[R-sig-ME] Multi-level qualitative (fixed-effects) factors

[Peter Francis]

Hi David and Ben, thanks for your help - 

I was worried this would make little sense!

I have set out my candidate models

A+B+C+D
B+C+D
A+C+D
etc etc

And am running through them in lmer. Factor A for instance is Habit, which takes 3 forms - aquatic, terrestrial or epiphyte. 

When i run the model with A as a factor i get the breakdown of the individual levels habitat 1, habitat 2 and habitat 3 and a corresponding AIC score. However if i just run it with habitat 3 - aquatic - i get a lower AIC score, therefore the model fits the data better?

I am unsure how to, without splitting my factors into their constituent levels at the beginning - A1+A2+A3 + B1 + B2 etc, arrive at the model with the lowest AIC?

Thanks

Peter

On 3 Aug 2010, at 00:16, David Duffy wrote:

On Mon, 2 Aug 2010, Peter Francis wrote:

> I have many multi level factors i.e habit - aquatic, terrestrial, epiphyte etc
> 
> I ran the model with habit as a factor
> 
>> model111 <-lmer(threatornot~1+(1|a/b) + habit, family=binomial)
> 
>> Generalized linear mixed model fit by the Laplace approximation
>> Formula: threatornot ~ 1 + (1 | order/family) + habit
>>  AIC  BIC logLik deviance
>> 1406 1436 -696.9     1394
>> Random effects:
>> Groups       Name        Variance   Std.Dev.
>> family:order (Intercept) 6.9892e-01 8.3602e-01
>> order        (Intercept) 4.2292e-14 2.0565e-07
>> Number of obs: 1116, groups: family:order, 43; order, 9
>> 
>> Fixed effects:
>>            Estimate Std. Error z value Pr(>|z|)
>> (Intercept) -0.04803    0.19174  -0.250  0.80219
>> habit2       1.10627    0.41607   2.659  0.00784 **
>> habit3       0.92578    0.78141   1.185  0.23611
>> habit4       0.14383    0.38477   0.374  0.70856
> 
> ---
> Which had a AIC of 1406
> 
> I then re-ran the model with only aquatic and got a lower AIC value - which i guess is to be expected as aquatic is highly significant and aquatic species are more prone to threat ( my response).
> 
> 
>>> model112 <-lmer(threatornot~1+(1|a/b) + aquatic, family=binomial)
>>> model112
>> Generalized linear mixed model fit by the Laplace approximation
>> Formula: threatornot ~ 1 + (1 | order/family) + aquatic
>>  AIC  BIC logLik deviance
>> 1395 1415 -693.4     1387
>> Random effects:
>> Groups       Name        Variance Std.Dev.
>> family:order (Intercept) 0.60007  0.77464
>> order        (Intercept) 0.00000  0.00000
>> Number of obs: 1116, groups: family:order, 43; order, 9
>> 
>> Fixed effects:
>>            Estimate Std. Error z value Pr(>|z|)
>> (Intercept)   0.1572     0.1827   0.860 0.389613
>> aquatic      -0.6683     0.1737  -3.847 0.000119 ***
> 
> My question is - when i developed the candidate models i thought using multilevel factors would be OK and i would be able to tease out the individual levels. If i split the factors into levels from the beginning then i am left with a huge amount of candidate models? This would not be a problem in stepwise regression as i could just remove the habit with the least significant P Value.
> 
> If i remove habits i "feel" are unimportant from the beginning i feel i would be limiting the model too much.
> 
> I hope this makes sense!

I don't understand at all, I'm afraid.  Is aquatic the same as habit=2, or something?  If so, there is something funny about the model fits.

If family and order are "nuisance" variables, then a stepwise approach is quite reasonable (if you are someone who thinks stepwise regression is reasonable, of course ;)).

Just 2c, David Duffy.

-- 
| David Duffy (MBBS PhD)                                         ,-_|\
| email: davidD at qimr.edu.au  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v