Adjusted Odds Ratios with Multiple Covariates

Why Compare Odds at the Same Covariate Values?

In a logistic regression with several predictors, an outcome’s modeled probability depends on all of them. To describe the association with one predictor while accounting for the others, compare two covariate profiles that differ in that predictor but match on the remaining predictors. The resulting comparison is an adjusted odds ratio: it describes the model-based odds contrast with the other included predictors held fixed.

This interpretation is especially direct for an additive logistic model, where each predictor contributes a separate term to the log odds. The coefficient of a predictor then determines its adjusted odds ratio. The comparison is conditional on the specified values of the other predictors; it is not automatically a causal effect or a comparison averaged over a population.

Model and Definition

Let \(Y_i\) be a binary outcome for observation \(i\), and let \(p_i=P(Y_i=1\mid \mathbf{x}_i)\). An additive logistic regression with \(q\) predictors is

Y_i\mid \mathbf{x}_i \sim \operatorname{Bernoulli}(p_i), \qquad \log\!\left(\frac{p_i}{1-p_i}\right) = \beta_0+\sum_{j=1}^{q}\beta_j x_{ij}.

Here \(\mathbf{x}_i=(x_{i1},\ldots,x_{iq})\) is the vector of predictor values, \(\beta_0\) is the intercept, and \(\beta_j\) is the coefficient for predictor \(j\). The model is additive: it contains no products between predictors. The odds at profile \(\mathbf{x}\) are \(p(\mathbf{x})/[1-p(\mathbf{x})]\).

Definition: For two covariate profiles \(\mathbf{x}\) and \(\mathbf{x}^{*}\), the model-based odds ratio comparing \(\mathbf{x}\) with \(\mathbf{x}^{*}\) is the odds at \(\mathbf{x}\) divided by the odds at \(\mathbf{x}^{*}\). It is adjusted for the other predictors when the profiles have the same values for those predictors and differ only in the predictor being compared.

For an estimated model, the fitted adjusted odds ratio is obtained by replacing the model coefficients \(\beta_j\) with their estimates \(\hat{\beta}_j\). The formula is the same for a population model and its fitted version; the distinction is whether it uses parameters or estimates.

The Coefficient-to-Odds-Ratio Result

The key calculation follows from the logistic model’s log-odds form. If two profiles differ only in predictor \(j\), by an amount \(\Delta\), then their linear predictors differ by \(\beta_j\Delta\). Exponentiating this difference gives their odds ratio.

Theorem: In the additive logistic model, compare profiles that have the same values of every predictor except \(x_j\), which changes from \(a\) to \(a+\Delta\). The odds at \(a+\Delta\), divided by the odds at \(a\), equal \(\exp(\beta_j\Delta)\). In particular, for a one-unit increase, the adjusted odds ratio is \(\exp(\beta_j)\).

Proof. Write the common values of the other predictors as \(c_k\), for \(k\ne j\). The linear predictor at the first profile is \(\eta_a=\beta_0+\beta_j a+\sum_{k\ne j}\beta_k c_k\). At the second profile it is \(\eta_{a+\Delta}=\beta_0+\beta_j(a+\Delta)+\sum_{k\ne j}\beta_k c_k\). Subtraction cancels the intercept and every shared predictor term, leaving \(\eta_{a+\Delta}-\eta_a=\beta_j\Delta\). Since the odds at any profile are the exponential of its linear predictor, the ratio of the odds is \(\exp(\eta_{a+\Delta})/\exp(\eta_a)=\exp(\eta_{a+\Delta}-\eta_a)=\exp(\beta_j\Delta)\). Setting \(\Delta=1\) gives \(\exp(\beta_j)\). \(\square\)

The same cancellation shows why this is a covariate-adjusted comparison. The values of the other predictors can change the odds at both profiles, but when they are held equal they do not change the ratio. Thus, in this additive model, the adjusted odds ratio for the same predictor contrast is constant across the fixed values of the other predictors.

Result: For a fixed change \(\Delta\) in predictor \(j\), the odds ratio \(\exp(\beta_j\Delta)\) does not depend on the common values of the other predictors. This invariance is a property of the stated additive model and contrast; it does not say that the odds or probabilities themselves are constant across covariate profiles.

Proof. For any chosen values \(c_k\) of the other predictors, the log-odds difference between the two profiles is \(\beta_j\Delta\), as shown above. This expression contains neither the \(c_k\) nor the intercept. Exponentiation therefore gives the same odds ratio for every choice of the shared values \(c_k\). \(\square\)

Worked Comparisons

Worked Example: Treatment Adjusted for Age and Smoking

Suppose a fitted model for a binary health outcome is

\log\!\left(\frac{p}{1-p}\right) = -2.00-0.70T+0.04A+0.80S,

where \(T=1\) denotes treatment and \(T=0\) denotes no treatment, \(A\) is age in years, and \(S=1\) denotes current smoking while \(S=0\) denotes no current smoking. Compare the modeled odds for treated and untreated people of the same age and smoking status.

