Skip to content

Logit mixed model power analysis

6 messages · Evelien Heyselaar, paul debes, Thierry Onkelinx +3 more

#
Hi,

Firstly, I'm sorry if this question has been asked (and answered) before
(although I couldn't find it). I'm doing my analysis using logit mixed
models (family = binomial(link = "logit"), glmer model) and I was wondering
if there was a way to calculate power for my final model. I have looked
over the web and all I can find are programs and simulations for linear
mixed models with a continuous outcome, but not if the outcome is only 0's
and 1's. Is there a way for me to calculate power for this model?

Thank you very much,

Evelien
#
Hi Evelien & List,

A very recent publication might answer your question:

http://onlinelibrary.wiley.com/doi/10.1111/2041-210X.12306/pdf

Best,
Paul


On Mon, 14 Sep 2015 10:58:57 +0300, Evelien Heyselaar <ev.heys at gmail.com>  
wrote:

  
    
1 day later
#
Dear Evenlien,

Post-hoc power tests are not very informative. You will get a high
power when the signal is significant and low power when not
significant.

You can always use a brute force approach to estimate the power.
Simulate a dataset with know effect size. Analyse that dataset with
your model and store the relevant p-values. Repeat this so you get N
simulated datasets for that specific effect size. Then power = mean(p
< alpha).

Best regards,
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature
and Forest
team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
Kliniekstraat 25
1070 Anderlecht
Belgium

To call in the statistician after the experiment is done may be no
more than asking him to perform a post-mortem examination: he may be
able to say what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does
not ensure that a reasonable answer can be extracted from a given body
of data. ~ John Tukey


2015-09-14 9:58 GMT+02:00 Evelien Heyselaar <ev.heys at gmail.com>:
#
Hi Evenlien,

I second what Thierry said about the pointlessness of post-hoc power analysis when using the observed effect size.

This tutorial by Ben Bolker shows how to do power analysis for a binomial GLMM:
http://rpubs.com/bbolker/11703
NB you don?t need the first 3 lines any more (assuming you have a reasonably up-to-date version of lme4).

I also recommend Chapter 5 of Ben?s book Ecological Models and Data in R as an introduction to simulation, including power analysis.

Our paper on simulation-based power analysis for GLMMs 
http://dx.doi.org/10.1111/2041-210X.12306
which Paul Debes mentioned, has a supplementary online R tutorial, including a binomial example. 

Best wishes,
Paul
#
On Wed, 16 Sep 2015, Paul Johnson wrote:

            
A few people defend it as an estimate of the reproducibility 
probability. Gelman and Carlin suggest plugging in a "sensible" 
(Bayesian) estimate of the true effect based on your prior knowledge, so 
you can estimate the probability of a "Type S" (sign of true effect 
could actually be opposite to your study point estimate) or "Type M" 
error.

Cheers, David.

| David Duffy (MBBS PhD)
| email: David.Duffy at qimrberghofer.edu.au  ph: INT+61+7+3362-0217 fax: -0101
| Genetic Epidemiology, QIMR Berghofer Institute of Medical Research
| 300 Herston Rd, Brisbane, Queensland 4006, Australia  GPG 4D0B994A
#
I understand post-hoc power as being a function of the p-value, and
therefore redundant.  When asked for the kind of information that post-hoc
power is used to represent, I prefer to provide confidence intervals on the
parameters of interest - if possible.

Best wishes,

Andrew


On Wed, Sep 16, 2015 at 11:29 AM, David Duffy <
David.Duffy at qimrberghofer.edu.au> wrote: