Test results adjusted for 17 confounders
Hello everyone
Consider the following. You have 1 continuous primary input variable (IV). You want to use this variable to declare or predict 3 primary output variables (OVs): 1 continuous OV (OVc), 1 categorical OV (2 categories) (OV2), and 1 other categorical OV (> 2 categories) (OV+). You have identified 17 variables, other than the IV, that may potentially influence either the IV, or the OVs, and which or may not thus be identified as confounders (potential confounders, PCs).
My question in itself is easy: how to do this, while adjusting for any confounding variable?
Normally, you could:
Correlate IV with OVc using Pearson or Spearmans's (rank) correlation coefficient
Compare the IV between the 2 categories in OV2 using an unpaired Student's t test or a Mann-Whitney U test
Compare the IV between all categories in OV+ using a one-way ANOVA or a Kruskal-Wallis test
Test choice naturally depending on the satisfaction of the parametric assumptions (most notably Gaussian-like distribution and homoscedasticity; assessed with a Shapiro-Wilk and Levene's test) regarding the values of the variables.
And upon revelation of any statistically significant result, you could further analyse the value of the IV to predict the OVs using receiver-operating characteristic (ROC) analyses, linear regression, bi-/multinomial logistic regression models ...
However. These would be results for values not adjusted for any of the 17 PCs. How could I:
Identify true confounders amongst the PCs? Is it sufficient to perform basic tests such as the ones mentioned above, to determine whether any of the PCs has a significant influence on any of the OVs, or on the IV? Is it sufficient to do so, as such? E.g. consider one PC to be continuous (let's say, for the sake of ease, age in years). Would it suffice to perform the same tests as mentioned above to determine whether it has a significant influence on the OVs? Would it suffice to determine Pearson or Spearman's (rank) correlation coefficient to determine whether this PC has a significant influence on the IV (or the other way round)? Would any significant outcome here be a sufficient condition to label the PC as a statistically significantly true confounder?
How to I adjust values and test results above for any true confounder in SPSS? I understand that ANOVA, GLM, and logistic regression have built-in options to implement the (potential/true) confounders as covariates, but: (a) am I correct if I say that they will not give any confounder-adjusted result, but only show the influence of a (potential/true) confounder on the model? (b) What if not all assumptions for any of these methods are met? How could I then adjust values and test results for any true confounder? I don't know all necessary assumptions for each model, but I'm certain that I could (and would, if I wouldn't take it into account) stumble upon problems and "illegal" statistics since there most certainly will be variables interfering with the assumptions.
Thank you very much in advance for offering any clarification ...
Regards
Michael