Population-Level Summaries and Treatment Contrasts

From Treatment Conditions to Population Summaries

A treatment contrast is not fully described by saying that one treatment is being compared with another. The outcome, the population over which outcomes are summarized, and the treatment conditions all matter. For example, comparing mean symptom scores and comparing the probabilities of clinical improvement answer different questions, even if both use the same trial participants.

Here, let \(a=1\) denote one treatment condition and \(a=0\) the comparator. The outcome under each condition is written \(Y(a)\). It is understood to be the outcome at the specified time point, defined according to the chosen handling of any intercurrent events. The two treatment conditions and that handling must be specified before calculating a summary. Let \(F\) denote the distribution of individuals in the target population. A population-level summary averages or otherwise summarizes outcomes under a treatment condition across this same population.

Definition: The marginal mean under treatment condition \(a\) is \(\mu_a=E_F[Y(a)]\), provided the mean exists. For a binary outcome coded 1 for the event and 0 otherwise, the marginal risk is \(p_a=P_F(Y(a)=1)\). “Marginal” means that the summary averages across the target population’s distribution of characteristics rather than fixing those characteristics at particular values.

The subscript \(F\) emphasizes that the population distribution is part of the definition. The means \(\mu_1\) and \(\mu_0\), or risks \(p_1\) and \(p_0\), summarize outcomes under two conditions for the same target population. Comparing summaries that average over different populations can combine treatment differences with population-composition differences.

These are population quantities, not calculations from a particular sample. Sample means and proportions can be used as estimates of such quantities, but the definitions themselves do not say how to estimate them or how uncertain an estimate is. Defining the population summary also does not, by itself, establish that available data identify it.

Three Common Treatment Contrasts

A contrast compares the two treatment-specific summaries. For a quantitative outcome, the mean difference is

\Delta_{\text{mean}}=\mu_1-\mu_0.

Its units are the units of the outcome. A negative value means the mean under condition 1 is lower than under condition 0; whether that is better depends on the outcome’s direction. For a binary outcome, the risk difference is

\Delta_{\text{risk}}=p_1-p_0.

The risk difference is an absolute contrast, expressed as a proportion or percentage-point difference. A value of \(-0.08\), for example, is an eight-percentage-point lower risk under condition 1. It is not an eight percent relative reduction.

A ratio compares the summaries multiplicatively. Common choices are the ratio of means and the risk ratio:

R_{\text{mean}}=\frac{\mu_1}{\mu_0}, \qquad R_{\text{risk}}=\frac{p_1}{p_0}.

A ratio is dimensionless, and a value of 1 indicates equal summaries when the ratio is defined. A risk ratio of 0.60 means that the risk under condition 1 is 60% of the risk under condition 0, or 40% lower relative to the comparator risk. A ratio requires a nonzero denominator; the risk ratio also requires \(p_0>0\). Ratios of means are usually most interpretable when the outcome has a meaningful zero and both means are positive. A difference and a ratio can therefore tell complementary things: the difference gives an absolute change, while the ratio gives a relative change.

Worked Example: A Mean Difference in Symptom Score

Suppose a hypothetical target population has a mean symptom score of 8.1 points under the comparator and 6.8 points under the new treatment, with lower scores indicating fewer symptoms. These are stipulated population means, not sample estimates. The population-level mean difference is

\Delta_{\text{mean}}=\mu_1-\mu_0 =6.8\text{ points}-8.1\text{ points} =-1.3\text{ points}.

Under the specified treatment conditions, the population mean is 1.3 points lower with the new treatment. The units remain points because the contrast is a difference of means. The sign is favorable for this outcome because lower scores indicate fewer symptoms. This calculation alone does not say how precisely the means are estimated or whether the difference is statistically distinguishable from zero.

Worked Example: Absolute and Relative Risks

Suppose, hypothetically, that the population risk of an event is \(p_0=0.25\) under the comparator and \(p_1=0.15\) under treatment 1. The risk difference is

\Delta_{\text{risk}}=p_1-p_0 =0.15-0.25 =-0.10.

Thus, the event risk is 0.10 lower, or 10 percentage points lower, under treatment 1. The risk ratio is

R_{\text{risk}}=\frac{p_1}{p_0} =\frac{0.15}{0.25} =0.60.

The risk under treatment 1 is 0.60 times the comparator risk, corresponding to a 40% relative reduction because \(1-0.60=0.40\). Both statements describe the same pair of risks on different scales. In particular, “10% lower” would be ambiguous and would not correctly state the absolute difference: the absolute change is 10 percentage points.

Marginal Means Average Over the Population

When outcome means differ across population subgroups, a marginal mean combines subgroup-specific means using the proportions of those subgroups in the target population. If \(S\) indexes a finite set of strata, with target-population weights \(w_s=P_F(S=s)\), then

\mu_a=\sum_s w_s E_F[Y(a)\mid S=s], \qquad \sum_s w_s=1.

The same weights must be used for both treatment conditions when the goal is to compare their marginal means in one target population. The weights are not automatically one-half each, and they need not equal the proportions in a particular study sample if that sample does not represent the target population.

Worked Example: Standardizing to a Shared Population Mix

Consider a hypothetical target population that is 75% mildly affected and 25% severely affected. The following values are the stratum-specific population mean outcomes under each treatment condition, in points.

