The manual page for sigma() claims that, for a linear model at least, the result is the same as summary(fit)$sigma. This is untrue if the lm fit has NA coefficients. The problem is that sigma.default uses the length of coef as a surrogate for degrees of freedom, it should use the number of non NA coefficients. Simple example: temp <- data.frame(x1 = LETTERS[c(1,2,3,1,2,3,1,2,3)], x2 = letters[c(1,2,1,2,1,1,1,2,1)], y = 1:9) fit <- lm(y ~ x1*x2, data=temp) sigma(fit) # 4.24 summary(fit)$sigma # 3.67

You are right. I tend to agree that probably length(coef(object)) should be changed to sum(!is.na(coef(object))) but I do not start any action right now...

fixed in both R-devel and (R 3.4.1)-patched