I have been asking a number of separate questions as each are unique in themselves but connected to my learning process in running mixed effects models. I apologise if i am becoming a nuisance.

Basically I learned how to run a model correctly using lme4 (an lmer) and as I had many factors I was then able to use the MuMIn package to identify those models that best explained y rather than having to run them all by hand. I would like to better understand Effect and then 95% CI using the model.avg command.

I have 9 factors (including the random but not y). After using the dredge command I ran:

get.models(dd, subset=delta<4)

where dd is the data.frame from the dredge command (i.e. all possible combinations of factors in the global model). I ended up with 14 models listed (out of over 200), although only the top two had a delta<2.

I know that papers will often report the importance of these factors, so would it be advisable to find the Estimate and 95% CI for all the factors for all 14 models or should I just look at those factors within the top two models (ie with a delta<2)? I think the two models is what other papers have done.

Could someone tell me how to do this? Do I run all models that are in the top (either 2 or 14) and then use the avg.model command? Could someone let me know what the code is for this?

In addition, once I do this, would it be okay to then list my results below so that someone could help me decipher what they mean? I don't have a deep understanding of the intricacies, but it would be helpful for me to be able to further explain the magnitude of the effects of the most important factors rather than just the weights of the models themselves. I understand AICc and weights, but not the t values and such that come after each model.

I appreciate everyone's patience with me in this.

  • 1
    $\begingroup$ I'm sure some will disagree with me, but I think you're well out of your depth and you need more help than can reasonably be provided here. I mean no disrespect, but you appear to be like someone who wants to learn to run without first learning how to crawl and walk. I'd encourage you to get advice from a local statistician, if at all possible. $\endgroup$ – mark999 Mar 24 '12 at 20:13
  • $\begingroup$ I appreciate your point of view and you are correct, however I do not have those resources available to me. I have learned a lot and grasp things quickly. Three months ago I had never even heard of GLMs. I did a research project, and had submitted a project proposal using simple stats, however my sample size ended up being so small that these became obsolete. I understand how models are used and what they show, but do not understand the inside out, but as my project advisor is 3000 miles away and said he can't help me online, then I don't have any options and I need to run these tests. $\endgroup$ – Dragonwalker Mar 24 '12 at 21:03
  • $\begingroup$ Fair enough. I hope it works out for you. $\endgroup$ – mark999 Mar 24 '12 at 21:07
  • 1
    $\begingroup$ I agree with @mark999. The tools you'd like to use are very sophisticated - which is why, for example, there are many health warnings in Doug Bates' lme4 documentation. Using any statistical or scientific method that you (or your couthors) cannot defend leaves a large credibility gap. So, while I sympathize with the need to get something out, reporting the results you get from simple tools is likely better science than reporting GLMM output, if the model formulation and fitting process are going to be unexplained "black boxes". $\endgroup$ – guest Mar 24 '12 at 22:24
  • 1
    $\begingroup$ I finally figured it out so no need to answer this question. $\endgroup$ – Dragonwalker Mar 25 '12 at 13:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.