What Is Regression Analysis

By Kevin Fernando · Sep 30, 2026 · 11 min read

Regression analysis is a statistical method that measures the relationship between one variable and one or more others. It produces a mathematical equation that lets you predict one variable's value based on others and tells you how strong that relationship is.

Regression analysis is a statistical method that quantifies the relationship between one variable and one or more others. It answers questions like: how does advertising spending relate to sales? How do interest rates affect bond prices? What is the connection between years of experience and salary?

Regression does two things at once. It produces a mathematical equation that lets you predict one variable's value from another, and it measures how strong that relationship is. The method appears everywhere: in academic research, in financial analysis, in policy evaluation, and in any field where someone needs to understand how things move together.

The power of regression lies in its clarity. Instead of saying "higher education correlates with higher earnings," regression lets you say "each additional year of education is associated with a specific increase in earnings, and this relationship explains this much of the variation in earnings we observe." That precision is why regression has become a standard tool for making sense of data.

What regression analysis is

Regression analysis finds the mathematical relationship between variables. You start with data on two or more things: perhaps hours of study and exam scores, or years of experience and salary, or advertising spend and revenue. Regression draws a line through those data points that best represents the overall pattern. That line becomes your equation.

The equation has a simple form. If you want to predict one variable (called the outcome or dependent variable) from another (called the predictor or independent variable), the equation is: outcome equals a starting value plus a slope times the predictor. The slope tells you how much the outcome changes when the predictor increases by one unit. The starting value tells you what the outcome would be if the predictor were zero.

Regression also produces numbers that tell you whether the relationship is strong, whether it is real rather than random noise, and how much of the variation in your outcome the predictor actually explains. Those measures let you judge whether the relationship is worth paying attention to.

How regression works: a worked example

A concrete example shows what regression actually produces. Suppose you have data on sales staff at a retail company and want to understand how staff size relates to store revenue. You collect data from ten stores: each store's number of staff members and its annual revenue.

Setting up the relationship

Your question is straightforward: does a store with more staff generate more revenue? You plot the data on a graph. The horizontal axis is the number of staff members; the vertical axis is annual revenue in thousands of dollars. Each store appears as a single point on the graph. The points do not fall on a perfect line, some stores with the same staff size have different revenues, and some stores with fewer staff earn more than others. But the cloud of points tilts upward: stores with more staff tend to have higher revenue.

Regression takes this messy reality and finds the line that best represents the trend. That line becomes your model.

Finding the line of best fit

Regression calculates the line by minimizing the distance between the actual data points and the line itself. The goal is to find the line that comes closest to all the points at once, balancing errors above and below.

Suppose the regression equation turns out to be: revenue equals 50 plus 3 times the number of staff. The 50 is the intercept, it means a store with zero staff would theoretically generate 50 thousand dollars in revenue (perhaps from online sales or other channels). The 3 is the slope, it means each additional staff member is associated with 3 thousand dollars more in revenue on average.

With this equation, you can predict revenue for a store you have not yet seen. If a new store has 15 staff members, the equation predicts revenue of 50 plus (3 times 15), or 95 thousand dollars. That prediction is your best guess based on the pattern in the data. It will not be exact, actual revenue will vary around this prediction, but it gives you a starting point.

Reading the output: R-squared and statistical significance

Regression produces two critical numbers that tell you whether the relationship matters. The first is R-squared, which measures how much of the variation in an outcome is explained by your predictor variables. R-squared ranges from zero to one. The remaining variation comes from other factors not included in the model: location, local competition, marketing, management quality, or random chance.

The second is statistical significance, often reported as a p-value. This number answers the question: could this relationship have occurred by random chance? If you had no real relationship and just drew random data, how likely is it that you would see a slope this steep? A p-value below 0.05 is the standard threshold in most fields. It means there is less than a 5 percent probability that a relationship this strong would appear in random data. You conclude that the relationship is real, not a fluke.

Together, R-squared and the p-value tell you two things: the relationship is strong enough to matter (R-squared is high) and it is unlikely to be random noise (p-value is low). Without both, the regression result may not be trustworthy.

