How to read regression results in plain English
Read a regression in five steps: how much the model explains (R² and the F-test), how much each factor moves the outcome and how sure you can be (the coefficients, confidence intervals and p-values), what “not significant” really means, which factor matters most, and whether the model can be trusted (the checks). The coefficient is the most useful number in the table, and the p-value the most misread.
By Gevorg Petrosyan, founder of StatEVN · Updated 20 November 2026 · 9 min read
The example
We'll use one of StatEVN's sample files: 300 weeks of sales for a small online shop. Each row is one week, with the region it sold in, the ad spend, the price, the number of website visitors, the number of emails opened, and the revenue. The question is the one most people start with: what drives revenue?
A linear regression answers it by finding, for each factor, how revenue differs when that factor is higher, with all the other factors held the same. That last part is what makes regression more useful than looking at one factor at a time, and it's behind most of the surprises below.
How much does the model explain?
Start with the model summary, not the coefficients. It tells you whether there is anything worth reading further.
| Measure | Value |
|---|---|
| R² | 0.908 |
| Adjusted R² | 0.905 |
| F-test | F(7, 292) = 409.26, p < 0.001 |
| R² on 60 weeks the model never saw | 0.900 |
| Weeks (N) | 300 |
- R² = 0.908: together, the factors explain about 91% of the differences in revenue from week to week. The other 9% comes from things not in the file.
- Adjusted R² is the same idea with a small penalty for each factor added, so it doesn't rise just because you added more columns. When it's close to R², as here, no factor is padding the model.
- The F-test asks whether all the factors together do better than simply predicting the average revenue every week. With p < 0.001 they clearly do. If this test fails, stop: the individual coefficients aren't worth reading.
- The R² on unseen weeks is the most honest number here. StatEVN fits the model on 240 weeks and tests it on the other 60. An R² of 0.90 on weeks the model never saw, against 0.91 on the ones it did, means it isn't just memorising the data.
What counts as a good R² depends on the field. Sales data like this often reaches 0.7 to 0.9; models of people's behaviour can be useful at 0.2. A low R² doesn't make an effect unreal, it means much of the outcome depends on other things.
Read the coefficient table
This is the table every regression produces, in SPSS, Excel, R or StatEVN's Advanced mode. Each row is one factor.
| Factor | Coefficient | Std. error | p-value | 95% confidence interval | VIF |
|---|---|---|---|---|---|
| Intercept | 7,009.04 | 509.40 | < 0.001 | 6,006.47 to 8,011.61 | |
| Ad spend | 3.23 | 0.11 | < 0.001 | 3.02 to 3.44 | 2.91 |
| Price | −183.99 | 14.65 | < 0.001 | −212.83 to −155.16 | 1.03 |
| Region = North | 829.69 | 221.63 | < 0.001 | 393.50 to 1,265.87 | 1.44 |
| Region = South | 2,431.81 | 221.22 | < 0.001 | 1,996.43 to 2,867.19 | 1.42 |
| Region = West | 1,527.50 | 251.31 | < 0.001 | 1,032.89 to 2,022.12 | 1.37 |
| Website visitors | −0.03 | 0.10 | 0.771 | −0.22 to 0.16 | 2.89 |
| Email opens | 0.08 | 0.16 | 0.625 | −0.24 to 0.40 | 1.01 |
Revenue per week, linear regression (OLS), 300 weeks. Region is compared with East, the region left out.
The coefficient: how much, and in which direction
A coefficient is the change in the outcome for one more unit of the factor, with the other factors held the same. It's in the outcome's own units, here revenue.
- Ad spend 3.23: each extra 1 spent on ads comes with about 3.23 more revenue that week. In more useful terms, 1,000 more ad spend goes with about 3,230 more revenue.
- Price −183.99: a price 1 higher goes with about 184 less revenue per week. The minus sign is the direction.
- The intercept, 7,009, is the predicted revenue when every factor is zero: no ads, a price of zero, no visitors, in the East region. That week never happens, so the intercept is needed for the calculation but usually has no meaning of its own. Don't report it as a finding.
Yes/no factors: one group against another
Region is a text column, so the regression turns it into yes/no columns, one for each region except a reference group, here East. The coefficient for South, 2,431.81, does not mean "revenue rises 2,432 for each unit of South". It means: in the South, weekly revenue is about 2,432 higher than in the East, for the same ad spend, price and the rest. Every region is compared with East, not with each other. South against West is 2,432 − 1,528, about 904.
The confidence interval: how sure
The 95% confidence interval is the range of values the true effect could reasonably take, given this data. For ad spend it runs from 3.02 to 3.44: narrow, so the estimate is precise. For price, from −213 to −155: wider in absolute terms, but clearly below zero.
This is often more useful than the p-value, because it shows both the size and the uncertainty. If the interval contains zero, the data can't tell you the direction of the effect.
The p-value: is it different from zero?
The p-value answers one narrow question: if this factor had no effect at all, how often would we see a coefficient this far from zero just by chance? Below 0.05 is the usual line for calling it statistically significant.
It does not tell you how big or how important an effect is. With enough data, a tiny effect gets a tiny p-value. With little data, a large effect can miss the line. Read the p-value together with the coefficient, never instead of it.
