Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Insurance Distribution Channel Causal Effect Evaluation

Abstract. The causal evaluation of insurance distribution channels studies how changing the institutional route through which eligible people and organizations encounter, choose, purchase, maintain and use insurance changes their economic protection. Direct insurer distribution, exclusive and independent agencies, brokerage, bancassurance, digital marketplaces, employer arrangements and embedded insurance differ in their product menus, access costs, information, advice, incentives and subsequent service. These differences also determine the population selected into each channel, creating a central identification problem. This monograph develops a comparative analysis of causal estimands, eligibility and treatment versions, experimental and observational identification, outcome construction, risk composition, heterogeneity, international transportability and reproducibility. It synthesizes selected primary field experiments from India, Nicaragua and the United States, institutional evidence from international and European supervision, UK market studies, and established econometric scholarship. The analysis distinguishes assignment from receipt, the completed transaction from continuing protection, and causal outcome changes from risk selection, accounting allocation and aggregate market trends. Hypothetical numerical examples demonstrate how channel comparisons can change when the population, denominator or observation horizon changes. The resulting argument is that a meaningful channel evaluation must specify an actionable intervention and a shared counterfactual, establish credible identification and trace outcomes beyond acquisition. The study contributes an original interpretive synthesis of existing causal methods and insurance economics. It does not estimate a new empirical effect or claim a universally superior channel.

Keywords: Insurance distribution; insurance intermediation; causal inference; counterfactuals; distribution channels; randomized encouragement; instrumental variables; difference-in-differences; risk selection; consumer outcomes; brokerage; bancassurance; embedded insurance; insurance inclusion; persistency; external validity.

Introduction

Insurance is a promise of financial performance under specified contingencies. Distribution determines whether that promise becomes available, intelligible, affordable and enforceable for a particular eligible population. The economic importance of distribution begins before a contract exists and continues through the period in which its obligations can be realized. A household that receives several accurate quotations, understands exclusions and obtains appropriate cover has experienced a different distribution process from an otherwise comparable household that receives only one quotation through a lender. A firm whose broker negotiates contingent business interruption insurance encounters a different institutional process from one that purchases a standardized commercial package directly. The scientific problem is the consequence of changing such institutional arrangements while allowing for the numerous reasons why customers reach them in the first place.

The phrase channel effect is frequently used to describe differences between portfolios assigned to distribution categories. A distribution channel can be profitable because its customers are wealthy, because its products have low expected losses, because it serves people in relatively affluent regions, because its contracts are more expensive or because it provides useful advice at modest cost. These mechanisms have different economic and policy implications. The distinction matters equally for consumer protection: high cancellation rates might reflect coercive selling, poor product design, loss of income, superior opportunities to switch, or a portfolio exposed to deteriorating economic conditions. An observed channel coefficient encompasses these processes unless a design separates them.

This dossier belongs to the Science of Insurance Distribution and Intermediation. Its independent object is the identification and interpretation of causal consequences of institutional distribution arrangements. Aggregate acquisition efficiency, experimental choice presentation, particular intermediary remuneration structures and distributor data-quality practices are relevant because they shape causal estimation; their broader institutional treatment remains the subject of adjacent research. The analysis is international, with jurisdiction-specific evidence preserved and no assumption of uniform licensing, supervision or welfare institutions.

The documentary research uses a purposive, question-driven selection of empirical studies, official texts, regulator reports and established causal methodological works. It is a critical narrative synthesis rather than a systematic review with exhaustive database screening or a meta-analysis. Sources are prioritized when they establish an intervention, identify a comparison group or clarify the institutional meaning of an outcome. The research ledger records the version and accessible content for each source. Abstracts and official summaries are used only for propositions warranted by them. Original microdata, code, insurer ledgers and unpublished field records were not obtained; numerical exercises that follow use explicitly constructed inputs. No jurisdiction-wide or global causal coefficient is inferred from the literature. The evidentiary gap for controlled institutional comparisons involving employer, broker, banker and embedded routes is itself analytically important.

