Six tracks through RWE & HEOR methods

The temporal skeleton every claims analysis is built on — start here before touching any code field.

Diagnosis codes define your outcome and comorbidities — understand hierarchy before building phenotypes.

Procedures anchor treatment exposure in professional claims; pairs directly with ICD codes on the same claim line.

Drug exposure lives in NDCs on pharmacy claims; required before tackling dispensing-based cohort entry.

Maps NDCs onto the NCPDP transaction format — needed to extract days supply, quantity, and refill data.

Inpatient and outpatient facility claims travel on UB-04 — understanding its structure is required for any hospital-based outcome.

Physician services ride the CMS-1500; links CPT codes to place of service and billing provider.

Revenue codes on UB-04 distinguish emergency, observation, and ICU encounters — critical for site-of-care phenotyping.

Failure to handle adjustment/void claim lines creates ghost utilization — must be cleaned before any analysis.

Gaps in coverage = gaps in data; defining eligible observation time is the gatekeeper for all downstream denominators.

Washout windows ensure new-user and drug-naïve cohorts — the prerequisite for credible exposure classification.

Combines every prior step into a validated cohort-entry rule — the capstone of claims-based cohort construction.

Mean, median, IQR — the vocabulary every table-1 and every subsequent test assumes you already know.

The central limit theorem is why most parametric tests work on large RWE samples even with skewed raw data.

p-values, confidence intervals, and the null-hypothesis framework underpin every comparison you will run.

MLE is the engine inside every regression model you will use — understanding it demystifies model fitting.

Knowing when normality assumptions break guides whether you reach for t-tests or rank-based alternatives.

The workhorse for comparing continuous baseline covariates between treatment arms in Table 1.

The analogous test for categorical variables — essential for comparing demographic and binary covariate distributions.

The two fundamental measures of comparative effect — every HEOR model uses one or both as its currency.

Binary outcomes (hospitalisation, event) follow the binomial; the logit link maps probabilities onto the real line for regression.

The intuition and matrix algebra behind OLS transfers directly to GLMs and survival models.

GLMs unify logistic, Poisson, and gamma regression under one framework — the workhorse of outcomes research.

Residual plots, leverage, and VIF reveal whether your model assumptions hold before you trust its estimates.

Right censoring, left truncation, informative censoring — the defining quirks of time-to-event data in RWE.

How to accumulate and truncate at-risk time in claims data — the denominator every incidence rate depends on.

Events per person-year is the universal rate measure; practice computing it before fitting any model.

Non-parametric survival curves — the first visual every reviewer expects and the baseline every model is judged against.

Tests equality of KM curves across groups — needed before deciding whether a Cox model is warranted.

HR ≠ relative risk; understanding the instantaneous-rate framing prevents the most common HTA submission error.

The semi-parametric workhorse of time-to-event analyses in pharma RWE — assumes proportional hazards, which must be verified.

Death before the outcome of interest invalidates standard KM; Fine-Gray subdistribution HR is the HTA-preferred alternative.

Parametric alternative when proportional hazards fails — and the basis for survival extrapolation in economic models.

HR-free summary of survival benefit over a defined horizon — increasingly required by HTA bodies when PH fails.

Directed acyclic graphs make confounding assumptions explicit — the foundation before any adjustment strategy.

The dominant bias in pharmacoepidemiology — sicker patients get treated, which inflates or deflates effect estimates.

Restricting to treatment initiators eliminates prevalent-user bias — the modern standard for drug comparisons in RWE.

Pairing active comparators with new-user entry balances measured confounders and reflects real prescribing decisions.

Misaligned index dates introduce immortal time; correct anchoring is non-negotiable for valid follow-up.

Time between cohort entry and treatment start cannot be spent having the outcome — mis-handling this inflates efficacy.

Matching and weighting on PS is the primary tool for reducing measured confounding in large claims databases.

Overlap weights outperform ATE/ATT weights in sparse-overlap settings common in specialty-drug RWE.

Standardisation via outcome regression — avoids PS trimming losses and handles time-varying confounders.

Formalises the hypothetical RCT your observational study is trying to answer — the FDA/EMA preferred framing.

Quantifies how strong unmeasured confounding must be to explain away your result — required for any regulatory submission.

Goes beyond E-values to model misclassification and selection bias jointly — the advanced bias-analysis capstone.

Per-member cost metrics are the lingua franca of payer communications and the raw input to budget impact models.

Healthcare costs are right-skewed and strictly positive — gamma GLMs are the appropriate parametric model.

Zero-inflated cost distributions need a logistic first part (any utilisation?) and a gamma second part (how much?).

Non-parametric CIs for cost differences and NMB estimates when distributional assumptions can't be verified.

Mapping condition-specific PROs to EQ-5D index scores is the bridge between RWE data and NICE/HTA cost-utility models.

How to derive state-transition parameters from observed data — the central plumbing of Markov cohort models.

Parametric extrapolation beyond the trial horizon is required by every NICE TA — RWE informs the long-run tail.

The decision metric for cost-effectiveness: ICER vs. threshold, and NMB as the linear transformation for regression.

Monte Carlo PSA propagates joint parameter uncertainty to the cost-effectiveness plane — required in every HTA dossier.

Payer-facing affordability analysis; complements the cost-utility model and is required for most formulary submissions.

Frames the prediction vs. causal distinction before diving into specific algorithms — avoids conflating the two goals.

k-fold CV is the standard guard against overfitting in high-dimensional claims data — required before tuning any model.

Penalised regression handles thousands of ICD/CPT indicators without overfitting; LASSO gives automatic variable selection.

Random forests and gradient boosting capture non-linear interactions that linear models miss in heterogeneous patient populations.

C-statistic measures how well the model ranks patients — the primary metric for algorithm validation studies.

Calibration (does predicted probability match observed frequency?) is as important as discrimination for clinical decision support.

External validation and recalibration ensure the model performs in a new population — required before clinical deployment.

Structured claims miss nuance captured in physician notes; NLP unlocks smoking status, severity, and disease course from EHR text.

Large language models as chart-abstraction tools — covers prompt design, hallucination risk, and regulatory validation requirements.

Measures inter-rater reliability between NLP/LLM output and human abstraction — the final gate before algorithm acceptance.