lr test likely invalid for models with robust vce

[Martin Imelda Borg]

Why does stata give me this error following:

stset survt [pw=weight], fail(status==1) scale(12) exit(time 120)

// estimate the time-varying hazard ratio using flexible parametric models
stcox var1 var2 var3, tvc(var3) texp(_t) 
estimates store nonph1

// estimate the time-varying hazard ratio using flexible parametric models
stpm2 var1 var2 var3, scale(h) df(5) eform tvc(var3) dftvc(3)

estimates store stpm2_nonph
lrtest nonph1 stpm2

Question 2 if I may
Also, all tests above since including [pw=weight] due to IPTW take at least 3hr to load as I have 600,000 observations, 16,000 failures.
Is this normal? there are 74000 iterations if it makes sense to anyone

[Luca Calianno]

Hi,

Stata gives that error warning message because when clustering both likelihoods need iid observations. I would suggest using a wald test. (More on this here https://www.stata.com/support/faqs/s...od-ratio-test/)

Regarding question 2, you did not provide much info. I don't think is normal but I believe it should be okay as long as there are not many backed up iterations. Try simplifying the model and see if something improves.

[Martin Imelda Borg]

Thanks Luca

the thing is , as mentioned I have a large dataset with weights
with 600,000 observations

A stcox model with 2 Tvc variables

and

stpm2 with 2 tvc variables

take a total of 3hr to run

i tried

stcox var1 var2 var3, tvc(var1 var2) texp(ln(_t)) hr
Estinates store m1

stpm2 var1 var2 var3,scale(n) df(5) eform tvc(var1 var2) dtvc(3) 
estimates store m2

test m2 m1
lrtest m2 m1

the lrtest gives me the result above in post1

the wald test gives me - stata error:

m2 not found
r(111)

Does this have to do with memory capacity perhaps ?

[Luca Calianno]

That is not the way to use test with different estimating equations. I intended something like this https://www.stata.com/support/faqs/s...-coefficients/ but this does not work in your case.

Currently, I don't know if there is an easily implementable way to test your hypothesis on Stata and any suggestion/correction from other members could be more meaningful.
If this has not to be a formal test but just a check you could:

1. use lrtest without vce and pweights (not a formal test but just a way to see which test works better with stronger assumptions)
2. run models on a sample and predict results on an other sample and see which model predicts better
3. compare the models testing "by hand" that your coefficients are not statistically different (using coefplot to show it visually may help).