Research Design for Insurance Channel Causality

Defining the counterfactual

A channel is a distribution arrangement defined at an identifiable decision point by its responsible organizations, accessible product set, assistance, communication medium, pricing arrangements and authority to conclude or service a contract. These attributes must be recorded as part of treatment definition. In a field experiment, assigning an eligible applicant to a broker-assisted pathway can mean offering one call, providing a series of advisory meetings, allocating access to a particular insurer panel or delegating the entire placement. These are materially different treatments. A label such as broker versus direct can describe an intervention only after its versions and fallback options are identified.

Let A denote an assigned distribution strategy, D the actual channel received, X preassignment attributes, Y a prespecified outcome, and t the observation horizon. For an eligible unit i, Y_i(a,t) is the outcome that would be observed at horizon t if strategy a were assigned under a specified implementation. For two strategies a and b, the average assignment effect is E[Y(a,t) − Y(b,t)] in the eligible population. Its interpretation depends on the type of unit. A person may hold several policies, a household may make a shared coverage decision, an employer may enroll thousands of workers, and a broker may handle several correlated commercial exposures. Assignment, observation and aggregation units must therefore be distinguished at the outset.

The counterfactual cannot usually be observed for the same person and moment. Causal identification uses randomization or an argument establishing comparability under explicit assumptions. Hernán and Robins' Causal Inference: What If (2020) organizes this methodological problem around causal contrasts, observational adjustment and longitudinal strategies. For insurance, the target population should typically be people or risks eligible for both alternatives before distribution begins. If risk eligibility is itself the outcome of the intervention, the target should instead be the population eligible to be contacted before underwriting, with underwriting acceptance measured as a downstream event.

The estimand and the policy question

The effect of offering an additional agent appointment may be directly actionable even where few prospective customers use it. An intention-to-treat contrast answers whether deploying that offer improves outcomes over the existing pathway. A per-protocol effect concerns following a specified strategy; it requires defensible handling of deviations and censoring. A treatment effect among actual users is a different parameter: the users of personal advice may have greater preexisting insurance need or motivation than those who declined advice. An average effect among compliers with randomized encouragement adds still another target. Each of these estimands has a legitimate role, but they cannot share a single numerical label.

An intervention comparing channel institutions with their existing product menus estimates a package effect. This includes the product opportunities created by the intermediary relationship. An intervention holding the same products and premiums fixed while varying counseling estimates a narrower service effect. Both matter: an insurer choosing a distribution contract may care about the package, whereas a supervisor investigating mis-selling under a common product design may care mainly about the advice process. Contract standardization can aid identification while removing an important mechanism of the broader institutional intervention. The research question determines which approach is appropriate.

The causal outcome can be defined at several horizons. Initial quote availability and binding within 30 days are process outcomes; premium payment and continuous coverage at 12 months capture later events; reimbursement and claim satisfaction may require longer observation. An intervention can increase binding without increasing the probability of effective protection when the underlying event occurs. For indemnity products, calendar time must also be aligned with exposure to risk. A hurricane season or concentrated epidemic changes the economic significance of identical quantities of person-months or property-months.

Identification and the design specification

Consistency requires that observed outcomes for a person receiving a defined strategy correspond to the potential outcome for that version of the strategy. Exchangeability concerns comparability of potential outcomes across assignments, conditional on specified baseline variables where appropriate. Positivity requires nonzero probability of the relevant alternatives within the comparison population. Interference arises when treatment of one unit changes the outcome of another, for example through service queues, product quotas, referrals or household decisions. A trial with fixed broker capacity may affect customers outside the treated sample through changes in waiting times. Consequently, a well-specified experiment defines the treatment allocation and market environment as well as the individual route.

A prospective protocol should record the research population, exclusions determined before assignment, alternatives, time zero, assignment mechanism, service and product versions, periods of exposure, outcome definition, missing-data rules, estimation strategy, expected precision and applicable consumer safeguards. The event time for an invitation precedes quote, underwriting, acceptance, premium collection and activation. If customers enter the database only after activation, a study loses the unsuccessful eligible population and cannot recover the total effect on uptake from policy records alone.

The choice of data system follows the target estimand. An insurer can compare profitability among its acquired customers, but without information about eligible people who did not buy or chose another insurer it cannot estimate a general population coverage effect. A marketplace may observe competing products but not continuing insurance at another marketplace. A regulator may link carrier records to complaints, but measures observed complaints jointly reflect detriment and reporting access. A claim against a channel's causal effectiveness thus begins with the construction of a credible observation system.

