How to Interpret Linear Regression Results: Step-by-Step Guide

Q: What is a good R² value in medical research?

There is no universal cutoff for a 'good' R² — it depends heavily on the field and outcome. In controlled laboratory or physiological studies, R² above 0.60-0.70 is often achievable. In studies of complex human health outcomes (quality of life, chronic disease progression, behavior-linked outcomes), an R² of 0.20-0.30 can be considered respectable, because so much biological and social variability is never fully captured by any finite set of measured predictors. Judge R² relative to similar published studies on the same outcome, not against an arbitrary universal threshold.

Q: What is the difference between the coefficient B and the standardized Beta?

B (the unstandardized coefficient) is expressed in the original units of the predictor and outcome — it tells you how much the outcome changes, in its own units, for a one-unit increase in the predictor, in its own units. Standardized Beta re-expresses every predictor and the outcome in standard deviation units, removing the original units entirely, which allows you to directly compare the relative strength of predictors measured on completely different scales (like age in years versus a biomarker in ng/mL) within the same model. Report B with its confidence interval for real-world clinical magnitude; report Beta only when comparing predictors against each other.

Q: Why is my R² low even though several predictors are statistically significant?

Statistical significance and R² answer different questions. A predictor can have a real, statistically significant, precisely estimated effect on the outcome while still explaining only a small fraction of the outcome's total variability, because many other unmeasured factors also contribute to that variability. A low R² with significant predictors is common and expected in most real-world clinical research — it means your predictors reliably move the outcome in a predictable direction, just not enough, on their own, to explain most of why the outcome varies from patient to patient.

Q: Do I need to check regression assumptions every time?

Yes. Regression assumption checks are not optional formalities — a model that meaningfully violates linearity, homoscedasticity, independence, or normality of residuals can produce standard errors, confidence intervals, and p-values that are simply wrong, even though the software will still calculate and display them without any warning. Checking assumptions takes only a few minutes (generating and reviewing two or three diagnostic plots) and is one of the most common items statistical peer reviewers specifically check for.

Q: What should I do if my regression model shows multicollinearity?

First confirm it with the Variance Inflation Factor (VIF) for every predictor — a VIF above 5 to 10 is generally considered concerning. If confirmed, address it by removing one of the highly correlated predictors on clinical or measurement-quality grounds, combining the correlated predictors into a single composite variable, or using a penalized regression method (ridge or LASSO regression) that is more robust to correlated predictors. Simply ignoring a high VIF and interpreting the individual coefficients as if they were reliable is a common and serious interpretation error.

Q: Can I use linear regression for a binary (yes/no) outcome?

No, not appropriately — linear regression assumes a continuous outcome and will produce nonsensical predicted values outside the 0-1 range, along with residuals that systematically violate the model's assumptions. For a binary outcome, use logistic regression instead, which is specifically designed to model the probability of a binary event and produces odds ratios rather than a raw coefficient. See our logistic regression guide for the correct approach to binary outcomes.

SPSS hands you a table with a coefficient, a standard error, a t-value, a p-value, a confidence interval, and an R² — and expects you to know, without being told, which of those numbers actually belong in your Results section. This guide walks through every part of a linear regression output, in plain English, from the coefficients themselves through assumptions, multicollinearity, and residual diagnostics, ending with a complete worked medical example and a practical reporting checklist.

What Is Linear Regression?

Linear regression is a statistical method that models the relationship between a continuous outcome and one or more predictor variables, estimating how much the outcome is expected to change as each predictor changes. It draws the "best-fitting straight line" (or, with multiple predictors, a best-fitting flat plane) through the data, chosen specifically to minimize the total distance between the observed values and the line's predicted values.

In medical research, linear regression answers questions like: how much does systolic blood pressure change per additional year of age? How much does a new drug change LDL cholesterol after adjusting for baseline levels and BMI? Any time your outcome is a continuous, roughly normally distributed measurement, linear regression is likely the tool you reach for.

