Treating 'Don't know/Refused' levels of categorical variables I am modeling Diabetes Prediction using Logistic Regression. The dataset used is the  Behavioral Risk Factor Surveillance System (BRFSS) of the Center for Disease Control (CDC). One of the independent variables is High Blood Pressure. It is categorical with the following levels 'Yes', 'No', 'Don't know/Refused'. Should I remove those rows with 'Don't know/Refused' while building the model? What difference does it make to keep or remove those rows from the model?
 A: I was just wondering about exactly the same question when analyzing the latest National Hospital Discharge Survey data.  Several variables have substantial missing values, such as marital status and type of procedure.  This issue came to my attention because these categories showed up with strong (and significant) effects in most logistic regression analyses I was running.
One is inclined to wonder why a missing code is given.  In the case of marital status, for instance, it is plausible that failure to provide this information could be linked to important factors such as socioeconomic status or type of disease.  In your case of high blood pressure, we should ask why would the value not be known or refused?  This could be related to practices at the institution (perhaps reflecting lax procedures) or even to the individuals (such as religious beliefs).  Those characteristics in turn could be associated with diabetes. Therefore, it seems prudent to continue as you have, rather than to code these values as missing (thereby excluding them from analysis altogether) or attempting to impute the values (which effectively masks the information they provide and could bias the results).  It's really not any more difficult to do: you merely have to make sure this variable is treated as categorical and you'll get one more coefficient in the regression output. Furthermore, I suspect the BRFSS datasets are large enough that you don't have to worry about power.
A: First you have to think over if the missing data are missing completely at random (MCAR), missing at random (MAR) or missing not at random (MNAR) as deletion (in other words complete-case analysis) may lead to biased results. Alternatives are inverse probability weighting, multiple imputation, the full-likelihood method and doubly-robust methods. Multiple imputation with chained equations (MICE) if often the easiest way to go.
A: Do you have any reason to think that study subjects with diabetes were more likely or less likely to end up with the DK/R response? If not (and I'd be pretty surprised to find out you did), including this predictor in the model w/o excluding these cases will result in noise. That is, you'll end up with less precision in your assessment of how "yes" vs. "no"  influences the estimated probability of diabetes (because you'll be trying to model the influence of either "yes" or "no" vs. random DK/R responses as opposed to just "yes" vs. "no"). The most straightforward option is to exclude the cases with DK/R responses. Assuming that their "yes/no" responses were indeed missing at random, excluding them will not bias your estimate of the influence of "yes" vs. "no." That approach, however, will reduce your sample size and thus reduce statistical power with regard to the remaining predictors. If you have a lot of DK/R on this variable, you might want to impute "yes"/"no" responses by multiple imputation (arguably the most, maybe only, defensible missing-value imputation strategy).
