Why Reporting a Fitted Mean Requires Context
A fitted mean is easy to calculate and easy to misdescribe. A regression model may produce a number such as 58, but that number does not stand alone: it estimates an average outcome for a specified combination of predictor values. Without those values, the number may be ambiguous; without the word “mean,” readers may mistake it for the outcome of one particular person. And if the predictor combination is far from the observed data, the calculation may look more dependable than it is.
Careful reporting makes clear what the model predicts, which covariate values define the prediction, how categorical predictors are coded, and whether the profile is represented in the data. The goal is not to add qualifications mechanically. It is to let readers understand exactly what the fitted number means and where it should be treated cautiously.
What a Fitted Mean Represents
A profile \(x\) includes the values of all predictors used to define the prediction. Some may be quantitative, such as age or baseline score; others may be categorical, such as clinic or treatment group. A fitted mean therefore answers a conditional question: what mean outcome does the fitted model estimate at these stated predictor values?
The distinction between a mean and an individual outcome matters. People with the same recorded predictor values can have different outcomes. The fitted mean describes the model’s estimated average at that profile; it does not say that every person at the profile will have that outcome. A confidence interval for a fitted mean, when reported, concerns uncertainty in that mean estimate. It should not be described as a range expected to contain an individual outcome. Individual-outcome prediction is a separate task.
For a point prediction, the fitted mean comes from the model’s estimated conditional-mean function. The reporting principles here presume that the model’s form is being used as a reasonable description of the conditional mean at the profile of interest, and that predictors and their units have been recorded and coded correctly. A numerical calculation alone cannot establish that the model is appropriate. Nor does a fitted mean, by itself, establish that changing a predictor would cause an outcome to change.
Factor Coding Must Be Translated Into Words
A categorical predictor is often represented in a regression model using an indicator variable. With two groups, let \(I(T=1)\) equal 1 for the named group and 0 for the reference group. In this coding, the coefficient multiplying the indicator describes a fitted difference between those groups, with other predictors held fixed. It is not automatically the fitted mean for the group coded 1.
For example, suppose a fitted mean model is
where \(x\) is baseline score in points, \(T=1\) denotes the program group, and \(T=0\) denotes the usual-care reference group. The outcome is a follow-up score in points. For the reference group at \(x=50\), the indicator is 0 and the centered baseline term is 0, so the fitted mean is 48 points. For the program group at the same baseline score, the indicator is 1, so the fitted mean is 52 points. The coefficient 4 is the difference between those fitted means, not the program group’s fitted mean by itself.
This reading follows directly from subtracting the two fitted means at the same baseline value:
The calculation justifies calling 4 points the model’s fitted program-minus-reference difference at a fixed baseline score. It does not justify calling the difference causal unless the study design and assumptions support a causal interpretation.
Worked Example: Reading a Group Indicator Correctly
Suppose a hypothetical fitted model for follow-up score, in points, is \(\hat m(x,T)=48+0.60(x-50)+4I(T=1)\). Here \(x\) is baseline score in points, \(T=1\) means program group, and \(T=0\) means usual-care group. The hypothetical data include both groups at baseline scores near 60. What fitted means should be reported for the two groups at \(x=60\)?
Check the coding and profile. The requested profile fixes baseline score at \(x=60\). The program group has indicator 1; the reference group has indicator 0. The model estimates a conditional mean at that profile, not an individual follow-up score.
Program group. Substitute \(x=60\) and \(I(T=1)=1\): \[ 48+0.60(60-50)+4(1)=48+6+4=58. \] The fitted mean is 58 points.
Usual-care group. Substitute \(x=60\) and \(I(T=1)=0\): \[ 48+0.60(60-50)+4(0)=48+6=54. \] The fitted mean is 54 points.
A clear report is: “At a baseline score of 60 points, the fitted mean follow-up scores are 58 points for the program group and 54 points for the usual-care group.” The fitted difference is 4 points, program minus usual care, at that baseline value. Because both groups are represented near the stated profile, this example does not rely on extending the model to a distant predictor value. The comparison is an adjusted association from the model, not necessarily a causal effect.
Do Not Omit the Values That Define the Prediction
In a model with several predictors, a fitted mean generally depends on all of them. If a report gives a single predicted number but omits the values of the other predictors, a reader cannot tell which conditional mean was estimated. Calling it a “group mean” does not solve the problem: the fitted mean for a group may differ across ages, baseline scores, or other covariate values.
This is especially important when reporting adjusted group means. “Adjusted” means that the comparison is made at specified covariate values or under a specified averaging scheme. For a conditional mean at a fixed profile, state those fixed values. Do not imply that a prediction at one profile is automatically the overall mean for everyone in the group.
Worked Example: Naming the Fixed Age and Clinic
A hypothetical model estimates follow-up symptom score, in score points, as \(\hat m(a,C)=12+0.8(a-50)+3I(C=B)\), where \(a\) is age in years, clinic A is the reference category, and \(I(C=B)=1\) for clinic B and 0 for clinic A. The observed data include patients from both clinics around ages 50 and 60. Calculate and interpret the fitted mean for a 60-year-old patient at clinic B.
