How to Read a Regression Table, Step by Step

A regression table condenses a statistical model into rows and columns. It can show how an outcome, such as test scores or sales, is associated with one or more explanatory variables, while accounting for other variables included in the model. Reading it carefully helps you understand the direction and size of estimates, how uncertain they are, and what the analysis does—and does not—support.

Before interpreting a result, find the table notes or accompanying methods section. They explain variable definitions, units, the sample, and how uncertainty was calculated. Without that context, a coefficient or significance marker can be easy to misread.

What You’ll Need

The full regression table, including its title, column headings, and notes.
A description of the study’s question, outcome variable, and data sample.
A basic understanding that a coefficient is an estimated relationship, not necessarily a causal effect.

Step-by-Step

Identify the outcome. Read the table title and column headings, then locate the dependent variable—the outcome the model aims to explain. Each column often represents a different model or outcome. Check whether the outcome is measured in dollars, points, years, a percentage, or another unit. In a logistic regression table, the reported estimates may be log-odds or odds ratios rather than changes in the outcome’s original units; use the table notes to tell which.
Understand what each column changes. Compare model specifications before comparing results. One column may include only a main predictor, while another adds control variables, fixed effects, or a different sample. Read the notes for definitions: “controls” are additional variables included in the model, and fixed effects account for certain group or time differences. A coefficient that changes across columns may reflect those modeling choices, not a contradiction.
Find the predictor row. Locate the variable you care about, using the row label and notes to confirm its definition and reference group. A categorical variable may be compared with an omitted baseline category. For example, an estimate for “Group B” could represent a difference relative to “Group A,” not an independent value for Group B. If the baseline is unclear, look for a note such as “reference category.”
Read the coefficient and its units. The coefficient is the estimated change in the outcome associated with a one-unit increase in the predictor, holding the model’s other included variables constant. If a score’s coefficient is 2, the model estimates a two-point difference for a one-unit predictor increase, assuming the outcome is measured in points and the model is linear. Check whether variables were standardized, logged, or rescaled; these transformations change what “one unit” means. A negative coefficient indicates an estimated decrease, not necessarily a harmful effect.
Check uncertainty. Look for standard errors, confidence intervals, and significance markers. A standard error reflects the estimate’s statistical uncertainty; it is not the same as the coefficient and is not a measure of the variable’s importance. A confidence interval gives a range of values compatible with the model and its assumptions at the stated confidence level. If an interval includes zero, that is commonly described as not statistically distinguishable from zero at the corresponding two-sided level, but it does not establish that the true relationship is zero.
Interpret significance markers cautiously. Asterisks often mark p-value thresholds, but their meaning varies by table. Read the legend rather than assuming what one, two, or three stars mean. Statistical significance is not the same as practical importance: a small estimate may be statistically precise yet matter little in context, while a larger estimate may be uncertain. Avoid treating a threshold as a yes-or-no verdict on whether a finding is real.
Review the sample and model fit. Find the number of observations, often labeled N, and check whether it differs across columns. Different samples can affect comparisons. Measures such as R-squared describe how much variation the model accounts for under a particular definition; they do not show that the model is correct, unbiased, or causal. For models with different outcomes or estimation methods, fit measures may not be directly comparable.
Read the notes, then state the result precisely. Check for details about standard errors, clustering, weighting, missing data, transformations, and included controls. Then summarize the estimate with its direction, units, comparison, uncertainty, and limits. For example: “In this model, a one-unit increase in X is associated with an estimated two-point higher Y, holding listed controls constant; the interval is wide, so the estimate is uncertain.” Use “associated with” unless the research design justifies a causal claim.

Common Mistakes

Calling association causation: Including controls does not automatically rule out confounding, reverse causality, or other biases. A causal interpretation depends on the study design and assumptions.
Ignoring the measurement scale: A coefficient may refer to a percentage point, percent change, log unit, or standardized unit. These are not interchangeable.
Reading stars as effect size: Stars communicate a threshold for statistical evidence, not how large or important an estimate is.
Comparing unlike columns: Columns may use different samples, outcomes, or methods. Check these differences before saying a result grew or shrank.
Confusing standard errors with estimates: The coefficient is the estimated relationship; the standard error describes uncertainty around it.
Assuming a non-significant estimate means no relationship: It may instead be too imprecise to distinguish from zero with the available data.

Tips

Translate each coefficient into a sentence with the predictor’s unit, outcome’s unit, and relevant comparison.
Use confidence intervals where available; they show a range of plausible estimates more clearly than stars alone.
Look for a table legend before interpreting abbreviations, symbols, or parentheses.
Compare estimates across models only after checking sample sizes, controls, and outcome definitions.
Separate three questions: How large is the estimate? How uncertain is it? What assumptions would support a causal interpretation?

Troubleshooting

You cannot tell what the coefficient means: Find the variable definition, units, and any log or standardization notes in the methods section. If those are missing, do not guess; describe the reported estimate as unclear.
Parentheses are unexplained: Check the note below the table. Parentheses may contain standard errors, t-statistics, or another quantity, depending on the author’s convention.
Stars have no legend: Do not infer exact p-value thresholds. You can still discuss the coefficient and any reported interval, but flag the missing definition.
Results conflict across columns: Check whether controls or samples change, and whether the estimates’ uncertainty ranges overlap. Ask whether the difference itself was tested; separate significance in one column and not another does not prove the estimates differ.
There is no sample size or uncertainty measure: Consult the paper’s methods or supplementary tables. If neither reports them, state that the table alone does not provide enough information for a confident interpretation.

Frequently Asked Questions

What does a positive regression coefficient mean?

It means the model estimates that higher values of the predictor are associated with higher values of the outcome, given the model specification. The units and comparison depend on how both variables are measured.

Does a significant coefficient prove that a variable causes the outcome?

No. Statistical significance describes evidence under a model’s assumptions, not proof of causation. Causal claims require a suitable research design and credible assumptions.

What is the difference between a coefficient and a standard error?

The coefficient is the estimated relationship. The standard error indicates its sampling uncertainty; it helps assess precision but is not the size of the relationship itself.

What does R-squared tell me?

It summarizes how much outcome variation the model accounts for under the measure’s definition. It does not establish that the model is correct or that its coefficients are causal.

Why do regression tables show several columns?

Columns commonly show alternative model specifications, such as adding controls or changing the sample. Read each heading and note to see exactly what differs.

Further reading: UCLA Institute for Digital Research and Education. Details can change, so check the official page for anything that matters to you. Spotted a mistake? Tell us. How we write an answer.