RESEARCH DOSSIER
Causal Inference versus Predictive Association in Insurance
Abstract. Insurance institutions predict future losses, select risks, price coverage, allocate claims resources, and evaluate interventions. The statistical associations that support these activities answer different questions from the counterfactual effects of changing deductibles, underwriting rules, safety investments, insurance access, or claims procedures. This monograph establishes the distinction at the level of the research question, target population, data-generating process, mathematical estimand, identifying assumptions, estimation strategy, and validation design. It examines policy selection, confounding, collider conditioning, post-treatment variables, dynamic feedback, interference, and distribution shift in insurance portfolios. It reconstructs the inferential content of randomized coverage offers, exogenous premium variation, observational claims databases, longitudinal policies, and staggered interventions. The Oregon Health Insurance Experiment and empirical research on insurance demand, selection, and moral hazard provide substantive illustrations of what a particular identification strategy can establish. A transparent numerical example shows a strong predictive association that reverses sign as an estimated causal effect under disclosed assumptions. An additional counterfactual portfolio example separates forecasting losses under an existing policy regime from forecasting losses after a change in that regime. The original contribution is a systematic analytical synthesis of the relation between insurance prediction, causal identification, policy choice, and external validity; no new named theory or doctrine is proposed. The principal conclusion is that high predictive accuracy establishes performance against an observational forecast target, while a causal claim additionally requires a specified intervention, credible identification, and a target-population interpretation. Forecasting and causal reasoning may use overlapping statistical methods and variables, but their validation targets and decision consequences remain distinct. The scope is analytical and documentary; no proprietary insurer dataset, new field experiment, systematic meta-analysis, or comprehensive national regulatory survey is asserted.
Keywords: Insurance; causal inference; predictive association; statistical learning; counterfactuals; potential outcomes; causal estimand; loss frequency; claims severity; underwriting; pricing; risk selection; confounding; selection bias; collider bias; treatment effect; causal diagrams; propensity scores; instrumental variables; natural experiments; randomized insurance offers; difference-in-differences; causal machine learning; heterogeneous effects; calibration; transportability; moral hazard; adverse selection; insurance welfare; algorithmic governance; actuarial modeling.
Introduction: Object, Questions, and Evidentiary Method
Insurance reasoning frequently begins with an observational record: an insured characteristic, a coverage contract, exposure measured over time, and a subsequently observed claim or loss. A regression, generalized additive model, or machine-learning system may recover highly stable relationships among those recorded quantities. The resulting predictions can be economically useful for setting reserves, forecasting workload, allocating capital, estimating claim probability, and detecting anomalies. The same record may also be consulted when a decision-maker asks whether changing a contract term, offering a safety subsidy, or expanding coverage will change behavior or financial outcomes. That second question concerns a different mathematical object. Shmueli's explanation and prediction distinction shows why a modeling purpose must be specified before model design and evaluation, while the Causal Inference: What If framework of Hernán and Robins (2020) organizes causal analysis around explicit hypothetical interventions.
The research object is the inferential relationship between predictive association and causal effect in insurance. Prediction concerns a distribution of outcomes for defined observed information under a specified data environment. A causal effect concerns contrasts between outcomes under alternative, well-defined interventions. Both require careful data measurement and uncertainty analysis. Their distinct identifying requirements are consequential for claims forecasting, premium design, incentive effects, wellness or loss-prevention programs, regulatory assessment, and distributional welfare analysis.
Six substantive questions organize the study. What exactly can be learned from a highly accurate insurance prediction model? When does observed correlation between a rating variable and future loss estimate an intervention effect? How do purchase, renewal, filing, settlement, and survival into the observed portfolio distort comparisons? Which research designs can identify causal effects of actual insurance policies? How can identified effects be transported across markets and portfolios? What scientific and editorial evidence warrants a causal rather than merely predictive claim?
The method combines formal causal identification, critical reconstruction of published empirical designs, selective examination of peer-reviewed statistical and insurance-economics literature, comparison with published actuarial and supervisory documents, and hand-checkable hypothetical calculations. Research sources were checked against publisher records, accessible full text, author-maintained material, or official regulatory pages, with access limitations recorded in EVIDENCE.json. Mathematical illustrations are explicitly hypothetical. Bibliographic metadata, rather than the existence of a DOI alone, is used to establish what a source actually studied. Regulatory texts are interpreted in their issuing jurisdiction and institutional form. Scientific review assesses whether each claim follows from an estimator, design, or documentary finding; editorial review separately assesses consistent terms, link specificity, attribution, and readability.
An important vocabulary distinction governs the entire work. A variable may be a predictor because it improves expected prediction accuracy, a confounder because it causes both exposure assignment and outcome, an effect modifier because the response to an intervention differs by its value, or a mediator because it lies on a causal path. The same recorded variable can occupy different roles in different questions. A policyholder's prior claims, for example, may predict later claim frequency while simultaneously affecting underwriting, premium, retention, reporting, and opportunities for further claims. Statistical importance therefore has no universal causal interpretation.
Prediction and Causal Estimands
Observational forecasts and their targets
Let Y denote a future observed claim outcome and X denote covariates known at the forecast origin. A predictive model estimates a conditional distribution P(Y | X=x, S=1), where S=1 denotes presence in the observed dataset. For a numeric aggregate claim amount, a squared-error-optimal point forecast is E[Y | X=x, S=1]; for claim occurrence, the analogous quantity is P(Y=1 | X=x, S=1). The choice of loss function changes which feature of the distribution is optimal. Predicting a conditional mean, upper-tail quantile, claim count, default-like lapse indicator, or probability of severe loss entails different outputs and operational uses. Breiman's two cultures of statistical modeling supplies an influential account of algorithmic prediction and data-modeling approaches; it is not, by itself, a claim that flexible algorithms identify interventions.
The sample restriction S=1 matters. An insurer usually observes accepted policies and may have incomplete information on rejected applicants, uninsured exposures, off-system claims, losses below deductibles, or events never reported. A model that accurately predicts paid claims within the insured book may fail to predict all physical losses among the applicant population. Paid claims are functions of exposure, damage, coverage wording, deductibles, claim filing, insurer decisions, and payment procedures. A change in any of these can alter the meaning of Y without changing physical hazard. A historical claims target therefore has an institutional definition, not just a mathematical scale.
