Effect Size Calculation (Cohen's D) on adjusted mean differences

[JG28]

Hi there,

I try to quantify the results from a test between a treatment and
placebo group using Cohen's d evaluated as below;

d=((μt-μp )) / ((pooled sd) )

Pooled SD=√ ((( (nt-1)(SEt √(nt ))^(2 )+ (np-1)(SEp √(np ))^(2 ) ))
/ ((nt+np-2)) )

Where the ‘t’ denotes active treatment condition and ‘p’ denotes
placebo.μt ad μp are the treatment and placebo group mean scores
respectively, and SEt and SEp are the respective standard errors.

In the case where my numerator is based on adjusted mean differences
e.g. I adjust the the placebo and treatment group means for the
average baseline value of the group (the score they received prior to
any treatment), should my pooled sd calculation also change? Or should
the difference still be quantified in terms of the original unadjusted
scale of the data?

(The adjustment to the means is done by modelling baseline as a
covariate in proc mixed in sas, and then using the LSMEANS statement)

Any help would be greatly appreciated!

Jo

[Frank Harrell]

As discussed near the top of http://biostat.mc.vanderbilt.edu/ManuscriptChecklist
there are many serious problems with using standardized effect sizes.

The SAS LSMEANS approach muddies the water. I suggest writing out the
predicted values you want to obtain, computing differences between
them, then getting confidence limits for the difference. The R rms
package's contrast function makes this easy.

Frank