From a Coefficient to a Useful Sentence
A regression table may contain a coefficient such as \(-0.42\), but the number alone does not tell a reader what changes, how much it changes, or what is being compared. A useful interpretation connects the estimate to the predictor and outcome in the units in which they were recorded. It also identifies the comparison clearly: for example, a five-minute difference in travel time, or one group compared with a named reference group.
This tutorial focuses on writing that interpretation, not on deciding whether a coefficient is statistically significant. A coefficient describes the fitted model’s estimated relationship. The wording should preserve that meaning: it should identify the estimate as a model-based association unless the study design and assumptions justify a causal conclusion.
A Basic Model and the Interpretation Target
For a continuous outcome \(Y\) and one continuous predictor \(X\), consider the linear model
Here, \(\beta_0\) is the intercept, \(\beta_1\) is the slope, and \(\varepsilon_i\) is the error for observation \(i\). The model assumes that the conditional mean of \(Y_i\) is linear in \(X_i\), so \(E(Y_i \mid X_i)=\beta_0+\beta_1X_i\), and that the errors have conditional mean zero. These assumptions describe the model’s mean relationship; they do not by themselves make an observational association causal.
The fitted slope \(b_1\) estimates the change in the model’s conditional mean of \(Y\) for a one-unit increase in \(X\). If the outcome is measured in points and the predictor in hours, then the slope is in points per hour. The estimate does not say that every individual’s outcome changes by exactly that amount. It describes the fitted mean relationship.
In a multiple regression model, a coefficient for one predictor is interpreted with the other included predictors held fixed. This is an adjusted comparison, defined by the model; it does not mean that the groups being compared are identical in every possible respect. If the model includes an interaction, a coefficient may apply only at a particular value of another predictor, so the interpretation must reflect that qualification.
Write the Predictor Change Explicitly
A coefficient is calculated per one unit of its predictor, but that unit may not be the most informative one for a report. If a coefficient is per one year of age, a report could describe a five-year difference by multiplying the coefficient by five. This is a change in how the result is expressed, not a new model estimate. The calculation and the units must both be shown correctly.
Worked Example: Expressing a Slope for a Meaningful Change
Suppose, in a hypothetical analysis of 120 adults, a fitted linear model relates weekly exercise time \(X\), measured in hours, to a fitness score \(Y\), measured in points. The estimated slope is \(b_1=1.8\) points per hour. The model includes age and smoking status as additional predictors, with no interaction involving exercise time. Write an interpretation for a three-hour difference in weekly exercise time.
The model’s slope is for a one-hour increase. For a three-hour increase, the estimated difference in the conditional mean fitness score is
A concise interpretation is: Among adults with the same age and smoking status, a three-hour-per-week difference in exercise time was associated with an estimated 5.4-point difference in mean fitness score in the fitted model. The phrase “same age and smoking status” states what the model adjusts for. “Was associated with” describes the estimate without claiming that increasing exercise time would cause a score increase.
The wording concerns a difference in modeled means, not a guaranteed 5.4-point difference for each pair of adults. It also does not claim that the difference is statistically distinguishable from zero; that would require reporting and interpreting inferential results separately.
When choosing a larger predictor increment, preserve the direction of the original coefficient. A positive slope gives a positive modeled outcome difference for an increase in the predictor; a negative slope gives a negative difference. State the increment in the predictor’s units rather than saying only “a higher value.”
Comparisons Between Groups
A categorical predictor in a linear model is commonly represented by an indicator variable. For a two-group predictor, suppose \(G_i=1\) identifies the group of interest and \(G_i=0\) identifies the reference group. In a model with other predictors \(Z_i\), the relevant part can be written as
At a fixed value of \(Z_i\), the modeled difference between \(G_i=1\) and \(G_i=0\) is \(\beta_1\). Thus, a categorical coefficient is a difference in outcome units between named groups, not a change per one unit of the category. State which group is the reference and which group is being compared with it.
Worked Example: Naming an Adjusted Group Difference
In a hypothetical sample of 86 patients, a linear model predicts a six-month symptom score in points from program participation and baseline symptom score. Participation is coded \(1\) for the program group and \(0\) for the usual-care group. The estimated participation coefficient is \(b_1=-3.2\) points. Assume the model includes no interaction between participation and baseline score. Interpret this coefficient.