The profiles differ only in \(T\), which changes from 0 to 1, so \(\Delta=1\) and the treatment coefficient is \(-0.70\). The adjusted odds ratio is

\frac{\text{odds}(T=1,A,S)}{\text{odds}(T=0,A,S)} = \exp(-0.70) \approx 0.4966.

At any fixed age and smoking status, the model therefore assigns the treated profile about \(0.50\) times the odds of the untreated profile, or about 50.3% lower odds because \(1-0.4966=0.5034\). This is a statement about odds, not a 50.3% decrease in probability. The age and smoking terms cancel in this comparison because those values are shared.

Worked Example: Changing a Continuous Predictor by Five Units

Suppose a logistic model includes exposure \(X\), measured in units of 1 mg/L, and other predictors such as age and sex. The fitted coefficient for \(X\) is \(\hat{\beta}_X=0.12\). Compare two otherwise identical profiles with exposure values 12 mg/L and 17 mg/L.

The exposure change is \(\Delta=17-12=5\) units. Therefore the fitted adjusted odds ratio is

\exp(\hat{\beta}_X\Delta) = \exp(0.12\times 5) = \exp(0.60) \approx 1.822.

At the same values of the other predictors, the model gives the 17 mg/L profile about 1.82 times the odds of the 12 mg/L profile. The coefficient itself corresponds to a one-unit increase: \(\exp(0.12)\approx1.128\). Applying that one-unit ratio across five units gives \(\{\exp(0.12)\}^{5}=\exp(0.60)\), the same result. The unit matters: an odds ratio per 5 mg/L is not the same numerical quantity as an odds ratio per 1 mg/L.

Worked Example: Comparing Two Nonreference Categories

Consider a model with hospital type coded using type A as the reference category. Let \(B\) and \(C\) be indicators for types B and C, and suppose

\log\!\left(\frac{p}{1-p}\right) = -1.20+0.55B-0.25C+0.03A_{\text{ge}}+0.60R,

where \(A_{\text{ge}}\) is age in years and \(R\) is an indicator for a risk factor. Compare type B with type C at the same age and risk-factor status. The type B profile has \((B,C)=(1,0)\), and type C has \((B,C)=(0,1)\). Their linear predictor difference is

(0.55-(-0.25)) + 0.03(A_{\text{ge}}-A_{\text{ge}}) + 0.60(R-R) =0.80.

Thus the adjusted odds ratio for type B versus type C is

\frac{\text{odds}(\text{type B},A_{\text{ge}},R)} {\text{odds}(\text{type C},A_{\text{ge}},R)} = \exp(0.80) \approx 2.226.

The type B profile has about 2.23 times the modeled odds of the type C profile at the same age and risk-factor status. Because neither category is the reference here, the relevant log-odds difference is the B coefficient minus the C coefficient, not either coefficient alone.

Using the Interpretation Carefully

“Adjusted for age and smoking,” for example, means that the model comparison holds age and smoking fixed. It does not mean that the two groups actually have identical age and smoking distributions, nor does the phrase by itself establish that treatment caused an outcome difference. A causal interpretation requires additional assumptions about study design and confounding that are not supplied by the odds-ratio formula.

The model also determines what contrast a coefficient represents. For a continuous predictor measured in years, \(\exp(\beta_j)\) is the adjusted odds ratio per year; a ten-year contrast uses \(\exp(10\beta_j)\). For a binary indicator, it compares the coded value 1 with the coded value 0 while other predictors are fixed. For a categorical predictor, a coefficient usually compares one category with its reference; comparing two nonreference categories requires subtracting their coefficients before exponentiating.

A frequent mistake is to report \(\exp(\hat{\beta}_j)\) as though it compared any two people, regardless of their other characteristics. In this model it describes the specified predictor change while the other predictors are held fixed. Another mistake is to describe an odds ratio as a probability ratio or a percentage-point change. Odds and probabilities are related, but they are not the same scale. The adjusted odds ratio alone does not give the probability difference between the profiles.

Key takeaway: In an additive logistic model, compare profiles that differ in one predictor and match on the others. A change of \(\Delta\) in predictor \(j\) has adjusted odds ratio \(\exp(\beta_j\Delta)\); for fitted comparisons, use \(\exp(\hat{\beta}_j\Delta)\).

Check Your Understanding

For each question, identify the comparison being made and state which other predictors are held fixed.

A model coefficient for a binary predictor is \(-0.4\). What is the adjusted odds ratio for coded value 1 versus coded value 0?
A continuous predictor has coefficient \(0.08\) per unit. What is the adjusted odds ratio for a four-unit increase?
A categorical predictor has coefficients \(0.3\) for category B and \(-0.2\) for category C, with category A as reference. What is the adjusted odds ratio for B versus C?
In the additive logistic model, why does changing the common values of the other predictors not change the odds ratio for a fixed contrast in predictor \(j\)?
Why does an adjusted odds ratio of 1.5 not, by itself, mean that one profile’s probability is 1.5 times the other’s probability?