Different ymodel within a single call

[huichen99]

Dear Developer,

I am using the gformula function to estimate intervention effects and would like to include multiple intervention scenarios. Specifically, I am working with four arms: one without intervention and three with different variables as interventions. The intervened variables are highly correlated—one may be the sum of the other two—so I need to use different ymodels for each intervention scenario.

For example.

Intervention 1: threshold intervention of fruit >2 unit/week
Intervention 2: threshold intervention of vegetable >2 unit/week
Intervention 3: threshold intervention of fruit+vegetable >4 unit/week

Because a new variable frt_vege = fruit+vegetable would be highly collinear with fruit and vegetable, they could not be placed in a single model. Is the package designed to accommodate this type of setup?

I also attempted to run separate gformula() calls for each intervention arm, but observed significant variation in the natural (no intervention) g-form risk across runs. While I'm not entirely sure of the cause, I assume that modeling multiple intervention arms within a single call may improve comparability.

Any guidance would be greatly appreciated.

[stmcg]

Hi,

Thanks for posting the question.

That's right, the same ymodel would need to be used across the different interventions. One workaround is to run the gformula function separately for the different interventions, as you did.

That's interesting that the natural g-form risk differed a lot across the different runs. I'd imagine that the cause of this is a model specification issue. It could alternatively be due to small sample sizes, in which case constructing CIs around those estimates could be insightful

If you're looking to include the different interventions in a single call to gformula, one approach could be to use a custom history function which would allow you to define frt_vege and use it in the outcome model in Interventions 1 and 2. Another idea to consider is implementing Intervention 3 as a custom, joint intervention on variables fruit and vegetable.

Best,
Sean

[huichen99]

Hi Sean,

Thanks for the helpful suggestions! I managed to perform the joint intervention.

I am wondering if there are any non-linear specifications available for the g-formula model. For instance, the health effect of alcohol intake is typically non-linear (e.g., J-curve). If I would like to perform a threshold intervention on this component, how should I specify the non-linearity in the ymodel or elsewhere?

Many thanks,
Hui

[stmcg]

Hi Hui,

That's good to hear

I'm not sure if I'm correctly understanding your question about non-linear specifications -- If you're talking about having a more flexible function of alcohol intake than just a linear term (e.g., such as including splines), you can directly do so by specifying the predictors in ymodel accordingly. If you're looking for a more flexible model besides a generalized linear model, you can use a custom outcome model. That would be able to support any type of outcome model, such as neural networks, provided you can supply a fit and predict function.

Best,
Sean