Why a Regression Table Needs Translating
A regression table reports coefficients for the terms in a model, not necessarily the group-specific lines that are easiest to interpret. In a model with a categorical group variable and a group-by-predictor interaction, one row may describe a difference at a particular predictor value, while another describes a difference in slopes. Neither row by itself generally gives the full difference between groups at every predictor value.
The key is to reconstruct the fitted line for each group from the model terms. Once those lines are written out, it becomes clear which coefficients describe intercepts, which describe slopes, and which combinations answer a particular comparison. All examples below use hypothetical data and software output.
The Reference-Coded Interaction Model
Begin with a categorical variable having two groups. Let \(G=0\) denote the reference group and \(G=1\) denote the other group. Let \(x\) be a quantitative predictor and \(Y\) the response. A reference-coded linear interaction model is
Here, \(\beta_0\) is the reference group’s intercept, \(\beta_1\) its slope, \(\beta_2\) the other group’s intercept difference from the reference group, and \(\beta_3\) the other group’s slope difference from the reference group. The intercept comparison is at \(x=0\); the slope comparison applies to the change in fitted response per one-unit increase in \(x\).
The mean structure is linear in the included terms. For standard coefficient tests and confidence intervals, the usual regression assumptions include independent errors with conditional mean zero and constant variance; normal errors are used for exact small-sample \(t\)-based inference. These assumptions concern the model and its errors, not whether the coefficient table looks tidy. They should be considered using study design and residual diagnostics. Reading the table correctly does not establish that the assumptions hold.
How to Reconstruct the Group Lines
For the reference group, substitute \(G=0\), so both terms containing \(G\) disappear. For the other group, substitute \(G=1\), so its indicator and interaction coefficients are added.
Thus the other group’s intercept is \(\beta_0+\beta_2\), and its slope is \(\beta_1+\beta_3\). Subtracting the reference line from the other group’s line gives
This last expression is especially important: the group difference depends on \(x\) whenever the interaction coefficient is nonzero. The main-effect coefficient \(\beta_2\) is the group difference specifically at \(x=0\), not a universal group difference.
Worked Example: Decoding Two Group-Specific Lines
A hypothetical analysis models a response \(Y\), measured in points, using \(x\), measured in hours. Group A is the reference group and \(x=0\) is meaningful in context. The software reports the following coefficient estimates:
| Term | Estimate |
|---|---|
| Intercept | 30.0 |
| \(x\) | 1.20 |
| Group B | -4.00 |
| \(x\) by Group B | 0.500 |
The fitted model is \(\hat{Y}=30.0+1.20x-4.00G+0.500xG\), where \(G=0\) for A and \(G=1\) for B. For Group A, set \(G=0\): its fitted line is \(\hat{Y}_A=30.0+1.20x\). Its intercept is 30.0 points and its slope is 1.20 points per hour.
For Group B, add the Group B coefficients: its intercept is \(30.0+(-4.00)=26.0\) points, and its slope is \(1.20+0.500=1.70\) points per hour. Therefore, \(\hat{Y}_B=26.0+1.70x\).
At \(x=4\) hours, direct substitution gives \(\hat{Y}_A=30.0+1.20(4)=34.8\) points and \(\hat{Y}_B=26.0+1.70(4)=32.8\) points. The A-minus-B difference is \(34.8-32.8=2.00\) points. Equivalently, the B-minus-A contrast is \(-4.00+0.500(4)=-2.00\) points. The group difference at this \(x\) is not the Group B coefficient alone.
This example assumes the model form is appropriate, observations are independent, and the conditional mean and variance assumptions are reasonable. In a real analysis, the design and residual diagnostics would need to support these conditions; this hypothetical table alone cannot verify them.
More Than Two Groups
The same translation works when a categorical predictor has more than two groups. With \(k\) groups, reference coding uses \(k-1\) indicator variables. If \(x\) interacts with the group factor, each nonreference group has an indicator term and an associated \(x\)-by-indicator term. The reference group’s line uses neither; each other line adds its own pair of coefficients.
A row labeled with a group indicator is still a difference from the reference group at \(x=0\). An interaction row is a slope difference from the reference group. To compare two nonreference groups, subtract their group-specific intercepts or slopes, rather than treating either one’s coefficient as the comparison.
Worked Example: Reading a Three-Group Table
In a hypothetical model, \(x\) is measured in days and \(Y\) in milligrams. Group A is the reference category, and the software reports:
| Term | Estimate |
|---|---|
| Intercept | 12.0 |
| \(x\) | 0.800 |
| Group B | -2.00 |
| Group C | 3.00 |
| \(x\) by Group B | -0.300 |
| \(x\) by Group C | 0.400 |
The reference group’s line is \(\hat{Y}_A=12.0+0.800x\). Group B’s intercept is \(12.0-2.00=10.0\) milligrams and its slope is \(0.800-0.300=0.500\) milligrams per day, so \(\hat{Y}_B=10.0+0.500x\). Group C’s intercept is \(12.0+3.00=15.0\) milligrams and its slope is \(0.800+0.400=1.20\) milligrams per day, so \(\hat{Y}_C=15.0+1.20x\).
To compare C with B, subtract B’s line from C’s line. Their intercept difference is \(15.0-10.0=5.00\) milligrams; their slope difference is \(1.20-0.500=0.700\) milligrams per day. At \(x=5\) days, the fitted values are \(\hat{Y}_B=10.0+0.500(5)=12.5\) milligrams and \(\hat{Y}_C=15.0+1.20(5)=21.0\) milligrams. Thus C minus B is \(21.0-12.5=8.50\) milligrams. The equivalent calculation from the coefficients is \(5.00+0.700(5)=8.50\) milligrams.
