Case study: asphalt

The asphalt data

• 31 asphalt pavements prepared under different conditions. How does quality of pavement depend on these?
• Variables:
  • pct.a.surf Percentage of asphalt in surface layer
  • pct.a.base Percentage of asphalt in base layer
  • fines Percentage of fines in surface layer
  • voids Percentage of voids in surface layer
  • rut.depth Change in rut depth per million vehicle passes
  • viscosity Viscosity of asphalt
  • run 2 data collection periods: 1 for run 1, 0 for run 2.
• rut.depth response. Depends on other variables, how?

Packages for this section

library(MASS, exclude = "select")
library(tidyverse)
library(broom)
library(leaps)

Make sure to load MASS before tidyverse (for annoying technical reasons), or to load MASS excluding its select (as above).

Getting set up

my_url <- "http://ritsokiguess.site/datafiles/asphalt.txt"
asphalt <- read_delim(my_url, " ")

• Quantitative variables with one response: multiple regression.
• Some issues here that don’t come up in “simple” regression; handle as we go. (STAB27/STAC67 ideas.)

The data (some)

asphalt

Plotting response “rut depth” against everything else

Same idea as for plotting separate predictions on one plot:

asphalt %>%
  pivot_longer(
    -rut.depth,
    names_to="xname", values_to="x"
  ) %>%
  ggplot(aes(x = x, y = rut.depth)) + geom_point() +
  facet_wrap(~xname, scales = "free") -> g

“collect all the x-variables together into one column called x, with another column xname saying which x they were, then plot these x’s against rut.depth, a separate facet for each x-variable.”

I saved this graph to plot later (on the next page).

The plot

g

Interpreting the plots

• One plot of rut depth against each of the six other variables.
• Get rough idea of what’s going on.
• Trends mostly weak.
• viscosity has strong but non-linear trend.
• run has effect but variability bigger when run is 1.
• Weak but downward trend for voids.
• Non-linearity of rut.depth-viscosity relationship should concern us.

Log of viscosity: more nearly linear?

• Take this back to asphalt engineer: suggests log of viscosity:

ggplot(asphalt, aes(y = rut.depth, x = log(viscosity))) +
  geom_point() + geom_smooth(se = FALSE) -> g

(plot overleaf)

Rut depth against log-viscosity

g

Comments and next steps

• Not very linear, but better than before.
• In multiple regression, hard to guess which x’s affect response. So typically start by predicting from everything else.
• Model formula has response on left, squiggle, explanatories on right joined by plusses:

rut.1 <- lm(rut.depth ~ pct.a.surf + pct.a.base + fines +
  voids + log(viscosity) + run, data = asphalt)

Regression output:

summary(rut.1)

Call:
lm(formula = rut.depth ~ pct.a.surf + pct.a.base + fines + voids + 
    log(viscosity) + run, data = asphalt)

Residuals:
    Min      1Q  Median      3Q     Max 
-4.1211 -1.9075 -0.7175  1.6382  9.5947 

Coefficients:
               Estimate Std. Error t value Pr(>|t|)   
(Intercept)    -12.9937    26.2188  -0.496   0.6247   
pct.a.surf       3.9706     2.4966   1.590   0.1248   
pct.a.base       1.2631     3.9703   0.318   0.7531   
fines            0.1164     1.0124   0.115   0.9094   
voids            0.5893     1.3244   0.445   0.6604   
log(viscosity)  -3.1515     0.9194  -3.428   0.0022 **
run             -1.9655     3.6472  -0.539   0.5949   
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 3.324 on 24 degrees of freedom
Multiple R-squared:  0.806, Adjusted R-squared:  0.7575 
F-statistic: 16.62 on 6 and 24 DF,  p-value: 1.743e-07

Comments

• R-squared 81%, not so bad.

• P-value in glance asserts that something helping to predict rut.depth.

• Table of coefficients says log(viscosity).

• But confused by clearly non-significant variables: remove those to get clearer picture of what is helpful.

