3 Reading odds ratios: risk ratios, standardization, and non-collapsibility
When you adjust with logistic regression, the result comes out as an odds ratio. In Chapter 1, the odds ratio for smoking adjusted for age and sex was 1.63. Does this mean “smokers have 1.63 times the risk of death”?
3.1 An odds ratio is not a risk ratio
Let us check the difference between risk (a proportion) and odds.
- Risk = number who died / everyone
- Odds = number who died / number who did not die = risk / (1 − risk)
When the risk is small (below about 10%), odds and risk are nearly the same. But when the risk is large, the odds become much larger than the risk. Therefore, for common outcomes (24-year death in this data is 35%), the odds ratio is further from 1 than the risk ratio.
3.2 Standardization: obtaining adjusted risk ratios and risk differences
Standardization (g-computation) is a method for obtaining risk ratios and risk differences from an adjusted model. The idea is simple.
- Build the adjusted model
- Predict each person’s probability of death as if everyone were a smoker, and average
- Predict and average in the same way as if everyone were a non-smoker
- Take the ratio and the difference of the two averages
The results adjusted for age and sex are:
- Odds ratio: 1.63
- Risk ratio: 1.28 (1.28 times the 24-year risk of death)
- Risk difference: 8.6 percentage points (about 9 more deaths per 100 people)
“1.63 times” and “1.28 times” leave quite different impressions.
3.2.1 Confidence intervals by the bootstrap
The confidence intervals of the standardized risk ratio and risk difference are easiest to obtain by the bootstrap (resample the same number of people from the data with replacement and repeat the whole calculation). It takes a little time.
The risk difference is 8.6 percentage points (95% CI 6.1–11.1) and the risk ratio is 1.28 (1.19–1.37). In a real analysis, increase the number of bootstrap replicates to around 1000–2000.
3.2.2 If the outcome is rare, odds ratio ≈ risk ratio
For 24-year stroke (about 9%), the difference between the odds ratio and the risk ratio becomes small.
3.3 Non-collapsibility: odds ratios change with adjustment even without confounding
Odds ratios have one more counter-intuitive property. Even with no confounding at all, the odds ratio changes just by adding a variable that strongly affects the outcome.
We simulate a randomized controlled trial. Treatment X is randomly assigned, so it is unrelated to severity Z (no confounding).
Although there is no confounding, the unadjusted odds ratio is 2.06, and adjusting for Z gives 2.70 (the true stratum-specific odds ratio is 2.72). This is a mathematical property of odds ratios (and hazard ratios) called non-collapsibility, and it is not a bias. The marginal (overall) odds ratio and the conditional (stratum-specific) odds ratio measure different things.
“The odds ratio changed before and after adjustment, so there is confounding” is not necessarily true. Odds ratios change even without confounding. (Standardized) risk differences and risk ratios do not have this property, so for questions like “how much would death decrease in the whole population if people stopped smoking?”, standardized risk differences and risk ratios are better suited.
- An odds ratio is not a risk ratio. When the outcome is common (10% or more), the odds ratio is further from 1
- Standardization (predict with everyone exposed and everyone unexposed, then average) gives adjusted risk ratios and risk differences. Use the bootstrap for confidence intervals
- Odds ratios are non-collapsible: their value changes with adjustment even without confounding
- In reports, showing not only the odds ratio but also the absolute difference (risk difference) conveys the size of the effect