survival analysis: cox v. parametric methods

[Caroline Wilson]

Hi all,

I am conducting a survival analysis of employee attrition (time to loss) using Stata and am trying to make a decision between using parametric methods or the cox proportional hazards model. The results (i.e., which variables are statistically significant) differ depending on whether I use parametric methods or the cox model. (Among the parametric models, I compared the different types and found the generalized gamma model had the lowest AIC.)

From what I have read, you should use parametric methods if you have an idea about the shape of the baseline hazard function. My question is: can one estimate the baseline cumulative hazard function and use this as information to determine whether to use parametric or cox? Or are there any plots that can be used to help make this decision?

thank you!
Caroline

[Carlo Lazzaro]

Caroline:
see Example 6 under -streg- entry in Stata .pdf manual (especially testing kappa parameter vs 0 and, separately, vs 1).
Comparing cumulative hazard functions from -stcox- and -streg- (with a given parametric form) can be a reasonable approach, that usually follows the comparison of the coefficients obtained from -stcox- and -streg- (for those parametric form that have a PH parameterization) regressions.
See for more on this topic: http://www.stata.com/bookstore/survi...-introduction/, pages 236-241.

[Marcos Almeida]

Carlo gave excellent advice as always.

I'm just wondering whether your time variable is continuous or discrete. Being the last case, you should consider a cloglog model.

[Carlo Lazzaro]

Caroline:
Marcos' wise advice implies, as requested by the FAQ, to post enough details about one's query, so that the interested listers can reply positively (and more efficiently) to the original poster.

[Caroline Wilson]

Thank you both for your helpful responses! The time variable is continuous.

[Carlo Lazzaro]

Caroline:
thanks for providing ione more detail about your query.
However, since generalized gamma model is implemented in AFT form only, I find difficult an easy comparison with -stcox- which is implemented in HR form.
That said:
- I would shy away from judging any statistical model by the significance of its coefficients;
- since -stcox- and -streg- machineries run differently, it may well be that the difference you came across are due to the featurs of the semi-parametric and parametric approaches;.
As an aside, you may want to post what you typed and what Stata gave you back (via CODE delimiters, please). That, in general boosts your chances of getting (more) helpful replies. Thanks.