Out-of-sample validation answers whether the forecast generalizes to observations drawn under the validation protocol. A randomly held-out set from the same period tests one form of interpolation. A temporal split tests a more relevant form of forward prediction, although it cannot guarantee performance during future shocks. A geographically held-out portfolio tests cross-market transport of predictive patterns. Strong discrimination and calibration can coexist with poorly understood causes; Steyerberg et al.'s prediction performance framework differentiates these evaluation dimensions and their decision-analytic extensions. In insurance, validation must also account for earned exposure, delay in claim emergence, inflation, censoring, portfolio composition, and changes in loss definitions.
Potential outcomes and interventions
Let A take values 0 and 1 for two well-defined insurance-related actions, such as offering a loss-prevention service or setting a specified deductible at renewal. For policyholder i, Y_i(1) is the outcome under action 1 and Y_i(0) the outcome under action 0. The unit-level contrast Y_i(1)-Y_i(0) cannot generally be observed for the same unit in the same period. The average treatment effect (ATE) in a defined population T is E[Y(1)-Y(0) | T]. A different target, the average treatment effect among those actually treated, is E[Y(1)-Y(0) | A=1, T]. The distinction descends from Rubin's potential-outcomes treatment analysis, which made the role of treatment assignment and experimental comparison explicit.
The treatment must be specified to support an interpretable counterfactual. 'Having insurance' might mean enrollment in a precisely stated public program, purchase of comprehensive private coverage, addition of a collision rider, waiver of a waiting period, or acceptance of an offer with partial compliance. These actions differ in premiums, services, coverage limits, incentives, and administrative institutions. Likewise 'safer driving' is a composite behavior rather than an operational intervention until speed control, training, monitoring, enforcement, or financial incentive is defined. The substantive definition of A is part of the claim, and it should be fixed before outcome inspection.
Under consistency, the observed outcome for a unit receiving A=a equals Y(a) for the specified treatment version. Under exchangeability, conditional on adequate pre-treatment covariates X, potential outcomes are independent of actual treatment assignment; under positivity, every relevant covariate stratum has a nonzero probability of receiving each compared option. If these assumptions and measurement conditions hold, the standardized contrast E_X[E(Y | A=1,X)-E(Y | A=0,X)] identifies the ATE. Hernán and Robins' identification and exchangeability treatment emphasizes that such equations express substantive assumptions and not merely computing instructions. When treatment assignment is randomized with sufficient adherence to the assigned protocol, some identification problems simplify, but nonadherence, attrition, and outcome definition still require examination.
Predictors, effects, and model parameters
Suppose an insurance dataset yields a coefficient β for a feature X in a fitted log-mean model, E[Y|X]=exp(α+βX). Then exp(β) measures a model-specific conditional mean ratio under the observed data and specified adjustment. It is a causal multiplicative effect only under an additional causal interpretation of X as an intervention and credible identifying conditions. In many cases X is age, prior claims history, household location, or a vendor-derived score. There is no well-defined intervention 'set chronological age backward' in the context of actuarial pricing; an apparent effect from changing its numeric value is a scenario imposed on a model, rather than an identified intervention on the human aging process. In contrast, a standardized safety installation program can be defined operationally and potentially evaluated causally.
Even a coefficient whose sign remains stable across many predictive models does not resolve its causal meaning. Omitted common causes may determine its magnitude; conditioning on variables affected by the treatment may remove or reverse part of a total effect; and the outcome might encode an insurer's own decision process. The distinction survives algorithm substitution. A random forest's importance score, a boosted model's feature attribution, and a neural network's local explanation all concern the fitted function unless backed by a causal design. The phrase 'the model found a driver of claims' is therefore ambiguous: it may designate a strong predictor, a mechanism of physical hazard, or an administrative variable that governs recording.
The decision problem and expected losses
Insurance decisions are interventions that alter subsequent information and incentives. Let d denote a proposed decision rule, and Y(d) the outcome under implementation of that rule. The policy question concerns E[Y(d)] and any additional costs or benefits, not simply E[Y | historical X=x]. For premium pricing, the historical expected loss can inform technical cost projections under stable contractual behavior. If a premium changes coverage demand, sorting, claims reporting, care seeking, fraud incentives, or prevention investment, the relevant future portfolio distribution also changes. The mean predicted loss conditional on the old contract cannot be substituted uncritically for mean loss under the changed contract. A robust evaluation defines the target insured or applicant population, decision rule, behavioral channels, market response, and horizon before estimating effects.
An insurer may responsibly use a variable for prediction while declining to claim causal influence. For example, a regional hazard score can improve fire-loss forecasts without demonstrating that changing the score label reduces wildfire severity. An analyst can separately study whether subsidizing vegetation clearance, modifying building codes, or changing emergency response alters losses. The first task values reliable information about risk; the second seeks the consequences of an action. Both can contribute to insurance governance when their meanings are preserved.
Confounding and Selection
Confounding from risk characteristics and behavioral choice
A comparison of claim frequency between policyholders with and without an optional prevention service illustrates a typical identification problem. Service adopters may have higher baseline hazard because insurers target high-risk properties, or lower hazard because preventive households are more attentive to safety. In causal-graph terms, an unmeasured factor U can cause both service adoption A and subsequent loss Y, creating the backdoor path A <- U -> Y. An observed difference E[Y|A=1]-E[Y|A=0] then combines treatment effect with composition. A high-dimensional predictive model might approximate E[Y|A,X] accurately while leaving U unmeasured. More predictors can improve fit without guaranteeing exchangeability.
Insurance selection is dynamic. Prior losses shape renewal offers; offers shape coverage choice; coverage choice shapes reporting and utilization; reporting shapes the recorded loss history used in the next renewal model. A policy variable that precedes one outcome may be downstream of prior treatment. A regression that mechanically includes all available variables can control for a mediator or for a common effect of treatment and latent risk, thereby changing the estimand. A credible model thus begins with a causal description of timing and institutions, not a spreadsheet's complete list of columns. The propensity-score theory developed by Rosenbaum and Rubin (1983) permits adjustment for observed covariates under appropriate assumptions; it never equates a measured propensity with proof that every relevant confounder was observed.
Consider insurance take-up as A and subsequent claims spending as Y. Higher anticipated illness can increase both enrollment and medical utilization, while insurance can itself change use and financial exposure. A comparison between insured and uninsured persons therefore mixes risk selection with potential insurance effects. The insurance economics of Einav and colleagues' selection on moral hazard specifically recognizes that people may choose plans partly because they anticipate different behavioral responses to coverage. Selection operates not only on fixed baseline costs but also on treatment-effect heterogeneity. This interaction makes simplistic adjustments using realized claims especially fragile.