Assignment, mediators and time

Distribution is dynamic. A consumer may start online, receive telephone help, move to an agency, bind through an insurer and subsequently manage the policy electronically. A single last-touch field has value for remuneration accounting; it does not identify this causal sequence. One may compare dynamic strategies such as online by default with assistance offered after 48 hours of inactivity versus immediate assistance. The strategy specifies what happens after observable intermediate states, so that a counterfactual is coherent even when people change their behavior.

For a total assignment effect, subsequent advice, quotes, price and acceptance are mediators. Adjusting for them can remove precisely the effects of interest or create bias through conditioning on a common consequence of treatment and unobserved risk. Separating effects through information, product choice or price requires additional mediation assumptions and often a separate experimental intervention. Causal inference cannot infer direct and indirect effects merely by comparing regression coefficients from models with and without a postassignment variable.

The authorial interpretation is that channel identity is scientifically useful when it denotes an implementable trajectory and an accountable institutional arrangement. A prospective allocation, a product-constrained service comparison and a long-term pathway comparison address different decisions. Each can be evaluated rigorously, provided the relevant target population and time horizon remain attached to the result.

Institutional Channels and Their Choice Mechanisms

Insurer direct distribution

Direct distribution includes contact with the insurer's sales staff, its branch facilities and remote services under its control. An automated quote environment can reduce the cost of acquiring standardized cover. It can also limit the visible product set to a single insurer and transfer effort to the applicant. Where product complexity is low, the burden of comparing coverage may be modest; where the risk depends on unusual exclusions, valuation or technical limits, a direct service can require considerable specialized underwriting assistance. Direct distribution is neither a uniform treatment nor necessarily synonymous with absence of human advice.

A trial might assign eligible inquiries to a direct self-service path or a human-supported direct route while preserving product availability. That comparison isolates assistance in a specific environment. Assigning customers instead to direct insurer routes or an independent broker estimates a different bundle, including cross-carrier search. The relevant counterfactual depends on the products genuinely available to each customer.

Agency and brokerage

Exclusive agents can combine local information and specialized knowledge of a particular carrier. Independent agents and brokers can search multiple insurers, collect risk information and coordinate service. The extent of market access and duties to customers depends on contract and local law. Distinguish agents representing several principals, brokers paid by customers, brokers remunerated by insurers, wholesale specialists and delegated underwriting operations. Their effects can vary with placement complexity, product quality and incentives.

Berger, Cummins and Weiss (1997), in their study of distribution systems, compare insurer cost and profit efficiency in property-liability insurance and report support for a product-quality explanation of coexistence between direct writers and independent-agency insurers. The source provides an economically important alternative to interpreting higher expenses as pure inefficiency. The design, as characterized by the accessible abstract, is not random assignment of identical customers to channels. Its reported results therefore motivate hypotheses rather than a causal channel ranking. A direct-to-broker intervention should measure both remuneration and the contract services made available to the customer.

Agency may lower search costs while increasing the effect of incentives on recommendations. Anagol, Cole and Sarkar's Indian life insurance experiments (2017, accessible author working-paper version) use trained prospective customers and examine recommendations from agents. Their research identifies problematic recommendations within the observed product and institutional setting, documents the importance of remuneration and reports substitution toward products with undisclosed commissions after a product-specific disclosure requirement. The experiments evaluate advice and responses to specific variations; they do not establish the effect of replacing all agents with direct digital distribution. Poor advice may reflect incentives, limited knowledge or both, so reform should be evaluated at the mechanism it changes.

Bank relationships and embedded coverage

Bancassurance leverages existing customer relationships, transaction records and branches. These capabilities can lower contact and enrollment costs. They may simultaneously bind the insurance decision to a loan, deposit or other primary transaction. Product choice, timing, consent, lender requirements and perception of obligation must be separately observed. A comparison between people purchasing credit protection from a bank and customers buying independent term insurance typically starts with dissimilar risk exposures and product obligations. Comparable causal questions include changing access to independent quotes at loan origination or varying whether a separate, genuinely optional offer is displayed after the loan decision.

