Why Standardize Across Periods?
In a platform trial, the participants enrolled in one period may differ from those enrolled in another. Their baseline risk factors may be distributed differently, and outcomes may also vary with calendar time as clinical practice or other conditions change. A treatment-effect estimate for each period can therefore reflect both the arm contrast and the population or time profile over which that contrast was estimated.
Model-based standardization makes the comparison explicit: predict outcomes under each arm in each period, then average those predictions over the same stated target distribution of baseline covariates and calendar time. The resulting contrasts answer a defined question about a common target, rather than comparing estimates that average over different observed participant mixes.
The choice of \(Q\) is part of the estimand, not a cosmetic calculation. It might represent the pooled eligible population in a prespecified set of periods, a clinically relevant population, or a clearly described covariate-and-time profile. A difference between \(\Delta_p(Q)\) and \(\Delta_{p'}(Q)\) compares period-specific effects standardized to that same target. It does not, by itself, establish why effects differ.
Model, Assumptions, and Calculation
A regression model supplies predictions for \(m(a,p,x,t)\). It should represent the outcome scale of interest and allow relevant relationships to vary. For example, a model may include treatment-by-period terms, baseline covariates, a flexible function of calendar time, and interactions when treatment effects plausibly vary by covariate or time. A treatment-by-period model without relevant covariate interactions may impose a common conditional effect across covariate profiles; standardizing such a model will not remove that restriction.
For each target profile, predict the outcome twice under each period: once setting \(A=1\) and once setting \(A=0\). Average each set of predictions using the target distribution \(Q\), then take the contrast. With a finite target sample of \(N\) profiles \((x_i,t_i)\) and target weights \(q_i\), where \(q_i\geq 0\) and \(\sum_iq_i=1\), the plug-in estimate is:
A causal interpretation requires more than a well-fitted regression. In each period and over the target support, the treatment groups must be exchangeable conditional on the measured baseline covariates and calendar time; both arms must have a positive probability of being observed for the profiles being standardized; treatment, period, covariates, and outcome must be defined consistently; and the outcome model must estimate the relevant conditional means adequately. In notation, the exchangeability requirement can be expressed as \(Y(a)\perp A\mid(P,X,T)\), for \(a\in\{0,1\}\), in the target population. If these conditions are not credible, the standardized contrast remains a model-based adjusted comparison, not necessarily a causal effect.
The calendar-time part of \(Q\) needs particular care. If periods occupy disjoint calendar ranges, a common distribution of actual calendar dates may have no support in both periods. A model prediction for an unsupported date is extrapolation, not alignment supplied by the data. A different time scale, such as time relative to an arm's opening, may be useful only if it represents the scientific target and is stated explicitly; it does not automatically account for secular changes in actual calendar time.
Why the Standardization Formula Works
The justification is the law of total expectation applied to the target distribution. Under the model, the expected outcome in arm \(a\) and period \(p\), for a population whose covariates and time follow \(Q\), is the average of the conditional mean over those profiles. Subtracting the two arm-specific averages yields the target-population mean difference:
The equality to a causal contrast uses the exchangeability, consistency, positivity, and model conditions just described. The formula also shows why the same \(Q\) matters: if each period instead uses its own distribution, the difference between period estimates can combine treatment-effect variation with a change in the profiles being averaged.
Worked Example: Standardizing Continuous Outcomes Across Two Periods
Suppose a hypothetical trial models a 12-week symptom score, in points, for a treatment arm and its reference arm. Lower scores are better. The model includes treatment-by-period and treatment-by-baseline-severity terms, as well as calendar time. For illustration, consider two baseline-severity profiles and a common target distribution assigning probability 0.60 to the lower-severity profile and 0.40 to the higher-severity profile. The model predictions below are assumed to be evaluated at the same prespecified target time profile in both periods.
| Period | Baseline profile | Target proportion | Predicted reference mean | Predicted treatment mean |
|---|---|---|---|---|
| 1 | Lower severity | 0.60 | 20 | 16 |
| 1 | Higher severity | 0.40 | 30 | 25 |
| 2 | Lower severity | 0.60 | 22 | 18 |
| 2 | Higher severity | 0.40 | 32 | 29 |
In period 1, the standardized reference mean is \(0.60(20)+0.40(30)=12+12=24\) points. The standardized treatment mean is \(0.60(16)+0.40(25)=9.6+10=19.6\) points. Therefore, \(\widehat{\Delta}_1(Q)=19.6-24=-4.4\) points.
