9
$\begingroup$

I am trying to produce a linear mixed model the R code is as follows.

lme(Average.payoff~Game+Type+Others.Type+Game:Type+Game:Others.Type+Type:Others.Type,random=~1|Subjects,method="REML", data=Subjectsm1)->lme1

The response term Average.payoff is continuous whilst all explanatory variables are all binary.

When I come to validation I can clearly see that the spread of the residuals decreases with larger fitted values. Although there seems to be alot of information on heterogeneity in the form where the residuals increase with larger fitted values I have read nothing about cases similar to my own.

I have plotted the residuals against each explanatory effect and can see that the spread decreases with larger fitted values for the variables Game and Type but increases for the variable Others.Type.

What is the cause of this and what should I do about it?

Should I look at adding quadratic terms or using additive modelling? Is there a transformation that should be applied?

Thanks,

Jonathan

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

Adding quadratic terms would help if the mean varied that way but the variability is in the variance in your case. Since it is the covariates that cause the change, a form of variance function estimation involving those covariates would be the approach I recommend.

Improve this answer
$\endgroup$
3
  • 1
    $\begingroup$ Hi @Michael - Can you give any more detail here? It may not be clear to some how you can incorporate a functional form for the error variance into a regression model. Did you have something specific in mind? $\endgroup$
    – Macro
    Commented Aug 21, 2012 at 14:01
  • $\begingroup$ @Macro I am think of the models that provide jointly equations for the conditional expectation and the conditional variance of the response variable Y given the covariates. These models are discussed in Wayne Fuller's "Measurement Error Models" and in Carroll et al "Measurement Error in Nonlinear Models: A Modern Perspective." The idea is to express the error component of the model as having a variance that is σ$^2$ multiplied by a nonnegative function g of the covariates in the model. $\endgroup$
    – Michael R. Chernick
    Commented Aug 21, 2012 at 14:36
  • 1
    $\begingroup$ See Carroll et al. pp. 79 to 85 and Fuller. $\endgroup$
    – Michael R. Chernick
    Commented Aug 21, 2012 at 14:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.