EIOPA's credit protection insurance review (2022) documents concerns involving remuneration, conflicts and sales practices within EU bank-distributed credit protection products. Its executive discussion emphasizes variation across products, countries and business models. The review identifies important conduct risks and supports the choice of outcome measures; it does not establish that all bank distribution is harmful or identify an average causal effect relative to independent brokerage.

Embedded distribution connects coverage with a noninsurance transaction, for example vehicle rental or travel. The point of enrollment can make the insured risk salient and simplify purchasing. It can also restrict shopping time, obscure total cost or generate duplicate protection. The UK's historical general insurance add-ons study (FCA, 2014) combines market information, consumer research and a simulated experiment. Its findings concern poor value and unsuitable purchases in observed add-on markets. A simulated choice experiment can estimate reactions to a presentation within the experiment; establishing effects on real contracts and claims requires field evidence. A checkout intervention should track the whole eligible transaction population, including purchasers who decline coverage and subsequently purchase independent protection.

Digital marketplaces and multichannel journeys

A marketplace permits a consumer to compare a set of offers and may lead directly or indirectly to a completed contract. A search engine, a regulated comparison distributor and an insurer-owned portal perform distinct functions. Product coverage, ranking order, sponsored placement, data sharing, commission, acceptance probability and servicing responsibilities determine what the consumer actually experiences. The geographic extension of a marketplace does not automatically extend underwriting appetite or reduce exclusions.

The UK FCA's price comparison websites review (2014, subsequently marked historical) concerned private motor, home and travel insurance and investigated the presentation of key characteristics, costs and exclusions. This provides a historical regulatory rationale for measuring quality of comparison along with speed. A randomized ranking intervention identifies a presentation effect; an intervention changing the institutional marketplace's product panel estimates an availability effect. Treating both as merely digital distribution conflates the mechanisms.

Employer, association and affinity distribution instead aggregates a group of eligible people. Employers may sponsor coverage, select providers, contribute to premiums or facilitate payroll payments. In comparing an employer offer with direct purchase, the product and subsidy may be inseparable institutional components. A causal intervention can hold contributions fixed to test assistance, or vary employer participation to estimate the total package. Employment transitions, dependent eligibility, portability, waiting periods and continuation of cover must be recorded. Where mandatory social insurance applies, assignment cannot ethically or legally withhold a statutory entitlement.

Supervision and functional channel classification

International and regional supervisory documents establish the institutional meaning and constraints of treatments. The IAIS ICP 18 on intermediaries describes professional and transparent intermediation and addresses insurer employees, wholesale participants and mere referrals differently. It is an international supervisory framework; domestic implementation determines operative legal obligations. In the EU, Directive (EU) 2016/97 on insurance distribution, in the January 2024 consolidated text, includes certain comparison activities enabling contract conclusion within its functional definition. Article 20 identifies demands-and-needs requirements and personalized recommendations where advice is provided. Member-state transposition, exemptions and later amendments matter to specific studies.

These distinctions imply that comparative designs should map actors by function rather than by commercial labels. Record the principal to whom the intermediary owes obligations, its authorization, its access to carriers, the stage at which it intervenes and who bears ongoing service obligations. A comparable treatment taxonomy may therefore identify assisted search, delegated choice, bundled offers, enrollment administration and post-sale advocacy across otherwise dissimilar markets. The scientific function of these categories is equivalence of treatment definition, not the creation of a new legal classification.

Selection into Channels and Confounding

Demand selection and underwriting selection

Channel portfolios are produced through selection by customers, insurers and intermediaries. Customers differ in income, liquidity, insurance literacy, access to technology, language, location, urgency, prior losses, risk tolerance, trust and product complexity. Insurers choose marketing populations, distribution territories, underwriting acceptance rules and product panels. Intermediaries select leads, specialize in risks and allocate limited staff time. These factors influence outcomes and the distribution channel observed in a dataset.

A directed graph can represent a minimal example: baseline risk R and financial resources W both affect channel choice D; both also affect eventual coverage Y. D affects advice M, which affects Y. Underwriting acceptance Q is caused by D (through accessible products and submission quality) and R. Conditioning on Q can thereby open an association between D and R even if D was initially randomized. A causal design must distinguish selecting a baseline eligible population from selecting accepted risks after treatment.