Collider conditioning and insurer-created data
A collider is a variable with two or more causal parents. Conditioning on a collider can induce association between its causes even when none existed before conditioning. Suppose both baseline hazard U and marketing outreach A influence policy purchase S. Observing only purchased policies means conditioning on S=1, thereby potentially associating A and U. A model of subsequent loss among customers then may find that a marketing variable forecasts risk, even if marketing had no direct effect on hazard. The association is partly created by portfolio entry. Similar processes occur if both event severity and willingness to claim determine whether a loss enters the insurer's claims database.
Paid claims introduce additional filters. An actual damaging event may be followed by a decision to notify, a coverage assessment, claims investigation, settlement, and payment. If a study analyzes only paid claims, conditioning on payment can select for both coverage design and damage severity. Comparisons of 'severity among paid claims' may change after a deductible adjustment even without any change in physical damage. For example, raising a deductible excludes smaller losses from recorded payments, raising the conditional average amount among payments while reducing the number of payments. Declaring an increase in loss severity from this conditional statistic alone mistakes an administrative truncation effect for an increase in underlying hazards.
Censoring and survival selection deserve equal attention. Life, disability, health, property, and casualty portfolios differ in eligibility, duration, claim development, and termination. Claims reported late may be disproportionately missing near the data cutoff. A hazard model that treats an unreported future claim as a true zero will misstate risk, and that error can correlate with distribution channel, jurisdiction, coverage type, or claims-processing capacity. Treating missing data as innocuous requires a separate argument about the missingness mechanism. A claims model's predictive success on mature years does not automatically validate an effect estimate for a recent policy change.
A transparent reversal example
Take two equally sized latent risk groups, H and L, with outcome risks measured as expected physical-loss events per 100 exposure units. Suppose an optional safety device reduces expected events by one in each group. In group H, the baseline is 10 without the device and 9 with it; in group L, the baseline is 2 without the device and 1 with it. Assume adoption is common among H: 90% of device users are H, but only 10% of nonusers are H. These proportions are illustrative, not estimates from insurer data.
The observed treated mean is 0.90×9 + 0.10×1 = 8.20 events per 100. The observed untreated mean is 0.10×10 + 0.90×2 = 2.80. Their raw association is +5.40 events per 100, even though the causal effect within each group is −1.00 and hence the standardized ATE is −1.00 under the stated constant-effect and positivity assumptions. The device is a strong positive predictor of high claims precisely because high-risk insureds adopt it. An intervention that increases device adoption among otherwise comparable risks would lower expected events by one per 100 units in this constructed world.
The arithmetic also shows why observational adjustment can recover the hypothetical effect when the true group variable is observed and adequately measured: (9−10) and (1−2) are both −1. In real portfolios, complete group membership is rarely available, and residual confounding, variable definitions, time ordering, selection into purchase, and measurement error must be assessed. The numerical example proves a logical possibility under disclosed numbers; it is not empirical proof that any actual device reduces insurance losses. Its value is that the same data pattern can yield excellent prediction and misleading causal interpretation.
Dynamic confounding and post-treatment variables
Longitudinal insurance contracts introduce treatment histories A_0, A_1, … and information histories L_0, L_1, …. An early deductible may influence loss-prevention behavior, which affects later risk assessment, subsequent deductible assignment, and future claims. If a time-varying confounder L_1 is itself affected by A_0, ordinary regression adjustment for L_1 may block part of the effect of A_0 while omitting it leaves confounding for A_1. Robins, Hernán, and Brumback's marginal structural model analysis addresses this causal structure with suitable weighting under sequential exchangeability and positivity conditions.
For insurance, the distinction applies to premium reminders, phased telematics incentives, annual deductible changes, repeated safety inspections, and claims-management interventions. A target intervention might be a sequence of eligibility-based actions over two renewals, not a single baseline yes/no treatment. The estimand becomes the difference in outcomes under two defined dynamic policies. Model specification must preserve the dates at which risk factors were observable, the actual treatment options then available, and the time before which outcomes could occur. Looking backward from a claim and labeling a person 'treated' if they eventually adopted a program can create immortal-time bias; Hernán et al.'s target-trial specification analysis clarifies this error and the role of aligning eligibility, assignment, and time zero.
Measurement, manipulation, and market equilibrium
Risk-factor measurement itself can change once it is used in pricing. If consumers know a telematics score affects premiums, some may modify driving behavior, temporarily alter measurement, or change vehicles and journeys. If insurers alter their reporting or investigation practices after receiving an anomaly score, recorded fraud and claim closures can respond to the model. A feature that predicted outcomes before deployment may carry a different meaning afterward. The data-generating process includes decision-makers reacting to forecasts, and its adaptation is part of insurance's institutional environment.
The same insight applies at the market level. Premium discrimination can change who buys insurance. A portfolio's average loss after a price revision can move because retained policyholders differ, because behavior changes, because the product's coverage changes, or because external hazards change. Aggregate insurer losses are equilibrium outcomes of selection and contract design. A within-portfolio statistical comparison must not be promoted to a population-wide welfare claim until these channels and their target populations are separated.
Predictive Features versus Interventions
Causal questions before modeling choices
A rigorous analysis first writes the proposed action in operational terms: offer a home retrofit subsidy before policy inception; increase a collision deductible from one named amount to another at renewal; give a random subset of eligible claimants access to a claims-navigation service; or introduce a loss-prevention inspection across designated territories. Next it states eligibility, assignment date, outcome period, competing actions, and target population. Only after those choices are fixed can researchers select data and an estimator. A predictor may be indispensable for risk stratification yet irrelevant to an intervention's identifying variation. Conversely, a variable used as an instrument may be a weak predictor of outcome while being crucial for causal identification.
This separation applies to claims algorithms as much as underwriting. A model that predicts which claims will be costly helps workload planning. To estimate whether assigning an adjuster with specialist expertise reduces final cost without reducing legitimate claimant benefits, one must compare outcomes under alternative staffing policies, account for claim complexity and allocation rules, and specify whose cost is being measured. An insurer's accounting expense and a claimant's achieved indemnity can move in opposite directions. A policy evaluation should state the perspectives and outcomes before declaring a 'benefit.'
Randomization, compliance, and estimand discipline
A randomized offer establishes comparability of groups at assignment in expectation. The intention-to-treat effect contrasts mean outcomes by assignment, including people who decline the offer. If the intervention is an offer of coverage rather than actual enrollment, the offer effect and coverage effect differ. Under valid instrumental-variable assumptions, random assignment can be used to infer a local treatment effect for individuals whose uptake changes because of that offer. Imbens and Angrist's local average treatment effect formulation shows why such effects concern specific compliers rather than necessarily every insured or applicant.
