[R-sig-ME] lmer, problem with dependent variables with a binomial distribution

[Ken Beath]

On 30/01/2009, at 1:56 AM, Liliana Martinez wrote:

> Hi,
>
> I am trying to use the lmer function from the lme4 package in R  
> 2.8.0. to fit a generalized mixed-effects model for a dependent  
> variable with a binomial distribution (for more info on my  
> experiment, look below). However, I encounter some problems due to  
> the nature of my data.
>
> First, some background info without which it is not possible to  
> explain my problem:
> I am a linguist trying to analyze the geometry of path encoded by  
> some verbs of motion. In my experiment the subjects drew sketches of  
> the trajectory of a moving car, prompted by stimulus sentences  
> containing certain verbs. The sketches then were categorized  
> according to the extension of the curve drawn by the subject. Thus,  
> my dependent variable was originally intended to be a factor – curve  
> extension – with 6 levels (45, 90, 135, 180, 270 and 360 degrees). I  
> preferred a factor instead of a numeric variable, because it was not  
> always possible to assign a precise angular value, and the only  
> solution was to assign the sketch to one of several cognitively  
> motivated categories. When I sought advice for appropriate means of  
> analysis, it became clear that no known (to my circle of contacts)  
> means exist to analyze this type of data. Therefore I had to  
> restructure my original dependent variable into 6 dependent  
> variables with a binomial distribution (occur
> – no-occur) – one for each of the curve-extension categories.
>

Commercial programs like Latent GOLD (probably requiring the Syntax  
Module) and MPlus will probably analyse your model, as they allow  
ordinal outcomes and multilevel modelling. If you have access to stata  
gllamm may also work.

The method you are using looks like it will be difficult to interpret.

>
> The problem:
> For some of these binary dependent variables some values of the  
> predictor do not elicit any occurrences, e.g., none of the subjects  
> drew a 360-degree curve when the stimulus contained verb A, and many  
> subjects drew loads of 360-degree curves when the stimulus contained  
> verb B, yet the model does not find any significant difference  
> between verb A and verb B (and there is a tremendously big value for  
> Standard Error.
>

The large value of the standard error is due to the estimate tending  
towards negative infinity, resulting from perfect relationship between  
the dependent and that level of the covariate.


>
> [13:43:20] Liliana Martinez sier : Fixed effects:
>                      Estimate Std. Error z value Pr(>|z|)
> (Intercept)            1.5476     0.4260   3.633 0.000280 ***
> v_prepobikaliam_ze    -0.1370     0.3691  -0.371 0.710548
> v_prepzaobikaliam_ze  -4.9991     0.5346  -9.352  < 2e-16 ***
> v_prepzavivam_kum    -19.6208   525.3207  -0.037 0.970206
> v_prepzavivam_pokrai  -6.6705     0.9123  -7.312 2.64e-13 ***
>
> Yet, the difference is striking, and is obvious both in the figures  
> (see attachment) and in crosstab statistics. How can I document and  
> report the difference between the predictor values in such cases? If  
> this method of analysis cannot provide a good solution, can you  
> suggest alternative ways of analysis?
>
>
> best regards
>
> Liliana
>
>
>      _________________________________________________________
> Alt i ett. F
> .mail.yahoo.com_______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models