Take-Home Points Linear regression estimates how a continuous outcome changes as one or more predictors change — the output is a set of coefficients describing exactly how much and in which direction.

When Should Linear Regression Be Used?

Use linear regression when your outcome variable is continuous (blood pressure, cholesterol, hospital stay in days, a symptom score treated as continuous) and you want to quantify its relationship with one or more predictors — whether those predictors are continuous (age, dose) or categorical (treatment group, sex).

Clinical Scenario	Why Linear Regression Fits
Predicting LDL change from a new statin dose, adjusted for baseline LDL and age	Continuous outcome, continuous + categorical predictors
Relationship between BMI and systolic blood pressure	Both variables continuous — a classic simple linear regression
Predicting hospital length of stay from comorbidity count and surgical complexity	Continuous outcome, multiple predictors — multiple linear regression

If your outcome is binary (event/no event), use logistic regression instead. If your outcome is time-to-event with censoring, use Cox regression. Linear regression is specifically for continuous outcomes.

Take-Home Points Match the outcome type to the model: continuous → linear regression, binary → logistic regression, time-to-event → Cox regression. Using linear regression on the wrong outcome type is one of the most fundamental modeling errors possible.

Simple vs Multiple Linear Regression

Simple linear regression has exactly one predictor and one outcome — for example, predicting weight from height alone. Multiple linear regression includes two or more predictors in the same model simultaneously, allowing you to estimate each predictor's effect on the outcome while statistically holding the others constant.

Practical Example

Simple: Systolic BP = f(Age) only — one predictor, describes the raw relationship between age and blood pressure with no adjustment for anything else.
Multiple: Systolic BP = f(Age, BMI, Sodium intake, Smoking status) — four predictors together, each coefficient now represents that variable's effect independent of the other three.

Multiple regression is far more common in medical research, because clinical outcomes are almost always influenced by several factors at once, and adjusting for known confounders is usually necessary to get a meaningful estimate of any one predictor's effect. See our full guide on choosing covariates for how to decide which predictors belong in a multiple regression model.

Take-Home Points Simple regression describes a raw, unadjusted relationship; multiple regression describes an adjusted relationship, isolating each predictor's independent contribution.

Dependent and Independent Variables

The dependent variable (also called the outcome or response variable) is what you are trying to predict or explain — it goes on the left side of the regression equation and must be continuous for linear regression. The independent variables (also called predictors, covariates, or explanatory variables) are what you use to predict or explain the outcome — they can be continuous or categorical.

A categorical predictor with more than two categories needs to be entered as a set of dummy (indicator) variables — a 4-category variable like disease stage (I–IV) becomes 3 dummy variables, each comparing one stage against a chosen reference stage, not a single variable coded 1–4 treated as if it were continuous.

Take-Home Points The dependent variable must be continuous; independent variables can be continuous or categorical, but categorical predictors with 3+ categories require dummy coding, never a single numeric code treated as continuous.

Regression Coefficients Explained (B)

The unstandardized coefficient, labeled B in SPSS output, tells you exactly how much the outcome is predicted to change, in its own original units, for a one-unit increase in that predictor, holding all other predictors in the model constant.

Practical Example

B = 0.71 for age (in years) predicting systolic blood pressure (in mmHg) means: for every additional year of age, systolic blood pressure is predicted to increase by 0.71 mmHg, holding all other model predictors constant. A 10-year age difference predicts a 7.1 mmHg difference in systolic BP, all else equal.

The sign of B tells you the direction: positive means the outcome increases as the predictor increases; negative means the outcome decreases as the predictor increases. The magnitude of B is only meaningful in the context of the predictor's own unit — a B of 0.71 per year of age is not directly comparable to a B of 1.12 per kg/m² of BMI, which is exactly why standardized Beta (covered next) exists.

Take-Home Points B tells you real-world clinical magnitude in the outcome's actual units — the number to quote when explaining what the model means for an actual patient.

