Pick one adult at random from the 5,107 adults examined in the National Health and Nutrition Examination Survey (NHANES) of 2017 to 2018 who answered the survey's questions on diabetes and smoking. The chance that this person has hypertension is 0.545. The chance that the person is obese is 0.411, and the chance that a doctor has told the person they have diabetes is 0.152. Every question in this unit is a question about such chances:
- How likely is it that the person has at least one of the three conditions?
- Does learning that the person is obese change the chance of diabetes, and by how much?
- Is diabetes more common among former smokers because they smoked, or because they are older?
- When a study can only collect people who already have a disease, what can it still say about the risk an exposure carries?
Probability is the set of rules for reasoning about uncertain events. Every other rule follows from three short statements, the axioms. This unit sets out the rules, uses them to compare the risk of disease between groups, and checks each one by counting real people.
The unit runs in three steps:
- First, the axioms and the rules proved from them, and the property called independence, a different idea from mutual exclusivity, with which it is often confused.
- Then, conditional probability, the probability of one event among the cases where another has occurred: how a probability changes when we learn something. Three results follow from it, and one of them explains how an association can appear, shrink, vanish or reverse when a population is split into groups.
- Finally, the two measures epidemiology uses to compare risks, and why one of them is the measure a case-control study can estimate.
The examples use one population throughout: the nonpregnant adults examined in NHANES 2017 to 2018 with blood pressure, body mass index, a diabetes answer and a smoking answer on record. The same adults appear in Inference on Proportions. Their conditions follow the definitions of the National Center for Health Statistics:
- hypertension: a mean systolic pressure of at least 130 mm Hg, a mean diastolic pressure of at least 80 mm Hg, or current use of blood pressure medication;
- obesity: a body mass index of 30 or more;
- diabetes: a doctor's diagnosis, with answers of "borderline" counted as no;
- smoking: adults who have smoked at least 100 cigarettes in their lives are current smokers if they now smoke on every day or some days, and former smokers if they do not smoke at all; the others never smoked.
We treat these adults as a finite population. Then every probability can be computed exactly by counting, and each rule can be checked against the data rather than illustrated with invented numbers. Every number comes from a script in the site's repository.
Articles
- The Rules of Probability: why can we not just add the three chances, and what replaces addition? Sample spaces, events and their operations. The classical and relative-frequency views of probability, and Kolmogorov's axioms. The complement rule, monotonicity, the addition rule and inclusion–exclusion for three and for $n$ events, each proved from the axioms and checked on hypertension, obesity and diabetes. Boole's inequality and its use in the Bonferroni correction. Independence of two and of several events, and why mutually exclusive events with positive probability can never be independent.
- Conditional Probability and the Law of Total Probability: how much does learning one fact change the chance of another? The definition, and why a conditional probability obeys the axioms. The difference between $P(D \mid O)$ and $P(O \mid D)$. The multiplication and chain rules, with draws without replacement, and independence restated. The law of total probability as a tree over age groups, and Bayes' theorem for reversing a conditional probability. The crude rate and direct standardization as total probability with different weights. Confounding by age and Simpson's paradox, worked on diabetes among former and never smokers.
- Relative Risk and the Odds Ratio: how should two risks be compared, and what can a study that starts from the cases estimate? Risk, risk difference, relative risk, odds and the odds ratio from a $2 \times 2$ table. The identity linking the two ratios, and why the odds ratio approximates the relative risk only for rare outcomes. Cohort, cross-sectional and case-control designs, with a proof that a case-control study estimates the odds ratio and a simulation showing that it cannot estimate the relative risk. The Katz and Woolf intervals derived on the log scale, with their exact coverage. Adjustment for a confounder with the Mantel–Haenszel estimators.
Related topics
This unit builds on the previous one and prepares several later ones.
- The previous unit, Rates and Standardization, adjusts rates for age. Its direct standardization is the law of total probability with a chosen set of weights.
- The next unit, Screening and Diagnostic Tests, applies conditional probability and Bayes' theorem to the accuracy of medical tests.
- Theoretical Probability Distributions builds random variables and their distributions on the rules developed here.
- The measures of this unit are proportions. The inference for them, including the test that two proportions are equal and the chi-square test of a $2 \times 2$ table, is in Comparing Two Proportions.
- The information retrieval wiki's Probability for Information Retrieval covers the same rules for ranking documents. The library's Logistic Regression estimates odds ratios that adjust for several variables at once.
References
- Breslow, N. E., and Day, N. E. (1980). Statistical Methods in Cancer Research, Volume I: The Analysis of Case-Control Studies. IARC Scientific Publications No. 32. International Agency for Research on Cancer. record · 3 pages
- Centers for Disease Control and Prevention (2006). Principles of Epidemiology in Public Health Practice: An Introduction to Applied Epidemiology and Biostatistics (3rd ed., updated 2011), Lesson 3: Measures of Risk. SS1978. Office of Workforce and Career Development, CDC. https://archive.cdc.gov/www_cdc_gov/csels/dsepd/ss1978/index.html record · 2 pages
- Grinstead, C. M., and Snell, J. L. (1997). Introduction to Probability (2nd rev. ed.). American Mathematical Society. record · 10 pages
- Kolmogorov, A. N. (1956). Foundations of the Theory of Probability (2nd English ed., N. Morrison, Trans.). Chelsea. Originally published in German in 1933. record · 2 pages
- Pagano, M., Gauvreau, K., and Mattie, H. (2022). Principles of Biostatistics (3rd ed.). CRC Press. The course text. record · 35 pages