atyre2 / nres803 Goto Github PK

Check non_linear does not work with transformed variables that are not specified as as columns in the data frame that was used to build the model. No idea why this is happening, anyone have any ideas?

E.g.
data("iris")
x <- lm(Sepal.Width ~ log(Sepal.Length), data = iris)
check_nonlinear(x)

Error in FUN(X[[i]], ...) : object 'Sepal.Length' not found

If there are few unique levels of .fitted, the scale-location and residual vs. fitted plots get whacky with lots of warnings.

library("tidyverse")
library("broom")
library("NRES803")
library("NRES803data")

data(RIKZ)
dim(RIKZ)

RIKZ$Richness <- rowSums(RIKZ[,2:76] > 0)
dim(RIKZ)

richness_week <- ggplot(RIKZ, aes(x = week, y = Richness)) +
  geom_point() +
  geom_smooth(method = "lm")
richness_week

RIKZ_model.2 = lm(Richness ~ week, data = RIKZ)
summary(RIKZ_model.2)

test_RIKZ2 <- augment(RIKZ_model.2, data=RIKZ)
ggplot(test_RIKZ2, aes(x = .fitted, y = .resid)) + 
  geom_point() + 
  geom_smooth()

# this is lifted out of base R qqline()
y <- quantile(test_RIKZ2$.resid, c(0.25, 0.75))
x <- qnorm(c(0.25, 0.75))
slope <- diff(y)/diff(x)
int <- y[1L] - slope * x[1L]

ggplot(test_RIKZ2, aes(sample = .resid)) + 
  stat_qq() + 
  geom_abline(slope = slope, intercept = int)

ggplot(test_RIKZ2, aes(x = .fitted, y = sqrt(abs(.std.resid)))) + 
  geom_point() + 
  geom_smooth() + 
  geom_hline(yintercept = 1)

ggplot(test_RIKZ2, aes(.hat, .std.resid)) +
 geom_vline(size = 2, colour = "white", xintercept = 0) +
  geom_hline(size = 2, colour = "white", yintercept = 0) +
  geom_point(aes(size = .cooksd)) + geom_smooth(se = FALSE)

I've noticed a strange issue with the check_assumptions() function. Check_assumptions() should return one (!) ggplot-object with four panels. During recent use however two ggplot-objects are returned with two panels each. This happens for lms, gams, etc.

I'm not sure why this is happening as the code seems fine. I assume it is an issue with how we add the individual plot objects together. Both ggpubr::ggarrange(rr0, rr1, rr2, rr3, nrow=2) and ggpubr::ggarrange(plotlist = list(rr0, rr1, rr2, rr3), nrow=2) have the same issue.

I have also found a workaround: instead of ggpubr::ggarrange(rr0, rr1, rr2, rr3, nrow=2) use ggpubr::ggarrange(rr0, rr1, rr2, rr3, nrow=2, ncol = 2).

I'll commit the changes to the code later but still this is strange behaviour as nrow = 2 should do the trick...

When there are few unique values on the x axis, the smooths in the standardized residuals and scale location plots get bizarre and unhelpful. Detect this condition and do something about it.

library(NRES803data)
RIKZ_model.2 = lm(Richness ~ week, data = RIKZ)
check_assumptions(RIKZ_model.2)

but doesn't happen all the time:

RIKZ$fweek <- factor(RIKZ$week)
RIKZ_model.3 <- lm(Richness ~ fweek, data = RIKZ)
check_assumptions(RIKZ_model.3)

I'd like to rewrite the check_assumptions function for gams using the evaluation plots in Zuur et al 2007 Analysing ecological data on p.118, chapter 7 Additive and generalised additive modelling.

Anyone interested in this/thinks its a good idea?

https://cran.r-project.org/package=gratia

but doesn't use the same sort of API as broom. Does some of the same things but in a different way.

fix with usethis::use_package()