Time-Varying Treatment Dose: When Recovery Gets Counted as the Intervention
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
Time-varying treatment dose looks simple after the fact. Add the minutes of mobilization, cumulative medication, delivered nutrition, dialysis intensity, or therapy sessions. Divide patients into high- and low-dose groups. Compare outcomes. The trouble is that the patient has been changing while the dose was accumulating.
A patient who stabilizes early can tolerate more treatment. A patient who deteriorates may receive less, stop treatment, transfer, or die before accumulating it. The final dose therefore records both the care delivered and the clinical path that made delivery possible. If we treat that final summary like a baseline assignment, recovery starts masquerading as the intervention.
The Clean Metaphor: Do Not Grade a Journey With Its Finish-Line Label
A longitudinal dose is a film, not a photograph.
Collapsing the whole film into “high dose” or “low dose” uses later frames to label the opening scene. A fair analysis must preserve when each treatment decision became possible, what clinicians knew then, and which patients were still at risk.
Interactive design audit
Build the Treatment-Dose Comparison
Configure a hypothetical longitudinal dose study. The audit does not score a paper; it shows which causal questions appear when treatment accumulates while patients are changing.
Treat the result as an association until the timeline and target strategy are repaired.
4 of 4 structural warnings active. A lower count is not proof of no bias; it means the protocol has stopped creating these particular problems by definition.
Future information enters the exposure label
Dose depends on treatment and survival accumulated after time zero. Earlier recovery can create more opportunity to enter the high-dose group.
Evolving clinical state remains a common cause
Readiness, shock, ventilation, delirium, and weakness can affect both tomorrow’s dose and the outcome. Baseline matching cannot balance changes that happen later.
Selection changes the population being compared
Removing early deaths or transfers can preferentially retain stable patients and make the dose contrast conditional on post-baseline survival or observation.
The intervention is still descriptive
“High dose” is not assignable until the protocol says how dose responds to instability, intolerance, weekends, and missed sessions.
Why Baseline Matching Cannot Balance Tomorrow
Propensity-score matching can balance measured baseline variables between exposure groups. It cannot, by itself, balance clinical changes that occur after baseline and influence later treatment. In an ICU, today's hemodynamics, ventilation, delirium, weakness, and clinician assessment may determine both tomorrow's mobilization and eventual function. In medication studies, symptoms, laboratory values, and adverse effects can play the same role.
This is time-varying confounding. It becomes especially difficult when earlier treatment also changes the confounder. Standard regression adjustment for that updated variable can block part of the treatment pathway or introduce other bias. G-methods—including the parametric g-formula and inverse-probability weighted marginal structural models—were developed for longitudinal treatment strategies under explicit assumptions. Their names are not the quality gate. The protocol, timing, measurement, positivity, and diagnostics are.
Three Biases That Often Travel Together
Reverse causation
Improving health permits more treatment, so recovery predicts dose before dose is credited for recovery.
Guarantee-time bias
Entering a high cumulative-dose group may require surviving and remaining observable long enough to accrue it.
Treatment-confounder feedback
Earlier treatment changes a later variable that influences both subsequent treatment and outcome.
These are related but not interchangeable. Restricting dose to an early window may reduce reverse causation, yet still grant guarantee time inside that window. Matching baseline severity does not solve daily treatment-confounder feedback. Excluding early deaths may make the timeline look cleaner while changing the target population through post-baseline selection.
A Current Example: Mobilization After Cardiovascular Surgery
A 2026 single-center retrospective cohort classified 610 postoperative cardiovascular patients by their average daily dose of out-of-bed mobilization during ICU stay. The authors reported more functional decline in the lower-dose group after baseline propensity-score matching, while carefully describing the result as an association. They also acknowledged that dose accumulated over time and reflected evolving clinical status.
A subsequent methodological commentary sharpened the causal concern: whole-stay average dose was analyzed as a between-group exposure even though stability and recovery influenced the opportunity to accumulate it. The commentary also questioned baseline-only adjustment, exclusions of death and early transfer, and whether the stated target trial was specified completely. This does not prove the association is false. It changes what the study can safely claim: a useful hypothesis about mobilization, not yet a clean effect of assigning one longitudinal dose strategy rather than another.
Write the Strategy Before Choosing the Estimator
| Protocol element | Question to specify | Common shortcut |
|---|---|---|
| Decision times | When can dose be assigned or updated? | Use one final exposure label |
| Dynamic rule | How should dose respond to readiness or safety? | Compare observed high versus low |
| Time-varying state | What predicts both later dose and outcome? | Adjust baseline severity only |
| Early events | How do death, discharge, transfer, or intolerance enter? | Delete or relabel them |
| Positivity | Could each strategy occur at each relevant patient state? | Report a baseline propensity overlap plot |
Reviewer Red-Flag Checklist
- The exposure is calculated over the entire follow-up but analyzed as if assigned at baseline.
- Patients must survive, remain admitted, or avoid transfer to qualify for the higher-dose category.
- Only baseline covariates are adjusted despite repeated treatment decisions.
- Post-baseline recovery markers enter an ordinary regression without a causal role being stated.
- “High dose” describes observed care but not a feasible rule clinicians could implement.
- An early-window sensitivity analysis is presented as solving every timing problem.
- Death and functional decline are merged after seeing inconvenient missing outcomes.
- No treatment or censoring weights, overlap diagnostics, or longitudinal measurement schedule are shown.
Why This Matters for Aqrab
Longitudinal studies often contain sophisticated models and familiar balance tables while the exposure definition quietly uses the patient's future. A useful methods critique reconstructs the clinical decision sequence before judging the estimator. That is where causal credibility is usually won or lost.
Use Aqrab Try to test whether a study's treatment clock, eligibility, confounder updates, and outcome handling support the claim being made. Teams building structured review workflows can explore the developer tools.
Methods Sources
- Hirakawa K, et al. Dose–response relationship of early out-of-bed mobilization with functional decline and adverse events after cardiovascular surgery. Journal of Intensive Care. 2026;14:59.
- Amolakhji, Kumar P. Time-varying exposure and incomplete target-trial emulation in the dose–response study of early mobilization after cardiac surgery. Journal of Intensive Care. 2026;14:63.
- Hernán MA, et al. The target trial framework for causal inference from observational data: why and when is it helpful? Annals of Epidemiology. 2025;101:14–22.
- Robins JM, Hernán MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology. 2000;11(5):550–560.
Keep reading
Don't stop at one method.
Good methods judgment comes from contrast. Read the neighboring guides, see where the assumptions diverge, and avoid treating every observational problem like it needs the same hammer.
Risk-Set Matching: Why the Comparison Group Must Follow the Clock
A practical risk-set matching guide for time-varying treatment. Learn why later-treated patients can be valid comparators now, how fixed exposure groups create timing bias, and what reviewers should demand.
Causal Readiness: When a Huge Linked Dataset Still Cannot Identify an Effect
A practical guide to causal readiness in linked health and administrative data. Learn why scale and propensity-score overlap are not enough when treatment, need, comparators, or outcomes are poorly measured.
Weak Instruments and Physician Preference IVs: When Treatment Movement Is Not Yet Causal Credibility
A practical guide to weak instruments and physician-preference IVs for clinical researchers. Covers first-stage weakness, exclusion leakage, local interpretation, and what reviewers should demand before trusting an IV claim.