In period 2, the standardized reference mean is \(0.60(22)+0.40(32)=13.2+12.8=26\) points. The standardized treatment mean is \(0.60(18)+0.40(29)=10.8+11.6=22.4\) points. Thus, \(\widehat{\Delta}_2(Q)=22.4-26=-3.6\) points. The standardized effects differ by \(-3.6-(-4.4)=0.8\) points: in this hypothetical example, the estimated treatment benefit is 0.8 points smaller in period 2 on the common target. This is a descriptive comparison of estimated effects; uncertainty and the assumptions needed for causal interpretation still matter.
The calculation assumes that the model predictions are adequate for these profiles and that the target time profile is meaningful and supported in both periods. If the target includes time points outside a period's observed support, these averages depend on extrapolation.
Outcome Scale and Effect Modification
Standardization can produce contrasts on several scales. For a binary outcome, predict each arm's probability and average probabilities over \(Q\); their difference is a standardized risk difference. A standardized risk ratio is the ratio of the two standardized risks. These are population-average contrasts, and they need not equal a model coefficient. In particular, averaging probabilities from a logistic model is not the same operation as averaging linear predictors and converting the resulting predictor to a probability.
Worked Example: Standardized Risk Difference for a Binary Outcome
Consider a hypothetical binary outcome: experiencing a specified adverse event by 30 days. A model provides the following predicted risks for two baseline-risk profiles. Suppose the common target assigns equal proportions to the profiles, and the predictions are evaluated at a supported, prespecified time profile.
| Period | Profile | Target proportion | Reference risk | Treatment risk |
|---|---|---|---|---|
| 1 | Lower baseline risk | 0.50 | 0.10 | 0.08 |
| 1 | Higher baseline risk | 0.50 | 0.40 | 0.32 |
| 2 | Lower baseline risk | 0.50 | 0.15 | 0.13 |
| 2 | Higher baseline risk | 0.50 | 0.45 | 0.41 |
For period 1, the standardized treatment risk is \(0.50(0.08)+0.50(0.32)=0.20\), and the standardized reference risk is \(0.50(0.10)+0.50(0.40)=0.25\). The risk difference is \(0.20-0.25=-0.05\), or 5 fewer events per 100 people in the target population.
For period 2, the standardized treatment risk is \(0.50(0.13)+0.50(0.41)=0.27\), and the standardized reference risk is \(0.50(0.15)+0.50(0.45)=0.30\). The risk difference is \(0.27-0.30=-0.03\), or 3 fewer events per 100 people. The standardized risk-difference contrast between periods is \(-0.03-(-0.05)=0.02\), a 2-per-100 smaller estimated reduction in period 2. The calculation averages risks, not odds ratios, and the interpretation is tied to the chosen target distribution.
These predicted probabilities must be calibrated well enough for the target profiles, and the profiles must be supported by data from both arms in each period for a data-based causal interpretation. If a profile is absent from one treatment group in a period, the model may still produce a number, but that number relies on extrapolation.
Reading the Comparison and Checking Its Weak Points
A treatment-by-period coefficient describes how a model's conditional treatment contrast changes between periods at the model's reference covariate values (or under its parameterization). It is not generally the same as the difference between two standardized effects. When treatment effects vary with baseline covariates or time, standardization averages those conditional contrasts over \(Q\); changing \(Q\) can change the answer. Even without such effect modification, the outcome scale matters, especially for nonlinear models.