Beta Coefficients: Standardized vs Unstandardized

Standardized Beta re-expresses every variable in the model — predictors and outcome alike — in standard deviation units instead of their original units. This removes the original scale entirely, which means standardized Beta coefficients from different predictors, even ones measured in completely different units, become directly comparable to each other in terms of relative strength.

Situation	Use
Explaining real-world clinical magnitude ("each year of age adds X mmHg")	Unstandardized B
Comparing which predictor has the strongest relative effect	Standardized Beta
Predictors measured in wildly different units (age in years vs a biomarker in pg/mL)	Standardized Beta, for fair comparison

❌ Incorrect Interpretation

"Age (B = 0.71) is a stronger predictor of blood pressure than BMI (B = 1.12) because 1.12 > 0.71" — comparing unstandardized B across two predictors measured in different units.

✅ Correct Interpretation

Compare standardized Beta instead: if age's Beta = 0.47 and BMI's Beta = 0.24, age is actually the stronger relative predictor, despite its smaller unstandardized B.

Take-Home Points Never compare unstandardized B coefficients across predictors with different units — use standardized Beta for that comparison, and B for describing real-world magnitude.

Confidence Intervals for Coefficients

The 95% confidence interval around a coefficient B gives the plausible range for the true population value, based on your sample. If the CI does not cross zero, the coefficient is statistically significant at the conventional threshold — consistent with the p-value, but the CI additionally conveys the precision and plausible range of the effect, which a p-value alone cannot.

Practical Example

B = 0.71, 95% CI [0.48, 0.94] for age predicting SBP. The true effect of age is estimated to lie somewhere between 0.48 and 0.94 mmHg per year, with 95% confidence — a fairly precise, clearly positive estimate. Compare this to B = 0.71, 95% CI [-0.12, 1.54] from a smaller study: same point estimate, but this CI crosses zero, meaning the effect is not statistically significant and much less precisely estimated.

Always report the 95% CI alongside every coefficient you discuss, not only the primary predictor of interest. See our confidence interval guide for the general concept behind this.

Take-Home Points A CI crossing zero means "not statistically significant," equivalent to p > 0.05 — but the CI's width also tells you how precisely the effect was estimated, which the p-value alone does not.

P Values in Regression Output

The p-value (Sig. column in SPSS) tests the null hypothesis that a given predictor's true coefficient is zero — no relationship with the outcome, after adjusting for the other predictors in the model. A p-value below your threshold (almost always 0.05) means the observed coefficient is unlikely to have arisen by chance if the true coefficient were really zero.

Report the exact p-value (e.g., p = 0.007), not just "p < 0.05," and convert SPSS's ".000" to "p < 0.001" rather than reporting it literally. See our p-value interpretation guide for the general logic behind this threshold.

Take-Home Points The p-value tells you whether a coefficient is distinguishable from zero — it says nothing about how large or clinically important that coefficient is. That's what B and its CI are for.

R²: How Much Variance Is Explained

R² (R-squared) is the proportion of the total variability in the outcome that is explained by all the predictors in the model combined, expressed as a value from 0 to 1 (or 0% to 100%). An R² of 0.32 means the model's predictors together explain 32% of why the outcome varies from person to person in your sample — the remaining 68% is due to factors not captured by the model, plus random variability.

Practical Example

A model predicting LDL cholesterol change from diet, baseline LDL, and statin dose reports R² = 0.41. This means 41% of the variability in LDL change across patients is explained by these three predictors together — a moderately strong model for a clinical outcome, though the majority of variability still comes from factors outside the model (genetics, adherence, other unmeasured factors).

Take-Home Points R² describes how much of the outcome's variability the whole model explains — it is a property of the model as a whole, not of any single predictor.

