two part model for health expenditure

[Lina Massou]

Hello everybody,

I have health expenditure as DV and demographics as IV. I want to run two part model with logit plus OLS log-trasnformed y and Duan's non-parametric smearing factor. I have tried the command I write below but I don't know how to insert the the Duan's smearing factor. Could you help me please.

 twopm Healthexpend i.gender i.AGEGROUP i.region i.Household_size i.marital_size  [pweight = WGHT], firstpart(logit) secondpart(regress, log)

[Carlo Lazzaro]

Lina:
you may want to replace OLS with GLM, as suggested in Andrew Briggs, Mark Sculpher, and Karl Claxton. Decision Modelling for Health Economic Evaluation. Oxford: OUP, 2006: 105-107.

[Lina Massou]

Thank you Mr. Lazzaro,
I will compare the results of 3 different models, one of them is this with GLM, but the supervisors of my research proposed also this model that I was asking for and that has been used in a relevant paper. Any idea for the Duan's smearing factor?

[Carlo Lazzaro]

Lina:
you may be interested in http://stats.stackexchange.com/quest...two-part-model

[Lina Massou]

Many thanks, I have seen it as well as the suggested book of Cameron and Trivedi's Microeconometrics Using Stata, . I think that I have understood the theory but I don't know how to write the commands in STATA.

[Joao Santos Silva]

Following up on Carlo's comment, I suggest you use the modified two-part model suggested by John Mullahy in 1998; see here.

[Carlo Lazzaro]

Lina:
as in http://stats.stackexchange.com/quest...two-part-model, you can calculate the -predicted- after -regress- and then apply the Duan's smearing factor.
I would also consider Joao's good advice about the approach followed in John Mullahy' famous paper.

[Stephen Jenkins]

Have you read the substantial article on two-part models in the Stata Journal, 2015(1)? It has accompanying software. (Lina: is this where you got twopm from? Remember, as per the Statalist FAQ, you should tell us where user-written commands come from. search twopm) According to the post-estimation help, twopm can implement Duan's smearing estimator:

Retransformation after ordinary least-squares regression of ln(depvar) is
    needed to obtain consistent predictions of depvar.  twopm implements this
    using normal theory and smearing retransformations, but both assume that
    the errors in the regression are homoskedastic.  Retransformation with
    heteroskedastic errors is conceptually complex, and we have not
    implemented it in twopm.  We suggest the gamma generalized linear model
    (glm) with log link as an alternative to a regression of ln(depvar).

The last sentence is consistent with Carlo's and Joao's suggestions. Alternatively, what is it about your use of the twopm post-estimation commands that is problematic for doing the "smearing"?

[Lina Massou]

Thank you all of you for your help. Yes, that's where I got twopm.If I have understood correctly, when we use regress for the log-transformed DV - for the second part of two-part model- we need to restransform, and in case of homoskedastic errors it is done automatically through the STATA calculations, but in case of heteroskedasticity it should be done by the user. Right?
I don't have any information about the errors. So I use the twopm as I wrote it above and that's all or I have to do something else by hand?
I will compare this model with others, so the case of GLM will be included in the rest models.