1  Do smokers live longer?

TipYou can run R in the browser in this chapter

Code cells can be edited and re-run by pressing “Run Code” (R is running inside your browser). The first time, loading the packages takes several tens of seconds.

First, let us experience what happens when you skip descriptive statistics and “jump straight into the analysis”.

1.1 Data: 1314 women in Whickham

In 1972–74, the smoking habits of women were surveyed in the town of Whickham, UK. Twenty years later, the women were followed up to see whether they were still alive.

  • smoker: whether she smoked at the time of the survey (Yes / No)
  • age: age at the time of the survey
  • outcome: alive (Alive) or dead (Dead) 20 years later
NoteMake a prediction

Between smokers and non-smokers, which group had a higher proportion of women who had died 20 years later?

1.2 First, a naive comparison

The result is that the proportion who died is lower among smokers (23.9%) than among non-smokers (31.4%).

“Just to be sure”, let us also run a test with logistic regression.

Odds ratio 0.68 (95% confidence interval 0.54–0.88), p = 0.003. Statistically significantly, smokers are less likely to die …

Can we report this result as “smoking lowers the risk of death”?

1.3 What we should have done before the analysis: look at the characteristics of the two groups

We should first have looked at how the two groups being compared differ in ways other than smoking. The only other variable we have in this dataset is age.

The mean age differs by only 4 years, but the proportion aged 65 or over differs greatly. Let us look at the distribution in a figure.

For each age group, let us line up the proportion of smokers and the proportion who died within 20 years.

  • Only 20.6% of those aged 65–84 were smokers (women of this generation smoked little)
  • 86% of those aged 65–84 had died 20 years later (naturally, given their age)

In other words, the “non-smoker” group contained many older women who were likely to die within 20 years.

1.4 Compare at the same age

Within the same age group, compare the proportion who died between smokers and non-smokers.

In every age group, the proportion who died is higher among smokers (or about the same). This is the opposite conclusion from the overall comparison.

The phenomenon in which an association seen overall reverses when the data are split into strata is called Simpson’s paradox. Here, age is a confounder that is related both to smoking (older women did not smoke) and to death (older women are more likely to die).

When age is added to the model, the odds ratio exceeds 1.

Warning“Just adjust for age” is not the solution

This time we happened to notice the confounder age because it was at hand. However, you only find out what to adjust for after looking at the differences between the two groups with descriptive statistics. If you run only the model without making a table that lines up the characteristics of the two groups (Table 1), you will not notice that you have missed variables that should be adjusted for either (Chapter 7).

1.5 What did we miss?

The first analysis was not wrong at all as a calculation. What was wrong was the interpretation. The cause was that, before comparing, we did not describe “what kind of people are in each group”.

This “misinterpretation from not describing” pattern is not limited to confounding. From the next chapter on, using real epidemiological data, we will experience in turn:

Chapter What happens if you don’t describe
2 Counting the same person many times; averaging code values as numbers
3 Missing skewed distributions, anomalous values, or rounding, so summary statistics diverge from reality
4 After dropping missing values, you were analysing a different population from the original
5 Assuming that proportions in the sample are the proportions in the population
6 Drawing conclusions only from the “healthy people” who remained in follow-up
7 Comparing groups with different backgrounds; being swayed by p-values with huge N
ImportantSummary of this chapter
  • A crude (overall) comparison alone mixes in differences in background between groups
  • Before comparing, describe the characteristics of each group (age, sex, etc.)
  • If descriptive statistics reveal differences in background, you can see how to deal with them (stratification, adjustment, etc.) and the limits of interpretation