Randomization does not by itself solve missing outcomes, interference, attrition, changing policy conditions, or ambiguity over treatment. In a property-insurance loss-prevention trial, safety measures on one building can influence fire spread to neighboring buildings, violating a simple no-interference assumption. For health coverage, one person's insurance can affect household spending and access to care. For a claims service, adjusters may transfer knowledge from treated files to control files. Cluster or market-level assignment may be preferable where spillovers are central; its causal interpretation and uncertainty must reflect that level of assignment.
Adjustment, propensity estimation, and double robustness
In observational studies with credible measured confounding control, regression standardization, inverse-probability weighting, matching, and doubly robust estimators offer different implementations of identification formulas. Their common foundation is an assumption about treatment assignment, not a guarantee supplied by a fitted algorithm. Propensity-score weights can be unstable when some policyholders almost always receive one treatment, violating practical overlap. Trimming unsupported strata changes the target population; failing to acknowledge that change exaggerates generality. Outcome-model adjustment can extrapolate across covariate combinations where treatment variation is absent. Reporting both overlap diagnostics and covariate balance is therefore essential.
Propensity scores have multiple legitimate uses, yet matching on a scalar estimated propensity is not automatically a safe preprocessing step. King and Nielsen's critique of propensity-score matching documents settings in which the popular matching procedure increases imbalance, dependence on models, or bias. It does not invalidate propensity weighting or adjustment under their proper assumptions. In applied insurance work, the choice among methods should be guided by the estimand, overlap, sample size, covariate distribution, and sensitivity to unmeasured confounding.
Modern statistical learning can help estimate nuisance functions such as propensity and conditional outcome regressions. Chernozhukov and colleagues' double/debiased machine learning approach uses orthogonal estimating equations and cross-fitting to reduce sensitivity to certain nuisance-estimation errors under stated regularity conditions. This is a methodological advance in estimating causal parameters from suitable data, not an automatic conversion of predictive accuracy into causal identification. An orthogonal score cannot repair a false exclusion restriction, a missing intervention definition, or unmeasured common causes. Confidence intervals based on large-sample approximations should also be examined for small samples, rare severe claims, and dependence among insured units.
Natural experiments and timing variation
Many insurance policies change by jurisdiction, employer, renewal cohort, or product category. Researchers may use differential timing for difference-in-differences (DiD), a threshold for regression discontinuity, or plausibly exogenous assignment for instrumental variables. Each design has its own estimand and identifying assumptions. In a simple two-group two-period DiD, the contrast (E[Y_post|treated]−E[Y_pre|treated])−(E[Y_post|control]−E[Y_pre|control]) is causal only under conditions including an appropriate parallel-trends counterfactual. Observed similar pretrends can inform credibility but cannot logically prove an unobserved post-treatment counterfactual. Roth et al.'s difference-in-differences synthesis examines complications involving staggered adoption, heterogeneous treatment effects, and alternative inference strategies.
For an insurance reform, changes to policy wording, distribution rules, claims definitions, weather exposure, and insurer entry can coincide with the reform itself. The study must identify whether these are components of a joint intervention, potential confounders, or post-treatment consequences. Simple before-after comparisons are especially vulnerable when reforms follow a catastrophic loss year or a change in economic conditions. Regression discontinuity can identify a local effect at a cutoff only when assignment near the threshold is suitably comparable and manipulation is addressed; it does not automatically recover effects far from the threshold. An instrument must shift treatment, satisfy an exclusion condition for the outcome, and meet the remaining assumptions appropriate to its estimand. Insurance regulators should therefore ask what event created identifying variation and why alternative pathways are implausible.
Heterogeneous effects and causal learning
Risk factors that strongly predict outcomes may also modify treatment effects, but predictive and causal heterogeneity are different objects. Let τ(x)=E[Y(1)-Y(0)|X=x]. Estimating τ(x) requires treatment comparisons within relevant covariate support; ranking baseline risk E[Y(0)|X=x] alone does not rank expected benefit. A high-risk insured may be highly preventable or almost entirely exposed to uncontrollable hazard. A low-risk insured may still have a large relative response to a targeted intervention. A risk-optimized allocation rule and an effect-optimized allocation rule can therefore identify different participants.
Wager and Athey's causal forest methodology develops estimation of heterogeneous effects under conditions including unconfoundedness. The methods can prioritize an insurance intervention when properly identified treatment variation exists and when treatment costs and effect uncertainty are accounted for. A model's strongest splits, however, must not be read as mechanisms without further evidence. A subgroup selected after observing many candidate effect estimates faces multiplicity and overfitting; honest estimation, independent evaluation, pre-specified subgroup hypotheses, and interval reporting are relevant safeguards.
Counterfactual Prediction and Transport
Forecasts after a changed decision rule
An insurer normally asks how much a defined portfolio will cost next year under existing contracts and current operating arrangements. The corresponding predictive target may be E[Y_future | portfolio, existing policy]. A policy reform asks for the outcome under a counterfactual assignment rule d, conventionally E[Y(d) | target portfolio]. If d changes both individual behavior and the composition of insured exposures, the historical conditional loss model needs a separate representation of participation and responses. Merely replacing a treatment value in a fitted prediction function while keeping all other features fixed is a well-defined computational operation, but it is a causal forecast only when its counterfactual assumptions are justified.
Consider a simplified portfolio of 10,000 independent one-year exposures with stable expected claim cost of $800 each under the current program. A pure baseline forecast gives $8.0 million expected cost. Suppose an intervention would reduce losses for existing participants by 5% but also attract 2,000 new exposures with an expected cost of $1,200 per exposure after intervention. Expected paid losses become (10,000×$800×0.95)+(2,000×$1,200)=$10.0 million. The total increases despite beneficial effects on the original participants, because the insured population changes. Comparing $8.0 million and $10.0 million does not show that the program harmed incumbent insureds. If administrative cost, premiums, subsidies, new coverage benefits, and consumer risk reduction matter to the decision, the relevant economic outcome must include those quantities on a stated accounting and welfare basis.
For a portfolio that experiences multiple claim outcomes, counterfactual forecasts may require a joint distribution of frequency, severity, reporting delay, retention, and capital usage. Changing a deductible affects the insurer's payment function, the propensity to file, possibly prevention behavior, and the distribution of retained risks. The counterfactual mean paid cost under deductible d can be written E[1{reported(d)}×max(L(d)−d,0)] for a simplified per-loss contract, where L(d) is the physical loss under the regime and reported(d) denotes reporting. Treating L, reporting, and portfolio selection as constant is a substantive restriction, not an algebraic identity. It may be serviceable for tightly limited short-term scenario analysis, but it needs to be stated as such.
