Hi Statalisters,
I have encountered an issue when comparing conditional Poisson and stratified stcox regression results, and could use some help. I am working with matched cohort data from a national register. (The actual data cannot be shared, so I have pasted an example below.) The exposed group ("exposed") are individuals who have been injured; the comparison group (exposed==0) are birthdate- and sex-matched non-injured individuals who have been sampled from the underlying population on the date of their matched exposed-individual's injury ("casedate"). Several comparison individuals are matched to each exposed person; the matched groups are identified in "group". The outcome is death. All individuals are followed until 1 Jan 2019, unless they die earlier; end-of-follow-up date is recorded in t1.
I first tried to use a bivariate Cox PH regression using the following commands:
This yields a HR of 3.71 (95% CI: 3.63, 3.80).
I was also considering using conditional Poisson regression, as I had read that this model was functionally equivalent to a stratified Cox regression but sometimes easier to implement in complex cases. I used the following code (where exposure_t is total time at risk, calculated as [t1 - casedate]) to estimate the bivariate IRR using conditional Poisson regression:
This yields an IRR of 4.14 (95% CI: 4.03, 4:26). These estimates are clearly not the same; even their confidence intervals don't overlap. I am not sure why they are so different, and I am not sure which result to "trust".
Can anyone point me to how I might be misspecifying one of these models?
Many thanks,
I have encountered an issue when comparing conditional Poisson and stratified stcox regression results, and could use some help. I am working with matched cohort data from a national register. (The actual data cannot be shared, so I have pasted an example below.) The exposed group ("exposed") are individuals who have been injured; the comparison group (exposed==0) are birthdate- and sex-matched non-injured individuals who have been sampled from the underlying population on the date of their matched exposed-individual's injury ("casedate"). Several comparison individuals are matched to each exposed person; the matched groups are identified in "group". The outcome is death. All individuals are followed until 1 Jan 2019, unless they die earlier; end-of-follow-up date is recorded in t1.
I first tried to use a bivariate Cox PH regression using the following commands:
Code:
stset t1, fail(outcome) scale(365.25) origin(casedate) stcox exposed, strata(group) vce(robust)
I was also considering using conditional Poisson regression, as I had read that this model was functionally equivalent to a stratified Cox regression but sometimes easier to implement in complex cases. I used the following code (where exposure_t is total time at risk, calculated as [t1 - casedate]) to estimate the bivariate IRR using conditional Poisson regression:
Code:
xtpoisson outcome exposed, fe irr i(group) vce(robust) exposure(exposure_t)
Can anyone point me to how I might be misspecifying one of these models?
Many thanks,
Code:
* Example generated by -dataex-. For more info, type help dataex clear input str5 id float(group exposed exposure_t outcome casedate t1) "A001" 1 0 2181 0 19369 21550 "A002" 1 0 2181 0 19369 21550 "A003" 1 0 2181 0 19369 21550 "C001" 1 1 1553 1 19369 20922 "B001" 2 0 682 0 20868 21550 "B002" 2 0 568 1 20868 21436 "B003" 2 0 682 0 20868 21550 "C002" 2 1 682 0 20868 21550 "D001" 3 0 1555 0 19995 21550 "D002" 3 0 1555 0 19995 21550 "D003" 3 0 1555 0 19995 21550 "C003" 3 1 462 1 19995 20457 "E001" 4 0 2580 0 18970 21550 "E002" 4 0 2580 0 18970 21550 "E003" 4 0 1974 1 18970 20944 "C004" 4 1 1559 1 18970 20529 "F001" 5 0 1327 0 20223 21550 "F002" 5 0 1327 0 20223 21550 "F003" 5 0 1327 0 20223 21550 "C005" 5 1 1327 0 20223 21550 end format %td casedate format %td t1