Set the full profile. The requested prediction is for age 60 years and clinic B. Omitting either value would leave the fitted mean unspecified.
Substitute and calculate. At clinic B, \(I(C=B)=1\): \[ 12+0.8(60-50)+3(1)=12+8+3=23. \] The fitted conditional mean is 23 score points for patients aged 60 years at clinic B.
The statement “clinic B has a fitted mean of 23” is incomplete because it leaves out age. At age 50, for example, the same model gives \(12+0.8(50-50)+3=15\) points for clinic B. The fitted value 23 is tied to age 60, not a universal clinic B mean. A suitable report is: “At age 60 years, the fitted mean symptom score is 23 points for clinic B.” The model’s clinic comparison is conditional on age; it should not be presented as a causal clinic effect on the basis of this calculation alone.
Check Whether the Profile Is Supported by the Data
A regression equation can return a number even for predictor values far outside those observed. That arithmetic does not mean the data provide strong support for the prediction. Predicting within the range and combinations represented in the data is often called interpolation; predicting outside them is extrapolation. Extrapolation relies on the model continuing to describe the conditional mean beyond the region where it was informed by observations.
Support is not always determined by checking each predictor’s range separately. In a model with multiple predictors, a profile can combine individually familiar values in a combination rarely or never observed—for example, a high age paired with a particular clinic or treatment group absent at that age. Factor categories also matter: a model cannot supply a directly estimated group-specific mean for a category that was not observed in the fitted data. Even inside the observed range, a prediction may be weakly supported if there are few relevant observations nearby.
Worked Example: Qualifying a Prediction Beyond the Observed Range
In a hypothetical class dataset, study hours range from 1 to 10. A fitted model predicts exam score, in points, using \(\hat m(h)=40+2.1h\), where \(h\) is study hours. Calculate the model’s fitted mean at 15 hours and decide how it should be reported.
Check the predictor range. The requested value, 15 hours, exceeds the largest observed value of 10 hours. The calculation is therefore an extrapolation.
Calculate the model output. \[ \hat m(15)=40+2.1(15)=40+31.5=71.5. \] The equation returns 71.5 points.
That arithmetic does not show that the data support a mean of 71.5 points at 15 study hours. A careful report is: “Extending the fitted linear equation to 15 study hours gives 71.5 predicted points, but 15 hours is beyond the observed range of 1 to 10 hours, so this extrapolated fitted mean is not directly supported by the data.” The model’s behavior beyond the observed range may differ from the true relationship. By contrast, at 10 hours the same equation gives \(40+2.1(10)=61\) points, a prediction at the observed upper boundary. It is still a model estimate, but it does not extend beyond the observed range.
A Practical Reporting Check
Before publishing or interpreting a fitted mean, check each of the following. These questions are useful for prose, tables, and plots alike.
State that the number is a fitted conditional mean, rather than an observed outcome for one individual.
List the values of the quantitative predictors and the categories of the factor predictors that define the fitted mean.
Identify the reference category and explain what an indicator coefficient means in the context of the comparison.
Compare the requested profile with the data used to fit the model, qualify extrapolation or sparse support, and give the outcome units and appropriate precision.
A table should make the same information visible. Useful columns include the predictor profile, the fitted mean, and the outcome unit; factor labels should be written in words rather than left as unexplained codes. If an interval for the mean estimate is shown, label it as a confidence interval for the conditional mean. Do not imply that the interval describes the range of individual outcomes.
Common Mistakes
- Calling a fitted mean an individual prediction. The wrong interpretation is “a person with this profile will have exactly this outcome.” The correction is that the fitted mean estimates the average outcome for cases at the stated profile.
- Reporting a number without its covariate values. A fitted mean at age 60 is not necessarily the fitted mean at age 50. State the fixed age, baseline score, or other predictor values used.
- Treating an indicator coefficient as the coded group’s mean. With reference coding, the indicator coefficient is a group difference conditional on the other predictors. Calculate the group’s fitted mean by substituting the group’s indicator value and the other profile values into the full model.
- Reporting extrapolation as if it were data-supported. A model can produce an output outside the observed predictor region. Identify that extension and explain that the data do not directly support it; do not present the number as equally grounded as an estimate within the data’s support.
- Calling an adjusted comparison causal. Adjustment and a fitted mean difference alone do not establish that group membership caused the difference. Use language describing a model-based association unless a causal interpretation is supported separately.
Check Your Understanding
For each question, focus on what a reader would need in order to interpret the reported fitted mean correctly.
- A model uses clinic A as the reference category and includes an indicator for clinic B. What does the indicator coefficient represent, and what extra information is needed to report a fitted mean for clinic B?
- A report says, “The fitted mean score is 23 points for clinic B,” but the model also includes age. What is missing from this statement, and why might it matter?
- Explain why a fitted mean is not an individual patient’s expected exact outcome, even when the patient’s predictor values match the stated profile.
- A predictor was observed only from 2 to 12 hours, but the model is evaluated at 18 hours. What should the report say about this prediction?
- Why is it not enough to check whether each predictor value falls within its separate observed range when deciding whether a multivariable profile is supported?