Transport from trial to market and between insurance lines
A causal effect identified within an experiment applies directly to its defined experimental population and treatment versions. External application requires additional information about how effect modifiers, implementation, compliance, and baseline conditions differ. The transport problem is acute in insurance because risks depend on local hazard, health systems, legal coverage, claims infrastructure, socioeconomic distribution, and premium responses. Dahabreh and collaborators' transportability methodology studies extending randomized-trial causal inferences to a new target population using trial and target-population information under explicit assumptions.
If exchangeability across trial participation holds conditional on measured effect modifiers X, a target-population effect can be reconstructed by averaging trial conditional treatment effects over the target distribution of X. Such transport also needs compatible outcome definitions, relevant overlap, and comparable intervention versions. A preventive intervention evaluated among urban owner-occupied homes cannot simply be assigned the same effect in remote rental properties with different maintenance practices and emergency services. Similarly, an insurance expansion result from a particular public eligibility system does not automatically identify an effect for a voluntary private product with premiums, exclusions, networks, and renewal incentives. Evidence should report whether it supports the trial-population effect, a reweighted target effect, or a bounded qualitative inference.
The difference between transport of predictions and transport of causal effects is worth emphasizing. Prediction transport asks whether P_target(Y|X) remains close to the learned conditional distribution. Causal transport asks whether contrasts between Y(a) under interventions, possibly conditional on X, carry to the target. A predictive model can fail because baseline loss frequency shifts while a preventive action's effect remains stable; a causal effect can change because a key effect modifier shifts even when overall predictive calibration appears acceptable. Validation protocols should test the quantity relevant to the research objective.
Insurance pricing, participation, and welfare
Insurance prices allocate financial risk, alter demand, and affect adverse-selection margins. The observed claims of purchasers at a given premium describe people who elected coverage, not the cost distribution of everyone who might purchase at a different premium. Einav, Finkelstein, and Cullen's welfare analysis using price variation exploits variation in insurance prices to infer demand and the cost composition of marginal participants in a specific employer setting. Their approach uses identifying variation to move from observed choice and cost patterns toward welfare analysis; its substantive conclusions are tied to the institution and sample studied.
A predictive risk score may be excellent at classifying existing policyholders yet inadequate for assessing a subsidy. Subsidies can change who enrolls, utilization, the degree of financial protection, insurer costs, and fiscal expenditures. Welfare depends on benefits and costs beyond the insurer's own claim payments. The analytical comparison should state whether the objective is financial soundness, underwriting profit, total social surplus, consumer welfare, access, or distributional effects. These objectives may overlap, yet they imply different treatment outcomes and policy rankings. An intervention producing higher insurer claim expenditures can simultaneously create valuable coverage or reduce catastrophic financial exposure for households.
The object 'risk' itself has several levels. Insurable physical risk can be approximated by expected real-world loss; contractual risk borne by the insurer depends on coverage and retention; financial risk for households depends on uncovered cost, premiums, and income; and systemic risk depends on correlated institutions and shocks. A causal claim must identify which outcome changes. A policy might reduce the insurer's paid losses by increasing deductibles while shifting financial volatility to households. Calling this an improvement in risk without specifying whose risk obscures its distributional structure.
Prediction after deployment and feedback loops
Underwriting and claims models can alter the environments from which their training data came. Suppose a claims triage algorithm flags difficult files and directs them toward intensive review. Historical 'paid within 30 days' is then partly caused by prior triage rules, staffing, and claim characteristics. Retraining on post-deployment files might learn the consequences of the organization's own decisions rather than a stable inherent feature of claimants. A feature's association with claim outcomes can strengthen because the insurer allocates staff or investigation resources based on it. The research analyst should reconstruct the decision-feedback path, retain time-stamped treatment and outcome definitions, and examine how model deployment changes selection and measurement.
This is not unique to artificial intelligence. Traditional underwriting manuals, agent discretion, manual fraud referrals, and reinsurance appetite have long affected who enters the portfolio and how events become documented. Automated systems increase scale and can change update frequency, but the causal question remains institutional: what decision regime generated the data, and what would happen under another regime? The ASOP No. 56 modeling guidance provides a relevant professional expectation for assessing purpose, model risk, assumptions, validation, and documentation in its applicable U.S. actuarial context. It is not an international statute requiring causal identification in every pricing model.
Evaluation with Real Policy Changes
The Oregon insurance lottery
The Oregon Health Insurance Experiment provides a particularly informative insurance illustration because it distinguishes randomized access from realized coverage. Finkelstein and colleagues' first-year experimental results examined a lottery that gave selected low-income adults the opportunity to apply for Medicaid. Lottery selection created an assignment contrast; actual insurance coverage differed because not every person offered the opportunity enrolled or qualified. The study reported higher health-care utilization and lower measures of financial strain for the selected group in the first-year window. The intervention was an offer under a particular public insurance system, and the first-year results therefore should not be restated as universal effects of all insurance products.
In a later clinical follow-up, Baicker et al.'s two-year outcome study reported important differences among outcomes: Medicaid coverage increased service use, diabetes detection and medication, reduced a depression-screening measure, and substantially reduced catastrophic out-of-pocket spending, while the trial did not find statistically significant improvements in the measured blood-pressure, cholesterol, and glycated-hemoglobin endpoints over the stated horizon. The finding 'no statistically significant improvement' is a statement about the tested outcomes, precision, and design; it does not prove a structural zero for every potential health effect. The researchers distinguished lottery assignment from enrollment with instrumental-variable reasoning, including a first-stage change in coverage probability.
This design makes three inferential layers visible. The offer has an intention-to-treat effect in the lottery population. The enrollment effect inferred with the lottery depends on instrumental-variable assumptions and refers to the people whose coverage is altered by the offer. A prediction model fitted to the study participants could forecast utilization conditional on personal characteristics, but its predictive discrimination would not recreate random assignment or reveal the treatment effect. Nor could the observed benefits for this population establish an identical effect in a different insurance system with different eligibility, providers, and baseline health risks.
Premium variation and endogenous demand
The price-variation welfare study by Einav, Finkelstein, and Cullen (2010) illustrates a different design: variation in prices is used to estimate demand and the cost of consumers at the margin of purchase. This approach responds to the essential insurance problem that average observed costs combine heterogeneous underlying risk with endogenous selection. The same individual may respond to price, coverage generosity, employer contribution, and personal anticipated needs. Observational correlations between cost and plan type can therefore reflect both selection and responses to insurance.