"Not significant" doesn't mean "doesn't matter"
Look at website visitors: coefficient −0.03, p = 0.771, interval −0.22 to 0.16. On its own, that reads as "visitors have nothing to do with revenue". But in the same data, visitors and revenue are strongly correlated (r = 0.74). What happened?
Visitors move together with ad spend (r = 0.80): weeks with more ads bring more visitors and more sales. Once ad spend is in the model, visitors add almost nothing that ad spend hasn't already explained. So the right reading is: beyond what ad spend already explains, the number of visitors adds no clear extra revenue. It is not "the website doesn't matter".
Email opens are a different case: they're barely related to revenue even on their own (r = −0.04), and the interval is wide around zero. Here the honest conclusion is that this data shows no link, which is still not proof that there is none.
When two factors carry the same information, the VIF column warns you. Values above 5 deserve a look, and above 10 the coefficients of those factors become unstable. Here the highest is 2.91, so the model can still separate ad spend from visitors.
Which factor matters most?
Not the one with the biggest coefficient. Ad spend's 3.23 and price's −183.99 are in different units, and you can't compare one unit of ad spend with one unit of price. Two fair ways to compare:
- Standardized coefficients (beta) put every factor on the same scale: the change in revenue, in standard deviations, for a one standard deviation change in the factor. Here ad spend is 0.91, price −0.23 and South 0.23. Ad spend is by far the strongest.
- A realistic change. Ad spend in the file ranges from about 500 to 5,000 a week; that range is worth about 14,500 in revenue. Price ranges from 15 to 35; that's worth about 3,700.
Also stay inside the data. The model has seen prices from 15 to 35. What it says about a price of 60 is a guess.
Can you trust it?
A regression table always looks confident, even when its assumptions are broken. Before acting on it, check these:
| Check | Here | What it means |
|---|---|---|
| Overlapping factors (VIF) | All below 3 | The factors are different enough to be separated. |
| Even spread of errors (Breusch–Pagan) | p = 0.88 | The model is about as accurate for big weeks as for small ones, so the standard errors are fine. |
| Errors following each other in time (Durbin–Watson) | 2.0 | No pattern from week to week. Near 2 is what you want. |
| Rows the model never saw | R² 0.90 | It predicts new weeks about as well as the ones it learned from. |
And the check no test can do: is something important missing? If a big promotion ran in some weeks and isn't in the file, part of its effect may show up in ad spend instead. A coefficient describes the data you gave it. It is a cause only if nothing important is left out.
How to write it up
For a report or a thesis, give the model summary first, then the effects with their intervals:
A linear regression of weekly revenue on ad spend, price, region, website visitors and email opens explained 91% of the variance, F(7, 292) = 409.26, p < .001, R² = .91. Revenue rose with ad spend (b = 3.23, 95% CI [3.02, 3.44], p < .001) and fell with price (b = −183.99, 95% CI [−212.83, −155.16], p < .001). Website visitors (p = .77) and email opens (p = .62) showed no clear effect once the other factors were included.
For someone who doesn't read statistics, drop the brackets: "Ad spend drives revenue: each extra 1,000 goes with about 3,200 more a week. Higher prices cost about 180 a week per unit of price. The South sells the most, about 2,400 a week more than the East."
The same result in StatEVN
StatEVN runs the checks from Step 5 on every model and starts the result in words. For this file, the lines read:
- Region = South: Revenue is about 2,431.81 higher than for East.
- When Ad_Spend increases by 1 unit, Revenue increases by approximately 3.23
- When Price increases by 1 unit, Revenue decreases by approximately 183.99
Advanced mode adds the coefficient table above, with the VIF column, and the model summary with R², adjusted R², the F-test, AIC, BIC and N. Both can be exported to Excel or PDF.
Questions
What is a good R² in regression?
It depends on the field and the purpose. Sales or physical measurements often reach 0.7 to 0.9, while models of people's behaviour are useful at 0.2 or 0.3. A low R² can still show a clear effect; a high one can come from a factor that is almost the outcome itself. Compare it with the R² on rows the model never saw.
What does p < 0.05 mean in a regression table?
If the factor truly had no effect, a coefficient this far from zero would turn up less than 5% of the time. It says the effect is unlikely to be zero, not that it is large or important. The coefficient and its confidence interval tell you how big it is.
Why does a factor that clearly matters come out as not significant?
Usually because another factor in the model moves together with it and carries the same information, so the model cannot separate them. Check the correlations between the factors and the VIF column.
Does a regression coefficient show cause and effect?
Not on its own. It shows how the outcome differs when the factor differs, with the other factors held the same. It is causal only if nothing important is missing from the model and the factor was not itself changed by the outcome, which is why experiments are the strongest evidence.
How do I compare which factor matters most?
Not by the raw coefficients, because they are in different units. Use the standardized coefficients (beta), or compare what a realistic change in each factor does to the outcome.
Sources and further reading
- STAT 501: Regression Methods, Penn State, a free full course on regression.
- Model validation, NIST/SEMATECH e-Handbook of Statistical Methods.
- The ASA statement on p-values, American Statistical Association, 2016.
- Coefficient of determination (R²), Variance inflation factor and Omitted-variable bias, Wikipedia.
Read next
Run this example yourself
The demo runs a real model on this same sample file and explains it. No sign-up.