Observed claims severity among policies issued through brokers may exceed direct-channel severity because brokers reach complex risks. That finding is consistent with better access to protection among high-risk firms, a difference in underwriting appetite, poorer underwriting information or several simultaneous effects. A composition difference in a portfolio is important for capital and pricing, but cannot identify a treatment effect on loss occurrence without a relevant design. Conversely, controlling for risk categories recorded only following intermediary advice may introduce a posttreatment problem if advice changes the accuracy of reported exposures.

An illustrative reversal in conversion rates

Consider an invented eligible population consisting of simple and complex risks. The direct channel has 800 simple inquiries with 400 purchases and 200 complex inquiries with 40 purchases. The assisted channel has 200 simple inquiries with 120 purchases and 800 complex inquiries with 240 purchases. Overall direct conversion is 440/1,000 = 44%, while assisted conversion is 360/1,000 = 36%. Yet assistance has a higher observed conversion rate within both simple risks (60% versus 50%) and complex risks (30% versus 20%). When both channels are standardized to a population with equal shares of simple and complex risks, their rates become 35% and 45% respectively.

The example illustrates Simpson's reversal and the role of a target distribution. The standardized difference of ten percentage points becomes a causal estimate only if risk complexity and whatever further baseline variables are necessary suffice for exchangeability, there is overlap and the treatment is consistently measured. In actual data complex customers may differ in urgency, eligibility across carriers and unobserved exposures. Standardization reveals the consequences of composition; it does not itself resolve unmeasured confounding.

Adjustment and the limits of predictive models

For a binary channel comparison let e(X) be the conditional probability of receiving a. Under conditional exchangeability and positivity, one may standardize predicted outcomes over a prespecified population, match on baseline characteristics, or estimate inverse probability weights. An augmented weighting estimator combines an outcome regression and a model of assignment probability. Its usual double-robust property concerns correct specification of at least one of two nuisance functions under the underlying identification conditions. It provides no protection against an unmeasured common cause or inconsistent treatment definitions.

Overlap is a substantive insurance matter. A commercial property risk rejected by every direct writer and eligible only through a specialist broker cannot contribute to a like-for-like direct-versus-broker comparison of successful placement. A direct route with no available product may nevertheless be part of an actionable strategy contrast for all inquiries, where placement failure is a legitimate outcome. In either case, extrapolating a fitted treatment effect across unsupported covariate regions obscures what is identified. Report propensity distributions, extreme weights, effective sample sizes, the restricted target population and the number excluded by overlap rules.

Rich predictive data may make measured confounding easier to address, but prediction and causal estimation have different objectives. A model predicting purchase well can rely on variables available after advice or purchase, making it unsuitable for baseline causal adjustment. A model trained on prior underwriting decisions can encode the channel's historical selection. Time-index every feature, preserve its provenance and specify whether it is a preexisting characteristic, a proxy for an unobserved cause, a downstream event or an outcome-derived field.

Attribution and missingness

Digital marketing attribution systems commonly divide credit among interactions or assign it to the first or last recorded contact. That convention is useful for reconciling expense and remuneration. A consumer might click a search advertisement after receiving a broker recommendation; last-touch attribution would credit the final medium even if the recommendation induced purchase. Without an intervention or a defensible observational design, fractional attribution does not establish how many contracts would disappear in the absence of a touchpoint.

The observation of channel histories is itself selective. A broker may retain detailed records for clients with completed contracts while unsuccessful requests remain in separate systems. A marketplace may lose records of people who complete a transaction at a carrier site. Insurers may observe their own lapses but not new cover elsewhere. Missing outcomes can be jointly related to treatment and prognosis. Define the data-generating process, report follow-up by assignment, use linkage when lawful, and provide sensitivity bounds where missingness assumptions remain uncertain.

Confounding analysis should distinguish variables that affect both channel and outcome from variables that merely improve prediction, and should account for data collected in different institutions at different times. This is a prerequisite for deciding whether a causal interpretation can be supported or the estimate must remain descriptive.

