Doll and Hill’s 1950 article studying the association between smoking and lung cancer contains one of the most important 2 x 2 tables in history.
Here’s their data:
smoking <- as.table(rbind(c(688, 650), c(21, 59)))
dimnames(smoking) <- list(has.smoked = c("yes", "no"),
lung.cancer = c("yes","no"))
smoking
## lung.cancer
## has.smoked yes no
## yes 688 650
## no 21 59
(a) Use fisher.test()
to test if there’s an association between smoking and lung cancer.
smoking.fisher.test <- fisher.test(smoking)
smoking.fisher.test
##
## Fisher's Exact Test for Count Data
##
## data: smoking
## p-value = 1.476e-05
## alternative hypothesis: true odds ratio is not equal to 1
## 95 percent confidence interval:
## 1.755611 5.210711
## sample estimates:
## odds ratio
## 2.971634
(b) What is the odds ratio? Interpret this quantity.
smoking.fisher.test$estimate
## odds ratio
## 2.971634
This says that the odds of having lung cancer are 2.97 times higher among smokers than non-smokers.
(c) Are your findings statistically significant?
smoking.fisher.test$p.value
## [1] 1.476303e-05
The findings are highly statistically significant.
Using Doll and Hill’s smoking data and, construct a bar graph with accompanying error bars showing the proportion of study participants with lung cancer in each smoking status group.
To succeed in this exercise, you’ll want to follow along careful with the lecture notes. Please read the section titled “Plotting the table values with confidence”.
# Reshape the count table into a tibble:
smoking.tbl <- as_tibble(smoking)
# Add a column showing total count in each (smoking status) group
smoking.tbl <- smoking.tbl %>%
group_by(has.smoked) %>%
mutate(total = sum(n)) %>%
filter(lung.cancer == "yes") # Retain only the lung.cancer = yes rows
smoking.toplot <- smoking.tbl %>%
group_by(has.smoked) %>%
summarize(prop = n / total,
lower = prop.test(n, total)$conf.int[1],
upper = prop.test(n, total)$conf.int[2])
# Here's our summary table
smoking.toplot
## # A tibble: 2 x 4
## has.smoked prop lower upper
## <chr> <dbl> <dbl> <dbl>
## 1 no 0.262 0.173 0.375
## 2 yes 0.514 0.487 0.541
# Plotting routine
smoking.toplot %>%
ggplot(aes(x = has.smoked, fill = has.smoked, y = prop)) +
geom_bar(position="dodge", stat="identity") +
geom_errorbar(aes(ymin=lower, ymax=upper),
width=.2, # Width of the error bars
position=position_dodge(0.9)) +
ylab("Proportion with lung cancer")
Let’s form our favourite birthwt data set.
# Import data, rename variables, and recode factors all in one set of piped
# commands
birthwt <- as_tibble(MASS::birthwt) %>%
rename(birthwt.below.2500 = low,
mother.age = age,
mother.weight = lwt,
mother.smokes = smoke,
previous.prem.labor = ptl,
hypertension = ht,
uterine.irr = ui,
physician.visits = ftv,
birthwt.grams = bwt) %>%
mutate(race = recode_factor(race, `1` = "white", `2` = "black", `3` = "other")) %>%
mutate_at(c("mother.smokes", "hypertension", "uterine.irr", "birthwt.below.2500"),
~ recode_factor(.x, `0` = "no", `1` = "yes"))
(a) Create a new factor that categorizes the number of physician visits into four levels: 0, 1, 2, 3 or more.
birthwt <- birthwt %>%
mutate(physician.visits.binned = recode_factor(physician.visits,
`0` = "0",
`1` = "1",
`2` = "2",
.default = "3 or more"
))
Hint: One way of doing this is with recode
by specifying .default = "3 or more"
. Have a look at the help file for recode
to learn more.
(b) Run an ANOVA to determine whether the average birth weight varies across number of physician visits. Interpret the results.
summary(aov(birthwt.grams ~ physician.visits.binned, data = birthwt))
## Df Sum Sq Mean Sq F value Pr(>F)
## physician.visits.binned 3 2259057 753019 1.426 0.237
## Residuals 185 97710599 528165
We find that there is no statistically significant variation in average birthweigh across different levels of the number of first trimester physician visits.
Below is figure showing how Price varies with EngineSize in the Cars93, with accompanying regression lines. There are two plots, one for USA cars, and one for non-USA cars.
qplot(data = Cars93, x = EngineSize, y = Price, colour = Origin) +
facet_wrap("Origin") +
stat_smooth(method = "lm") +
theme(legend.position="none")
(a) Use the lm()
function to regress Price on EngineSize and Origin
cars.lm <- lm(Price ~ EngineSize + Origin, data = Cars93)
summary(cars.lm)
##
## Call:
## lm(formula = Price ~ EngineSize + Origin, data = Cars93)
##
## Residuals:
## Min 1Q Median 3Q Max
## -13.474 -4.362 -1.051 2.743 34.626
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -3.0892 2.5419 -1.215 0.227
## EngineSize 7.0637 0.7617 9.274 9.24e-15 ***
## Originnon-USA 7.7596 1.5725 4.935 3.66e-06 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 6.948 on 90 degrees of freedom
## Multiple R-squared: 0.4939, Adjusted R-squared: 0.4826
## F-statistic: 43.91 on 2 and 90 DF, p-value: 4.925e-14
(b) Run plot()
on your lm
object. Do you see any problems?
par(mfrow = c(2,2))
plot(cars.lm)
The residual plot shows a clear sign of non-constant variance. (The plot looks like a funnel, with variance increasing with fitted value.) One can also see this from the upward slope evidence from the the scale-location plot.
(c) Try running a linear regression with log(Price)
as your outcome.
cars.lm.log <- lm(log(Price) ~ EngineSize + Origin, data = Cars93)
summary(cars.lm.log)
##
## Call:
## lm(formula = log(Price) ~ EngineSize + Origin, data = Cars93)
##
## Residuals:
## Min 1Q Median 3Q Max
## -0.62164 -0.18968 -0.01582 0.18488 0.93083
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.74712 0.11122 15.708 <2e-16 ***
## EngineSize 0.35790 0.03333 10.739 <2e-16 ***
## Originnon-USA 0.33803 0.06881 4.913 4e-06 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.304 on 90 degrees of freedom
## Multiple R-squared: 0.5627, Adjusted R-squared: 0.553
## F-statistic: 57.9 on 2 and 90 DF, p-value: < 2.2e-16
(d) Run plot()
on your new lm
object. Do you see any problems?
par(mfrow = c(2,2))
plot(cars.lm.log)
The variance now looks pretty constant across the range of fitted values, and there don’t appear to be any clear trends in the plots. All of the diagnostic plots seem pretty good. It looks like the log transformation helped.