Why regression matters in research and analysis

Regression enables three kinds of work that matter in practice. First, it lets you predict. Once you have an equation, you can estimate an outcome for new cases you have not yet observed. A bank uses regression to predict loan default risk based on borrower characteristics. A hospital uses it to predict patient readmission risk. These predictions inform decisions: which loans to approve, which patients need extra follow-up care.

Second, regression lets you measure influence. How much does each factor matter? If you are modeling house prices, regression tells you how much price changes when square footage increases by one unit, holding other factors constant. That measure of influence is often what researchers and analysts actually want to know. Investment managers use regression to understand the relationships between factors such as commodity prices and the stocks of businesses dealing in those commodities, helping them understand what drives asset values.

Third, regression lets you control for other variables. Real-world relationships are rarely simple. If you want to know whether a new drug improves outcomes, you need to account for the fact that sicker patients might receive it more often. Regression lets you ask: holding severity constant, does the drug still help? That is a much harder question to answer with raw data, but regression makes it possible.

What regression can and cannot tell you

The most common mistake with regression is treating a strong statistical relationship as proof of causation. Regression is powerful, but it has a hard limit: it measures association, not causation. Understanding that limit is essential to reading research and analysis correctly.

Correlation versus causation

A strong regression result means two variables move together. It does not mean one causes the other. Consider a classic example: ice cream sales and drowning deaths are strongly correlated. Regression would show a steep positive slope. But ice cream sales do not cause drowning. Both are driven by a third variable: warm weather. When summer arrives, people buy more ice cream and spend more time in the water.

Regression cannot distinguish between these scenarios. It sees that two variables move together and quantifies that relationship. It cannot tell you why. Three patterns produce the same regression output: X causes Y, Y causes X, or a third variable Z causes both X and Y. Regression alone cannot separate them.

This matters enormously in research. A study might show that people who take a vitamin have fewer colds, with a strong regression result. But the regression cannot tell you whether the vitamin prevents colds, whether healthier people are more likely to take vitamins, or whether people who take vitamins also exercise more and that exercise prevents colds. You need additional reasoning or experimental design to answer that question. Regression is a tool for measuring relationships, not for proving causation.

When regression breaks down

Regression produces reliable results only when certain conditions hold. Understanding these conditions helps you judge whether a regression result is trustworthy.

First, regression assumes a linear relationship. It draws a straight line through the data. If the true relationship is curved or has a threshold effect, regression will miss it or misrepresent it. If you plot the data and the points form a curve rather than a cloud around a line, linear regression is the wrong tool.

Second, regression is sensitive to outliers. A single extreme data point can pull the line away from the bulk of the data and change the slope dramatically. If your data contains one store with an unusually high or low revenue, it can distort your estimate of how staff size relates to revenue.

Third, regression requires that important variables are included in the model. If you omit a variable that strongly influences the outcome, the slopes of the variables you did include become biased. They pick up some of the effect of the missing variable, and you misinterpret their influence. This is called omitted variable bias.

Fourth, regression assumes that the relationship is the same across all parts of your data. If the relationship between staff size and revenue is different in urban stores than rural stores, fitting one line to all the data will mislead you about both groups. The line will be a compromise that fits neither well.

These limitations do not make regression useless. They make it essential to think carefully about whether the conditions hold before trusting the results.

Types of regression and when to use each

Regression comes in several varieties, each suited to different questions and data structures.

Simple linear regression uses one predictor variable to predict one outcome. This is the most basic form and the one most commonly taught. It answers questions like: how does age predict salary?

Multiple linear regression uses several predictors at once. You might predict house price from square footage, number of bedrooms, age of the house, and location. Multiple regression lets you measure the influence of each factor while holding the others constant. It is far more common in practice than simple regression, because real outcomes depend on multiple factors.

Logistic regression predicts a yes-or-no outcome rather than a continuous number. It answers questions like: what factors predict whether a loan will default? The output is a probability rather than a predicted value. Logistic regression is standard in credit risk, medical diagnosis, and any field where the outcome is binary.

