I have been using streg to fit predictive survival models (for time from diagnosis to death) to some data on some fiendishly unpredictable patients, of which (according to their Kaplan-Meier curve) about 15 percent are dead after the first month, another 15 percent are dead after the first year, and the rest die at a much more relaxed pace afterwards over the following 17 years. To me, this seems to suggest choosing a Weibull model., which seems to predict 1-year survival a bit better than a Gompertz model in these patients. And my gut response is to choose the hazard rate parameterization, for comparability with other people's Cox regressions, although I am open to persuasion in favour of the accelerated failure time parameterization (which of course is an alternative parameterization of the same model with the same predicted k-year survival probabilities).
The main problem with a Weibull model seems to be how to interpret the intercept parameter to my colleagues. Assuming that we use an eform option, the intercept parameter of a logistic model is the baseline odds if all covariates are zero, the intercept parameter of a linear regression model for log-transformed data is the geometric mean when all covariates are zero, and the intercept parameter of a Gompertz model is the hazard rate at zero time if all covariates are zero. But, in a Weibull model with decreasing hazard rate (0<p<1 in the notation of [ST] streg where p is the power parameter), the hazard rate tends to infinity as time tends to zero.
One idea I can think of is to define the exponentiated intercept of a Weibull model as h(1)/p, where h(1) is the hazard rate at time 1 (in the units used which in my case are years). Another possibility is -log(S(1)), where S(1) is the 1-unit survival time (which would be a 1-year survival time in my case) and log() is the natural log. However, both of these quantities are expressed in counter-intuitive units. (I suppose I could end-point transform a confidence interval for the exponentiated intercept to get a confidence interval for the 1-year survival.)
Can anybody think of a better interpretation?
Best wishes
Roger
Related Posts with How do we explain intercept parameters for a Weibull survival model?
looping twoway graph on panel datasetDear Sirs, I have a panel of 21 countries over 45 years. I am trying to graph this: foreach i of n…
Combining dichotomous variablesHello, I have a large dataset which is already programmed into STATA of which I want to combine 8 va…
xtmixedDear stata users, I am trying to analyze the effect of a intervention where we have used a cluster r…
ARDL Error Correction Model InterpretationI am trying to estimate the relationship between real currency per capita (dependent variable) and d…
New version of invdesc on SSCThanks as always to KIt Baum, a new version of the invdesc package is now available for download fro…
Subscribe to:
Post Comments (Atom)
0 Response to How do we explain intercept parameters for a Weibull survival model?
Post a Comment