Dose Modification and Selection Bias in Safety Models

Why Dose Changes Complicate Safety Exposure-Response

In clinical development, a safety signal can lead to a dose reduction, interruption, or discontinuation. The change may lower a patient's subsequent exposure, but the signal may also predict a later adverse event. Consequently, patients observed at lower exposure can have a higher underlying risk than patients who remain at the original dose. Conversely, patients who remain at a high dose may be a selected group who have so far tolerated treatment. Comparing outcomes by observed dose alone can therefore mix the effect of exposure with the reasons that exposure changed.

A key safeguard is to define exactly who is eligible to experience the endpoint at each time. This is especially important for a time-to-first-event analysis: a patient who has already experienced the event is no longer at risk for a first event. The dose change, the event definition, and the denominator must all refer to the same time interval and estimand.

Exposure must also be time-indexed. Let \(A_k\) denote the dose or exposure category that applies during interval \(k\), assigned using information available before that interval begins. Let \(Y_k\) indicate a first event during interval \(k\). A signal measured before the dose decision can help explain both \(A_k\) and \(Y_k\). A signal measured after the event, by contrast, cannot be treated as a pre-event reason for the dose decision. Temporal ordering is part of the analysis, not a detail to resolve after fitting a model.

How Selection Enters the Comparison

Suppose a measured safety signal \(L\) influences the dose decision and predicts the endpoint. The event risk among patients observed at dose category \(a\) is a mixture over the signals in that group:

The signal distribution \(P(L=l\mid A_k=a,R_k)\) can differ across dose groups because the signal helped determine the dose. Thus, even if risks are compared correctly within signal strata, the crude comparison can differ simply because the groups contain different proportions of higher-risk patients. This is selection through a dose decision based on evolving patient information; it is often described as confounding by indication.

A useful causal interpretation requires stronger conditions than a correctly calculated crude rate. In particular, relevant predictors of both the dose decision and the endpoint must be measured and adjusted for; each dose option being compared must occur with positive probability for the relevant patient histories; and the exposure and outcome models must be adequately specified. Under these conditions, consistency (the observed outcome under a patient's received dose corresponds to the outcome under that dose) and sequential exchangeability (no unmeasured common causes of dose and outcome at each decision time, conditional on measured history) support causal comparisons. These assumptions are not guaranteed by randomization at baseline: later dose changes are not necessarily randomized.

Adjustment for Measured Safety Signals

One transparent sensitivity analysis is direct standardization. First calculate an event risk within each dose group and each level of a pre-decision signal. Then average those risks using the same signal distribution for both dose groups. This removes differences in the measured signal mix from the comparison, while leaving other potential sources of bias unresolved.

Weighting as a Sensitivity Analysis

Inverse-probability weighting is another way to create a dose comparison less dependent on measured differences in patient history. At each decision time, estimate the probability of the dose category actually received, conditional on measured pre-decision history. A stabilized weight for a patient receiving category \(a\) can be written as:

where \(L_{ik}\) denotes measured history available before the decision. The numerator is the marginal probability of that dose category in the risk set; the denominator is the probability conditional on history. In practice, weights may be estimated from a model, and an analysis may also need to address informative censoring. The weighted comparison is meaningful as a causal sensitivity analysis only if the histories capture important dose-outcome predictors, there is adequate overlap in treatment probabilities, and the probability models are suitable.

Diagnostics, Alternatives, and Limits

Before interpreting an exposure-response model with dose changes, check the data construction and the model behavior. Verify that each dose category applies before the events counted in its interval; that the risk-set denominator includes only patients eligible for the endpoint; and that pre-decision signals are distinguished from measurements taken after an event. For a first-event endpoint, any event after a patient's first event must not be counted as another first event. If the scientific question concerns subsequent events, it requires a recurrent-event definition and corresponding risk-set and estimand choices rather than relabeling those events as first events.

Useful diagnostics include tabulating dose decisions by signal history, checking how many patients in each history stratum receive each dose option, inspecting estimated probabilities and weight distributions, and comparing crude with adjusted results. Extreme weights indicate that some histories have little support for one of the dose options; a fitted model cannot create information where there are few or no comparable patients. Weight truncation can reduce instability but changes the analysis and does not repair a lack of overlap. A zero-event cell can also make a stratum-specific estimate unstable, as in the standardization example.

The main alternatives answer different questions. A treatment-policy analysis based on original randomized assignment preserves the randomized comparison but estimates the effect of assignment under the dose changes and discontinuations that occurred; it does not directly estimate the effect of maintaining a particular subsequent exposure. An observed-dose analysis describes outcomes by dose actually received, but is vulnerable to the time-varying selection described here. Standardization and weighting can adjust for measured, time-ordered predictors of dose and outcome, but cannot correct unmeasured confounding merely by being more elaborate. A lagged-exposure analysis can help prevent an event from being assigned to exposure measured after its onset, but lagging alone does not remove selection by prior safety history.

Competing events and censoring also need explicit handling because they determine who remains eligible or observable. The chosen approach should match the endpoint and estimand; changing the denominator without stating what happened to patients who died, discontinued, or were lost to follow-up changes the question. Finally, a signal-triggered dose reduction may be followed by lower exposure and altered risk, but distinguishing these effects requires adequately measured histories and sufficient data. These analyses are sensitivity tools for understanding the observed association, not a substitute for design or a guarantee of causal identification.