Skip to main content

How to Interpret Regression Results

Reading a regression output comes down to five things: the model's R² (what percentage of variance in the outcome the predictors explain together), adjusted R² (the version penalized for the number of predictors), each predictor's standardized beta (its relative effect size), the p-value (whether that predictor's effect could plausibly be chance), and VIF (how much the predictors overlap with each other). Looking at only one number — usually "the p-value was significant" — is the most common way this output gets misread.

This guide walks through the table you get from SPSS, R or Python row by row. Send us your own output and we'll explain what each number means for your specific data, delivered with APA-formatted interpretation sentences ready to drop into your report.

Who is this guide for?

  • Graduate students asked to explain a regression table at their thesis defense

  • Researchers reporting regression findings in a journal article

  • Anyone staring at SPSS/R output wondering what the numbers actually mean

  • Authors told by a reviewer or advisor that multicollinearity is a concern

Five steps to reading the output

  1. 01

    Check the model summary first

    R² and adjusted R² show how much variance in the outcome the model explains overall. The model's F-test p-value tells you whether the model as a whole beats a no-predictor baseline — check this before looking at individual coefficients.

  2. 02

    Read each coefficient's B and p-value

    The unstandardized B shows how much the outcome changes per one-unit increase in that predictor, holding others constant (meaningful only together with its unit). p < .05 means that predictor's effect is distinguishable from chance.

  3. 03

    Compare relative strength with standardized beta

    If predictors are on different scales (age, income, a survey score), their B values aren't comparable. Which predictor matters more is answered only by the standardized beta (β).

  4. 04

    Check the assumptions

    VIF flags multicollinearity, Durbin-Watson checks independence of residuals, and residual plots check linearity and homoscedasticity. If an assumption is violated, the coefficient interpretations above become unreliable.

  5. 05

    Translate the numbers into a sentence

    A report isn't complete until the numeric output becomes a sentence a reader can follow: which predictor, in which direction, how large an effect, and with how much confidence.

What the model summary tells you

Before touching individual coefficients, look at the model as a whole:

StatisticWhat it means
Percentage of variance the predictors explain together (0-1 scale, e.g. .35 = 35%)
Adjusted R²Corrects R² for the artificial inflation caused by adding unnecessary predictors; the expected figure to report with more than one predictor
F-test (model p-value)Is the model as a whole significantly better than a baseline with no predictors? p < .05 means yes
Standard error of the estimateHow far the model's predictions typically fall from actual values, in the outcome's own units

Reading the coefficients table

The columns you get for each predictor's row:

ColumnInterpretation
B (unstandardized)Expected change in the outcome per one-unit increase in this predictor, holding the others constant
Std. ErrorHow stable the B estimate is — smaller means a more precise estimate
β (standardized beta)B converted to standard-deviation units; this is where you compare relative effect strength across predictors
t and pWhether B is statistically distinguishable from zero (no effect); p < .05 is the common threshold
95% confidence intervalThe range the true B likely falls in; if it crosses zero, the effect isn't considered reliable

Common interpretation mistakes

  • Assuming "higher R² = better model": in the social sciences, an R² of .20-.30 is meaningful in many contexts — the right threshold depends on the field, so calling it "low" on its own is misleading.

  • Confusing p < .05 with "a strong or important effect": the p-value tells you whether an effect exists, not how large it is — a tiny effect can be significant in a large enough sample.

  • Comparing predictors using unstandardized B: B values on different scales can't be compared directly — use β for that.

  • Skipping the VIF check: severe multicollinearity can even flip a coefficient's sign; this is the most commonly omitted check in student reports.

  • Confusing correlation with causation: regression on non-experimental data shows association, not causation — "X is associated with Y" is more defensible than "X affects Y."

An example interpretation sentence (APA style)

"The regression model explained a significant proportion of variance in job satisfaction, R² = .34, adjusted R² = .32, F(3, 116) = 19.84, p < .001. Managerial support (β = .41, p < .001) and workload (β = -.22, p = .012) were significant predictors, while tenure (β = .09, p = .241) did not contribute significantly."

That single paragraph covers the model summary, each predictor's direction and size, which one didn't reach significance, and why (the p-value) — exactly what a committee or reviewer is looking for.

Frequently asked questions

How high should R² be — is there a "good" value?

There's no universal threshold — it depends on the field. Natural-science models are often expected above .60, while social-science models predicting human behavior commonly report .20-.30 as acceptable. Comparing against similar published studies in your own field is a better reference than any single rule of thumb.

What's the difference between R² and adjusted R², and which should I report?

R² always increases as you add predictors, even useless ones. Adjusted R² corrects for that inflation based on the number of predictors. With more than one predictor, reporting both is standard practice; with a single predictor the difference is negligible.

Why is standardized beta used for comparison?

Unstandardized B depends on each predictor's own unit (age in years, income in thousands, a scale score, etc.), so comparing one predictor's B against another's is meaningless. β converts every variable to standard-deviation units, making relative effect size comparable across predictors.

What VIF value signals a multicollinearity problem?

The common threshold is VIF > 10; some social-science sources are more conservative and flag VIF > 5 as a warning. If there's a problem, options include dropping a redundant predictor, combining correlated scales, or alternative methods like ridge regression.

My p-value is significant but R² is low — what does that mean?

They answer different questions: the p-value asks "could this effect be chance?" while R² asks "how well does this model explain the outcome overall?" In large samples, even a small, practically limited effect can be significant (p < .05) — a low R² shows that effect alone isn't a sufficient explanation, and both figures should be reported together.

Can I send you my own regression output for interpretation?

Yes — send your SPSS/R/Python output (or your raw data) and we'll walk through the table variable by variable, tell you which interpretation is defensible, and deliver APA-formatted sentences you can drop straight into your report.

Let's interpret your regression output together

Send your table or raw data — we'll reply with a plain explanation of what each number means, written as report-ready interpretation text.

Last updated: July 10, 2026