Plot fitted values against residuals

ggplot(rut.1, aes(x = .fitted, y = .resid)) + geom_point()

Normal quantile plot of residuals

ggplot(rut.1, aes(sample = .resid)) + stat_qq() + 
  stat_qq_line()

Plotting residuals against \(x\) variables

• Problem here is that residuals are in the fitted model, and the observed \(x\)-values are in the original data frame asphalt.
• Package broom contains a function augment that combines these two together so that they can later be plotted: start with a model first, and then augment with a data frame:

rut.1 %>% augment(asphalt) -> rut.1a
rut.1a

What does rut.1a contain?

names(rut.1a)

 [1] "pct.a.surf" "pct.a.base" "fines"      "voids"      "rut.depth" 
 [6] "viscosity"  "run"        ".fitted"    ".resid"     ".hat"      
[11] ".sigma"     ".cooksd"    ".std.resid"

• all the stuff in original data frame, plus:
• quantities from regression (starting with a dot)

Plotting residuals against \(x\)-variables

rut.1a %>%
  mutate(log_vis=log(viscosity)) %>% 
  pivot_longer(
    c(pct.a.surf:voids, run, log_vis),
    names_to="xname", values_to="x"
  ) %>%
  ggplot(aes(x = x, y = .resid)) +
  geom_point() + facet_wrap(~xname, scales = "free") -> g

The plot

g

Comments

• There is serious curve in plot of residuals vs. fitted values. Suggests a transformation of \(y\).
• The residuals-vs-\(x\)’s plots don’t show any serious trends. Worst probably that potential curve against log-viscosity.
• Also, large positive residual, 10, that shows up on all plots. Perhaps transformation of \(y\) will help with this too.
• If residual-fitted plot OK, but some residual-\(x\) plots not, try transforming those \(x\)’s, eg. by adding \(x^2\) to help with curve.

Which transformation?

• Best way: consult with person who brought you the data.
• Can’t do that here!
• No idea what transformation would be good.
• Let data choose: “Box-Cox transformation”.
• Scale is that of “ladder of powers”: power transformation, but 0 is log.

Running Box-Cox

From package MASS:

boxcox(rut.depth ~ pct.a.surf + pct.a.base + fines + voids +
  log(viscosity) + run, data = asphalt)

Comments on Box-Cox plot

• \(\lambda\) represents power to transform \(y\) with.
• Best single choice of transformation parameter \(\lambda\) is peak of curve, close to 0.
• Vertical dotted lines give CI for \(\lambda\), about (−0.05, 0.2).
• \(\lambda = 0\) means “log”.
• Narrowness of confidence interval mean that these not supported by data:
  • No transformation (\(\lambda = 1\))
  • Square root (\(\lambda = 0.5\))
  • Reciprocal (\(\lambda = −1\)).

Relationships with explanatories

• As before: plot response (now log(rut.depth)) against other explanatory variables, all in one shot:

asphalt %>%
  mutate(log_vis=log(viscosity)) %>% 
  pivot_longer(
    c(pct.a.surf:voids, run, log_vis),
    names_to="xname", values_to="x"
  ) %>%
  ggplot(aes(y = log(rut.depth), x = x)) + geom_point() +
  facet_wrap(~xname, scales = "free") -> g3

The new plots

g3

Modelling with transformed response

• These trends look pretty straight, especially with log.viscosity.
• Values of log.rut.depth for each run have same spread.
• Other trends weak, but are straight if they exist.
• Start modelling from the beginning again.
• Model log.rut.depth in terms of everything else, see what can be removed:

rut.2 <- lm(log(rut.depth) ~ pct.a.surf + pct.a.base +
  fines + voids + log(viscosity) + run, data = asphalt)

• use tidy from broom to display just the coefficients.

Output

tidy(rut.2)

Taking out everything non-significant

• Try: remove everything but pct.a.surf and log.viscosity:

rut.3 <- lm(log(rut.depth) ~ pct.a.surf + log(viscosity), data = asphalt)
tidy(rut.3)

Check that removing all those variables wasn’t too much

anova(rut.3, rut.2)

• \(H_0\) : two models equally good; \(H_a\) : bigger model better.
• Null not rejected here; small model as good as the big one, so prefer simpler smaller model rut.3.

Find the largest P-value by eye:

tidy(rut.2)

• Largest P-value is 0.78 for pct.a.base, not significant.
• So remove this first, re-fit and re-assess.
• Or, as over.

Get the computer to find the largest P-value for you

• Output from tidy is itself a data frame, thus:

tidy(rut.2) %>% arrange(p.value)

• Largest P-value at the bottom.

Take out pct.a.base

• Copy and paste the lm code and remove what you’re removing:

rut.4 <- lm(log(rut.depth) ~ pct.a.surf + fines + voids + 
              log(viscosity) + run, data = asphalt)
tidy(rut.4) %>% arrange(p.value) %>% select(term, p.value)

• fines is next to go, P-value 0.32.

“Update”

Another way to do the same thing:

rut.4 <- update(rut.2, . ~ . - pct.a.base)
tidy(rut.4) %>% arrange(p.value)

• Again, fines is the one to go. (Output identical as it should be.)

