I'm hoping for some help with conceptual issues related to G-computation/standardisation following the use of stpm2 or streg for time-varying covariate data (one tvc - higheff2; the rest baseline covariates). I have tried following Paul Lambert's really good tutorials:
https://pclambert.net/software/standsurv/
but a few ideas still elude me with these models and what standsurv actually does.
Am I correct in thinking that the following model allows a 1 df rcs (i.e. forces a straight line) on the cumulative baseline hazard:
It also allows a 1df spline (again a linear effect) for the tvc. This allows non-proportional hazards on the tvc, right?
A LR test shows that I can then remove the dftvc term without affecting the fit (p=0.49). This is that model:
I note that then comparing this to a standard Weibull model gives the exact same results:
I then want to use standsurv to calculate a marginal hazard ratio (averaged over the individual-specific covariates) - we have been asked by a reviewer to include an additional causal inference approach to our multivariable Cox model.
I understand that a 'summary' HR is a weighted average of time-varying HR's (this paper was an interesting read):
https://jamanetwork.com/journals/jam...stract/2763185
It also seems that standsurv produces marginal estimates based on specified time points of interest. What I don't fully appreciate is why when one removes the dftvc term (or equivalently uses streg to produce the same results), the marginal HR's remain time-varying? I would have expected proportional hazards were assumed and the marginal HR estimated over time to be constant. But clearly this is not the case and clearly I don't understand something.
Thanks for any enlightening.
https://pclambert.net/software/standsurv/
but a few ideas still elude me with these models and what standsurv actually does.
Am I correct in thinking that the following model allows a 1 df rcs (i.e. forces a straight line) on the cumulative baseline hazard:
Code:
stpm2 higheff2 vaccdum2 smokdum2 smokdum3 ocpdum2 ocpdum3 seifadum2 seifadum3 seifadum4 seifadum5, df(1) scale(hazard) eform nolog tvc(higheff2) dftvc(1)
A LR test shows that I can then remove the dftvc term without affecting the fit (p=0.49). This is that model:
Code:
stpm2 higheff2 vaccdum2 smokdum2 smokdum3 ocpdum2 ocpdum3 seifadum2 seifadum3 seifadum4 seifadum5, df(1) scale(hazard) eform nolog
Code:
streg higheff2 vaccdum2 smokdum2 smokdum3 ocpdum2 ocpdum3 seifadum2 seifadum3 seifadum4 seifadum5, dist(weibull)
I understand that a 'summary' HR is a weighted average of time-varying HR's (this paper was an interesting read):
https://jamanetwork.com/journals/jam...stract/2763185
It also seems that standsurv produces marginal estimates based on specified time points of interest. What I don't fully appreciate is why when one removes the dftvc term (or equivalently uses streg to produce the same results), the marginal HR's remain time-varying? I would have expected proportional hazards were assumed and the marginal HR estimated over time to be constant. But clearly this is not the case and clearly I don't understand something.
Thanks for any enlightening.
