Dear R community, In the excellent nlme package the default diagnostic plot graphs the innermost residuals against innermost fitted values. I recently fit a mixed-effects model in which there was a very clear positive linear trend in this plot. I inferred that this trend occurred because my fixed effect was a two-level factor, and my random effect was a 12-level factor. The negative residuals were associated with negative random effects (because of shrinkage, I assume), and the positive with positive. The fixed effects explained little varaition. Therefore plotting the innermost residuals against the innermost fitted values had the negative residuals to the left and the positive residuals to the right, occasioning a trend. My questions are: is it (as I suspect) harmless, or does it suggest that the model is lacking? And, is this effect likely to compromise the interpretation of any of the other standard diagnostic plots (eg qqnorm)? Thanks much for any thoughts, Andrew -- Andrew Robinson Ph: 208 885 7115 Department of Forest Resources Fa: 208 885 6226 University of Idaho E : andrewr at uidaho.edu PO Box 441133 W : http://www.uidaho.edu/~andrewr Moscow ID 83843 Or: http://www.biometrics.uidaho.edu No statement above necessarily represents my employer's opinion.
Is the fixed effect estimated at the innermost level? If not, plots of residuals at that level are surely of limited interest. qqplots, to be relevant, surely need to assess normality of effects (rather than residuals) at the level that matters for the intended inferences. If the fixed effect is estimated at the level of the random effect, then of course there are just 12 effects that should appear in any qq or suchlike plot. John Maindonald email: john.maindonald at anu.edu.au phone : +61 2 (6125)3473 fax : +61 2(6125)5549 Centre for Bioinformation Science, Room 1194, John Dedman Mathematical Sciences Building (Building 27) Australian National University, Canberra ACT 0200. On 19 Apr 2005, at 8:03 PM, r-help-request at stat.math.ethz.ch wrote:> From: Andrew Robinson <andrewr at uidaho.edu> > Date: 19 April 2005 12:41:24 PM > To: r-help at stat.math.ethz.ch > Subject: [R] Odd diagnostic plots in mixed-effects models > > Dear R community, > > In the excellent nlme package the default diagnostic plot graphs the > innermost residuals against innermost fitted values. I recently fit a > mixed-effects model in which there was a very clear positive linear > trend in this plot. > > I inferred that this trend occurred because my fixed effect was a > two-level factor, and my random effect was a 12-level factor. The > negative residuals were associated with negative random effects > (because of shrinkage, I assume), and the positive with positive. The > fixed effects explained little varaition. Therefore plotting the > innermost residuals against the innermost fitted values had the > negative residuals to the left and the positive residuals to the > right, occasioning a trend. > > My questions are: is it (as I suspect) harmless, or does it suggest > that the model is lacking? And, is this effect likely to compromise > the interpretation of any of the other standard diagnostic plots (eg > qqnorm)? > > Thanks much for any thoughts, > > Andrew > -- > Andrew Robinson Ph: 208 885 7115 > Department of Forest Resources Fa: 208 885 6226 > University of Idaho E : andrewr at uidaho.edu > PO Box 441133 W : http://www.uidaho.edu/~andrewr > Moscow ID 83843 Or: > http://www.biometrics.uidaho.edu
John, thanks for your response. The fixed effect is estimated at the innermost level. Andrew -- Andrew Robinson Ph: 208 885 7115 Department of Forest Resources Fa: 208 885 6226 University of Idaho E : andrewr at uidaho.edu PO Box 441133 W : http://www.uidaho.edu/~andrewr Moscow ID 83843 Or: http://www.biometrics.uidaho.edu No statement above necessarily represents my employer's opinion. ----- Original Message ----- From: John Maindonald <john.maindonald at anu.edu.au> Date: Tuesday, April 19, 2005 4:32 am Subject: [R] Odd diagnostic plots in mixed-effects models> Is the fixed effect estimated at the innermost level? If not, > plots of residuals at that level are surely of limited interest. > qqplots, to be relevant, surely need to assess normality of > effects (rather than residuals) at the level that matters for > the intended inferences. > > If the fixed effect is estimated at the level of the random > effect, then of course there are just 12 effects that should > appear in any qq or suchlike plot. > > John Maindonald email: john.maindonald at anu.edu.au > phone : +61 2 (6125)3473 fax : +61 2(6125)5549 > Centre for Bioinformation Science, Room 1194, > John Dedman Mathematical Sciences Building (Building 27) > Australian National University, Canberra ACT 0200. > > On 19 Apr 2005, at 8:03 PM, r-help-request at stat.math.ethz.ch wrote: > > > From: Andrew Robinson <andrewr at uidaho.edu> > > Date: 19 April 2005 12:41:24 PM > > To: r-help at stat.math.ethz.ch > > Subject: [R] Odd diagnostic plots in mixed-effects models > > > > Dear R community, > > > > In the excellent nlme package the default diagnostic plot graphs > the > > innermost residuals against innermost fitted values. I recently > fit a > > mixed-effects model in which there was a very clear positive > linear > > trend in this plot. > > > > I inferred that this trend occurred because my fixed effect was > a > > two-level factor, and my random effect was a 12-level factor. > The > > negative residuals were associated with negative random effects > > (because of shrinkage, I assume), and the positive with > positive. The > > fixed effects explained little varaition. Therefore plotting the > > innermost residuals against the innermost fitted values had the > > negative residuals to the left and the positive residuals to the > > right, occasioning a trend. > > > > My questions are: is it (as I suspect) harmless, or does it > suggest > > that the model is lacking? And, is this effect likely to > compromise > > the interpretation of any of the other standard diagnostic plots > (eg > > qqnorm)? > > > > Thanks much for any thoughts, > > > > Andrew > > -- > > Andrew Robinson Ph: 208 885 7115 > > Department of Forest Resources Fa: 208 885 6226 > > University of Idaho E : andrewr at uidaho.edu > > PO Box 441133 W : > http://www.uidaho.edu/~andrewr> Moscow ID 83843 > Or: > > http://www.biometrics.uidaho.edu > > ______________________________________________ > R-help at stat.math.ethz.ch mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide! http://www.R-project.org/posting- > guide.html