--
Chelsea-
I would not expect the two kinds of estimates to align.
The ratio of Deaths1/PY1 to Deaths2/PY2 is guaranteed to estimate
the hazard ratio only in special circumstances. One is that the data
are generated bya two-group exponential model with no time-dependent
covariates and constant hazards.
There could be many reasons for the dissimilar estimates. Kieran has
described an important one. A related issue with time dependent
covariates is that the risk sets and exposure for different levels of
time-dependent covariates may not cover the same time-periods, whereas
the Cox model automatically compares people at risk at the same
follow-up times. Confounding could also contribute to the
discrepancy, even in the ideal exponential case, because you are
computing the PY-based hazard ratios one factor at a time.
You can explore distributional shapes for time-dependent groups if you
designate the groups as strata. There are also some good ideas for
plotting in Applied Longitudinal Data Analysis: Modeling Change and
Event Occurrence by Judith D. Singer and John B. Willett
-Steve
On Wed, May 6, 2009 at 7:35 PM, Polis, Chelsea B. <[email protected]> wrote:
> I apologize for the formatting of my table, it looked ok when I sent it. Please let me try again.
>
>
> Variable Deaths PY at risk IR HR 95% CI p-value
>
> HC use 0.07
> No 91 1262.7 7.21 1.00
> Yes 13 293.0 4.44 0.58 0.32-1.04
> Current age 0.38
> 15-24 20 394.0 5.08 1.00
> 25-34 49 711.8 6.88 0.73 0.43-1.24
> 35+ 35 449.9 7.78 0.68 0.38-1.20
> Sex partners in past year 0.01
> None 18 241.2 7.46 1.00
> One 76 1204.6 6.31 1.31 0.78-2.21
> Two+ 10 109.9 9.10 3.40 1.54-7.54
>
>
> Dear statalisters,
>
> I am doing survival analysis on time to death with time-varying covariates on an open, population
> -based cohort study. The base sample is essentially a census of individuals in 56 villages, and
> I am utilizing information from all female incident HIV seroconverters.
>
> I computed incidence rates based on the raw data ((number of deaths/person-time at risk)*100 - I obtained
> time at risk using the stdes command), but the IRs don't seem to echo trends in the univariate Cox HRs.
>
> In the sample data below, things appear reasonable for HC use (deaths per 100 person years is lower
> if HC=yes, and the HR reflects this). But for current age, deaths are higher in the 25-34 category
> than in the 15-24 category, but the HR trends suggests that being 25-34 is protective (though not
> significantly). Also, the magnitude seems off, for example, in the variable "Sex partners in past
> year" - having two or more seems to more than triple the hazard in the Cox regression, but merely
> increases from 7.46 to 9.10 in the deaths per 100 p-y.
>
> Am I missing something in expecting these numbers to echo trends in each other? Is this just a
> matter of non-significance within individual categories? Or a difference in time-to-event versus
> person-time analysis? Or because I am doing an analysis with time-varying covariates? Should I not
> expect these to align? Any help is appreciated!
>
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