The related selection-on-moral-hazard evidence of Einav et al. (2013) used employee-level panel data from one firm and a model of plan choice and utilization to investigate heterogeneous behavioral responses to coverage. Its central warning for decision-makers is conditional: policy effects can differ across people, and the choice to enroll can depend on anticipated responsiveness. A predictive model trained to rank high spenders does not by itself rank people by their response to deductible changes. This distinction is central to estimating the net effect of changing an employer or insurer's benefit design.
These studies also have bounded external validity. An employer's menu of plans, tax treatment, contribution policies, and employee distribution influence available variation and the cost-demand relationship. Their observed mechanisms may be informative across insurance settings, but direct quantitative transfer to another country, non-employer market, or line of insurance requires new arguments and data. A literature review should separate the authors' actual design and reported results from an encyclopedia author's interpretation of implications for general insurance reasoning.
Evaluation of underwriting and risk-prevention changes
A property insurer testing a rooftop reinforcement incentive should specify whether treatment means a financial offer, completed installation, verified compliance, or long-term maintenance. Claims records can identify some insured outcomes, but reliable conclusions also require accounting for weather exposure, catastrophic event clustering, geographic differences, building type, and spatial spillovers. Randomizing incentives may be possible; when it is not, geographical rollout could support carefully specified differences-in-differences if parallel counterfactual trends and event measurement are defensible. Negative-control outcomes, placebo dates, pre-intervention outcome patterns, and alternative exposure normalizations can reveal design weaknesses, although they do not alone establish full identification.
Auto-insurance telematics creates related issues. Safer drivers may voluntarily enroll, monitoring may alter driving, pricing incentives may change trip exposure, and privacy preferences may shape retention. A program's lower observed claim frequency among enrolled drivers could reflect any combination of these. To evaluate behavioral effects, a study might randomize the offer or pricing schedule, measure driving exposure consistently, and distinguish an intention-to-offer effect from adoption and behavior effects. An insurer seeking to use these findings for pricing must additionally examine whether improved claims predictions are lawful, fair, and contractually appropriate within the jurisdiction at issue.
In claims administration, one can evaluate automated triage by randomized or plausibly exogenous assignment of comparable files to the new and old procedures. Suitable outcomes include timeliness, warranted payments, complaint reversals, administrative cost, claimant experience, and recovery quality. A shorter time to closure alone may indicate legitimate efficiency or premature denials. The correct outcome set reflects both institutional resources and fulfillment of contract obligations. Studies that measure only accounting expense or number of closed claims should make clear which important outcomes they have not tested.
Comparative strength of research designs
Randomization provides a powerful assignment mechanism, but its relevance depends on treatment, compliance, attrition, interference, and target population. Natural experiments can offer credible identifying variation in real insurance institutions but need a detailed exclusion, threshold, or parallel-trends argument. Observational adjustment may support causal interpretation when covariates and assignment processes are sufficiently understood, with sensitivity analysis for remaining confounding. Pure prediction studies can identify important forecasting relationships and generate hypotheses, but their metrics alone do not identify the effects of an action. The relative strength of designs is therefore conditional on the precise causal claim, not a single blanket ranking of study labels.
A mature evidence synthesis reports the estimand before the point estimate, the outcome and comparison period before the significance level, the assignment mechanism before the estimator, and the population before the extrapolation. The publication record should distinguish randomized results from structural-model estimates and observational associations even when all are reported numerically. A regulator, actuary, or insurer using such work can then see what the evidence actually authorizes: improved prediction, internally identified causal effects, externally transported estimates, or a still-hypothetical policy scenario.
Causal Interpretation and Scientific Boundaries
Distinguishing statistical validation from identification
A predictive model should be evaluated on the outcome distribution and decision task for which it is intended. Loss functions, calibration, discrimination, tail behavior, and portfolio stability can demonstrate that a forecast meets a stated requirement. A causal design should be evaluated on a different set of grounds: clarity of treatment and outcome, consistency, timing, identification assumptions, overlap, effect estimation, uncertainty, sensitivity to confounding, and the target population. A model can perform very well on one axis and poorly on another. The prediction performance framework of Steyerberg et al. (2010) provides an established vocabulary for examining several components of predictive quality; the causal inference text by Hernán and Robins (2020) provides a corresponding language for interventions and identifying conditions.
Data-generating-process awareness should determine the validation split. A claims-frequency forecast trained on data from policies already selected under a particular risk-scoring model can show good calibration among current insureds. It cannot be taken as proof of how formerly excluded applicants would behave if the acceptance rule changed. A causal evaluation of an expanded eligibility rule must observe or credibly reconstruct outcomes for newly eligible people and account for selective enrollment. Where outcomes are fundamentally unobserved for excluded applicants, no computationally sophisticated adjustment identifies them without additional information or assumptions.
For both forecasting and causal analysis, data quality should be understood as a relationship between measurement and the estimand. Duplicate claims, missing exposure, incorrect policy inception dates, changes in reserving conventions, case-definition drift, delayed settlement, and unrecorded cancellations can cause severe distortion. A claims payment may serve as an accurate indicator of insurer cash outflow yet an incomplete proxy for underlying physical injury. Whether this matters depends on whether the research target is incurred loss, cash expense, compensated harm, or physical risk. Source data should be accompanied by a dictionary defining exposure, claim occurrence, claim report, claims-paid development, policy changes, and observation horizons.
Validation of causal estimates
Causal estimation needs falsification-oriented diagnostics that match the research design. Randomized evaluations should report allocation integrity, baseline comparability, attrition, noncompliance, clustering, and outcome ascertainment. Observational comparisons should present the causal diagram or an equivalent assignment narrative, observed confounder balance, positivity, sensitivity to unobserved bias, and the effect of alternative specifications. Instrumental-variable studies should disclose the first stage, substantive exclusion rationale, monotonicity or related assumptions where relevant, and the population identified. Difference-in-differences studies should show credible counterfactual trend evidence, address treatment timing and anticipation, and use uncertainty procedures appropriate to grouped policy rollouts.
These tests have a fundamental asymmetry: they can detect certain contradictions, but their success cannot logically establish every assumption that concerns missing counterfactuals. For example, a zero pre-treatment gap in average claims does not show that two insured populations would have maintained identical unobserved trends in the next hurricane season. Balance on observed demographics does not show balance on unrecorded willingness to claim. An instrument can strongly predict enrollment and still have a direct effect on care access or financial stress. Researchers should present the substantive institutional case for each identification restriction and reserve causal language for conclusions that that case can support.
