Time-Varying Covariates

A time-varying covariate is a predictor whose value is not fixed at baseline but can change during follow-up. Unlike time-fixed covariates (e.g., sex, birth year), these variables capture dynamic patient states such as repeated lab measurements, medication adherence patterns, or disease progression markers. In survival analysis, they require data restructuring into a counting process format where each subject contributes multiple rows corresponding to intervals where covariate values remain constant. The hazard at time t is modeled as a function of the covariate value at that exact moment, not its baseline value.

Time-varying covariates are classified into two fundamental types that dictate analytical approach:

• Internal (Endogenous) Covariates: Generated by the subject themselves and require the subject to be alive to be measured. Examples include CD4 count in HIV patients, tumor size in cancer studies, or blood pressure readings. Their existence is conditional on survival, creating complex dependencies.

• External (Exogenous) Covariates: Determined independently of the subject's survival status. Examples include air pollution levels, calendar time, or treatment assignment in a crossover trial. These can be further subdivided into fixed, defined, and ancillary external covariates.

Immortal time bias is a critical error that occurs when time-varying exposures are mishandled as fixed baseline variables. Consider a study comparing survival between patients who receive a transplant versus those who do not. Patients must survive long enough to receive the transplant—this waiting period is immortal time. If the transplant variable is coded as a fixed baseline characteristic, the survival time before transplantation is incorrectly attributed to the transplant group, artificially inflating its apparent benefit. Proper handling using time-dependent Cox models or landmark analysis eliminates this bias by classifying person-time before and after the exposure separately.

The extended Cox proportional hazards model accommodates time-varying covariates by modifying the hazard function:

h(t|X(t)) = h₀(t) × exp(β₁X₁ + β₂X₂(t))

Where X₂(t) represents a covariate that changes over time. The model assumes that the effect of the covariate is instantaneous and that the proportional hazards assumption holds for the time-varying effect. Implementation requires data in counting process (start, stop] format, where each row represents an interval with constant covariate values. Software implementations in R (survival::coxph) and Python (lifelines) support this structure natively.

Landmark analysis is an alternative to time-dependent Cox models that avoids complex data restructuring. The method:

• Selects a specific landmark time (e.g., 6 months after diagnosis)
• Excludes patients who experienced the event or were censored before the landmark
• Uses covariate values measured at the landmark to predict subsequent survival
• Treats all predictors as fixed from that point forward

This approach is intuitive for clinicians and avoids immortal time bias by conditioning on survival to the landmark. However, it discards information from patients who die early and is sensitive to the choice of landmark time. Dynamic landmarking extends this by updating predictions at multiple landmarks.

When time-varying covariates are measured with error or at irregular intervals, joint models for longitudinal and survival data provide a superior framework. These models simultaneously estimate:

• A longitudinal submodel (e.g., linear mixed effects) describing the trajectory of the time-varying biomarker
• A survival submodel (e.g., Cox model) linking the true unobserved trajectory to the hazard

This approach accounts for measurement error and handles endogenous covariates naturally. The association is typically parameterized through the current value, current slope, or cumulative effect of the longitudinal process. Software includes R packages JM and joineRML.

Serial biomarker values such as CD4+ T-cell counts, serum creatinine, or tumor marker levels are classic time-varying covariates. A patient's risk profile changes with each new lab result. Using only the baseline value ignores disease progression or treatment response. In the extended Cox model, these are formatted in counting process style with (start, stop] intervals, updating the hazard at each measurement time. This allows the model to capture the dynamic relationship between biomarker trajectory and event risk.

In pharmacoepidemiology, cumulative dose or duration of treatment changes continuously. A patient on a 6-month course of a cardiotoxic agent has a different risk at month 1 versus month 6. Treating exposure as a baseline binary variable creates immortal time bias—patients must survive long enough to receive the full course. Time-varying analysis segments follow-up into pre-exposure, during-exposure, and post-exposure intervals, accurately attributing events to the correct exposure window.

In oncology trials, disease progression is both a time-varying covariate and a potential endpoint. When analyzing overall survival, progression status acts as an internal time-dependent covariate—it only exists because the patient survived to experience it. Standard Cox regression fails here. Landmark analysis or joint models for longitudinal and survival data are required to avoid conditioning on the future. The covariate is updated at the progression date, splitting the patient's timeline into pre- and post-progression intervals.

In randomized trials where patients in the control arm may cross over to the experimental treatment upon progression, the actual treatment received becomes time-varying. An intent-to-treat analysis preserves randomization but underestimates the treatment effect. Rank-preserving structural failure time models or marginal structural models with inverse probability weighting are used to estimate the causal effect of treatment while accounting for the time-dependent confounding introduced by crossover decisions.

A solid organ transplant or surgical procedure occurring during follow-up fundamentally alters the hazard. In survival analysis of end-stage disease, receiving a transplant is a time-varying covariate. The challenge is that patients must survive to the transplant date, creating waiting-time bias. The covariate is coded as 0 before the procedure and 1 after, with the timeline split at the surgery date. This allows estimation of the post-transplant hazard ratio relative to the pre-transplant baseline.

Medication possession ratio (MPR) or proportion of days covered (PDC) are dynamic adherence measures that fluctuate over a study period. A patient with 90% adherence in month 1 may drop to 40% by month 6. Treating adherence as a fixed baseline average misclassifies exposure. Time-varying analysis updates the adherence metric at each refill or assessment window, allowing the model to capture the true relationship between current adherence behavior and clinical outcomes.

A critical source of systematic error that occurs when a period of follow-up time is misclassified or excluded, during which the event of interest cannot occur. In studies with time-varying covariates, subjects who survive long enough to receive a treatment or reach a certain exposure level accrue 'immortal' time. If this time is incorrectly assigned to the treatment group rather than the control group, the treatment appears artificially beneficial. Proper handling via extended Cox models or landmark analysis is essential to prevent this distortion.

A straightforward alternative to complex joint modeling for handling time-varying effects. The method selects a specific landmark time point and analyzes only the subset of subjects still at risk at that moment. Baseline covariates and updated measurements at the landmark are used to predict future survival. This avoids complex mathematical integration but requires careful selection of the landmark and discards events occurring before it. It is particularly useful for evaluating the predictive power of a biomarker measured mid-study.

The specific data structure required to fit an extended Cox model with time-varying covariates. Instead of one row per subject, the data is split into multiple intervals defined by start and stop times. Each interval captures the constant covariate values during that specific window. This format allows the model to 'check' the status of predictors at every event time, ensuring that the risk assessment is based on the most current information. It is the fundamental engineering prerequisite for dynamic survival analysis.

A fundamental distinction in survival analysis. Internal covariates are generated by the subject themselves and require the subject to be alive to be measured (e.g., CD4 count, blood pressure). Their existence implies survival. External covariates are defined independently of the subject's survival status (e.g., air pollution level, a randomized treatment assignment). This distinction dictates the analytical approach; standard survival models struggle with internal time-varying covariates without specialized techniques like joint modeling to handle their stochastic nature.