Why the Effect Scale Matters in Transport
When effects are transported to a target population, the target distribution of baseline covariates determines how subgroup risks are combined. The resulting marginal effect depends not only on the conditional effects within covariate strata, but also on the effect scale used for comparison. This matters especially for odds ratios: even if the conditional odds ratio is the same in every stratum, the marginal odds ratio can differ from it and can change when the target population’s covariate distribution changes.
The key is to distinguish two questions. Effect modification asks whether the treatment effect differs across strata on a specified scale. Noncollapsibility describes a difference between conditional and marginal measures that can occur even without confounding and even when the conditional odds ratio is constant. Neither concept says that a particular effect scale is universally preferable. The scale should match the scientific question and the decision being informed.
Standardized Marginal Effects
Let \(A\) be a binary treatment, \(Y\) a binary outcome, and \(X\) a set of baseline covariates. For a target population, write \(q(x)=P_{\text{target}}(X=x)\). Let \(r_a(x)=P(Y^a=1\mid X=x)\) be the outcome risk under treatment level \(a\) among people with covariates \(x\). The target-population risk under treatment level \(a\) is the average of these stratum-specific risks using the target distribution.
For continuous covariates, the sum is replaced by an integral. The marginal causal risk difference, risk ratio, and odds ratio are then defined from the two standardized risks:
These are contrasts of target-population risks, not averages of stratum-specific contrasts. The risks \(r_a(x)\) must be identified from suitable causal assumptions—for example, appropriate exchangeability, consistency, and positivity conditions. Once the stratum-specific risks are identified, standardization combines them using \(q(x)\). A difference between marginal and conditional effects is not, by itself, evidence of confounding.
A Constant Conditional Odds Ratio Can Produce a Different Marginal Odds Ratio
Suppose the conditional odds ratio is \(\theta\) in every stratum. If the untreated risk in a stratum is \(p_0(x)\), the treated risk implied by that odds ratio is
The transformation from \(p_0(x)\) to \(p_1(x)\) is nonlinear. Consequently, averaging the stratum-specific risks first and then calculating an odds ratio generally does not give \(\theta\). In contrast, a constant conditional risk difference or a constant conditional risk ratio does pass through the corresponding averaging operation, as shown below.
Worked Example: Same Conditional Odds Ratio, Different Target Marginal Odds Ratios
Consider two baseline-risk strata, \(X=0\) and \(X=1\). Suppose their untreated risks are 0.10 and 0.50, respectively, and the conditional odds ratio comparing treatment with control is 2 in both strata. These are hypothetical risks. The treated risks are \(2(0.10)/(0.90+2(0.10))=0.1818\) and \(2(0.50)/(0.50+2(0.50))=0.6667\). Thus the conditional odds ratio is 2 in each stratum.
First consider a target population in which 20% are in stratum \(X=1\), so \(q(1)=0.20\) and \(q(0)=0.80\). Its standardized control risk is
The standardized treatment risk is
Therefore, the target risk difference is \(0.2788-0.1800=0.09879\), or 9.879 percentage points. The target risk ratio is \(0.2788/0.1800=1.549\). The target odds ratio is
Now change the target distribution so that 80% are in stratum \(X=1\). Then
The target risk difference is \(0.5697-0.4200=0.1497\), the risk ratio is \(0.5697/0.4200=1.356\), and the odds ratio is
The conditional odds ratio remains 2 in both strata and under both target distributions, but the marginal odds ratio changes from 1.761 to 1.828. The risk difference and risk ratio change too: here the conditional risk differences (0.08182 and 0.1667) and conditional risk ratios (1.818 and 1.333) vary across strata. Their marginal changes therefore also reflect effect modification on those scales. This example shows why a change in a marginal effect across target populations should be interpreted with both the effect scale and the stratum-specific risks in view.
Noncollapsibility Is Not Confounding
The odds ratio’s conditional-to-marginal difference can be seen even when the covariate distribution is identical in the treated and control groups. In a causal calculation, both standardized risks use the same target weights \(q(x)\); that is exactly what the example below does. Thus the discrepancy between the conditional and marginal odds ratios is not created by treatment groups having different covariate compositions.
