How do I deal with data with 50% missing value for a specific dimension? I have a dataset with 50 thousand records. 50% percent of the records don't have gender and birth dates.
I wish to analyze age-group/gender item purchase preference. 
Values are missing completely at random.
Should I delete, impute, ignore, or just create a different model? 
Lastly, can I still draw a valid conclusion using a model with 50% of missing values removed?
 A: Can your conclusions be considered valid if you listwise-delete MCAR data? Certainly - especially if your coefficients prove significant and relationship direction is all you care about. On the other hand, if you have other variables on-hand (which may or may not be missing values) and those variables are likely to correlate with age and/or gender - which I bet is the case - then it begs the question of why you didn't do imputation. Or worse yet, one may wonder: did they do imputation but the results were less favorable for their argument and so they're hoping no one asks about it?
Because we know the data are MCAR, we would not "expect" the imputation to change the coefficients - but we also don't "expect" the estimated effects from any random sample to be higher or lower than the true effect, so that's expected values for you. The imputation may very well change the coefficient because it's correcting random sampling error in the available cases. And even if the coefficient doesn't change, then the imputed cases will increase your statistical power and tighten your confidence intervals - legitimately so, if you do the imputation correctly with a number of imputed datasets and pooled results.
