What a Fitted Conditional Mean Represents

Introduction

A regression prediction is often written as a single number, but that number can answer different questions. It might describe the average outcome among cases with particular predictor values, or someone might mistakenly treat it as the outcome that one particular person will have. These are not the same quantity.

The conditional mean is the average outcome for the population of cases whose predictor values match a specified profile. A fitted conditional mean is the value a regression model estimates for that population average. Understanding this distinction helps keep a prediction in its proper role: it summarizes an expected outcome at a profile, not a promise about any individual case.

The Conditional Mean for a Predictor Profile

Let \(Y\) be an outcome and let \(\mathbf X\) denote one or more predictors. A predictor profile, written \(\mathbf x\), specifies particular values of those predictors. For example, a profile might specify a patient’s age and treatment group, or a home’s floor area and number of bedrooms. The conditional mean of \(Y\) at that profile is the population average outcome among cases with those predictor values.

The word conditional indicates that the average is taken within the group defined by the predictor values. The word mean indicates an average outcome. If \(Y\) is measured in days, the conditional mean is also measured in days; it is not a probability unless the outcome and model make it one.

For continuous predictors, an exact profile such as age 47.3 years may not occur repeatedly in a finite data set. The conditional mean still has a clear population meaning: it describes the mean of the outcome distribution at that predictor value. In practice, a regression model uses the observed data and its assumed mean structure to estimate that quantity. The quality of the estimate depends on whether the profile is supported by the data and whether the model represents the mean pattern adequately.

An Observed Outcome Is Not a Conditional Mean

An individual case has one observed outcome, say \(Y_i\), along with its observed predictor profile. That outcome can be above or below the population conditional mean for the profile. A useful way to express the distinction is to write the outcome as the conditional mean plus a departure from that mean:

Here, \(\varepsilon\) represents the case-specific difference from the conditional mean. The equation does not say that each case equals the mean. It says that, at a given profile, the departures average to zero in the population. A particular observed outcome may differ substantially from the mean.

This decomposition also clarifies what a fitted mean estimates. A regression procedure uses a sample and a chosen model to produce an estimate, written \(\widehat m(\mathbf x)\), of the population quantity \(m(\mathbf x)\). The fitted value is not itself a newly observed outcome. Nor is it necessarily the average of only the observations that happen to have exactly that profile; a regression model may use information across many profiles to estimate the mean at the requested one.

Worked Examples

What Makes a Fitted Mean an Estimate?

A fitted mean depends on both the observed data and the regression model used to summarize them. The model specifies how the conditional mean is represented as predictor values change. Fitting the model then uses the sample to estimate the unknown features of that mean pattern. For a requested profile, the resulting fitted value \(\widehat m(\mathbf x)\) is the model’s estimate of \(m(\mathbf x)\).

The meaning of the target, \(m(\mathbf x)\), does not depend on choosing a particular regression method. It is defined by the population distribution of outcomes at that profile. The fitted estimate does depend on the selected model, the sample, and the estimation method. Two models fitted to the same data can therefore give different fitted means at a profile.

The basic definition of a conditional mean requires that the outcome have a finite mean at the profile. It does not, by itself, require normally distributed outcomes. However, using a fitted mean responsibly requires attention to the model: its mean structure should be suitable for the data, and the requested profile should be represented well enough for the sample to inform the estimate. A prediction at a profile far outside the observed data requires extrapolation and may be unreliable.

Why the Distinction Matters

The distinction matters whenever a fitted value is communicated or used. Saying “the predicted mean stay is 6.8 days for this profile” describes an estimated population average. Saying “this patient will stay 6.8 days” incorrectly turns that average into an individual guarantee. Many patients with the same profile may have shorter stays, and many may have longer stays.

It also matters for interpretation. A conditional mean is an association between predictor values and an average outcome in the population described by the model. By itself, it is not a causal effect. A fitted mean does not establish what would happen if a predictor were changed; that conclusion would require additional design and assumptions beyond the meaning of a regression prediction.

Finally, the profile must be described precisely. A fitted mean is conditional on the predictor values supplied to the model. If relevant predictors are omitted from the description, or if the profile is outside the range of cases that informed the fit, a reader may misunderstand what the estimate applies to. Always report the outcome’s units and enough of the profile to make the target clear.

Common Mistake

Incorrect reasoning: “The fitted mean is 142.6 mmHg, so every person with this profile should have a blood pressure of 142.6 mmHg. A person with 151.0 mmHg shows the model failed.”

Correction: The fitted value estimates a population average at the profile, not each individual outcome. A value of 151.0 mmHg is 8.4 mmHg above the fitted mean, but individual departures are expected. Whether the model estimates the population mean well is a separate question; one individual outcome alone does not answer it.