Regression Analysis in Public Health

Regression analysis is the statistical backbone of modern epidemiological research. It allows you to quantify the relationship between exposures—like air pollution or a new drug—and health outcomes, while untangling the web of other influencing factors. Mastering these techniques is essential for moving from observing associations to understanding potential causes and informing evidence-based policy and clinical decisions.

The Foundation: What is Regression Modeling?

At its core, regression analysis is a set of statistical methods used to model and analyze the relationships between a dependent variable and one or more independent variables. In public health, the dependent variable is typically a health outcome (e.g., blood pressure, disease status, number of hospital visits), while the independent variables include exposures of interest (e.g., smoking status, treatment group) and confounders (e.g., age, socioeconomic status) that you need to control for. The primary goal is to estimate the independent effect of an exposure on an outcome, holding other factors constant. This process of statistical adjustment is your primary tool for mitigating confounding, which occurs when an extraneous variable distorts the true relationship between exposure and outcome.

Types of Regression Models

Linear Regression: Modeling Continuous Outcomes

Linear regression is used when your health outcome is a continuous variable, such as birth weight, cholesterol level, or life expectancy. It models the relationship as a straight line, defined by the equation:

$$Y={\beta }_0+{\beta }_1{X}_1+{\beta }_2{X}_2+...+{\beta }_k{X}_k+ϵ$$

Here, $Y$ is the outcome, ${\beta }_0$ is the intercept, ${\beta }_1$ is the coefficient for the primary exposure $X_1$, and ${\beta }_2$ through ${\beta }_k$ are coefficients for other controlled variables. The term $ϵ$ represents the random error. The interpretation is straightforward: for a one-unit increase in $X_1$, the outcome $Y$ changes by ${\beta }_1$ units, assuming all other variables in the model are held constant. For example, in a model for systolic blood pressure (SBP), a coefficient of ${\beta }_1=2.5$ for "years smoking" would mean that for each additional year of smoking, SBP increases by 2.5 mmHg, after adjusting for age, diet, and other covariates.

Logistic Regression: Modeling Binary Outcomes and Odds Ratios

When your outcome is binary (e.g., disease present/absent, alive/dead), logistic regression is the appropriate tool. Instead of modeling the outcome directly, it models the log-odds of the outcome occurring. The log-odds is transformed via the logistic function to produce a probability between 0 and 1. The coefficients from a logistic regression model are reported as odds ratios (ORs), which are the exponentiated values of the coefficients ($e^{\beta }$).

An odds ratio compares the odds of the outcome in an exposed group to the odds in an unexposed group. An OR of 1.0 means no association. An OR of 2.0 means the odds of the outcome are twice as high in the exposed group. For instance, if a study on lung cancer yields an OR of 3.5 for "smoker vs. non-smoker," it means smokers have 3.5 times the odds of developing lung cancer compared to non-smokers, after adjusting for confounders like age and asbestos exposure. It is critical to remember that an odds ratio is not the same as a relative risk, especially when the outcome is common.

Poisson Regression: Analyzing Count and Rate Data

Poisson regression is designed for outcome variables that are counts (e.g., number of asthma attacks in a year, number of injuries in a factory) or rates (e.g., disease incidence per 100,000 person-years). It assumes the outcome follows a Poisson distribution, where the mean equals the variance. The model uses a log link function, meaning it models the log of the expected count or rate.

The coefficients are interpreted as incidence rate ratios (IRRs) after exponentiation. An IRR tells you how many times higher the rate of the event is for a one-unit change in the predictor. For example, in a study of emergency room visits for asthma, an IRR of 1.2 for a 10 ppb increase in ozone pollution means the visit rate is 20% higher for each 10 ppb increase. Poisson regression is fundamental in analyzing count data from cohort studies and surveillance systems.

The Process: Model Building and Diagnostics

Selecting variables and checking your model's validity are essential epidemiological analysis skills. Model building often involves a thoughtful mix of including known confounders (based on subject-matter knowledge) and using statistical criteria (like AIC or p-values) for other variables. A common strategy is to start with a simple model containing the main exposure, then add potential confounders one group at a time, observing how the exposure coefficient changes.

After building a model, you must perform diagnostics. For linear regression, this includes checking for linearity (the relationship is truly a straight line), homoscedasticity (constant variance of errors), normality of residuals, and identifying influential data points. For logistic and Poisson regression, you check for overdispersion (where variance exceeds the mean, often requiring a negative binomial model) and assess model fit with tests like the Hosmer-Lemeshow test. Ignoring diagnostics can lead to biased, inefficient, or completely misleading results.

Interpretation in Context

The final step is translating statistical output into meaningful public health conclusions. This goes beyond stating a p-value or an OR. You must interpret the magnitude and precision of the estimate (using its confidence interval), its clinical or public health significance, and the limitations of your model. For example, a highly precise OR of 1.05 for a medication side effect might be statistically significant but clinically negligible. Conversely, a wide confidence interval (e.g., OR = 3.0, 95% CI: 0.9–10.0) indicates too much uncertainty for a firm conclusion, despite a large point estimate. Always interpret your findings within the biological and social context of the health question.

Common Pitfalls

1. Misinterpreting Correlation for Causation: A significant regression coefficient, even after adjusting for many variables, does not prove causation. It indicates an association. Causal inference requires careful study design (like randomized trials) and consideration of biases like residual confounding or reverse causality.
2. Ignoring Model Assumptions: Fitting a linear regression to binary data or using Poisson regression on overdispersed count data violates core assumptions and produces invalid results. Always choose your model based on the outcome variable's structure and perform diagnostic checks.
3. Overfitting or Underfitting the Model: Overfitting occurs when you include too many variables, especially relative to your sample size, making the model fit your specific sample perfectly but perform poorly on new data. Underfitting happens when you omit important confounders, leaving residual confounding and biased estimates of your exposure's effect.
4. Confusing Odds Ratios and Risk Ratios: In case-control studies or when the outcome is common (exceeding 10%), the odds ratio from logistic regression does not approximate the relative risk. Presenting an OR as if it were a risk ratio can dramatically overstate the effect.

Summary

• Regression analysis is the key statistical method for modeling relationships between exposures and health outcomes while controlling for confounders to approximate causal effects.
• Linear regression models continuous outcomes, logistic regression models binary outcomes and produces odds ratios, and Poisson regression is used for analyzing count data and rates, producing incidence rate ratios.
• Proper model building is a strategic process that combines subject-matter knowledge with statistical guidance to specify which variables to include.
• Conducting diagnostics is non-negotiable; you must verify that your data meet the assumptions of your chosen regression model to ensure the results are valid.
• Interpretation requires assessing both statistical significance (p-values, confidence intervals) and public health significance (magnitude, context), avoiding the trap of mistaking association for causation.