mi impute chained: dealing with binary dependent variable?

[danishussalam]

Hello everyone,

I'm using the following commands to impute missing values.

mi set mlong
mi register imputed treatment $continous
mi impute chain (logit) treatment (regress) $continous , add(10) rseed (54321)
mi estimate: regress treatment $continous

The problem is with the last command since treatment is a binary variable with no missing values, is it okay to use regress with it? The command doesn't work when I replace regress with logit.

Also how can i be sure that the missing values are replaced by mi impute chain rather than being stored in stata memory?

* $continous is simply a list of variables which are continuous in nature. They are basically the variables that I will be using (after imputation) for matching via PSM.

[daniel klein]

Many things not clear here.

If treatment has no missing values, why do you register it as imputed? If it has missing values you should consider deleting the cases with missing values from the estimation sample after imputation (cf. von Hippel 2007).

Please review the FAQ on statements like

The command doesn't work when I replace regress with logit.

What does this mean? Do you get an error message? Which? Does the model give you unexpected results? Which? Please show what you typed and what Stata did in response.

In general using a linear probability model would require correction for the estimated covariance matrix because the model implies heteroscedasticity.

Best
Daniel


von Hippel, Paul T. (2007).Regression with Missing Ys: An Improved Strategy for Analyzing Multiply Imputed Data. Sociological Methodology, 37, pp. 83-117.

[danishussalam]

Hey Dainel,

Thanks for replying. I was able to answer most of the questions raised above. I however, am still confused on how to replace the missing values with the imputed values? I believe the mi impute chain doesn't really do that. It simply impute values which can later be used with mi estimate and the preferred regression model. I on the other hand want to use the variables with imputed observations form Propensity Score Matching.

Any help on that will be appreciated.

[daniel klein]

mi impute creates new observations (or variables in mi set wide format) in your dataset where the imputed values are stored.

Best
Daniel

[danishussalam]

Thanks Dainel. You're right. I was not able to see them only because I was using mi set long.

[Weiwen Ng]

Hello everyone,

I'm using the following commands to impute missing values.

mi set mlong
mi register imputed treatment $continous
mi impute chain (logit) treatment (regress) $continous , add(10) rseed (54321)
mi estimate: regress treatment $continous

The problem is with the last command since treatment is a binary variable with no missing values, is it okay to use regress with it? The command doesn't work when I replace regress with logit.

Also how can i be sure that the missing values are replaced by mi impute chain rather than being stored in stata memory?

* $continous is simply a list of variables which are continuous in nature. They are basically the variables that I will be using (after imputation) for matching via PSM.

Just a quick note. I believe that if you want to include treatment in the MI equation, but there are no missing values, this should be the "proper" syntax:

mi set mlong
mi register imputed $continous
mi register regular treatment
mi impute chain treatment (regress) $continous = i.treatment , add(10) rseed (54321)
mi estimate: regress treatment $continous

You can add any other variables you think predict missingness on the right hand side of the imputation specification.