library(survival) aeq <- function(x, y, ...) all.equal(as.vector(x), as.vector(y), ...) fit1 <- lm(skips ~ Opening + Solder + Mask + PadType + Panel, data=solder) y1 <- yates(fit1, "Opening") temp <- levels(solder$Opening) tpred <- matrix(0., nrow(solder), 3) for (i in 1:3) { tdata <- solder tdata$Opening <- temp[i] tpred[,i] <- predict(fit1, newdata=tdata) } all.equal(y1$estimate[,"pmm"], colMeans(tpred)) # This fit is deficient: there are no Opening=L and Mask=A6 obs # The MPV for Mask=A6 and Opening L will therefore be NA, as well # as for all levels of Solder, but we can compute the others. # Solder will be NA for all levels fit2 <- lm(skips ~ Opening*Mask + Solder, data=solder) y2a <- yates(fit2, "Mask", population="factorial") y2b <- yates(fit2, "Opening", population="factorial") y2c <- yates(fit2, "Solder", population="factorial") # The predict.lm function gives correct predictions for estimable # functions (all but L,A6) and nonsense for others. It knows that # some are not estimable due to the NA coefficients, but not which ones, # so always prints a warning. Hence the suppressWarnings call. tdata <- do.call(expand.grid, fit2$xlevels[1:3]) temp <- levels(solder$Mask) tpreda <- matrix(0., nrow(tdata), length(temp), dimnames=list(NULL, temp)) for (i in seq(along=temp)) { tdata$Mask <- temp[i] suppressWarnings(tpreda[,i] <- predict(fit2, newdata=tdata)) } tpreda[,"A6"] <- NA # the A6 estimate is deficient aeq(y2a$estimate[,"pmm"], colMeans(tpreda)) tdata <- do.call(expand.grid, fit2$xlevels[1:3]) temp <- levels(solder$Opening) tpredb <- matrix(0., nrow(tdata), length(temp), dimnames=list(NULL, temp)) for (i in seq(along=temp)) { tdata$Opening <- temp[i] suppressWarnings(tpredb[,i] <- predict(fit2, newdata=tdata)) } tpredb[,"L"] <- NA aeq(y2b$estimate[,"pmm"], colMeans(tpredb)) # Solder should be all NA all(is.na(y2c$estimate[,"pmm"])) # Tests for Solder are defined for a non-factorial population, however. # the [] below retains the factor structure of the variable, where the # runs above did not. R gets prediction correct both ways. y2d <- yates(fit2, ~Solder) temp <- levels(solder$Solder) tdata <- solder tpredd <- matrix(0, nrow(tdata), length(temp), dimnames=list(NULL, temp)) for (i in seq(along=temp)) { tdata$Solder[] <- temp[i] suppressWarnings(tpredd[,i] <- predict(fit2, newdata=tdata)) } aeq(y2d$estimate$pmm, colMeans(tpredd)) # # Verify that the result is unchanged by how dummies are coded # The coefs move all over the map, but predictions are unchanged fit3 <- lm(skips ~ C(Opening, contr.helmert)*Mask + C(Solder, contr.SAS), data=solder) y3a <- yates(fit3, ~Mask, population='yates') equal <- c("estimate", "test", "mvar") all.equal(y3a[equal], y2a[equal]) tdata <- do.call(expand.grid, fit2$xlevels[1:3]) # use orignal variable names temp <- levels(solder$Mask) cpred <- matrix(0., nrow(tdata), length(temp), dimnames=list(NULL, temp)) for (i in seq(along=temp)) { tdata$Mask <- temp[i] suppressWarnings(cpred[,i] <- predict(fit3, newdata=tdata)) } aeq(cpred[, temp!="A6"], tpreda[, temp!= "A6"]) # same predictions all.equal(y3a$estimate, y2a$estimate) y3b <- yates(fit3, ~Opening, population='yates') # column names will differ all.equal(y3b$estimate, y2b$estimate, check.attributes=FALSE) y3d <- yates(fit3, ~Solder) for (i in 1:3) { print(all.equal(y3d[[i]], y2d[[i]], check.attributes=FALSE)) } # Reprise this with a character variable in the model sdata <- solder sdata$Mask <- as.character(sdata$Mask) fit4 <- lm(skips ~ Opening*Mask + Solder, data=sdata) y4a <- yates(fit4, ~ Mask, population= "yates") y4b <- yates(fit4, ~ Opening, population= "yates") y4d <- yates(fit4, ~ Solder) equal <- c("estimate", "tests", "mvar", "cmat") all.equal(y2a[equal], y4a[equal]) # the "call" component differs all.equal(y2b[equal], y4b[equal]) all.equal(y2d[equal], y4d[equal])