Target trial emulation reconstructs a randomized experiment from observational data. The method is powerful: define the hypothetical trial you wish you had run, then use the observational data to approximate it. Align the entry criteria, the treatment assignment, the follow-up, the outcome measurement. If done correctly, the observational estimate converges to the experimental effect.
The method requires discretizing time. Continuous-time data must be sliced into periods: baseline, treatment windows, follow-up intervals. The choice of time grain — daily, weekly, monthly, quarterly — is typically treated as a technical convenience, chosen for computational tractability or data availability.
It is not a convenience (arXiv:2603.15924). The time grain changes the causal estimate. Slice too finely, and you inflate the model's dimensionality — each time point becomes a separate treatment-confounder layer, requiring separate adjustment, and the statistical precision collapses. Slice too coarsely, and the causal structure within each period is lost — a treatment that affects the outcome within the same period cannot be captured by a model that assumes treatment precedes outcome across periods.
The failure mode is specific and sharp: standard cloning-censoring-weighting methods — the workhorse of target trial emulation — can produce biased estimates when treatments influence outcomes within the same study period. The bias is not random noise; it is systematic, directional, and proportional to the mismatch between the true causal timing and the imposed time grid.
This means the analyst's choice of calendar — not the data, not the method, not the model — can determine whether the treatment appears beneficial, harmful, or null. The resolution of the measurement is load-bearing.
The clock is part of the instrument. Choose the wrong clock, and you measure the wrong effect.