Fracreg: elaborating "conditional mean"

[Hang Vu]

Dear Statalist,

My dependent variable is continous from zero to one and contains a large amount of zero and show the non-normal distribution. Therefore I am going to apply fractional regression.
However, I do not get one point when reading the manual of Stata : Fractional regression is a model of the mean of the dependent variable y conditional on covariates x, which we denote by μx. Because y is in [0,1], we must ensure that μx is also in [0,1]
My independent variables are all categorial variables. How could I calculate and know if all μx is in [0,1]?
As I am very new with statastics, I was not able to find the definition of conditional mean but only conditional expectation. Would these two phrase be the same meaning?
Hope to get your explaination

Thank you so much,
Hang Vu

[Richard Williams]

fracreg basically does logit (or probit) without the constraint the DV must be 0 or 1 (it can range from 0 to 1). So, like estimated probabilities in logit and probit, estimated ux will range between 0 and 1.

In other words, so long as you use fracreg there is no problem with this. If you used regress, predicted values could be less than 0 or greater than 1.

Incidentally, given that have a lot of zeros, fracreg may not be the best option. For other options see

https://www3.nd.edu/~rwilliam/stats3...onseModels.pdf

[Hang Vu]

Richard Williams Dear Professor,
Thank you so much for your answer as well as your paper, I already read your paper several days ago and saw that the zoib might be the better option for many zero.
However, given my research questions are:
- How do household socio-demographic characteristics influence their probability to buy the product?
- If consumers buy this product, how do consumers’ household socio-demographic characteristics influence their share of expenditure for that product?
I think that for the first research quesition, I will use logistic regression and then fractional regression (take out all of zeros in the model) for the second question. I already applied two models and saw that the result of logistic and fracreg regression produced more number of coeffecient for independent variables with statistic significant than the result of the zoib. Therefore I go with fracreg
Would you mind giving your comment about my procedure?

Thank you so much and best regards,
Hang Vu

[Richard Williams]

It sounds like you want to run a heckprobit model except you want to use fractional probit instead of probit. I'm not sure if there is software or procedures out there for that or if it could otherwise easily be done. My guess is heckprobit could be tweaked. But maybe someone knows an easy way to do this (or has opinions on why you shouldn't).