Because the outcome is measured in points and usual care is coded as the reference group, the coefficient estimates the program-minus-usual-care difference in conditional mean score at the same baseline score. The negative sign means the program group’s fitted mean is lower:
A suitable sentence is: At the same baseline symptom score, program participation was associated with an estimated 3.2-point lower mean six-month symptom score than usual care in this fitted model. Naming the reference group makes the sign interpretable, and naming the adjustment variable explains what “adjusted” means here.
The estimate alone does not establish that the program caused lower scores. If participation was not assigned in a way that supports a causal comparison, unmeasured differences between the groups may remain. The sentence therefore reports an adjusted association, not a treatment effect.
When a Coefficient Depends on Another Predictor
A model with an interaction allows the relationship between a predictor and the outcome to vary with another predictor. In that case, a coefficient attached to the first predictor is not necessarily a single slope that applies to everyone. The full model determines the relevant comparison. For a report, give the value of the other predictor at which the interpretation applies, or calculate a contrast for a stated range of values.
Worked Example: Interpreting a Slope at a Stated Value
Consider a hypothetical model for monthly energy use \(Y\), in kilowatt-hours, using average outdoor temperature \(X\), in degrees Celsius, and an indicator \(G\) for whether a home has electric heating. The fitted model is
Here \(G=1\) means electric heating and \(G=0\) is the reference group. For homes without electric heating, the fitted slope for temperature is \(-8\) kilowatt-hours per degree Celsius. For homes with electric heating, the slope includes both the \(X\) term and the interaction term:
Thus, for electric-heating homes, a one-degree Celsius increase in average outdoor temperature is associated with an estimated 5-kilowatt-hour decrease in mean monthly energy use, according to this model. The interaction changes the slope; reporting only “the temperature coefficient is \(-8\)” would incorrectly present the reference-group slope as applying to both groups.
For a two-degree increase among electric-heating homes, the fitted mean difference is \(2 \times (-5)=-10\) kilowatt-hours. The interpretation should retain the group and units: Among electric-heating homes, a two-degree Celsius increase in average outdoor temperature was associated with an estimated 10-kilowatt-hour decrease in mean monthly energy use. This is an association described by the fitted model, not a claim that changing the weather causes a particular household’s use to change by exactly that amount.
A Short Check on the Wording
A reliable interpretation can be assembled by answering four questions before writing the sentence. What is the outcome, and in what units? What predictor change or group comparison does the coefficient represent? Which other predictors are held fixed? Does the wording match the study design, or does it imply causation without support?
Name the recorded outcome and its units, such as points, minutes, or kilowatt-hours.
For a continuous predictor, state the increment and its units. For a categorical predictor, name both groups and the reference group.
If this is a multiple regression coefficient, name the included predictors held fixed when that is relevant to the interpretation.
Use “associated with” for an association unless the design and assumptions justify a causal interpretation.
Why Precision Matters
Small wording choices can change a statistical claim. “A 1.8-point increase per hour” may sound like an individual-level guarantee; “an estimated 1.8-point higher conditional mean per additional hour” describes the model more accurately. Likewise, “the program lowered the score” makes a causal assertion, while “program participation was associated with a lower adjusted mean score” reports what an observational regression estimates.
Coefficient prose should also distinguish the predictor’s scale from the outcome’s scale. A coefficient of \(-0.4\) is not simply “a 0.4 decrease.” If the predictor is measured in days and the outcome in minutes, it is a decrease of 0.4 minutes per day. If the predictor is a group indicator, it is a difference in outcome units between the specified groups. Include enough detail that a reader could reconstruct what the number compares.
A frequent pitfall is to interpret a coefficient as a causal effect simply because other predictors were included in the model. For example, “holding age and baseline score constant” describes an adjusted model comparison; it does not guarantee that all relevant confounding has been removed. The correction is to report the adjusted association and reserve causal wording for a design and assumptions that support it.
Check Your Understanding
For each item, focus on the comparison and wording, not on statistical significance.
- A fitted model estimates a slope of \(0.6\) quiz points per hour of study. Write an interpretation for a two-hour difference, including the outcome units.
- A model codes rural residence as \(1\) and urban residence as \(0\). The estimated coefficient is \(4.1\) minutes for commute time. Which group is the reference, and how should the sign be described?
- Why is “holding the included covariates fixed” not, by itself, enough to justify saying that a predictor causes a change in the outcome?
- In an interaction model, why may the coefficient for a continuous predictor fail to describe the slope for every value of a categorical predictor?
- Identify the missing information in the sentence: “The coefficient was \(-2.5\), so the outcome decreased.”