Adjusted R²

Plain R² has a mechanical flaw: it always increases, or at worst stays the same, every time you add another predictor to the model — even a completely irrelevant one — simply because more predictors give the model more flexibility to fit the specific sample's noise. Adjusted R² corrects for this by applying a penalty based on the number of predictors relative to the sample size, giving a more honest estimate of how well the model would generalize.

❌ Incorrect Interpretation

"Adding a fifth predictor raised R² from 0.41 to 0.43, so it clearly improved the model" — reporting plain R² alone as evidence a predictor is worth keeping.

✅ Correct Interpretation

Check adjusted R² too: if it dropped from 0.39 to 0.37 despite plain R² rising, the fifth predictor added essentially no real explanatory value and may not be worth keeping in the model.

Always report adjusted R² alongside plain R² once your model has more than one or two predictors — it is the more honest number for judging whether your model is genuinely informative.

Take-Home Points Plain R² can only go up as you add predictors; adjusted R² can go down, which is exactly why it's the better indicator of whether an additional predictor is truly earning its place in the model.

Regression Assumptions

Linear regression rests on four key assumptions. SPSS will calculate a full set of results whether or not these assumptions hold — checking them is your responsibility, not something the software warns you about automatically.

Assumption	What It Means	How to Check
Linearity	The relationship between predictors and outcome is a straight line, not curved	Residuals vs. fitted values plot — no clear curve pattern
Homoscedasticity	The spread of residuals is roughly constant across all predicted values	Residuals vs. fitted values plot — no funnel/cone shape
Independence of residuals	Residuals are not correlated with each other (important for repeated/clustered data)	Durbin-Watson statistic (for time-ordered data); study design review
Normality of residuals	Residuals are approximately normally distributed	Q-Q plot of residuals; Shapiro-Wilk test on residuals

If linearity or homoscedasticity is meaningfully violated, consider transforming the outcome (e.g., a log transformation for a right-skewed outcome) or adding a non-linear term. If normality of residuals is violated in a large sample, this is often a minor concern due to the Central Limit Theorem; in a small sample, it is a more serious issue.

Take-Home Points Check assumptions with a residuals-vs-fitted plot and a Q-Q plot every time you fit a model — this takes only a couple of minutes and is one of the most common things a statistical reviewer checks for.

Multicollinearity

Multicollinearity occurs when two or more predictors in the same model are strongly correlated with each other, making it statistically difficult for the model to separate their individual, independent effects on the outcome — coefficients become unstable, standard errors inflate, and a genuinely important predictor can appear falsely non-significant.

Practical Example

A model predicting cardiovascular risk includes both waist circumference and BMI, which are typically correlated above r = 0.85. VIF for both predictors comes out above 8, confirming problematic collinearity — the fix is to retain only one of the two, chosen on clinical grounds, not to keep both and interpret their individual coefficients as if they were reliable.

Check the Variance Inflation Factor (VIF) for every predictor — a VIF above 5 to 10 is generally treated as concerning. See our full guide on choosing covariates, which covers multicollinearity and overadjustment bias in depth.

Take-Home Points Always check VIF before interpreting individual coefficients in a multi-predictor model — a high VIF means the individual coefficients cannot be trusted even if the overall model fits well.

Residual Analysis

A residual is the difference between an observed value and the value the model predicted for that same case — residual analysis is how you actually verify the assumptions above, rather than just assuming they hold. Two plots do most of the work.

Residuals vs Fitted Values — What Good and Bad Look Like

Outliers and Influential Points

A large standardized residual (typically beyond ±3) flags a case the model predicts poorly. Cook's Distance identifies influential points — cases that, if removed, would substantially change the model's coefficients. Investigate flagged cases for data entry errors before deciding whether to exclude them; see our guide on outlier detection for the full decision process.

Take-Home Points A residuals-vs-fitted plot showing random scatter with no pattern and no widening/narrowing funnel shape supports both linearity and homoscedasticity at once — it is the single most useful diagnostic plot to generate.

