Hi, I am currently performing survival analysis using SPSS. I wanted to determine which of the following factors: (expression of several proteins from immunohistochemistry and several clinicopathologic parameters) influence the survival of cancer patients. However, during univariate analysis, most of the factors did not fulfill the proportional hazard assumption (the Kaplan-Meier curve have multiple overlaps). Questions:
1) Can I still include the factors (especially the protein expressions) in Cox Regression analysis? It is a fairly small study with about 30 patients.
2) If not allowed, are there any potential solutions? Some of the curves overlap at multiple time points. (I have attached an example of the curve)
3) During Cox Regression, which method is preferred? Can I use 'Enter' or 'Backward LR'? If I use Backward LR, do I still need to perform univariate analysis before and filter the variable to be included into the regression? How about the 'Enter' method?
4) When I tried including all factors into Cox Regression, there are a couple of variables where the hazard ratio is < 1 during univariate analysis and non-significant, but becomes > 1 and significant during Cox regression. What are the possible explanations?
Thank you for your help.
Youi might try this approach it worked for me. Best wishes David Booth
Further here's where you can find more information on the PH assumptions. Best wishes, David Booth
Dear Andy,
1) in order to check how strongly the PH assumption is violated, some p-value should be reported beyond visual inspection. One way to do so would be to calculate the chi-square test of the Schoenfeld residuals, that way you can get a feel of if the factors are all equally bad for specifying the model or not.
If there are factors that do not fit in the assumptions, they should not be included as the model is badly specified for them.
2) The solution depends on the relationships between your variables. Are those variables correlated to others that do satisfy the PH assumption so they can just be safely omitted, is that level of correlation expected from the theory? Are they biologically relevant and need to be included, so you might try a parametric survival model like a Weibull? Can you constrain your survival data to a range where there are not many overlaps between curves?
3) While I am not a big fan of stepwise selection and would rather go with a LASSO approach, within this paper it is suggested that the backwards stepwise might be the best method out of those used for the selection. However, take into account that by selecting features based on an algorithm, you are estimating coefficients and performing statistical tests under the hood, therefore your final model should have the reported p-value corrected for the number of tests done. I would not do univariate analysis first, as the significance and values of the coefficients change based on the other features and you want a multivariable model (also this increases the number of tests to correct for afterwards). Rather, I would drop variables based on correlation first with each other. Article Five myths about variable selection
4) The explanation for this is simple, the variables are not independent between themselves so adding more variables to the model alters the coefficients that are estimated. Another way to look at it is that, if a target variable´s variation is due to two (or more) variables that are correlated, in a univariate model all that variation is assigned to just one variable, biasing the importance and the coefficient of the variable (the omitted variable bias).
Hope it can help a bit! :)
- Badges
- Science topic
- Similar topics
- Correction