Randomization, Instruments and Natural Experiments

Field experiments and channel encouragement

Random assignment is especially informative when the intervention changes an option rather than mandates a purchase. A trial may offer eligible people an appointment with an adviser, additional access to a marketplace or assisted enrollment. Every arm should retain required disclosures, existing entitlements and ordinary complaint channels. The design should preserve customer autonomy, offer legally permissible products and have safeguards for vulnerable customers. Assignment probabilities should be known and recorded. Nonparticipation and switching are part of the assignment effect and should remain in the assigned denominators.

Where treatment is delivered by a branch, agent or employer, individual randomization may contaminate the comparison because staff adopt the intervention in other interactions. Cluster assignment can reduce this risk. With equal cluster size m and intracluster correlation rho, the conventional approximate design effect is 1 + (m − 1)rho. For 40 observations per branch and rho = 0.05, that factor is 2.95; an individually calculated effective sample requirement would be multiplied by approximately 2.95 before allowing for other design features. This illustrative calculation emphasizes the importance of the number of independently randomized branches, variable cluster size, stratification and appropriate cluster-level inference. Many customers served by three branches cannot substitute for many independently randomized branches.

Capacity affects both treatment compliance and spillovers. A trial where assistance is scarce may allocate opportunities randomly, identifying the consequences under current capacity. Scaling the treatment may lead to longer calls, shorter consultations or changing recruitment. If saturation is the scientific object, variation in the proportion of eligible people offered service across clusters can help separate own assignment and local congestion or peer effects.

Instrumental variables

Let Z be a binary randomized invitation, D indicate actual receipt of the specified service and Y indicate a prespecified outcome. The reduced form is E[Y|Z=1] − E[Y|Z=0]; the first stage is E[D|Z=1] − E[D|Z=0]. Their ratio is a Wald instrumental-variable estimand. Angrist, Imbens and Rubin's instrumental variable identification (1996) gives the potential-outcome interpretation and conditions for a complier effect. In addition to relevance and well-defined treatment, the standard binary local effect interpretation requires independence of assignment, an exclusion restriction, monotonicity and an appropriate absence of interference.

Exclusion deserves special attention when the encouragement is itself informative. A mailed invitation may remind a consumer of a purchase deadline even if the call does not occur. If receipt D is defined as completed conversation, voicemail or unsuccessful contact may nevertheless affect purchase, violating exclusion. Changing the definition to receiving a placed call, including voicemail, changes the treatment under study. Distance to a broker office can be associated with income, internet access, population density and exposure; therefore it is not a generally valid instrument for broker choice. A bank assignment can affect loans and liquidity directly. Distributor compensation policies may affect both channel choice and the product terms offered by all channels. Instruments require institutional as well as statistical arguments.

For an invented randomized offer, suppose 70% of the treatment arm receive advice compared with 20% of controls, and coverage increases from 30% to 35%. The assignment effect is five percentage points. The first stage is 0.50, and the IV ratio is 0.05/0.50 = 0.10, or ten percentage points for compliers under the assumptions. That estimate does not imply all customers would benefit by ten points. Always users and never users are not separately identified in the same way. If the invitation changes coverage by one point through a pathway outside actual advice, a simple additive direct-effect sensitivity exercise gives (0.05 − 0.01)/0.50 = 0.08. This is a scenario, not a generally identified correction.

Report first-stage magnitude and uncertainty, assignment effects, the precise receipt definition and weak-instrument diagnostics appropriate to the estimator. Confidence intervals for a ratio are not obtained by dividing the endpoints of two unrelated intervals. An instrument with a negligible first stage can produce a numerically large yet practically uninformative estimate.

Evidence from real distribution-related experiments

Myerson and colleagues' personalized telephone outreach trial (2022) uses randomized assignment to outbound calling among 79,522 California Marketplace applicants who had not yet chosen a plan and had not delegated their cases to an agent or navigator. The reported control enrollment was 12%, with an assignment effect of 1.3 percentage points and an IV effect of receiving an outbound call of 2.7 points. These are different estimands; the latter is attached to consumers whose call receipt was induced by assignment. The study reports calls placed to 49% of assigned households. The approximate division 1.3/0.49 = 2.65 points is consistent with the rounded magnitudes, although it does not reproduce the adjusted regression. The measured outcome is enrollment in Covered California, not all-market coverage or realized claims protection.

