Interpreting Least Absolute Deviations Coefficients

Why Interpretations Depend on the Target

A regression coefficient describes a change in a particular feature of the response distribution, as represented by the model. In least squares, the usual interpretation concerns a conditional mean. In least absolute deviations (LAD) regression, the fitted line or surface targets a conditional median. That distinction affects what a slope, a group comparison, and a fitted value mean.

The interpretation is about the response distribution for observations with specified predictor values. It does not say that every individual’s response changes by the coefficient amount. Nor does a coefficient by itself establish that changing a predictor would cause a change in the response.

Definition: A conditional median of $Y$ given predictors $X=x$ is a value $m(x)$ such that at least half of the conditional probability is at or below $m(x)$, and at least half is at or above it. A linear conditional median model specifies that this median is a linear function of the predictors: $$ Q_{0.5}(Y\mid X=x)=\beta_0+\beta_1x_1+\cdots+\beta_px_p. $$ Here $Q_{0.5}(Y\mid X=x)$ denotes the conditional median response at predictor values $x_1,\ldots,x_p$. If a conditional distribution has more than one median, this notation refers to a selected median satisfying the defining condition.

In a fitted LAD model, the estimates $\hat{\beta}_0^{LAD},\ldots,\hat{\beta}_p^{LAD}$ give a fitted conditional median:

\widehat{Q}_{0.5}(Y\mid X=x)=\hat{\beta}_0^{LAD}+\hat{\beta}_1^{LAD}x_1+\cdots+\hat{\beta}_p^{LAD}x_p.

The interpretation below relies on the model providing a reasonable linear description of the conditional median over the predictor values being discussed. The coefficients are estimated from data, so they are estimates of the model’s median relationship, not exact facts about every possible population or predictor profile. A linear conditional median model does not require the whole conditional distribution of $Y$ to be symmetric or normally distributed.

Interpreting a Slope for a Continuous Predictor

When a predictor is continuous, its coefficient is the change in the modeled conditional median for a one-unit increase in that predictor, with the other predictors held fixed. The units are response units per predictor unit. The phrase “with the other predictors held fixed” is essential in a multiple-predictor model: it describes a comparison between predictor profiles that differ in the specified predictor but agree on the others.

\widehat{Q}_{0.5}(Y\mid x_j+1,\text{ other predictors fixed})- \widehat{Q}_{0.5}(Y\mid x_j,\text{ other predictors fixed}) =\hat{\beta}_j^{LAD}.

This follows directly by subtracting the two fitted values. Every term involving the other predictors, as well as the intercept, is the same in both expressions and cancels; the $x_j$ term increases by $\hat{\beta}_j^{LAD}$. For a change of $c$ units in $x_j$, the modeled median changes by $c\hat{\beta}_j^{LAD}$, provided the linear specification is appropriate over that change.

Worked Example: Age and a Fitted Median Response

Suppose a hypothetical LAD analysis models systolic blood pressure in millimeters of mercury (mmHg) using age in years and an indicator for a program: $P=1$ for participants in the program and $P=0$ otherwise. The fitted conditional median model is $$ \widehat{Q}_{0.5}(\text{blood pressure}\mid \text{age},P)=92+0.65(\text{age})-4P. $$

Age coefficient. The fitted age coefficient is $0.65$ mmHg per year. At a fixed program status, the model estimates that a one-year increase in age corresponds to a $0.65$ mmHg increase in the conditional median blood pressure. This describes the modeled median relationship, not the change in every person’s blood pressure.

Calculate a fitted median. For a 60-year-old in the program, substitute age $=60$ and $P=1$: $$ 92+0.65(60)-4(1)=92+39-4=127\text{ mmHg}. $$ The fitted conditional median blood pressure for that predictor profile is 127 mmHg.

Compare ages. At age 70 with the same program status, the fitted median is $92+0.65(70)-4=133.5$ mmHg. The difference is $133.5-127=6.5$ mmHg, equal to $10(0.65)$. The model therefore estimates a 6.5 mmHg difference in conditional medians between these ages at the same program status.

A slope is an adjusted comparison in a multiple-predictor model, not necessarily the same as the unadjusted difference between two groups in the observed data. The adjustment is defined by holding the included predictors fixed in the model. If the relationship is not adequately represented as linear, a single slope may not summarize the median change across the whole range.

Interpreting an Indicator-Predictor Coefficient

For a binary predictor coded 0 and 1, the coefficient is the difference in fitted conditional medians between the group coded 1 and the group coded 0, with other predictors held fixed. The sign indicates which group has the higher modeled median. It is not a difference in means unless a mean model is being fitted and interpreted.

Worked Example: Training and a Credential

Suppose a hypothetical LAD model predicts weekly earnings in dollars using training hours and an indicator $C$, where $C=1$ means the worker has a specified credential and $C=0$ means the worker does not: $$ \widehat{Q}_{0.5}(\text{weekly earnings}\mid \text{training},C)=500+18(\text{training})+240C. $$

The coefficient $240$ means that, at the same number of training hours, the fitted conditional median weekly earnings for workers with the credential are $240 higher than for workers without it. It is an adjusted median difference under this model; it does not show that earning the credential causes an individual’s earnings to rise by $240.