The same model and error assumptions described above apply. Here \(x=0\) is within the hypothetical predictor range, so the intercept comparisons are not extrapolations beyond that range. Whether the linear mean structure and constant variance are adequate would still require design knowledge and diagnostic checks.
Contrasts and Their Uncertainty
A contrast is a specified linear combination of model coefficients that answers a comparison question. For example, in the two-group model, the B-minus-A difference at a chosen value \(x_0\) is \(\beta_2+x_0\beta_3\). Its estimate is \(b_2+x_0b_3\). A standard error for this estimate requires the uncertainty of both coefficient estimates and their covariance; the two coefficient standard errors alone are not enough.
For any two estimated coefficients \(b_2\) and \(b_3\), the variance of their weighted sum follows from expanding the variance of a sum:
The standard error is the square root of this variance. Software can calculate a contrast and its standard error directly, or they can be calculated from the estimated covariance matrix. For a \(t\)-test of a contrast equal to zero, the test statistic is the estimate divided by its standard error, using the model’s residual degrees of freedom under the usual linear-model inference conditions.
Worked Example: Testing a Group Difference at a Chosen Value
Use the two-group estimates above, so the B-minus-A contrast at \(x=4\) hours is \(-4.00+4(0.500)=-2.00\) points. Suppose a hypothetical software contrast calculation reports \(SE(b_2)=0.800\), \(SE(b_3)=0.200\), and \(\operatorname{Cov}(b_2,b_3)=0.0200\), with 36 residual degrees of freedom. The covariance is needed because both estimated coefficients contribute to this contrast.
The variance of the contrast estimate is
Therefore its standard error is \(\sqrt{1.440}=1.20\) points, and the test statistic for \(H_0\): B-minus-A difference at 4 hours equals zero is \(t=-2.00/1.20=-1.67\), with \(df=36\). The two-sided \(p\)-value is approximately 0.104. At the 0.05 significance level, this is not sufficient evidence of a nonzero B-minus-A difference at 4 hours; it does not establish that the groups are equal.
A 95% confidence interval uses \(t^* \approx 2.028\) for 36 degrees of freedom: \(-2.00\pm2.028(1.20)=-2.00\pm2.43\), or approximately \((-4.43,\ 0.43)\) points. In repeated use of this interval procedure under the model assumptions, 95% of intervals constructed this way would cover the true contrast. This interval includes zero, consistent with the two-sided test.
The calculation assumes the coefficient covariance estimate and standard errors come from the specified regression fit, whose observations and error structure satisfy the conditions for the reported \(t\)-based inference. The example provides those quantities as hypothetical output; in practice, the analyst must also assess the study design and model diagnostics.
A Reliable Way to Read the Table
A useful sequence is to identify the reference categories first, then write the fitted equation from the table, and finally substitute the indicator values for each group. This prevents a common visual trap: a software table lists terms separately, even though the desired group line or comparison may involve a sum or difference of several terms.
Check the software’s factor coding or model specification. The omitted category is usually the reference; it may not appear as a coefficient row.
Copy the intercept, predictor, indicator, and interaction terms with their reported estimates and signs.
Set the reference indicators to zero and the selected group’s indicator to one. Collect the constant and \(x\) terms to obtain that group’s intercept and slope.
State which groups are compared and whether the question concerns intercepts, slopes, or their predicted difference at a particular \(x\).
For an inferential contrast, use its standard error from software or calculate it with the coefficient covariance terms included.
Common Mistakes and Why the Translation Matters
The most frequent error is to read the group-indicator coefficient as the group difference at every value of \(x\). In an interaction model, it gives the difference at \(x=0\); away from zero, the interaction contributes to the difference. If \(x=0\) is not meaningful or lies far outside the observed range, the indicator coefficient may have little direct practical value, even though the fitted lines can still be interpreted over their supported range.
- Wrong: “The Group B coefficient is the overall difference between B and A.” Correction: It is the B-minus-A difference at \(x=0\); at \(x\), the difference is \(b_2+xb_3\).
- Wrong: “The interaction coefficient is Group B’s slope.” Correction: It is the difference between B’s slope and the reference group’s slope; B’s slope is \(b_1+b_3\).
- Wrong: “A nonsignificant interaction proves the group slopes are identical.” Correction: It means the fitted data do not provide sufficient evidence against equal slopes at the chosen significance level; it is not proof of equality.
- Wrong: “A significant group-indicator row proves a group difference at every \(x\).” Correction: That row tests the difference at \(x=0\). A comparison at another value requires the corresponding contrast.
- Wrong: Treating a fitted association as evidence that group membership causes the response difference. Correction: Interpret the lines as model-based associations unless the study design supports a causal conclusion.
Correctly reading an interaction table turns a list of software terms into statements about fitted group patterns. Always report the group labels, the predictor value when relevant, and the units. If the question is about uncertainty or statistical evidence, identify the exact contrast and use its standard error rather than borrowing the \(p\)-value from a different coefficient row.
Check Your Understanding
Use the reference-coded model and distinctions in this tutorial to answer the following.
- In a two-group model, what does the group-indicator coefficient represent when the model includes a predictor-by-group interaction?
- A reference group has fitted line \(\hat{Y}=8+2x\); the other group has indicator coefficient \(-1\) and interaction coefficient \(0.5\). Write the other group’s fitted line.
- In a model with three groups, why does comparing two nonreference groups usually require more than reading one coefficient row?
- For a two-group contrast at \(x=3\), explain why the standard error generally depends on the covariance between the group and interaction coefficient estimates.
- A test of the interaction coefficient has \(p=0.20\). What conclusion is justified, and what conclusion would be too strong?