[R-sig-ME] completely nested glmer

[Troels Ring]

Thanks a lot!!! - but there is certainly no expected gradient here over 
the categories of TRT. The problem pertains to early diabetes 
hyperfiltration which is not a direct function of hyperglycemia but 
controversial and complicated. I refrained from looking more into the 
problem of the distribution whether log-normal or gamma, since 
everything seems buried in the problem with nesting, and I feel 
uncomfortable deciding on the needed structure based on statistical 
tests, mainly since I fear that the TRT effect is totally mixed up with 
the RAT effect - if I manage to put it reasonably correct. Everybody 
would be happy if I had just stayed with the TukeyHSD which provides me 
beautifully low P values - which cannot be correct under the structure 
of the data - right?

Best wishes

Troels

  


Den 31-08-2017 kl. 08:46 skrev David Duffy:
> Troels Ring asked:
>
>> I have 2013 measurements of capillary flow speed obtained from 204
>> glomeruli originating from 29 rats, 10 of whom are controls, 11 made
>> diabetic, and 8 made hyperglycemic otherwise.  Altogether I have 621
>> capillaries from controls, 964 from diabetics and 428 from hyperglycemic.
>> If I make a direct aov
>> summary(z2 <- aov(Speed~TRT,SCAN))
>> TukeyHSD(z2)
>> #$TRT
>> #                 diff         lwr         upr     p adj
>> #Diab-Ctrl  -0.4266708 -0.65092625 -0.20241526 0.0000255
>> #Hyper-Ctrl -0.2206938 -0.49449343  0.05310581 0.1416500
>> #Hyper-Diab  0.2059769 -0.04716959  0.45912348 0.1365243
>> summary(z3a <- glmer(Speed~TRT + (1|RAT)+(1|ind) + (1|RAT:ind)
>> ,data=SCAN,family=Gamma)) where ind is an indicator for each of the 204
>> glomeruli. And the fixed effect TRT is not significant.
> I would look at the permutation test for the Jonckheere-Terpstra  test statistic for Speed ~ TRT, where TRT is a quantitative
> variable 0=Ctrl 1=Hypergly 2=Frank diabetes, given I would have a strong prior hypothesis that an
> association will take this form. You should look to see if RAT:ind is required eg
> z3b <-  glmer(Speed~TRT + (1|RAT)+(1|ind),...)
> anova(z3a, z3b)
> and look at diagnostics for whether a gamma is really best (cf log-normal).
>
> Just 2c, David Duffy.