2
$\begingroup$

I have been reading quite a few papers and recently have started analyzing some data on a review that I am doing.

I am trying to examine the effect of two different treatment options on a group of patients. Using Kaplan-Meier analysis, the p-value is statistically significant at 0.03. However I believe that since this is an observational study that various confounding factors may be influencing my survival graphs in each arm. I have done a univariate analysis (using Kaplan-Meier and log rank testing) for three factors age, weight and height. My univariate analysis only identifies a statistically significant difference in survival among my patients in the age group (categorical age >45 or <45).

Now I am not sure what to do.....I believe that I should now fit a Cox proportional hazard model and choose 'age' as a variable to control for and see if this affects my survival curves for treatment modality......is this the right thing to do?

Or should I analyze each of the variables....weight, height, etc., with a different Cox regression model. And if I do, how do I interpret that data? For example, when placing 'age' in the variable column (I am using epi info) my p-value now is no longer significant in my analysis of survival between treatment modalities. Does that mean that age was one of the factors that caused my two groups to have differing survival times?

Is this a multivariate analysis?

$\endgroup$
2
$\begingroup$

First, you shouldn't use variable screening to pick variables. It's not a good method.

Second, you shouldn't dichotomize variables such as age.

It seems like your main interest is in the treatment group. Then the other variables are covariates. So, you should include them if they have an effect on the model, whether or not they are significant. Or, possibly, you want to do matching or propensity score analysis or something.

Finally, Cox proportional hazards makes fewer assumptions than other forms of survival analysis; that's one reason it is dominant in the literature.

$\endgroup$
  • 1
    $\begingroup$ Firstly @Peter Flom, thank you so much for your comment. I am not sure that I understand you right ( likely due to my limited knowledge in stats). I also think that I may not have given the full picture. I am doing a retrospective review of patients in my clinic. They were treated with one of two treatment options A vs B. Kaplan Meier analysis for survival between the two treatment options and the p value was 0.03. I did not expect this. I was only able to collect data on a few of the factors that others have identified as predictors of survival in other reviews. $\endgroup$ – Roz Oct 4 '12 at 21:11
  • 1
    $\begingroup$ Of these three factors,I did a 2 x 2 chi squared analysis of my sample proportions treatment modality (each arm) vs eg. histology ( 2 options) and the p value here suggested that my groups may have been weighted heavily in favour of one characteristic eg. age among patients. I then did a univariate analysis for the three factors including age and only age was signif. However when controlling for age the p value was no longer significant. $\endgroup$ – Roz Oct 4 '12 at 21:12
  • 1
    $\begingroup$ I am taking this to mean that since age was a factor on univariate that did affect survival and when i did Cox proportional hazard and controlled for it and this survival diff disappears.... that means my previous p value was due to one group being weighted more heavily with patients with an age which adversely affects survival. And my chi squared proportional analysis did seem suggest that my sample distribution was not equal. $\endgroup$ – Roz Oct 4 '12 at 21:13
  • 1
    $\begingroup$ You make a valid point about age not being used as a dichotonous variable. However, I do not know the first thing about propensity score analysis or matching sadly. And I am not sure whether my reasoning is even partially sound. Could you help? Excuse that I repeated some information $\endgroup$ – Roz Oct 4 '12 at 21:13
  • 1
    $\begingroup$ Your reasoning makes sense. $\endgroup$ – Peter Flom Oct 4 '12 at 21:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.