Saturday, August 7, 2010

Compare 3 groups using stats.5 variables are sig different between groups.How do I control for the variables?

I have 3 groups that are based on a clinical characteristic. I am looking at five outcome variables. There are 5 independent variables that are significantly different between the groups. I ran regressions on all of the outcome variables for the entire group and for the 3 subgroups based on my clinical characteristic to see which variables could be used to predict the outcomes. How do I control for the significantly different independent variables in the analyses or have I already done it by running the analyses for each of the subgroups. I am stumped as to whether I have done enough. I have run chi square for the categorical variables, ANOVA for the continuous variables, Pearson and Spearman correlations and regressions appropriate for the type of outcome variable I have. Does the fact that some of the independent variables are different across the groups mean anything now? HELP. I am using SPSS to analyze the data.Compare 3 groups using stats.5 variables are sig different between groups.How do I control for the variables?
I am a clinical statistician. Are the five outcome variables clinically, not statistically related. For example, five outcome variables from the same scale. If so you need to use statisticial methodology that simultaneously controls for multiple endpoints. Otherwise, you need to control for multiplicity. But this depends on the hypothesis(es) you have in mind. With more detail I can answer your question. If the outcome variables are dependent (say correlation %26gt; 0.66) then you need to use MANOVA. Another means is to use proc mixed in SAS using a dependent covariance structure. Cannot help with SPSS. I am a SAS user.Compare 3 groups using stats.5 variables are sig different between groups.How do I control for the variables?
i think you need a separate control group for the independent variables a group you do not apply the treatment to but test as a control
you can treat ';group'; as a discrete variable with three levels. introducing a main effect for 'group' into the regression will fit parallel regression surfaces for the three groups that differ by their intercepts. if you add also two-factor interactions between 'group' and each of the independent variables (including quadratic and higher terms if you used them), that will be equivalent to fitting regressions separately for the three groups. you can then test to see if the model with the 'group' two-factor interactions gives a significantly better fit than the model with just the main effect for 'group'. that some independent variables are significantly different between groups should not affect the preceding analysis.

No comments:

Post a Comment