shirley zhang
2007-Jun-05 15:29 UTC
[R] Can I treat subject as fixed effect in linear model
Hi, There are 20 subjects grouped by Gender, each subject has 2 tissues (normal vs. cancer). In fact, it is a 2-way anova (factors: Gender and tissue) with tissue nested in subject. I've tried the following: Model 1: lme(response ~ tissue*Gender, random = ~1|subject) Model 2: response ~ tissue*Gender + subject Model 3: response ~ tissue*Gender It seems like Model 1 is the correct one since my experiment design is nested design. However, can anybody tell me whether Model2 is completely illegal? Thanks
Spencer Graves
2007-Jun-07 04:16 UTC
[R] Can I treat subject as fixed effect in linear model
The short answer is that you could fit that fixed-effect model using 'lm', for example. That would make sense if you wanted to make inference only about that particular group of 20 subjects AND you thought it was inappropriate to consider that their contribution to a model would follow a normal distribution. If you want to make inference beyond that group of 20 subjects, then the fixed effect analysis is not appropriate. If you thought it was inappropriate to think that individual adjustments for each subject were normally distributed, then the question is how far from normal might they be. I don't think Model2 is "illegal" in the sense that you are not likely to be sent to prison for using it. However, I wouldn't do it. I'd make use your Model 1 and make all the plots that seem consistent with that model, as described in Pinheiro and Bates (2000) Mixed-Effects Models in S and S-Plus (Springer). If the plots (or something else) suggested that some of my assumptions were inappropriate, then I'd consider other alternative models. However, that could be a lot of work, and I wouldn't undertake such an effort without some pretty strong justification. Hope this helps. Spencer shirley zhang wrote:> Hi, > > There are 20 subjects grouped by Gender, each subject has 2 tissues > (normal vs. cancer). > > In fact, it is a 2-way anova (factors: Gender and tissue) with tissue > nested in subject. I've tried the following: > > Model 1: lme(response ~ tissue*Gender, random = ~1|subject) > Model 2: response ~ tissue*Gender + subject > Model 3: response ~ tissue*Gender > > > It seems like Model 1 is the correct one since my experiment design is > nested design. However, can anybody tell me whether Model2 is > completely illegal? > > Thanks > > ______________________________________________ > 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 > and provide commented, minimal, self-contained, reproducible code. >