Truncreg for censored data

[Gene Korku]

I have data on bi-weekly expenditures on drugs, I am interested in looking at the effect of some variables on these bi-weekly expenditures. As is the case with survey data with such short intervals a lot of respondents do not report spending on drugs. Is a Tobit regression suitable for such a case?. Also, if I am interested in estimating the model for only households that report positive expenditures on drugs is Truncreg the appropriate command to use compared to OLS on positive expenditures?

[Andrew Musau]

Sounds like a case of missing data and not censoring or truncation. See the definitions and difference here: https://blog.stata.com/2016/12/13/un...-and-censoring. Missingness in itself is not a problem if it is random.

Also, if I am interested in estimating the model for only households that report positive expenditures on drugs is Truncreg the appropriate command to use compared to OLS on positive expenditures?

We may say here that you are intentionally truncating the data and then asking about an estimation method for truncated data. Why if you already have all the data? If you want to focus on just the sample with positive expenditures, then that is your sample and no adjustment is required.

[Gene Korku]

Whether the data is missing is debatable. This is because these surveys typically collect information in expenditure diaries over two weeks. Hence households who do not spend anything over that time frame do not report any values so technically it is not "missing" data but rather zero expenditure, so I guess it's more or less a corner solution problem than censoring if I'm right. If this is the case of a corner solution problem, is it fine to estimate the model on only households with expenditures > 0 via OLS?

[Andrew Musau]

OK, that description is different from #1. If the story is something like poor households often report zero expenditures whereas for wealthier households, there is a large variance in expenditure, then you have censoring. If you can model the censoring process, look at heckman. Otherwise see tobit.

help heckman
help tobit

Ignoring the zero expenditures and using OLS on the continuous sample will lead to biased and inconsistent estimates.

[Gene Korku]

Thanks for the feedback. However, this still leads back to my first question assuming I want to run a regression on the sample that have expenditues>0, is truncreg the appropriate command to use since I am basically truncating my sample by ignoring the zero expenditures or there is some different way to model this? My apologies if this isn't so clear.

[Andrew Musau]

I already answered your question. You do not have truncated data, so no you should not use truncreg. You already know the distribution of the zero observations, so look at estimators for censored data as suggested in #4.