Worked Example: Marginalizing With One Fixed Target Population
Use the two strata and risks from the previous example, but let the target population be half in each stratum. The conditional odds ratio is 2 in both strata. The standardized risks are
The corresponding marginal odds ratio is
This marginal odds ratio is not 2, although the conditional odds ratio is exactly 2 in each stratum. The calculation averages treatment-specific risks over the same covariate distribution before taking the odds ratio. That order of operations is essential: averaging conditional odds ratios is not the definition of the marginal odds ratio.
When Risk Differences and Risk Ratios Collapse
A measure is collapsible over \(X\), under a stated homogeneity condition, if its common conditional value equals the marginal value formed using the same weights for both treatment levels. For the risk difference, if \(r_1(x)-r_0(x)=\delta\) for every stratum, then
For the risk ratio, if \(r_1(x)=\rho r_0(x)\) in every stratum and \(R_0^{\text{target}}>0\), then
These results also explain an important qualification: risk differences and risk ratios are not guaranteed to remain unchanged across target populations. If their corresponding conditional effects vary across strata, changing \(q(x)\) can change the marginal measure. The collapsibility results apply when the relevant conditional effect is homogeneous on that scale.
Worked Example: A Homogeneous Risk Ratio Across Two Targets
Suppose the control risks are 0.10 in stratum \(X=0\) and 0.40 in stratum \(X=1\), and treatment multiplies each risk by 1.5. The treated risks are therefore 0.15 and 0.60. The conditional risk ratio is 1.5 in each stratum, while the conditional risk differences are 0.05 and 0.20.
For a target with \(q(1)=0.20\), the standardized risks are \(R_0^{\text{target}}=0.80(0.10)+0.20(0.40)=0.1600\) and \(R_1^{\text{target}}=0.80(0.15)+0.20(0.60)=0.2400\). The marginal risk ratio is \(0.2400/0.1600=1.500\), and the marginal risk difference is \(0.2400-0.1600=0.08000\).
For a target with \(q(1)=0.80\), the risks are \(R_0^{\text{target}}=0.20(0.10)+0.80(0.40)=0.3400\) and \(R_1^{\text{target}}=0.20(0.15)+0.80(0.60)=0.5100\). The marginal risk ratio remains \(0.5100/0.3400=1.500\), but the marginal risk difference is \(0.5100-0.3400=0.1700\). The risk ratio is invariant here because it is homogeneous across strata; the risk difference changes because its conditional values differ.
Practical Interpretation in Transport
For transport, first specify the target population and the effect scale. Then standardize the treatment-specific risks to the target covariate distribution and calculate the chosen contrast. A conditional odds ratio from a regression model answers a conditional question; it is not automatically the marginal odds ratio for the target. Conversely, a marginal effect can change between targets even when the conditional odds ratio is stable. These are differences in estimands, not necessarily contradictions in the analysis.
The odds ratio also depends strongly on baseline risk. For the same conditional odds ratio, the implied risk difference and risk ratio can differ across strata, and a target with more people in high-risk strata can have different marginal risks and contrasts. In reporting results, state whether the measure is conditional or marginal, identify the population over which marginal risks were standardized, and give the scale and units. Risk differences are naturally expressed in percentage points; risk ratios and odds ratios are unitless.
Common Mistake
- Wrong reasoning: “The conditional odds ratio is 2 in every stratum, so the target-population odds ratio must be 2. If it is not 2, there must be confounding.”
- Correction: A common conditional odds ratio does not generally equal the marginal odds ratio. Standardize both treatment-specific risks to the same target distribution, then calculate the marginal odds ratio from those risks. The conditional and marginal odds ratios answer different questions, and their difference can arise from noncollapsibility without confounding.
Check Your Understanding
Use the definitions and calculations above to check your understanding.
- What two target-population risks are needed to calculate a marginal odds ratio?
- Why does a constant conditional odds ratio not generally equal the target-population marginal odds ratio?
- State the homogeneity condition under which a conditional risk difference equals the marginal risk difference.
- If the conditional risk ratio is constant across strata, what happens to the marginal risk ratio when the target covariate distribution changes? State the condition needed for this result.
- Why is a difference between conditional and marginal odds ratios not, by itself, evidence of confounding?