Interpreting SPSS Linear Regression Output

Run via Analyze → Regression → Linear. Three tables matter: Model Summary, ANOVA, and Coefficients.

Example Output — Model Summary & Coefficients

Model Summary

Value

R / R²

0.58 / 0.336

Adjusted R²

0.322

Coefficients

B (95% CI)

Beta

Sig.

Age (years)

0.71 [0.48, 0.94]

0.47

.000

BMI (kg/m²)

1.12 [0.31, 1.93]

0.24

.007

Regression ANOVA: F(2, 107) = 26.9, p < 0.001 (overall model is significant). Report: "Age (B = 0.71, 95% CI [0.48, 0.94], β = 0.47, p < 0.001) and BMI (B = 1.12, 95% CI [0.31, 1.93], β = 0.24, p = 0.007) were independent significant predictors of systolic blood pressure. The model explained 32.2% of the variance (adjusted R² = 0.322), F(2, 107) = 26.9, p < 0.001."

What to Interpret, What to Ignore

Interpret B, its 95% CI, Beta, and Sig. for every predictor, plus the overall F-test and adjusted R² for the model as a whole. Ignore the Constant (intercept) row's own p-value in most clinical write-ups — it rarely has a meaningful independent clinical interpretation. See our full SPSS output interpretation guide for how this compares across other tests.

Take-Home Points Report B with CI, standardized Beta, exact p, the overall model F-test, and adjusted R² — this five-part combination is what a complete linear regression results sentence requires.

Common Mistakes

Mistake 1: Comparing Unstandardized B Across Predictors

Treating a larger B as automatically "more important" when predictors are measured in different units produces a meaningless comparison.

✓ Fix: Use standardized Beta for comparing relative predictor strength.

Mistake 2: Never Checking Assumptions

Reporting coefficients, CIs, and p-values from a model that violates linearity or homoscedasticity produces numbers that look precise but may not be trustworthy.

✓ Fix: Generate and review a residuals-vs-fitted plot and a Q-Q plot before interpreting any coefficient.

Mistake 3: Ignoring Multicollinearity

Including two highly correlated predictors and interpreting both coefficients independently, when the model cannot actually separate their effects reliably.

✓ Fix: Check VIF for every predictor; address any VIF above 5-10 before interpreting individual coefficients.

Mistake 4: Using Plain R² Instead of Adjusted R² With Multiple Predictors

Plain R² mechanically rises with every added predictor, overstating how much a specific addition actually improved the model.

✓ Fix: Report and compare adjusted R² whenever the model has more than one or two predictors.

Mistake 5: Using Linear Regression for a Binary Outcome

Fitting a linear model to a yes/no outcome produces predicted values outside the 0-1 range and systematically violates model assumptions.

✓ Fix: Use logistic regression for binary outcomes instead.

Mistake 6: Claiming Causation From an Observational Regression Model

A significant, adjusted coefficient from an observational study shows an association, not proof of a causal effect, regardless of how many covariates were included.

✓ Fix: Use association language ("was associated with") unless the data come from a randomized design.

Complete Worked Medical Example

Question: What predicts fasting glucose level in adults — age, BMI, and physical activity level?

Check assumptions

Residuals-vs-fitted plot shows random scatter, no funnel shape; Q-Q plot shows residuals close to the diagonal — linearity, homoscedasticity, and normality all reasonably satisfied.

Check multicollinearity

VIF for age, BMI, and activity level all below 2 — no meaningful multicollinearity concern.

Run the model and read the ANOVA table

F(3, 196) = 22.4, p < 0.001 — the overall model is statistically significant.

Read the coefficients

Age: B = 0.42, 95% CI [0.21, 0.63], β = 0.24, p < 0.001. BMI: B = 1.08, 95% CI [0.71, 1.45], β = 0.38, p < 0.001. Physical activity (hours/week): B = -1.15, 95% CI [-1.79, -0.51], β = -0.21, p = 0.001.