Polynomial regression fits a curved line rather than a straight one. If your data shows a curved relationship, polynomial regression captures it better than linear regression. It is less common but essential when the relationship is nonlinear.

Time series regression predicts a variable over time, accounting for the fact that past values influence future ones. It is used in forecasting and in analyzing data collected repeatedly over time.

Each type answers a different question and produces output suited to that question. Recognizing which type is being used tells you what the results can and cannot tell you.

How to read regression results in published reports

When you encounter regression results in a research paper, a financial filing, or an analyst report, knowing what to look for helps you judge the quality of the work and understand what the authors actually found.

First, look at the sample size. Regression with a small sample can produce results that look strong but are actually unreliable. A regression based on ten data points is far less trustworthy than one based on a thousand. The p-value accounts for sample size, results are harder to achieve statistical significance with small samples, but size still matters for practical interpretation.

Second, examine the R-squared. A high R-squared means the model explains most of the variation in the outcome. A low R-squared means other factors not in the model are doing much of the work. Neither is inherently good or bad; it depends on the context.

Third, check the p-values for each predictor. A p-value below 0.05 is the conventional threshold for statistical significance. If a predictor has a p-value above 0.05, its effect is not statistically distinguishable from zero. It might still be important in practice, but the data do not provide strong evidence for it.

Fourth, look at the coefficients (slopes) and ask whether they make intuitive sense. If a regression predicts that higher education is associated with lower earnings, that contradicts your prior understanding and deserves scrutiny. The authors might have found something real, or they might have made an error or omitted an important variable.

Fifth, check whether the authors discuss the limitations. Do they acknowledge that their results measure association, not causation? Do they discuss potential omitted variables? Do they test whether their assumptions hold? Honest researchers acknowledge what their regression can and cannot claim.

Regression and related statistical methods

Regression sits within a family of statistical methods that measure relationships and make predictions. Understanding how regression relates to these methods helps you recognize them in reports and know when each applies.

Correlation measures the strength of a linear relationship between two variables, ranging from negative one to positive one. Correlation is simpler than regression, it does not produce an equation or a prediction, but it answers a narrower question: do these two variables move together? Regression goes further by quantifying the relationship and allowing multiple predictors.

Analysis of variance (ANOVA) tests whether groups differ on some outcome. If you want to know whether three different drug treatments produce different average outcomes, ANOVA tells you whether the differences are statistically significant. Regression can do this too, by treating group membership as a predictor variable. ANOVA is simpler when you have only groups; regression is more flexible when you have continuous predictors.

Machine learning prediction methods like random forests and neural networks are cousins of regression. They also predict one variable from others, but they do not assume a linear relationship and can capture complex patterns. The trade-off is that they are harder to interpret: you get a prediction, but not a simple equation that shows you how each factor influences the outcome. Regression prioritizes interpretability; machine learning prioritizes prediction accuracy.

Structural equation modeling extends regression to more complex relationships, including indirect effects and latent variables. It is used in psychology, sociology, and fields where you want to model how variables influence each other through chains of relationships.

Each method serves a different purpose. Regression remains the most widely used because it balances simplicity, interpretability, and power.

Do this automatically

Manually calculating regression lines across multiple datasets, plotting each data point, estimating distances, iterating to find the best fit, and computing R-squared values, consumes hours per analysis, multiplied across the dozens or hundreds of relationships a researcher might need to test. Software automates these arithmetic steps, executing the same regression calculations instantly and freeing analysts to focus on interpreting results and testing new hypotheses. Academia research runs this kind of analysis end to end and cites every figure back to its original source, across as many subjects as you need, with Mimir attaching a citation to every figure. It will not write your paper for you. It gathers and cites the evidence; the argument stays yours, and deliberately so. Every plan includes it, including the free one.

Start your first report and run it on something you already know well.

This article is general information and research commentary. It is not professional advice, and it does not account for your particular circumstances.