An explicit sensitivity analysis is more informative than a ritual claim of robustness. It should ask how large an unmeasured common cause or an exclusion violation would need to be to change the result, whether an extreme propensity stratum drives the estimate, and what happens under plausible changes in event timing or outcome definitions. Uncertainty should include appropriate sampling variation, with clear separation between measured statistical uncertainty and structural uncertainty about assumptions. Catastrophic losses, common weather systems, network effects, and rare claims challenge independence assumptions that might otherwise produce deceptively narrow intervals.
Documentation, reproducibility, and reproducible computation
A reproducible causal insurance study records the target estimand in advance, operational definitions, eligibility criteria, assignment rule, cohort entry, time zero, exposure, confounders, outcome window, exclusions, missingness handling, and estimation code. It should preserve source-data provenance and a versioned description of policy rules. Reproducibility can be achieved with privacy-preserving synthetic data or sufficiently detailed computational examples when real claims records cannot be shared. Such examples must be labeled synthetic and must not be presented as independent empirical validation.
A technical appendix for a quantitative study should contain the exact functional specification or estimator, design-dependent assumptions, standard-error procedure, sample accounting, and output definitions. Modern causal machine learning additionally needs cross-fitting folds, hyperparameter settings, regularization choices, software versions, and procedures for tuning without contaminating evaluation. The double/debiased machine learning research explains why orthogonalization and data splitting serve particular statistical properties; merely reporting that a system used a sophisticated estimator is insufficient. When algorithms are proprietary, an auditable description of the assignment mechanism, the uncertainty, and the outcome definition remains necessary even if code disclosure is limited.
The two hypothetical calculations in this dossier are fully reproducible by hand. In the safety-device example, 0.9×9+0.1×1=8.2, 0.1×10+0.9×2=2.8, and 8.2−2.8=+5.4, whereas the groupwise causal contrasts 9−10 and 1−2 are each −1. In the portfolio example, 10,000×800=8,000,000; 10,000×800×0.95=7,600,000; 2,000×1,200=2,400,000; total counterfactual payments equal 10,000,000 under the stipulated conditions. These are deterministic implications of stated inputs, not population estimates. A third-party reviewer can change the risk composition or program effect and inspect how the association and counterfactual result respond.
Supervisory and professional interpretation
Supervisory expectations for insurance algorithms often concern reliability, governance, unfair discrimination, and lawful treatment of consumers. These requirements should be read with precision rather than recast as a universal demand that every rating variable have a directly manipulable causal effect. The New York State Department of Financial Services' Circular Letter No. 7 (2024) addresses the use of artificial intelligence systems and external consumer data in underwriting and pricing. It refers to empirical relationships, actuarial validity, proxy assessments, quantitative analyses of potentially adverse effects, documentation, and governance. These are New York supervisory expectations within the circular's scope, including its stated exclusions; they do not prove that each predictor is a causal determinant of physical loss.
The National Association of Insurance Commissioners' 2023 model AI bulletin describes governance and consumer-protection expectations for insurers' use of AI systems, with model language intended for state-level application. It is a model bulletin and should not be represented as a self-executing federal statute. Within the professional actuarial sphere, the ASOP No. 56 Modeling standard covers a broad modeling lifecycle for actuaries in its scope, including purpose, assumptions, limitations, validation, and communications. Conformity with modeling governance may improve accountability for both prediction and causal studies, yet substantive causal identification remains a research-design matter.
The European Union's Artificial Intelligence Act identifies specified uses of AI systems for risk assessment and pricing relating to natural persons in life and health insurance among the Annex III high-risk categories, subject to the regulation's definitions, exemptions, application rules, and amendments. That classification reflects regulatory concerns about significant effects on individuals, not a statistical declaration that model features cause the predicted outcomes. The United Kingdom Information Commissioner's Office's automated decision-making guidance explains safeguards and restrictions within the UK data protection framework, while noting ongoing revisions in light of the Data (Use and Access) Act 2025. The legal analysis must therefore track jurisdiction, current commencement, decision context, and the distinction between explaining an automated decision and identifying causal mechanisms of insured losses.
Discrimination analysis needs its own estimands. A disparity in premiums, denial rates, or outcomes between protected groups is an observable distributional association under given decisions. It may warrant a legal or ethical assessment even when its individual causal pathways remain unidentified. Conversely, establishing a causal mechanism does not by itself determine whether a pricing practice is permitted under local law. A rigorous paper can present predictive calibration, group impacts, counterfactual effects of a changed rule, and the governing legal standard separately. This preserves scientific clarity and enables a regulator or reader to understand exactly which conclusion rests on evidence and which rests on a normative or legal judgment.
Limits of this research and the evidence available
This dossier synthesizes selected foundational statistical research, selected insurance-economics studies, identifiable experimental results, and selected official institutional documents as of November 14, 2026. It is not a registered systematic literature review and does not claim complete global coverage of causal applications in every insurance line. Some publisher access was limited to abstracts, bibliographic records, or available author versions; the detailed reading scope is documented in the evidence ledger. Insurance records with personally identifiable claims histories were neither accessed nor analyzed. The causal effects calculated in the illustrative examples are consequences of assumed numbers, not claims about measured interventions.
Substantive limitations are inherent to the inference tasks. Counterfactual outcomes cannot both be directly observed for a single unit at a single time. Identification assumptions often depend on institutional facts that datasets alone cannot prove. Severe claims may be rare and correlated; insured populations and claim reporting rules can change after pricing or intervention decisions; effects may vary across time and jurisdiction. A limited available empirical basis should produce a bounded conclusion rather than a universal numerical estimate. External peer review and original prospective data collection are separate activities and have not been undertaken for this manuscript.
Integrated analytical findings
The research supports five connected findings. First, an insurance prediction model estimates a well-defined observational quantity only when its target, sample selection, time origin, and outcome definition are clear. Second, a causal statement requires a specified action and contrast of potential outcomes, along with a credible identification argument. Third, insurer-controlled selection, contractual design, and claims administration can generate predictive features that reflect institutional practices rather than independent physical causes. Fourth, real experimental or quasi-experimental variation can connect particular insurance policies to identified effects, but the effect population and assumptions remain part of the result. Fifth, a decision rule must be evaluated on the outcomes and distribution relevant to its purpose, including consumers and market participation when those consequences are material.
These findings are an analytic organization of established statistical and insurance-economics reasoning applied to insurance institutions. They establish no new named causal theory, valuation doctrine, or universal regulatory test. The constructive research agenda is specific: pre-register target estimands for insurance interventions; record assignment mechanisms and policy versions; distinguish claim payment from physical harm; measure portfolio entry and exit; report overlap and sensitivity; and test transport when programs move across populations. The priority is traceable inference rather than terminological expansion.
