st: Thread-Index: Ac4AoFLb+vy4/foNSUKBOunsHojEbg==

[Clyde B Schechter <clyde.schechter@einstein.yu.edu>]

Francisco Lopez wrote:

...
Ive tried:
poisson deaths year sex age1 age2 age3 age4, exposure(population)
predict p_deaths
collapse (sum) deaths p_deaths population, by(year)
gen rate=deaths/population*100000
gen p_rate=p_deaths/population*100000
twoway (scatter rate year) (line p_rate year)

This way i get, what i think, are age and sex adjusted mortality
rates. However when I compare this rates with those obtained using a
direct or indirect standarization, are quiet different.

- - Is this model right?
- - Is suitable to use a mixed effect model (xtmepoisson) due to
multiple levels of my data i.e. xtmepoisson deaths year sex age1 age2
age3 age4, exposure(population) || sex: || age: ?


There is no reason to expect that the age-sex adjusted mortality rates obtained by a regression  method will match or even come close to those obtained by direct or indirect standardization.  The latter methods are ways of conforming your own data to the age-sex distribution of an external population (direct) or the age-sex-specific mortality rates (indirect) of an external population.  Regression methods do not involve any external population.  They are an internal approach that attempts to separate out the contributions of age and sex to mortality and model the rest of the variation as noise.  Your predicted value for each observation then is the expected number of bladder cancer deaths for a person with the corresponding age-sex variables conditional on your model being "correct."

Is your model "right?"  Models are almost never "right."  But they may be useful for particular purposes.  You don't say what you are trying to use the model for, so it's hard to know.  We also don't know what your data represent, how they were gathered, etc.

I will say this, it would be a highly unusual situation for group-aggregated mortality data to be such that adding random intercepts at the age and sex level to make sense.

One thing that might, generically, be helpful is to use a linear or cubic spline for age instead of five discrete categories.  It is seldom realistic that mortality rates jump discontinuously at the boundaries of age groups.  But even that depends on what you plan to use the model for.

Clyde Schechter
Department of Family & Social Medicine
Albert Einstein College of Medicine
Bronx, NY, USA








*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/