R-Squared and Adjusted R-Squared Explained

Ask what makes a regression good and most people who have taken one statistics course will reach for the same number. R-squared is the most quoted statistic in applied economics and the most routinely misread: it measures how much of the variation in the outcome a model reproduces in the sample at hand, and nothing else. It does not measure whether the model is true, whether its coefficients mean what the author claims, or whether the relationship would survive in new data, and some of the best empirical work in economics carries R-squared values that would embarrass a first-year homework assignment, while some of the worst carries values near one. Learning what the number actually says, what its adjusted cousin corrects, and when either deserves attention is one of the fastest upgrades available to anyone who reads regression output.

The Arithmetic of Explained Variation

The construction is a simple accounting split. Take the outcome’s total variation around its mean, the total sum of squares, and divide it into the part the fitted model reproduces and the residual part it misses. R-squared is the reproduced share:

so a value of 0.4 says the regressors account for 40 percent of the sample variation in the outcome, and the remaining 60 percent lives in the residual. In the two-variable case the number is simply the squared correlation between outcome and fitted values, which connects it to the geometry laid out in our guide to simple linear regression. Two properties follow directly from the formula and explain most of the trouble. First, R-squared never falls when a regressor is added, however irrelevant, because the fitting procedure can always use a new variable at least trivially; a model with enough junk variables can climb toward one while explaining nothing real. Second, the denominator is the outcome’s own variance, so the same true relationship yields a high R-squared where the outcome varies little and a low one where life is noisy, which makes cross-study comparisons of the number close to meaningless.

Adjusted R-squared exists to blunt the first problem. It reweights the two sums of squares by their degrees of freedom, imposing a penalty for every added regressor, so it rises only when a new variable earns more fit than its cost in flexibility, and it can fall when junk is added. That makes it the better default for comparing specifications of different sizes within the same data, a task treated more fully in our article on selecting an econometric model. It is a repair, not a redemption: adjusted R-squared inherits every deeper limitation of the original, because it is still, at bottom, a measure of in-sample fit.

Low Can Be Excellent, High Can Be Garbage

The number’s reputation rests on an assumption that fit and quality travel together, and applied economics breaks that assumption in both directions. In large cross-sections of individuals, outcomes like wages, health, or firm growth are driven overwhelmingly by idiosyncratic factors no dataset records, so even a perfectly identified causal effect leaves most variation unexplained; R-squared values of 0.1 or 0.2 are routine in careful microeconomic work, and the estimate of interest can still be precise, believable, and policy-relevant. What matters for a causal claim is whether the regressor is clean of the endogeneity problems treated in our article on instrumental variables, a property R-squared cannot see: adding a badly endogenous variable typically raises the fit while destroying the interpretation.

The opposite failure is the more dangerous one. Regress any trending series on any other trending series, national consumption on the stock of anything that grows, and the fit will be spectacular, with R-squared above 0.9 for relationships that are pure coincidence of direction. This is the classic spurious regression problem, and it is why time-series work tests for unit roots before believing any fit, a discipline explained in our article on why stationarity matters. A reader can adopt the working rule that a very high R-squared in levels time-series data is a warning to investigate, not a virtue to admire, while a modest R-squared in micro data is simply what honest variation looks like. The multivariate version of the fit-versus-meaning gap, where variables share their explanatory load, is part of the story of multiple regression models.

When the Number Deserves Attention

None of this makes the statistic useless; it makes it a tool with a job description. R-squared is genuinely the right lens when prediction is the goal: a forecasting model exists to reproduce variation, so out-of-sample versions of exactly this quantity are the currency in which forecasts are judged, and the honest evaluation compares predicted with realized values in data the model never saw. It is informative as a description of how much a set of factors collectively accounts for, provided nobody converts that share into a causal ranking. And its movement across specifications carries diagnostic value: a coefficient that survives while R-squared jumps as controls enter is telling a robustness story that formal tests, of the kind covered in our guide to hypothesis testing, will formalize. What the number can never do is referee between a model that explains much falsely and one that explains little truly, and empirical economics contains both in quantity. The question to ask of any regression is what the coefficient means and how it was identified; the question R-squared answers, how much sample variation the fitted line soaks up, is a different and smaller one.

MASEconomics Explains

Conclusion

R-squared answers exactly one question: what share of the outcome’s variation, in this sample, the fitted model reproduces. Its adjusted version repairs the mechanical inflation from added regressors and is the better habit when comparing specifications, but both are in-sample fit measures and nothing more. The number cannot certify a model true, cannot rank causes by importance, and cannot warn that a beautiful fit is two unrelated trends holding hands, which is why the same statistic decorates the best and the worst regressions in the literature with equal enthusiasm.

The mature reading habit is to place the number in its context before reacting to it: expect low values in noisy cross-sections where honest effects live, distrust very high values in trending time series where spurious ones do, reserve real weight for out-of-sample fit when prediction is the goal, and never let fit substitute for the identification question that decides what a coefficient means. R-squared is a useful clerk and a terrible judge, and most of the damage it has done in applied work comes from promoting it to the bench.

Frequently Asked Questions

Thanks for reading! R-squared is a useful clerk and a terrible judge; keep it off the bench. Happy learning with MASEconomics