At 10 training hours, the fitted median for a worker without the credential is $$ 500+18(10)+240(0)=500+180=\$680. $$ For a worker with the credential at the same training level, it is $$ 500+18(10)+240(1)=500+180+240=\$920. $$ The fitted difference is $\$920-\$680=\$240$, as given by the indicator coefficient. The units of the coefficient are dollars per comparison between the two credential categories, with training held fixed.

The intercept in a model is the fitted conditional median when every predictor equals zero. That value may have little practical meaning if zero is not a plausible or observed value for the predictors. The fitted median at a realistic predictor profile is often more useful to report than the intercept alone.

Fitted Values and Least-Squares Comparisons

A fitted LAD value is the model’s estimated conditional median response for the specified predictor values. It is not a claim that half the responses in a small sample must fall on each side of that fitted value, and it is not a guaranteed prediction for an individual observation. It describes the model’s estimate of the middle of the conditional response distribution at that profile.

By contrast, a least-squares coefficient is commonly interpreted through a linear conditional mean model. Under that model, the fitted value estimates the conditional mean response, and a slope describes the change in that mean for a one-unit predictor increase with other predictors held fixed. The two coefficients can differ because the conditional mean and conditional median are different features of a distribution.

Model target	Fitted value represents	Slope describes
LAD regression	Modeled conditional median of the response	Change in that conditional median per predictor unit, holding other predictors fixed
Least-squares regression	Modeled conditional mean of the response	Change in that conditional mean per predictor unit, holding other predictors fixed

Worked Example: Distinguishing Median and Mean Slopes

Suppose two separate hypothetical models relate weekly home energy use in kilowatt-hours (kWh) to average outdoor temperature $T$ in degrees Celsius. The fitted LAD model is $$ \widehat{Q}_{0.5}(Y\mid T)=30+2.0T, $$ and the fitted least-squares conditional mean model is $$ \widehat{E}(Y\mid T)=28+2.6T. $$

In the LAD model, the slope $2.0$ means that each one-degree increase in temperature corresponds to a 2.0 kWh increase in the modeled conditional median energy use. In the least-squares model, the slope $2.6$ means a 2.6 kWh increase in the modeled conditional mean energy use per degree.

At $T=4$, the LAD fitted value is $30+2.0(4)=38$ kWh, an estimated conditional median. The least-squares fitted value is $28+2.6(4)=38.4$ kWh, an estimated conditional mean. The numerical fitted values happen to be close at this temperature, but they describe different targets. Their slopes also differ: one summarizes a median relationship and the other a mean relationship. Neither difference alone indicates which model is preferable; the choice depends on the question being asked and the adequacy of each model.

Why the Distinction Matters

The median is often a useful summary when a response distribution has extreme values or is asymmetric, because the center represented by the median can differ from the mean. LAD estimates are therefore interpreted in terms of that center, rather than being described as if they were least-squares mean effects. This distinction is about the target of the fitted model; it does not mean that every LAD coefficient is immune to unusual observations or that the model has established a causal effect.

For clear reporting, name the target, the comparison, the units, and any adjustment. For example: “At the same training hours, the fitted conditional median weekly earnings were $240 higher for workers with the credential.” That sentence identifies a median comparison and the predictor being held fixed. “The credential increased earnings by $240” would overstate both the causal evidence and what the coefficient represents.

State “conditional median” when describing an LAD fitted value or coefficient.
Include units for continuous-predictor slopes and fitted responses.
Name the fixed predictors when interpreting an adjusted coefficient.
Keep association separate from causation: a regression coefficient alone does not show that intervening on a predictor would change the response.

Common mistake: “The LAD slope is 0.65, so every person’s response rises by 0.65 for each one-unit increase in the predictor.” This treats a conditional-median relationship as an individual-level rule. The correction is: “Under the fitted linear median model, the conditional median response is estimated to change by 0.65 response units per predictor unit, with the other predictors held fixed.”

Check Your Understanding

For each question, identify whether the statement concerns a conditional median or a conditional mean, and include the relevant comparison and units.

A fitted LAD model has a coefficient of $-1.4$ for a predictor measured in hours, with a response measured in kilograms. Interpret the coefficient when other predictors are held fixed.
A binary predictor is coded 1 for evening classes and 0 for daytime classes. Its LAD coefficient is 35, with a response measured in dollars. What adjusted comparison does this represent?
A fitted LAD model gives a fitted value of 72 for a particular predictor profile. What does that value estimate, and what does it not guarantee about an individual response?
Why is it incorrect to describe a least-squares slope as a change in the conditional median solely because the data include extreme responses?
Rewrite this statement more carefully: “The program caused every participant’s response to increase by 4 units,” given that the program indicator’s LAD coefficient is 4.