Skip to content

much different results for random effect vs simple lm.

2 messages · Brent Pedersen, David Duffy

Brent Pedersen
# 
Hi, I have a model like this:

    # for many y values/probes
    formula = y ~ concordant + age.proband + age.sibling + sex.proband
+ sex.sibling

I run this model and get p-values with the formula:


    model = lm(formula, data=df2)
    s = summary(model)
    p.cordant = s$coefficients["concordantT", "Pr(>|t|)"]


But, an proband can have multiple siblings, so I want to account for
family structure:
So, I use:

    library(lme4a)
    # for many y values.
    model = lmer(y ~ concordant + age.proband + age.other +
sex.proband + sex.proband + sex.other + (1| family_id.proband),
data=df)

    degrees.of.freedom = length(unique(df$family_id.proband)) - 1

    p.from.t = function(t){
        2*pt(-abs(t),df=degrees.of.freedom)
    }
    s = summary(model)
    concordant.t.score = s$coefficients['concordantT', 't value']
    pcordant = p.from.t(concordant.t.score)


Everything else between the 2 runs is the same. For the simple case, I
have unique 80 pairs (since I only use each proband once),
and for the latter, I have 98 pairs. I'm doing this test for millions
of probes and looking for regions of where the concordant
parameter is significant, I find much different regions between the 2
models--very little overlap.
Is this to be expected? Intuitively, I'd figure that using
the random effect via lme4a would just give more power. Are my p-value
calculations correct?

Thanks for any feedback,
-Brent
David Duffy
#
On Mon, 20 Jun 2011, Brent Pedersen wrote:

            

      
      
      
    

        
You need to look at just a few probes in detail.  Given you have such a 
small sample size (and how many concordant pairs?), you might expect a bit 
of shifting about.  The other model you should check (in a subset) is your 
first model fitted to all 98 pairs, using your conservative degrees of 
freedom from model 2 (this would be pretty similar to a GEE, AIUI).

Cheers, David Duffy.