Take out fines:

rut.5 <- update(rut.4, . ~ . - fines)
tidy(rut.5) %>% arrange(p.value) %>% select(term, p.value)

Can’t take out intercept, so run, with P-value 0.36, goes next.

Take out run:

rut.6 <- update(rut.5, . ~ . - run)
tidy(rut.6) %>% arrange(p.value) %>% select(term, p.value)

Again, can’t take out intercept, so largest P-value is for voids, 0.044. But this is significant, so we shouldn’t remove voids.

Comments

• Here we stop: pct.a.surf, voids and log.viscosity would all make fit significantly worse if removed. So they stay.
• Different final result from taking things out one at a time (top), than by taking out 4 at once (bottom):

coef(rut.6)

   (Intercept)     pct.a.surf          voids log(viscosity) 
    -1.0207945      0.5554686      0.2447934     -0.6464911 

coef(rut.3)

   (Intercept)     pct.a.surf log(viscosity) 
     0.9001389      0.3911481     -0.6185628 

• Point: Can make difference which way we go.

Comments on variable selection

• Best way to decide which \(x\)’s belong: expert knowledge: which of them should be important.
• Best automatic method: what we did, “backward selection”.
• Do not learn about “stepwise regression”! eg. here
• R has function step that does backward selection, like this:

step(rut.2, direction = "backward", test = "F")

Gets same answer as we did (by removing least significant x).

• Removing non-significant \(x\)’s may remove interesting ones whose P-values happened not to reach 0.05. Consider using less stringent cutoff like 0.20 or even bigger.
• Can also fit all possible regressions, as over (may need to do install.packages("leaps") first).

All possible regressions (output over)

Uses package leaps:

leaps <- regsubsets(log(rut.depth) ~ pct.a.surf + 
                      pct.a.base + fines + voids + 
                      log(viscosity) + run, 
                    data = asphalt, nbest = 2)
s <- summary(leaps)
with(s, data.frame(rsq, outmat)) -> d

The output

d %>% rownames_to_column("model") %>% arrange(desc(rsq))

Comments

• Problem: even adding a worthless x increases R-squared. So try for line where R-squared stops increasing “too much”, eg. top line (just log.viscosity), first 3-variable line (backwards-elimination model). Hard to judge.
• One solution (STAC67): adjusted R-squared, where adding worthless variable makes it go down.
• data.frame rather than tibble because there are several columns in outmat.

All possible regressions, adjusted R-squared

with(s, data.frame(adjr2, outmat)) %>% 
  rownames_to_column("model") %>% 
  arrange(desc(adjr2))

Revisiting the best model

• Best model was our rut.6:

tidy(rut.6)

Revisiting (2)

• Regression slopes say that rut depth increases as log-viscosity decreases, pct.a.surf increases and voids increases. This more or less checks out with out scatterplots against log.viscosity.
• We should check residual plots again, though previous scatterplots say it’s unlikely that there will be a problem:

g <- ggplot(rut.6, aes(y = .resid, x = .fitted)) + 
geom_point()

Residuals against fitted values

g

Normal quantile plot of residuals

ggplot(rut.6, aes(sample = .resid)) + stat_qq() + stat_qq_line()

Plotting residuals against x’s

• Do our trick again to put them all on one plot:

augment(rut.6, asphalt) %>%
  mutate(log_vis=log(viscosity)) %>% 
  pivot_longer(
    c(pct.a.surf:voids, run, log_vis),
    names_to="xname", values_to="x",
  ) %>%
  ggplot(aes(y = .resid, x = x)) + geom_point() +
  facet_wrap(~xname, scales = "free") -> g2

Residuals against the x’s

g2

Comments

• None of the plots show any sort of pattern. The points all look random on each plot.
• On the plot of fitted values (and on the one of log.viscosity), the points seem to form a “left half” and a “right half” with a gap in the middle. This is not a concern.
• One of the pct.a.surf values is low outlier (4), shows up top left of that plot.
• Only two possible values of run; the points in each group look randomly scattered around 0, with equal spreads.
• Residuals seem to go above zero further than below, suggesting a mild non-normality, but not enough to be a problem.

Variable-selection strategies

• Expert knowledge.
• Backward elimination.
• All possible regressions.
• Taking a variety of models to experts and asking their opinion.
• Use a looser cutoff to eliminate variables in backward elimination (eg. only if P-value greater than 0.20).
• If goal is prediction, eliminating worthless variables less important.
• If goal is understanding, want to eliminate worthless variables where possible.
• Results of variable selection not always reproducible, so caution advised.