Interpreting Logistic Propensity Model Coefficients

What a Logistic Propensity Coefficient Describes

A logistic propensity model expresses the probability of receiving treatment as a function of measured covariates. Its coefficients describe how those covariates relate to the conditional odds of treatment in the specified model. They do not directly report changes in treatment probability, and they are not treatment effects on an outcome.

Let \(A\) be a binary treatment indicator, with \(A=1\) for treatment and \(A=0\) otherwise, and let \(X=(X_1,\ldots,X_p)\) denote measured covariates. The propensity score is \(e(X)=P(A=1\mid X)\). In a logistic propensity model, the log-odds of treatment is modeled as a linear function of the covariates:

The model therefore gives a fitted linear predictor \(\eta(X)=\beta_0+\sum_{j=1}^p\beta_jX_j\). To obtain the fitted propensity score, convert that predictor back to a probability:

This interpretation assumes the stated logistic model is the model being used: its log-odds have the specified linear form, and the model includes no interactions that change the slope for different covariate values. As with any fitted model, the interpretation describes the model’s conditional relationship; it does not by itself establish that the model is correctly specified or that a covariate causes treatment assignment.

From Coefficients to Odds Ratios

For a one-unit increase in \(X_j\), with the other covariates held fixed, the linear predictor changes by \(\beta_j\). Because the logit is the logarithm of the odds, exponentiating this change gives the ratio of the two conditional odds:

where \(X_{-j}\) denotes the remaining covariates. If \(\beta_j>0\), the modeled conditional odds increase with \(X_j\); if \(\beta_j<0\), they decrease. An odds ratio is a multiplicative comparison of odds, not a difference in probabilities. When \(X_j\) is measured in years, the one-unit contrast is one year; when it is measured in decades, the contrast is one decade.

For a binary covariate coded 0 and 1, \(\exp(\beta_j)\) compares the conditional odds at value 1 with those at value 0, holding the other covariates fixed. The coding matters: reversing the reference category reverses the contrast, and the coefficient changes sign. For a continuous covariate, a \(k\)-unit increase corresponds to an odds ratio of \(\exp(k\beta_j)\), under the model’s linear log-odds specification.

Calculating Probabilities at Covariate Values

A coefficient is often easiest to communicate through a predicted probability for a clearly specified covariate profile. First substitute the covariate values into the linear predictor. Then apply the inverse-logit function. The intercept is the modeled log-odds when all covariates equal zero, so its practical interpretation depends on whether that covariate profile is meaningful.

The coefficient and odds-ratio interpretations require a contrast that is meaningful in the data and in the model. For example, if a continuous covariate is recorded in years, interpreting a one-unit change as a year is appropriate; interpreting it as a decade requires multiplying the coefficient by 10 before exponentiating. A large contrast may also extend beyond covariate values represented in the data, so its numerical interpretation should not be mistaken for evidence that the model is reliable there.

Why the Coefficient Is Not a Probability Change

The inverse-logit function is nonlinear. The same change in the linear predictor therefore produces different changes in probability at different starting values. To see this directly, let the linear predictor be \(\eta\), and let a one-unit increase in a covariate change it to \(\eta+\beta\). The corresponding probability change is

This difference depends on both \(\eta\) and \(\beta\); it is not generally equal to \(\beta\). For an infinitesimally small change in a continuous covariate, the derivative of the probability with respect to that covariate is

The factor \(e(X)\{1-e(X)\}\) varies with the covariate profile, reaching its largest value when \(e(X)=0.5\) and becoming small when the probability is near 0 or 1. This derivative is a local probability change per unit of \(X_j\); it is not the same as the finite probability difference for a full one-unit change.

Assumptions, Diagnostics, and Communication

The coefficient interpretations above are exact consequences of the specified logistic model. Applying them to a fitted propensity model requires care. The model must be understood as describing treatment probability conditional on the modeled covariates, with a linear logit and no interactions for the stated constant-slope interpretation. The estimated coefficients are estimates rather than known population values, and their precision depends on the data and fitting procedure. These interpretations do not establish that all relevant covariates have been measured or that the fitted functional form is adequate.

Useful checks when communicating or using coefficients include:

• Confirm the scale. State the covariate’s units and the contrast represented by one unit. Check whether a reported odds ratio is for one year, ten years, or another increment.
• Confirm the coding. For a binary covariate, identify which value is the reference and which is being compared. An odds ratio greater than 1 refers to the coded contrast, not inherently to a favorable or unfavorable direction.
• Check the probability calculation. Substitute all covariate values into the linear predictor before applying the inverse logit. Do not exponentiate the linear predictor and report that number as a probability; it is the odds.
• Keep contrasts within context. A coefficient gives the model-based contrast, but a prediction at covariate values poorly represented in the data may be unstable or rely on extrapolation.
• Separate model interpretation from causal interpretation. A treatment-assignment coefficient describes conditional treatment odds in the model. It is not an effect of treatment on an outcome.

An odds ratio can be useful for summarizing a conditional association, while predicted probabilities make the treatment-assignment implications concrete for specified profiles. Neither presentation should replace the other automatically: the odds ratio is a multiplicative odds contrast, whereas a probability is on the scale of the propensity score. Reporting a small set of probabilities alongside an odds ratio can reveal how different the probability changes are across starting profiles.

If a coefficient estimate has a confidence interval \([L,U]\), exponentiating both endpoints gives the corresponding interval for the odds ratio, \([\exp(L),\exp(U)]\), because the exponential function is increasing. This interval concerns the model coefficient or odds ratio under the inferential procedure; it is not an interval for an individual’s treatment status, nor does it alone assess whether the model is suitable for propensity-score use.

Common Mistakes

• Wrong: “The age coefficient is 0.5, so treatment probability rises by 50 percentage points per decade.” Correction: The coefficient is a 0.5 increase in conditional log-odds per decade. The odds ratio is \(\exp(0.5)\); the probability change depends on the starting probability.
• Wrong: “Exponentiating the coefficient gives a probability.” Correction: Exponentiating a coefficient gives an odds ratio. Convert a complete linear predictor to a probability with the inverse-logit function.
• Wrong: “An odds ratio of 2 means the treatment probability doubles.” Correction: It means the conditional odds double. The resulting probability depends on the initial odds and probability.
• Wrong: “A positive treatment-model coefficient means the covariate improves the outcome.” Correction: It describes the modeled conditional odds of receiving treatment, not an outcome effect.