whether missing values are attributed to some unobserved characteristics of respondents or not, once I construct panel data. But maybe it is better to keep attention on missing values while I am merging dataset. $\endgroup$ – Satoko Sep 27 '17 at 6:19
0
$\begingroup$

I would construct and maintain only the full (unbalanced) data set with many variables with many missing values. Why? We can always extract balanced panel data sets from the full one; we can't go the other way around. When a balanced set is needed (in order to reduce the file size or send it to others), I would use Stata or R and never do it manually. To err is human.

$\endgroup$

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.