# # This code verified some oddities about model frames # options(contrasts=c("contr.treatment", "contr.poly")) # clean slate tdata <- data.frame(y=1:10, x1=letters[c(1,2,3,1,2,3,1,2,3,1)], x2=LETTERS[c(1,2,3,4,1,2,3,4,1,2)], stringsAsFactors=TRUE) tdata$x3 <- as.character(tdata$x1) fit1 <- lm(y ~ x1 + x2, tdata, x=TRUE) m1 <- fit1$model t1 <- terms(fit1) # Lesson 1: xlev is ignored when the variable already has levels # as an attribute temp <- list(x1=c("a", "c", "b"), x2=LETTERS[4:1]) x2 <- model.matrix(t1, m1, xlev=temp) x3 <- model.matrix(t1, tdata, xlev=temp) all.equal(x2, fit1$x) !is.logical(all.equal(x2, x3)) # x2 and x3 do not agree attributes(m1$x1) # Lesson 2: character variables do not have their levels # remembered as attributes in the model frame, but these are # found in fit$xlevels and fit$contrasts. # However, the xlev argument is still ignored for a model frame! fit2 <- lm(y ~ x3 + x2, tdata, x=TRUE) m2 <- fit2$model x3 <- model.matrix(terms(fit2), m2, xlev=fit2$xlevels) x4 <- model.matrix(terms(fit2), m2, xlev=list(x3=letters[3:1])) x5 <- model.matrix(terms(fit2), tdata, xlev=list(x3=letters[3:1])) all.equal(fit2$x, x3) all.equal(x3, x4) all.equal(x3, x5) # FALSE # Lesson 3: contrasts.arg is relevant, even when the model frame # has a saved contrast ctemp <- list(x1="contr.SAS", x2= contr.helmert(LETTERS[1:4])) x4 <- model.matrix(t1, m1, contrasts.arg=ctemp) # no saved contrast fit3 <- lm(y ~ x1 + C(x2, contr.SAS), tdata) m3 <- fit3$model attr(m3[[3]], 'contr') # the contrast is saved c2 <- ctemp names(c2) <- names(fit3$contrasts) x5 <- model.matrix(terms(fit3), m3, contrasts.arg=c2) all.equal(x4, x5, check.attributes=FALSE) # Lesson 4: and this holds for a character variable as well fit4 <- lm(y ~ x3 + C(x2, contr.SAS), tdata, x=TRUE) c4 <- ctemp names(c4) <- names(fit4$contrasts) x6 <- model.matrix(terms(fit4), fit4$model, xlev=fit4$xlevels, contrasts.arg=c4) x7 <- model.matrix(terms(fit4), fit4$model, contrasts.arg=c4) all.equal(x7, x6) all.equal(x7, x5, check.attributes=FALSE)