I have analyzed several survival data using the KM technique, showing more than one variable significantly predictive of the outcome. Now, I think I need to construct a model to compare these variables, and find the truly independent ones, by their RR. Can anyone suggest how to do these on SPSS, or should I look elsewhere (R, Stata, etc)?
What is the goal of your analysis? Do you wish to construct a prediction model, or to evaluate the association of risk-factor associations that are adjusted for confounders or extraneous effects? (Fully) p-value driven interpretation and manipulation of data has been strongly criticized in the literature, and should be avoided whenever possible.
@Thomas: Thanks for your comments. I intend to evaluate the association of risk-factor associations that are adjusted for the possible confounding effects of other seemingly significant variables? Also, I would like some info on other options for data interpretation that exclude the p-value.
When it comes to adjustment of confounding, it is important to think about which variables may truly have an influence on the determinant and the outcome. The interpretation of adjusted RRs then depends on the included variables in the model (whether they are significant or not). Some options to select confounders are:
- Add (expected) strong confounders until the effect estimate of the risk factor converges. Unfortunately, it is possible that the combined effect of some confounders may be very strong, or that changes in effect estimates occur due to non-confounders.
- p-value driven (arbitrary, does not account for strength of confounding, and confounder-outcome relation is not of primary interest)
- model fit driven (arbitrary, model fit is not of primary interest)
All these approaches have disadvantages and may result into biased estimates of risk factor associations under specific conditions. If you have any knowledge about your data, it would be useful to think about potential confounders and try to combine this knowledge with statistical approaches (e.g. check to which extent what you expect and what you see agree). Then, you may apply some sensitivity analyses (e.g. try different sets of confounders, more simple to more complicated models) and see how conclusions change based on the underlying assumptions.
Some references:
- http://www.ncbi.nlm.nih.gov/pubmed/21796419
- http://www.sciencedirect.com/science/article/pii/S0883944106000062
Hi, I beta-test for SPSS. I may be able to help. What analyses have you tried?
I am uncertain from your description so far whether logistic regression or Cox proportional hazards is the best initial approach. Both can allow you to winnow-down the field of potential predictors.
Thanks Thomas & Howard. Your suggestions have been very helpful. I'm now using the Cox Regression function in SPSS (Analyze>Survival>Cox Regression), and it appears to do the job quite well. Though, I've also noticed that the effect of the "main" independent variable on the outcome, changes with the confounding variables being co-tested; and I now wonder which of the HR figures to report. Some suggestions in that direction would be appreciated.
Hi Anthony
I favor including those confounding factors in the modeling that you expect to be present in the situation to which you wish to generalize your findings, keeping those potential confounders explicitly present in the results tables you present, and commenting on them in the results section or discussion. You could also consider presenting the coefficients calculated with and without the confounders present, treating this as a sensitivity analysis of the stability of the model you present. This is the sensitivity analysis Thomas suggested above. With regards to p-values, you can set a threshold level for inclusion in the model as you're developing the analysis. I prefer using confidence intervals around the coefficients rather than (or in addition to) point significance levels in the final presentation, The CIs give a sense of the stability of the coefficients.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software