Conclusion
Predictive association and causal inference perform related but distinct functions in insurance science. Prediction maps observed information to expected outcomes under a defined data-generating environment. Causal inference compares outcomes under alternative actions and must make an intervention, target population, assignment mechanism, and identifying assumptions explicit. A highly accurate claims model can validly forecast losses without establishing what changing one of its predictors would do. A credible causal experiment can identify the effect of a coverage offer or prevention intervention even when its outcomes remain difficult to predict for particular individuals.
The distinction has practical consequences at every major insurance decision point. Underwriting prediction informs the anticipated cost of accepting risk, while a causal study of underwriting changes examines participation, behavior, and financial outcomes after a revised rule. Claims forecasting supports staffing and reserving, while evaluating claims-management changes requires a defensible counterfactual and claimant-relevant outcomes. Insurance price analysis must distinguish average cost among current purchasers from selection and welfare responses at new prices. Public coverage research must separate randomized eligibility offers from actual take-up and from external transfer to other insurance systems.
The hypothetical safety-device example demonstrates that an association can have the opposite sign from the treatment effect when adoption is selected on underlying risk. The counterfactual portfolio example shows that total claims expenditure can rise after a beneficial intervention because the insured population expands. Both results follow exactly from their stated assumptions and demonstrate why a research question must precede statistical interpretation. The reviewed literature supplies methods for treating confounding, longitudinal feedback, experiments, natural experiments, heterogeneity, and transportability, with explicit boundaries around their identifying power.
Insurance science advances when observational forecasts are assessed as forecasts and causal claims are assessed as claims about interventions. Well-designed predictive models and credible causal studies can jointly improve decisions, provided each is evaluated against the quantity it actually estimates. The central scientific obligation is to preserve this distinction from the initial question through the model, the evidence, the documented inference, and the eventual insurance decision.
References
Actuarial Standards Board. (2019; effective October 1, 2020). Actuarial Standard of Practice No. 56: Modeling. Official standard.
Baicker, K., Taubman, S. L., Allen, H. L., Bernstein, M., Gruber, J. H., Newhouse, J. P., Schneider, E. C., Wright, B. J., Zaslavsky, A. M., & Finkelstein, A. N. (2013). The Oregon experiment—Effects of Medicaid on clinical outcomes. New England Journal of Medicine, 368(18), 1713–1722. https://doi.org/10.1056/NEJMsa1212321.
Breiman, L. (2001). Statistical modeling: The two cultures (with comments and a rejoinder). Statistical Science, 16(3), 199–231. https://doi.org/10.1214/ss/1009213726.
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., & Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1), C1–C68. https://doi.org/10.1111/ectj.12097.
Dahabreh, I. J., Robertson, S. E., Steingrimsson, J. A., Stuart, E. A., & Hernán, M. A. (2020). Extending inferences from a randomized trial to a new target population. Statistics in Medicine, 39(14), 1999–2014. https://doi.org/10.1002/sim.8426.
Einav, L., Finkelstein, A., & Cullen, M. R. (2010). Estimating welfare in insurance markets using variation in prices. The Quarterly Journal of Economics, 125(3), 877–921. https://doi.org/10.1162/qjec.2010.125.3.877.
Einav, L., Finkelstein, A., Ryan, S. P., Schrimpf, P., & Cullen, M. R. (2013). Selection on moral hazard in health insurance. American Economic Review, 103(1), 178–219. https://doi.org/10.1257/aer.103.1.178.
European Parliament and Council of the European Union. (2024; consolidated text as accessed 2026). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Annex III and Article 6. Official EUR-Lex legislation.
Finkelstein, A., Taubman, S., Wright, B., Bernstein, M., Gruber, J., Newhouse, J. P., Allen, H., Baicker, K., & Oregon Health Study Group. (2012). The Oregon health insurance experiment: Evidence from the first year. The Quarterly Journal of Economics, 127(3), 1057–1106. https://doi.org/10.1093/qje/qjs020.
Hernán, M. A., & Robins, J. M. (2020; online updated edition). Causal Inference: What If. Boca Raton, FL: Chapman & Hall/CRC. Author-maintained book and materials.
Hernán, M. A., Sauer, B. C., Hernández-Díaz, S., Platt, R., & Shrier, I. (2016). Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses. Journal of Clinical Epidemiology, 79, 70–75. https://doi.org/10.1016/j.jclinepi.2016.04.014.
Imbens, G. W., & Angrist, J. D. (1994). Identification and estimation of local average treatment effects. Econometrica, 62(2), 467–475. Stanford publication record.
Information Commissioner's Office. (2026 access; guidance under review). Rights related to automated decision making including profiling. United Kingdom regulatory guidance. Official ICO guidance.
King, G., & Nielsen, R. (2019). Why propensity scores should not be used for matching. Political Analysis, 27(4), 435–454. https://doi.org/10.1017/pan.2019.11.
National Association of Insurance Commissioners. (2023, December 4). Model Bulletin: Use of Artificial Intelligence Systems by Insurers. Adopted model supervisory bulletin. Official model bulletin.
New York State Department of Financial Services. (2024, July 11). Insurance Circular Letter No. 7 (2024): Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing. Official circular letter.
Robins, J. M., Hernán, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550–560. https://doi.org/10.1097/00001648-200009000-00011.
Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1), 41–55. https://doi.org/10.1093/biomet/70.1.41.
Roth, J., Sant'Anna, P. H. C., Bilinski, A., & Poe, J. (2023). What's trending in difference-in-differences? A synthesis of the recent econometrics literature. Journal of Econometrics, 235(2), 2218–2244. https://doi.org/10.1016/j.jeconom.2023.03.008.
Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5), 688–701. https://doi.org/10.1037/h0037350.
Shmueli, G. (2010). To explain or to predict? Statistical Science, 25(3), 289–310. https://doi.org/10.1214/10-STS330.
Steyerberg, E. W., Vickers, A. J., Cook, N. R., Gerds, T., Gonen, M., Obuchowski, N., Pencina, M. J., & Kattan, M. W. (2010). Assessing the performance of prediction models: A framework for some traditional and novel measures. Epidemiology, 21(1), 128–138. https://doi.org/10.1097/EDE.0b013e3181c30fb2.
Wager, S., & Athey, S. (2018). Estimation and inference of heterogeneous treatment effects using random forests. Journal of the American Statistical Association, 113(523), 1228–1242. https://doi.org/10.1080/01621459.2017.1319839.