Roland, you must also consider the influence of period (calendar year)
and cohort (birth year), as well as the influence of age. As is
well-known, knowledge of any two of these quantities determines the
third. Because older members of your population experienced the
depression and war years between 1929 and 1945, cohort effects appear
to be a real possibility.
The only book in my library about this topic is: William Mason and
Stephen Fienberg, Eds. "Cohort Analysis in Social Research: Beyond the
Identification Problem", Springer-Verlag, 1985.
As a start, use attained age as your time dimension in -stpm2- and
try the other two quantities as covariates, but only one at a time.
Fit a flexible function to the covariate with one of Patrick Royston's
commands -mvrs- (multivariate restricted splines) or -mfp- (
multivariate fractional polynomials).
-Steve
On Mon, Dec 14, 2009 at 6:42 PM, Lambert, Paul C. (Dr.)
<[email protected]> wrote:
> Roland,
>
> I agree with Steve, you should be using age as the time-scale.
>
> If you do choose to use -stpm2- with age as the time scale then you can see an example using a large data set (>400,000 individuals) in my recent UK Stata Users Group presentation (slides 27-32).
> http://www2.le.ac.uk/Members/pl4/slides/Stata-UK-2009-handout.pdf
>
> This example is described in more detail in our recent Stata Journal article including extensing to multiple time-scales.
> http://www.stata-journal.com/article.html?article=st0165
>
> If the relative effect of socioeconomics status is a function of age then this is just a time-dependent effect and can be implemented in -stpm2- using the -tvc()- and -dftvc()- options. Another way to think of a time-dependent effect is that there is an interaction between your covariate of interest (socioeconomic status) and your time-scale, i.e. age.
>
> Paul
>
>
>
>
> ________________________________________
> From: [email protected] [[email protected]] On Behalf Of [email protected] [[email protected]]
> Sent: Monday, December 14, 2009 9:08 PM
> To: [email protected]
> Subject: Re: st: Choice of analysis time in survival analysis where the hazard is associated with age
>
> Definitely do use age as the time dimension. It is the natural
> dimension. The start-time in 1/1/1991 is an artifact of your data
> collection and has no natural meaning. I would, at least to begin,
> also ignore events before 1/1/1991 and operate on the true cohort:
> those alive and well without appendicitis on that date.
>
> For analysis, I recommend -stpm2- by Patrick Royston and Paul
> Lambert ("findit stpm2") That command fits flexible spline models to
> the underlying hazard in the Cox model, and allows interactions of the
> spline functions with other factors. The original -stpm- command is
> described in: Patrick Royston Flexible parametric alternatives to the
> Cox model, and more, The Stata Journal (2001) 1, Number 1, pp. 1–28.
>
> Fitting a model to most of the life-span is a challenge. lRather than
> fit a single model, stratified or not, l to the entire life span (ages
> 1-64), I suggest that initially, you divide age into "natural"
> intervals and fit a separate model to each, with observations censored
> at the interval end (this is not an analysis with age strata). Then
> pause and decide if and how you would like to do a combined analysis.
>
> You do not mention the goals of your analyses, but if you have a
> sufficient number of events, consider setting aside one portion of
> your data for choosing your best models, and another portion for
> testing the hypotheses and fits of these models.
>
> -Steve
>
>
> On Mon, Dec 14, 2009 at 7:38 AM, roland andersson
> <[email protected]> wrote:
>> I have a large national dataset and plan to analyse socioeconomic
>> differences in risk of having appendicitis. Persons born after 1949 to
>> end of 1990 and alive at 1/1/1991 (n=4.500.000) are included in the
>> dataset. The dataset contains the date when a person was operated for
>> appendicitis between 1/11/1990 and 31/12/2003 and date of death.
>>
>> The hazard of having appendicitis is agedependent and is increasing
>> till age 13 and then decreasing. Previous collaborators have made
>> Coxregression with time from 1/11/1990 to the appendicitisdiagnosis or
>> censoring as analysistime, and have entered age at operation as
>> covariates to control for the age-dependent differencies in hazard.
>>
>> Knowing that the hazard is age-dependent I think this may give biased
>> results. I therefore wonder if it would be better to use age as
>> analysis time (with late entries) instead of time from start of follow
>> up with adjustment for age. Am I right to think that this will be a
>> better way of controling for the age-dependent differences in hazard?
>> If that is the case may I also include age as a covariate in the
>> analysis or would that complicate things?
>>
>> I have also considered to use a stratified analysis with ageintervals
>> as stratification variable. Can you give some advice on how these
>> ageintervals should be chosen in view of the agedependent increasing
>> and decreasing hazard?
>>
*
* 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/