Ravel and colleagues (2022) conducted a four-arm Medicaid-to-Marketplace transition trial involving 2,173 households eligible for subsidized coverage after leaving Medicaid in 2017. The control enrollment rate was 14.5%. Assignment to email reminders alone changed enrollment by −0.5 percentage points (p = .82); phone-only assignment by 4.4 points (p = .08); phone plus email assignment by 6.9 points (p = .008). The reported difference between the two phone-containing arms was statistically inconclusive. Consequently the evidence supports the combined intervention against control in that population without establishing that email adds a benefit to telephone outreach. Neither California study is randomized selection of every institutional distribution channel. They identify defined assistance policies in regulated enrollment settings.

In a different institutional environment, Thornton and colleagues' Nicaraguan enrollment experiment (2010) randomized premium costs and enrollment locations in voluntary health insurance for informal-sector workers. The accessible abstract reports approximately 20% initial participation, a slight adverse enrollment effect associated with microfinance institution involvement and less than 10% of enrollees remaining after one year following subsidy expiry. The observations identify the importance of actual procedures and duration. A preexisting institutional network may be useful for accessing clients, but the effect depends on its enrollment process and surrounding conditions. The low retention result applies to the program described; a separate causal claim about the reason for every lapse would require additional evidence.

A comparison closer to credit-linked insurance distribution was randomized in Colombia. Bauchet, Damon, and Larsen's microfinance bundling experiment assigned voluntary crop microinsurance offers either with a loan application or at a delayed time. The publisher reports no significant overall effect on take-up or insured amounts, heterogeneous crop-level responses and low consumer understanding. The assigned treatment is timing of an offer inside a microfinance relationship; it is not random assignment to a bank rather than an independent broker, and it reports no identified claims-service effect.

Employer-based insurance offers supply a separate measurement example. Atanasov and Baker's employer plan-choice research combines a post-enrollment survey with a framed field experiment in a high-deductible plan setting. The publisher reports 3% observed take-up of the new plan in the survey, failure of added financial information alone to increase plan selection in the framed experiment, and sensitivity to physician-access perceptions and choice-menu description. These results concern observable and experimental components of an employer distribution experience; they do not identify a causal comparison of actual employee enrollment against insurer-direct or brokerage delivery. A future channel trial must randomize institutional assignment or another clearly defined mechanism and observe realized, continuing coverage to answer that broader question.

Cole and colleagues' rainfall insurance field experiments (2009 working-paper version; published 2013) report evidence from rural India that purchase decisions respond to price, liquidity and trust. Gaurav, Cole and Tobacman's randomized insurance education study (2011), involving 600 small-scale farmers, reports an increase in adoption from 8% to 16% following a targeted education intervention. These results address specified products and interventions rather than average effects of a general channel. Taken together, the studies support the importance of the components of a distribution service. A targeted education program, a trusted introduction, a reminder and a premium subsidy are different interventions even when delivered by the same agent. Their effects cannot be aggregated as if they were interchangeable treatments.

Comparative evidence from institutional distribution mechanisms

The institutional role of a broker is narrower than the economic role of a competitive brokerage market. Karaca-Mandic, Feldman, and Graven (2013) combined broker membership information with an employer survey in their study of broker-market competition. The reported associations link greater broker competition to a higher probability of small firms offering health insurance, lower premiums, and less dispersed premiums. This evidence concerns employer access and market structure; the abstract alone cannot establish a causal effect of individual broker assignment. A competitive market and a particular intermediary are distinct treatment units. A firm-level design must measure local competition before any intervention, and a market-level intervention requires attention to entry, exit, and spillovers.

A different institutional comparison is provided by Staiger et al. (2022), whose Medicaid enrollment broker evaluation contrasted changes after state contracts with brokers against contemporaneous changes in states that never contracted with them. The study reported no statistically significant increase in participation or evidence of improved application processing times. The measured intervention is a procurement and organizational arrangement for public coverage, not a randomized encounter between an applicant and a private insurance agent. Its result qualifies the inference from individual outreach experiments: personal outreach can remove a specific friction even when an institutional contract designed to improve enrollment produces no detectable aggregate improvement. Statistical non-significance is not evidence that all effects are exactly zero. Contract start dates, concurrent program reforms, comparator trends and implementation fidelity determine the strength of that inference.

