I'm a researcher on HIV resistance, I have a study population of 95
patients with 76% responders. I compare responders with
non-responders, and found six (continuous) variables that are
associated with, and thus predictive of, response. Now I want to
correct my six study variables for confounding. I have identified 1
confounder (p=0.011). I performed bivariate logistic regression
analysis (variable + confounder) (Enter method as well as
forward+backward analysis) for all six variables, using SPSS. In all
analyses, the confounder was thrown out of the equation, while all the
study-variables ended in the equation of the model. (1) Can I conclude
that after correction, the variables are still assoiated with respons?
When I look at the p-values of the variables in the end-model, for 3
variables, the p-values are higher than the uncorrected p-values I got
from the previous univariate log regression. (2) Does this mean that
after correction for confounding, the association becomes less
certain?
For one variable, the p-value ended even slightly above 0.05! (3) Is
this variable, although it is still in the equation, not significant
anymore after correction for confounding?
I thought correction for confounding strengthens associations, but now
I'm puzzeled... |
