Two time-point measure, subgroups

[Maria Kirsipuu]

Hi Statalist, I hope for some ideas - because I am stuck.

I have dyadic (child & caregiver) data with two time points (pre- and post-intervention). The dependent variable is zero-bound continuous (health service use) data.
Co-variates are clinical and demographical variables (both categorical and continuous, both constant and time-varying).
I am aiming to look at treatment effects in groups of caregivers (male vs female) in children (if there is a difference) according to treatment effect.

Which analysis makes the most sense here, when I am interested in the intervention effect in a subgroup of caregivers? I have been thinking about reg with interaction term .reg depvar1 i.treatment##i.cgsex##timepoint i.covariable c. covariable, vce(cluster id)

Does this make sense or am I on the very wrong track here?

Kind regards, Maria

[Clyde Schechter]

You are basically on the right track. Your representation of the interaction looks right to me. The space between c. and covariable is a syntax error: remove it.

What looks problematic to me, however, is the use of -regress- here. Health service utilization, in addition to being bounded below at 0 usually has an extremely skew distribution. Ordinary linear regression is likely to fit the data poorly. Consider using Poisson regression instead (and be sure to use cluster-robust standard errors, as you show in #1.)

[Maria Kirsipuu]

Clyde, thank you very much for your comprehensive and helpful answer. I appreciate it.

Bets regards, Maria

[Maria Kirsipuu]

I am sorry, I have one more question.

Is it suitable to use Poisson regression if the dependent variable is not counted data (not times service has been used), but how many hours per month (many have zero hours).
I will also add a picture of the histogram of the dependent variable, whereas zeroes are true zeroes.

Best regards, Maria

[Clyde Schechter]

Yes. The only absolute restriction on the use of a Poisson model is that the outcome variable must be non-negative. It does not have to be count data.