Missing data, multiple imputation alternatives?

[William Brewer]

I would greatly appreciate if you could give me a really quick bit of advice around missing values:

A bit of context - I have a panel dataset with 19 variables and approx 250k observations

My dataset has less than 10% missing values, however this leads to stata regressing using only 20% of the overall observations. I have written code for multiple imputation which seems like the best option, however our university computers simply aren’t powerful enough to compute this and would take weeks. Having looked through all alternatives, substituting the mean (and perhaps using a dummy variable indicating missing data) seems the most best solution but has been criticised for artificially decreasing the standard errors, therefore leading to invalid inference.

Are there any alternatives (or simplifications) to multiple imputation which will give unbiased estimates for the missing data without adversely affecting the variance?

Many thanks,

WB

[Marcos Almeida]

I gather the main concern here is identifying the pattern of missing data (MCAR, MAR, MNAR).

That said, I don't know whether there is an "alternative strategy" which provides "unbiased estimates" for all patterns of missing data (MI is not capable of that as well), but maybe using previous or following nonmissing values can be helpful.

There is a FAQ on this, here.

[William Brewer]

That is a bit of a tricky one, most of the missing data is MCAR however there are one or two variables which aren't, such as a dummy variable which is only present every other time period.

It may solve some of the problem copying over previous observations, but would not be appropriate a lot of the time (e.g. copying characteristics to other persons).


My model is:

ln(hourly wages) = b0 + b1exp +b2exp^2 + b3occupation + ... + u

[Weiwen Ng]

I think that full information maximum likelihood estimation, which is available in our linear sem command:

sem wages <- x1 x2 x3 ... , method(mlmv)

is another option for missing covariates, but it is also very processor intensive. That and MI are the only two options I know of to get (asymptotically) unbiased estimates from data missing at random.

Another problem is that efficient multiple imputation for panel/clustered data is a theoretical problem that doesn't appear to have been resolved very well. If you have relatively few clusters, or you're willing to assume multivariate normality among the missing variables, then the options described here may work. I somehow don't think your use case falls into any of the 3 approaches described.