Regression analysis is a statistical method researchers use to measure the relationship between variables, specifically how one or more factors predict or explain changes in an outcome. It’s one of the most widely used tools in scientific research, applied everywhere from clinical medicine to economics to psychology. At its core, regression fits a mathematical equation to observed data, allowing researchers to quantify how strongly a predictor influences a result and to forecast outcomes for new observations.
How Regression Works
Every regression model has two types of variables. The dependent variable (also called the outcome variable) is the thing you’re trying to explain or predict. The independent variables (also called predictors) are the factors you think influence it. In a study on smoking and lung cancer, for instance, lung cancer diagnosis is the outcome variable, while years of smoking and BMI are the predictors.
The model draws a line, or in more complex cases a curve, through the data points that best captures the pattern connecting predictors to the outcome. Each predictor gets a coefficient, which is a number representing how much the outcome changes when that predictor increases by one unit while everything else stays constant. A coefficient of 2.0, for example, means that for every one-unit increase in the predictor, the outcome increases by two units.
Common Types of Regression
The simplest form is simple linear regression, which examines the relationship between one predictor and one outcome. The equation takes the familiar form of a straight line: the outcome equals a starting value plus a slope multiplied by the predictor. That slope is the regression coefficient, telling you the direction and strength of the relationship.
Multiple linear regression extends this to include more than one predictor. A researcher studying walking speed after a medical intervention, for example, might include age, baseline fitness, and treatment type as separate predictors in the same model. This approach lets you isolate the contribution of each factor while accounting for the others. The equation simply adds more terms, one coefficient for each predictor, plus an error term that captures the variation the model can’t explain.
Logistic regression handles a different kind of outcome: yes-or-no questions. When the thing you’re predicting is binary (has lung cancer or doesn’t, experienced a relapse or didn’t, is a current smoker or isn’t), standard linear regression doesn’t work well. Logistic regression instead estimates the probability that someone falls into one category versus the other. It’s especially common in survey-based research and case-control studies in medicine.
Other specialized forms exist for other data types. Cox regression handles time-to-event outcomes, like how long until a patient relapses. Poisson regression is used for count data, such as the number of new brain lesions detected on an MRI over two years. The core logic is the same across all types: quantify how predictors relate to an outcome.
What Makes a Regression Model Valid
Linear regression relies on several assumptions, and violating them can produce misleading results. The first is linearity: the relationship between each predictor and the outcome should follow a roughly straight-line pattern. If the real relationship is curved, a linear model will miss it.
The second assumption is that the data points’ spread around the fitted line stays roughly constant across all values of the predictor. This property is called homoscedasticity. If the spread fans out or narrows as the predictor increases, the model’s estimates of precision become unreliable. The third assumption is that the residuals (the gaps between the predicted values and the actual values) follow a normal distribution. All three assumptions play an important role in producing trustworthy results.
One additional risk in multiple regression is multicollinearity, which occurs when two or more predictors are so closely correlated with each other that the model can’t reliably separate their individual effects. If exercise frequency and daily step count are both in the same model, they may overlap so much that neither appears significant, even though both genuinely predict the outcome. Researchers check for this using a metric called the variance inflation factor, or VIF. A VIF higher than 5 to 10 signals a problem that needs to be addressed, usually by removing one of the overlapping predictors or combining them.
Reading the Results
Regression output typically reports several key numbers for each predictor. The coefficient (sometimes labeled B) tells you the raw change in the outcome per one-unit change in the predictor. If you’re predicting income in dollars and the coefficient for years of education is 3,000, each additional year of education is associated with $3,000 more in income, holding other factors constant.
When predictors are measured in completely different units (say, years of education and hours of weekly exercise), comparing raw coefficients directly doesn’t make much sense. Standardized coefficients convert everything to the same scale, standard deviations, so you can compare which predictor has a stronger influence within that particular sample. However, standardized coefficients are sensitive to the variability in the sample. The same true relationship between two variables can produce different standardized coefficients in different populations simply because the spread of scores differs. For straightforward comparisons, unstandardized coefficients are often more reliable.
Each coefficient also comes with a p-value, which indicates whether the relationship is likely real or could have appeared by chance. The conventional threshold is p < 0.05, meaning there’s less than a 5% probability the result is due to random variation alone. Some fields set a stricter bar. Genetics research, for instance, often requires p-values below 0.00000001 to account for the massive number of comparisons being made simultaneously.
How Well the Model Fits
Beyond individual predictors, researchers evaluate how well the overall model explains the data. The most common measure is R-squared, also called the coefficient of determination. It ranges from negative infinity to 1, though in practice you’ll usually see values between 0 and 1. An R-squared of 0.70 means the model’s predictors collectively explain about 70% of the variation in the outcome. The remaining 30% is driven by factors not included in the model, measurement noise, or randomness.
An R-squared of 1 means perfect prediction, where every data point falls exactly on the fitted line. An R-squared of 0 means the model does no better than simply guessing the average outcome for every observation. Negative values, which are uncommon, mean the model actually performs worse than that flat-line guess, a sign something has gone wrong with the model’s specification. There’s no universal threshold for a “good” R-squared because it depends entirely on the field. Predicting physical measurements in engineering might yield R-squared values above 0.95, while predicting human behavior in social science research might produce values around 0.20 to 0.40, and that can still be meaningful.
How Researchers Use Regression in Practice
In medical research, regression models appear in nearly every observational study. A clinical team studying multiple sclerosis might use linear regression to analyze how quickly patients walk after treatment, logistic regression to determine whether patients relapsed within six months, and Poisson regression to count new lesions on brain scans over two years. Each outcome variable demands a different regression type, but the study might include the same set of predictors (age, treatment group, baseline severity) across all models.
Outside medicine, regression analysis drives decisions in business forecasting, educational research, environmental science, and public policy. Economists use it to estimate how minimum wage changes affect employment. Educators use it to identify which classroom factors predict student performance after controlling for socioeconomic background. Environmental scientists use it to model how temperature and precipitation predict crop yields.
What makes regression so central to research is its ability to separate the influence of one factor from everything else happening at the same time. Observational data is messy. People who exercise more also tend to eat better, earn more, and smoke less. Regression lets researchers statistically hold those other factors constant, isolating the specific contribution of the variable they care about. It doesn’t prove causation on its own, but it provides the quantitative foundation that most cause-and-effect arguments in science are built on.