7

I have a question relating to the checking for outliers and / or influential points in my dataset using a glmer model with 3 random variables. I'm investigating the detection rate (SumDetections) of receivers over increasing distance (sc.c.distance), and the effect of environmental influences on this (depth, temperature and wind) and how this differs between different transmitters used, controlling for random effects of receiver ID, replicate and area. I found that the influence.ME might be of help, so I checked it out.

In the manual I read that this package is only able to delete levels of 1 single grouping factor or 1 data point per time over the whole data set. Unless I read the package info incorrectly, this package cannot do what I'm looking for. I'm looking for a way to check for outliers nested within 4 grouping layers.

How my data is organized is as follows: First my data discerns between Areas. Within areas, multiple replicates were done. Each replicate consisted of 5 distances at which the detection rate was tested. For each distance, 20 receivers were tested.

My model looks like this:

m <- lmer(SumDetections ~ tm + sc.c.distance + tm:sc.c.distance + c.tm.depth + 
             c.receiver.depth + c.temp + c.wind + (1|replicate) + (1|SUR.ID) + (1|Area), 
             data = df3, family = poisson)

My questions are:

  • Is it possible to check for outliers of which the data is nested within 3 layers with use of influence.me?
  • If so, how should I specify the command to get what I'm looking for, and how should I interpret the returned data by the Cook's distance or dfbetas?
  • If not, is there another package that allows me to check for outliers?
mdewey
  • 16,541
  • 22
  • 30
  • 57
FlyingDutch
  • 253
  • 1
  • 4
  • 7
  • Anyone? I posted in the lme4 mailing list as well, but I still haven't got a solution to my problem. THanks – FlyingDutch Apr 10 '13 at 08:30
  • Which function in influence.ME are you trying to use? If I remember correctly 1 of the 3 has limits on the number of random effects it can handle, while the other 2 are more flexible. Bolker has a glmm walk through with code for diagnostic functions that do basic flagging of outliers http://rpubs.com/bbolker/glmmchapter You could use these to eyeball outliers and remove them by hand. I don't know the theory behind how influence.ME works but you could probably get an approximate solution by looping over your dataset and calculating influence by hand. – N Brouwer Feb 03 '15 at 02:40

1 Answers1

1

try the romr.fnc in the LMERConvenienceFunctions to remove outliers

df3.trimmed = romr.fnc(m, df3, trim = 2.5)
df3.trimmed = df3.trimmed$data

update initial model on trimmed data

mB = update(m1)
cconnell
  • 119
  • 1
  • 2
    I looked at the package, and found that this cannot be used for GLMMs. Sorry if I didn't specify this clearly, but i'm trying to fit a GLMM with a poisson distribution. – FlyingDutch Apr 15 '13 at 15:50