Tobit Model with FIML

[Qing Zhang]

Hi Stata experts,

I'm trying to handle my missing values using FIML, and the code I use is sem(y <- x), method(mlmv) nolog.

I also want to use tobit regression because one of my dependent variables is right-censored. However, I can't find the code to combine tobit regression with FIML. Is there such a code? Or are there other ways to handle missing values and adjust right-censoring of the data simultaneously?

I guess I can also use MI with tobit, but here in this thread it says tobit models might violate the assumption of MI, which is "the statistic that you are estimating must have an asymptotic normal distribution."

Any thoughts? Thanks in advance!

[Phil Bromiley]

You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output (fixed spacing fonts help), and sample data using dataex. Also, screen shoots and attached files are not encouraged.

I think you'll find that many of the fancier models require GSEM rather than SEM, and some of the nice things in SEM don't work in GSEM. If you look at the tobit examples in the sem/gsem documentation, I suspect they use gsem. Depending on what you're doing, you might look at the eregress and related estimators or cmp (user-written). It is also possible to use multiple imputation for the missing instead of maximum likelihood.

[Weiwen Ng]

Hi Stata experts,

I'm trying to handle my missing values using FIML, and the code I use is sem(y <- x), method(mlmv) nolog.

I also want to use tobit regression because one of my dependent variables is right-censored. However, I can't find the code to combine tobit regression with FIML. Is there such a code? Or are there other ways to handle missing values and adjust right-censoring of the data simultaneously?

I guess I can also use MI with tobit, but here in this thread it says tobit models might violate the assumption of MI, which is "the statistic that you are estimating must have an asymptotic normal distribution."

Any thoughts? Thanks in advance!

To clarify the thread you linked to: Tobit regression was excluded from Stata's supported MI commands. This could mean that they know that the Tobit parameters don't meet Rubin's rules. It could also mean that they have no idea if the Tobit parameters meet Rubin's rules. I have absolutely no clue which is correct.

To comment on Phil's suggestion, gsem will allow you to specify a Tobit regression, but it does not support maximum likelihood with missing values (aka full-information max likelihood). The sem command will support the latter, but it is restricted to linear models with uncensored data, as you know.

I don't know that there is an easy solution to the problem. It may be worth simulating missing values with one of the stock datasets, conducting multiple imputation, and seeing if you can recover the parameters through MI + Tobit regression.