Stratum	Target-population weight	Mean under comparator	Mean under treatment 1
Mild	0.75	4	3
Severe	0.25	12	9

The marginal mean under the comparator is

\mu_0=(0.75)(4)+(0.25)(12)=3+3=6\text{ points}.

The marginal mean under treatment 1 is

\mu_1=(0.75)(3)+(0.25)(9)=2.25+2.25=4.5\text{ points}.

Therefore, the marginal mean difference is \(4.5-6=-1.5\) points. The severity-specific differences are \(-1\) point in the mild stratum and \(-3\) points in the severe stratum. Weighting those differences by the target distribution gives

(0.75)(-1)+(0.25)(-3)=-0.75-0.75=-1.5\text{ points}.

A simple average of the two stratum differences would be \((-1-3)/2=-2\) points. That is not the marginal contrast for this target population because it gives the mild and severe strata equal weight, contrary to the specified 75%/25% population mix. The shared weights make clear which population the marginal summary describes.

Justification and Conditions for Interpretation

The weighted calculation follows from the law of total expectation. For a treatment condition \(a\), partition the population into strata \(S=s\). The population mean is the sum of each stratum’s conditional mean multiplied by its population proportion:

E_F[Y(a)] =\sum_s E_F[Y(a)\mid S=s]P_F(S=s) =\sum_s w_s E_F[Y(a)\mid S=s].

Subtracting the corresponding expression for condition 0 from that for condition 1 gives

\mu_1-\mu_0 =\sum_s w_s\left\{E_F[Y(1)\mid S=s]-E_F[Y(0)\mid S=s]\right\}.

This identity requires the means to exist and the strata to cover the target population. It does not require treatment effects to be equal across strata. It shows that the marginal mean difference is a target-population-weighted average of the stratum-specific mean differences, using the same weights for both conditions.

The population summaries are defined once the outcome, treatment conditions, target population, and treatment-strategy handling of intercurrent events are specified. Interpreting a contrast as causal requires a causal estimand: \(Y(1)\) and \(Y(0)\) must refer to outcomes under well-defined treatment conditions for the population of interest. Estimating that causal contrast from observed data requires additional identification and estimation considerations. A contrast’s definition should not be mistaken for evidence that the data identify it.

Choosing and Checking a Summary

A summary should fit the outcome and the question. The mean difference is natural for a quantitative outcome when additive change in its units is meaningful. The risk difference describes an absolute change in event probability, while the risk ratio describes the event risk relative to the comparator. Neither scale is universally preferable; they answer distinct questions. The introductory material on treatment-policy and other estimand strategies addresses how the treatment conditions incorporate intercurrent events; the summaries here apply after those conditions have been specified.

Before reporting a contrast, check the following:

Outcome and time point: Are both treatment-specific summaries for the same outcome definition, scale, and time point?
Population: Do both summaries use the same target-population distribution? If stratum weights are used, do they sum to 1 and represent the intended population?
Direction: Does a positive or negative difference indicate a favorable outcome, given how the outcome is coded?
Scale and units: Is the result an absolute difference with outcome units, a risk difference in percentage points, or a dimensionless ratio?
Denominator: Is the denominator of a ratio nonzero, and is a ratio a meaningful summary for this outcome scale?

A common mistake is to compare a treatment mean averaged over one subgroup mix with a comparator mean averaged over a different mix, then call the difference a treatment contrast for a shared population. The arithmetic may be correct, but the comparison no longer holds population composition fixed. The correction is to define the target population and apply its weights to both treatment conditions, as in the standardization example.

Another pitfall is to treat a population-level summary as if it described every individual or subgroup. A marginal mean difference does not imply that every subgroup has the same difference, and a risk ratio does not specify either risk on its own. Marginal summaries compactly describe outcomes across a population; they do not replace subgroup-specific information when that information is the question.

Key takeaway: Define both treatment-specific summaries in the same target population, then state the contrast on its chosen scale. A mean difference, risk difference, and ratio summarize different aspects of the same pair of treatment conditions, so report the scale, units, and direction explicitly.

Check Your Understanding

Answer each question using the definitions and calculations above. Solutions follow the questions.

A hypothetical outcome has population means of 14 points under the comparator and 11 points under treatment 1. What is the mean difference, with units and direction?
Suppose the event risks are 0.08 under treatment 1 and 0.20 under the comparator. Calculate the risk difference and risk ratio, and interpret both.
A target population has two strata with weights 0.60 and 0.40. Under treatment 1, their conditional means are 5 and 10. Calculate the marginal mean.
Why is an equally weighted average of stratum-specific effects not necessarily the marginal effect for a target population?

Solutions. (1) The mean difference is \(11-14=-3\) points, so the treatment 1 population mean is 3 points lower. (2) The risk difference is \(0.08-0.20=-0.12\), or 12 percentage points lower. The risk ratio is \(0.08/0.20=0.40\), so treatment 1 has 40% of the comparator risk, a 60% relative reduction. (3) The marginal mean is \((0.60)(5)+(0.40)(10)=3+4=7\) points. (4) A marginal effect weights stratum-specific effects by the target population’s stratum proportions. Equal weighting gives the target marginal effect only when those proportions are equal or when the particular effects make the two weighted averages coincide.