Survival analysis with Cox proportional hazard model

[Tim Hartmann]

Hi guys,

I am analysing a dataset which contains information on the return to work behaviour of mothers after giving birth. I am interested in the effects of a legislation change. My sample consists of 400 individuals which are observed over at least 48 months. In total, I have 280 failure events.
Now my questions:

(1) my treatment and control group differ by size: 160 control vs 120 treatment. Does this influence the stcox regression? Do i have to take groups of equal size?

(2) My regression results are a bit disappointing with respect to the reform effect. From a histogramm one can see a positive reform effect (moms return earlier) but the estimated hazard ratio implies the opposite. Can this be because of the sample size? Has anyone experience with sample sizes which are used for Cox regressions?

My code for the regression looks like this. Results dont change much when i use robust standard errors or cluster by id

stcox treated covariate1 covariate2 covariate3 covariate4

Covariates are time constant and in most cases dummy variables.
Im greatful for any interesting comments on my problem.

Cheers
Tim

[Marcos Almeida]

Welcome to the Stata Forum / Statalist.

As recommended in the FAQ, the best approach so as to entice helpful replies are: share command and data; share images accordingly.

That said, in spite of having scarcity of information, I'd say:

1) Not much. No, it seems fine.

2) Maybe it is disappointing to you on account of your previous expectations, in terms of study question, but not necessarily due to the Cox regression method itself. Anyway, you didn't show the - stset - as well as - stdes and log-rank test let alone the output, hence I'm just hazarding a guess. With regards to graphs, you may wish to present a plot concerning the survival function. That would probably be more informative than a description of a histogram.

Hopefully that helps.

P.S: Edited: how come you have a sample size of 400 individuals, 160 in control group and 120 in treatment group?

[Carlo Lazzaro]

Tim:
- you reported the number of failure in the compared groups (120+160), not the total sample size. I assume that the remaining observations (400-280=120) do not experience the event (return to work) or at lost at follow-up?
- May a mom have more than one birth during the selected time horizon?
- I'm not clear with your statement about a seeming conflict between HR and histogram. It may well be that the seeming conflict is due to the fact that HR is adjusted for the remaining predictors.
Anyway, as suggested by the FAQ, posting what you tyoed and what Stata gave you back is, as always, the best way to increase your chances of getting helpful replies. Thanks.