Real-World Data For Exposure-Response Analysis

From Electronic Records to an Exposure-Response Question

Electronic health records (EHRs) can connect medication histories, laboratory results, clinical outcomes, and patient characteristics across routine care. Their richness does not guarantee that an exposure-response analysis is valid. A medication order is not necessarily a dose taken, a laboratory result is observed only when someone orders and records the test, and patients receiving more treatment may differ systematically from those receiving less. The analysis therefore begins by defining what exposure and outcome mean over a specific time window.

A useful question specifies the population, exposure window, outcome, and time relationship. For example: among eligible patients, how is average predicted concentration during the 30 days before a laboratory assessment associated with the laboratory value at that assessment? This is more precise than asking whether “dose is related to response,” because dose, concentration, timing, and the recorded outcome are distinct quantities.

Define the analysis population and time origin before examining results. Specify the outcome source and unit, the medication and formulation, the exposure window, and whether the target is an association or a causal contrast. In routine records, distinguish prescriptions, dispensing or administration records, and evidence that a patient actually took a dose. These sources have different meanings and error profiles. An order can describe intended treatment; it does not by itself establish the timing or amount of drug received.

Reconstructing and Aligning Exposure

For each patient, assemble dose events with their timestamps, amounts, units, route, and relevant medication changes. Check that timestamps use a consistent time zone and that dose records refer to the intended patient and formulation. A pharmacokinetic model can then translate dose histories into an exposure measure, provided its structural assumptions and parameter values are appropriate for the population and data. When concentration measurements are available, they can inform or check that reconstruction; sparse measurements do not justify pretending that concentrations were continuously observed.

One simple illustration uses a one-compartment model with instantaneous intravenous bolus doses and first-order elimination. Let \(D_j\) be a dose at time \(t_j\), \(V\) the apparent volume of distribution, and \(k\) the elimination rate constant. If the concentration at the start of the window is zero, then for time \(t\) in the window:

This is a deliberately simple model, not a default description of oral dosing or every clinical drug. Its assumptions include the stated one-compartment disposition, instantaneous doses, constant \(V\) and \(k\) over the window, and a known dose history. For a different drug or route, use an appropriate pharmacokinetic model rather than applying this equation mechanically.

A time-aligned metric might be the area under the predicted concentration curve, \(AUC\), over the exposure window. If a dose \(D_j\) occurs within \([0,T]\), its contribution to the area from its dosing time to the end of that window is the integral of its exponential concentration contribution:

The equality follows by integrating the exponential and using \(CL=Vk\), where \(CL\) is clearance in this model. Summing the contributions gives the window AUC. AUC divided by window duration is the time-averaged concentration. This derivation assumes that dose history before the window contributes nothing; if residual drug is present at the start, its concentration must also be represented.

Confounding and the Meaning of an Association

A confounder is a variable related to both exposure and outcome that can distort their observed association. In EHR data, disease severity is a common concern: more severely ill patients may receive higher doses and may also have poorer outcomes regardless of treatment. Adjustment can reduce this distortion only if relevant confounders are measured adequately, the analysis handles them appropriately, and there is sufficient exposure overlap across their levels.

For a binary outcome, direct standardization illustrates the issue. Within each confounder stratum, estimate the outcome probability for each exposure group. Then average those stratum-specific probabilities using the same chosen stratum weights for both groups. The common weights remove differences in the stratum mix from this particular comparison; they do not remove unmeasured confounding.

A causal interpretation additionally requires assumptions such as adequate measurement and adjustment of common causes, comparable treatment definitions, positivity (patients with relevant covariate patterns could plausibly receive either exposure level), and correct analysis of the time ordering. If worsening disease changes treatment during the exposure window, ordinary adjustment for a single baseline severity value may not suffice. In all cases, state whether the result is an adjusted association or a causal estimate and identify the assumptions needed for the latter.

Uneven Observation and Data Capture

EHR measurements are usually irregular: patients have different numbers of visits, tests occur at different times, and records may be absent for reasons related to health or care. A row in the database is not necessarily an independent patient. If a patient with frequent tests contributes many rows, a simple row-level average gives that patient greater weight. Whether that weighting is appropriate depends on the intended estimand; it is not automatically the average patient experience.

For longitudinal outcomes, use actual measurement times and a model suited to the question, such as a repeated-measures model with patient-level dependence represented. A model that includes time can distinguish a changing outcome trajectory from a difference in measurement timing. If observation frequency itself depends on health status or treatment, simply adding more rows does not fix the problem; describe the capture process and consider sensitivity analyses or methods that explicitly address observation or selection.

Diagnostics, Alternatives, and Limitations

Before interpreting an estimate, check the data pathway and the model, not just the final coefficient. Useful checks include:

• Compare medication orders with available dispensing or administration records, and quantify how often dose timestamps, amounts, or formulations are missing or inconsistent.
• Plot dose histories, predicted concentrations, and outcome timestamps for a sample of patients. Check that exposure windows precede the outcomes they are meant to explain.
• Inspect exposure distributions across key covariates. If one severity group has almost no patients at a particular exposure, adjustment relies on extrapolation rather than direct comparison.
• Compare observed concentrations with model predictions when measurements exist, and investigate systematic deviations rather than assuming the model is calibrated.
• Summarize outcome measurement frequency and timing by exposure and health status. Large differences can indicate that the analyzed records represent different observation processes.
• Check whether results change under reasonable alternative exposure windows, definitions, and adjustment sets. Such checks reveal sensitivity; they do not identify a uniquely correct analysis by themselves.

A single dose category or the latest recorded concentration can be easier to calculate than a time-aligned model-based exposure, but may misrepresent exposure over the response window. Conversely, a detailed pharmacokinetic reconstruction uses more information but depends on model structure, parameter transportability, and the quality of dose records. Regression adjustment or standardization can address measured confounders; weighting approaches can also reweight observed patients, but require well-estimated probabilities and adequate overlap. No method repairs inaccurate timestamps, unmeasured confounding, or an outcome that is selectively recorded without additional assumptions.

Real-world exposure-response analyses can be valuable for checking whether clinical patterns are consistent with prior pharmacology, identifying data gaps, and generating hypotheses. Their limitations should remain visible: routine records are created for care rather than a prespecified analysis, key covariates may be absent, exposure and outcomes may be misclassified, and observation may depend on patient condition. Present the population, window, exposure definition, adjustment variables, missingness and capture checks, and uncertainty alongside the estimate. A precise number cannot compensate for an unclear estimand or weak data provenance.