A crude comparison instead averages within each period's observed participant mix. It can be useful descriptively, but it does not answer the common-target question if those mixes differ. Standardization uses an outcome model to transport predictions to the target profiles. Weighting is a different broad strategy: it changes the contribution of observed participants to represent a target distribution. Neither approach makes unsupported profiles appear in the data, and the two approaches rely on assumptions that must be assessed in context.
Useful diagnostics include comparing target profiles with the covariate and time ranges actually represented in each arm-period group; examining predicted versus observed outcomes across relevant profiles; checking calibration and functional forms for calendar time; and assessing whether treatment-effect interactions are clinically plausible. A summary fit statistic alone cannot establish that predictions are reliable for the target. Report the target distribution, prediction model, outcome scale, and period-specific estimates so that the comparison is interpretable.
Uncertainty must account for estimating the model and forming the standardized predictions. A bootstrap that resamples the relevant participants and refits the model is one practical method; a delta-method calculation is another. If \(Q\) is intended as a fixed target, hold its definition fixed across resamples. If the target itself represents a sampled population and its sampling uncertainty is part of the inferential goal, the procedure should also reflect that uncertainty. An interval for a standardized effect quantifies uncertainty under the chosen model and design assumptions; it does not account automatically for unmeasured confounding or unsupported extrapolation.
Worked Example: A Target That Is Not Supported in Both Periods
Suppose a hypothetical platform has observed calendar months 1–6 in period 1 and months 9–14 in period 2. A proposed common target assigns equal probability to months 4 and 11. Month 4 is represented in period 1 but not period 2; month 11 is represented in period 2 but not period 1. A model with a linear calendar-time term predicts standardized outcomes for both periods at both months, but this target requires extrapolation in each period.
For example, suppose the model predicts treatment-minus-reference differences of \(-4\) points at month 4 and \(-2\) points at month 11 for period 1, and \(-3\) points at month 4 and \(-1\) point at month 11 for period 2. The stated target gives \(\widehat{\Delta}_1(Q)=0.50(-4)+0.50(-2)=-3\) points and \(\widehat{\Delta}_2(Q)=0.50(-3)+0.50(-1)=-2\) points. The calculated difference is 1 point.
The arithmetic is correct, but neither period has observed support across both target months. The period comparison therefore depends on the assumed time-model form, not just on the observed data. A defensible response is to choose a scientifically relevant target supported in both periods, if one exists, and state the resulting narrower target. If no common supported time target exists, report that this standardized comparison is not identified by the observed support without stronger modeling assumptions; do not present the extrapolated contrast as though it were directly supported.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For a target with 75% lower-severity and 25% higher-severity profiles, predicted treatment means are 12 and 20 points and predicted reference means are 15 and 24 points. Calculate the standardized mean difference.
- Why should the same target distribution be used to compare period-specific effects?
- A binary-outcome model predicts treatment risks of 0.10 and 0.30 and reference risks of 0.15 and 0.35 in two equally represented target profiles. Calculate the standardized risk difference.
- What is the main concern if a target calendar month is absent from one period's observed data?
- Does including a treatment-by-period interaction in the model automatically make its coefficient equal to the difference between standardized effects? Explain.
Solutions. (1) The treatment mean is \(0.75(12)+0.25(20)=9+5=14\) points. The reference mean is \(0.75(15)+0.25(24)=11.25+6=17.25\) points. The standardized difference is \(14-17.25=-3.25\) points. (2) A shared target prevents a change in the covariate-and-time mix from being conflated with a change in the period-specific effect. (3) The treatment risk is \(0.50(0.10)+0.50(0.30)=0.20\); the reference risk is \(0.50(0.15)+0.50(0.35)=0.25\). The standardized risk difference is \(-0.05\), or 5 fewer events per 100 people. (4) Predictions at that month may require extrapolation, so the estimate is driven by model assumptions rather than direct support in that period. (5) No. The interaction coefficient is conditional on the model's parameterization and reference profile; standardized effects average predicted outcomes over the specified target distribution.