Read model fit

R² = 0.29, Adjusted R² = 0.28 — the model explains 28% of the variance in fasting glucose after adjustment.

Final Reported Result

"In a multiple linear regression model, age (B = 0.42, 95% CI [0.21, 0.63], β = 0.24, p < 0.001), BMI (B = 1.08, 95% CI [0.71, 1.45], β = 0.38, p < 0.001), and physical activity (B = -1.15, 95% CI [-1.79, -0.51], β = -0.21, p = 0.001) were all independent significant predictors of fasting glucose. BMI showed the strongest relative association (highest standardized β). The overall model was significant, F(3, 196) = 22.4, p < 0.001, explaining 28% of the variance in fasting glucose (adjusted R² = 0.28)." Run this analysis directly with the linear regression calculator or multiple regression calculator, and see our reporting-by-test guide for more worked examples.

Practical Reporting Checklist

State simple or multiple regression, and every predictor included

Named explicitly, not left implicit.

Report B with its 95% CI for every discussed predictor

Never a coefficient without its interval.

Report standardized Beta when comparing predictor strength

Not raw B, across predictors in different units.

Report the exact p-value for every coefficient

Three decimals, or "p < 0.001" — never "p = 0.000."

Report the overall model F-test

F(df, df) = value, p = value, alongside individual coefficients.

Report both R² and adjusted R²

Especially with more than one or two predictors.

Confirm assumptions were checked

Residuals-vs-fitted and Q-Q plots reviewed, not assumed.

Confirm VIF was checked for multicollinearity

Especially with clinically related predictors.

Frequently Asked Questions

What is a good R² value in medical research? +

There is no universal cutoff — it depends heavily on the field and outcome. In controlled laboratory or physiological studies, R² above 0.60-0.70 is often achievable. For complex human health outcomes (quality of life, chronic disease progression), R² of 0.20-0.30 can be respectable, since so much variability is never fully captured by any finite set of predictors. Judge R² relative to similar published studies, not an arbitrary universal threshold.

What is the difference between the coefficient B and the standardized Beta? +

B is expressed in the original units of the predictor and outcome — how much the outcome changes for a one-unit increase in the predictor. Standardized Beta re-expresses every variable in standard deviation units, removing original units, allowing direct comparison of relative predictor strength across variables measured on different scales. Report B with its CI for clinical magnitude; report Beta only when comparing predictors.

Why is my R² low even though several predictors are statistically significant? +

Statistical significance and R² answer different questions. A predictor can have a real, significant, precisely estimated effect while explaining only a small fraction of the outcome's total variability, because many other unmeasured factors also contribute. A low R² with significant predictors is common and expected in most real-world clinical research.

Do I need to check regression assumptions every time? +

Yes. A model that meaningfully violates linearity, homoscedasticity, independence, or normality of residuals can produce standard errors, confidence intervals, and p-values that are simply wrong, even though the software calculates and displays them without warning. Checking takes only a few minutes and is a common item statistical reviewers specifically check for.

What should I do if my regression model shows multicollinearity? +

First confirm it with VIF for every predictor — above 5 to 10 is generally concerning. If confirmed, remove one of the highly correlated predictors on clinical grounds, combine them into a composite variable, or use a penalized regression method (ridge or LASSO) more robust to correlated predictors. Ignoring a high VIF and interpreting coefficients as reliable anyway is a common, serious error.

Can I use linear regression for a binary (yes/no) outcome? +

No, not appropriately — linear regression assumes a continuous outcome and will produce nonsensical predicted values outside the 0-1 range, with residuals that systematically violate model assumptions. Use logistic regression instead, which is specifically designed for binary outcomes and produces odds ratios.

Run Your Linear Regression With Confidence

Use StatClinic's free linear regression calculator to get coefficients, confidence intervals, R², and a journal-ready results sentence generated automatically. Free, no registration required.

Try StatClinic Free →