Credit-linked and embedded insurance

Larsen's (2016) Colombian microfinance bundling experiment randomized whether voluntary crop insurance was offered during the loan application or several weeks later. The repository abstract reports no statistically significant differences in uptake, coverage amounts or understanding. This thesis evaluates sequencing within the credit-linked channel while preserving the insurance product; it does not identify the effect of replacing a bank or microfinance intermediary with a direct insurer. A delayed offer may also change opportunities for contact, liquidity and recollection, so the randomized assignment identifies the timing policy as implemented, rather than a pure psychological framing effect. Commercial bancassurance requires independent evidence on its own contract structure and population.

An Australian government experiment conducted by BETA with ASIC evaluated add-on insurance information statements in a randomized simulated shopping environment (2021). The final sample contained 6,243 participants, and the project summary reports a reduction in hypothetical purchases of nine percentage points, or 24% relative to control. The research changes information at a point of sale, not the entire distribution architecture. A decrease in purchases may indicate informed refusal of low-value coverage, but the experiment did not establish actual reductions in unsuitable sales, durable protection or claims outcomes. Consequently a distribution evaluation must preserve both the behavioral endpoint and the economic objective: the desirable direction of an enrollment effect depends on suitability, price, protection and the relevant counterfactual.

Together these institutional studies extend the evidence beyond individually delivered assistance. They distinguish competitive market structure, public contracting, voluntary credit-linked sequencing and embedded point-of-sale information. Their heterogeneous units, outcomes and identification conditions preclude a pooled channel coefficient. They support a research strategy in which the intervention specifies the actual organizational change and separates measurable acquisition outcomes from subsequent protection, consumer welfare and distributional consequences.

Natural experiments and longitudinal comparison

A regulator's staged introduction of distributor requirements, an insurer's sequential opening of sales offices or a marketplace's staged roll-out of live assistance can supply comparison variation when treatment timing is plausibly unrelated to prospective outcomes conditional on the design. In a two-period comparison, the difference-in-differences estimator subtracts the change in an untreated comparison group from the change in the treated group. Parallel trends in untreated potential outcomes, absence of anticipation and appropriate treatment timing are critical. A common geographic shock, change in employer contributions, competitor withdrawal or contemporaneous premium reform can undermine that interpretation.

With staggered adoption, earlier-treated groups may be inappropriate controls for later-treated groups when effects evolve. Callaway and Sant'Anna's multiple-period difference-in-differences (2021) provides a group-time treatment effect approach accommodating varying adoption dates and conditional parallel trends. For insurance research, define adoption at the point a distribution strategy becomes available to eligible units, distinguish it from announcement, and compare each cohort with suitable untreated observations. Estimate effects by event time and show the composition and weights of aggregated estimates. Differential anticipatory marketing can affect preadoption outcomes; graphical checks should accompany institutional examination.

For an illustrative branch reform, treated branches improve coverage from 40% to 55%, while comparison branches improve from 35% to 45%. The raw before-after increase is 15 points, whereas the difference-in-differences contrast is five. Under parallel untreated trends, this estimates the average effect for the treated comparison at that horizon. If the treated region simultaneously introduced a subsidy that itself raises coverage by five points beyond the change in control areas, the contrast could entirely reflect that subsidy. More observations alone cannot resolve the institutional ambiguity.

Regression discontinuity can be informative when channel eligibility changes at a known threshold, such as a prespecified contract-size criterion or enrollment priority score. Continuity of potential outcomes at the cutoff, lack of precise sorting and stable treatment definitions matter. A threshold that simultaneously changes contribution levels or underwriting terms identifies the combined discontinuity. A fuzzy discontinuity uses the discontinuous change in receipt and inherits local IV assumptions. The result pertains to eligible units close to the threshold; scaling to the overall population is a separate operation.

Coverage, Cost, Persistency and Consumer Outcome Denominators

Access, acquisition and exposure

Distribution effectiveness concerns both access and continuing protection. Access is the ability to encounter a suitable offer given a person's or firm's risk and eligibility. It can be measured as the proportion of an independently defined eligible population receiving a feasible quotation. Quotation receipt and successful underwriting are not equivalent; an indicative quote may be withdrawn or altered when complete data become available. Record offered, accepted, bound and effective contracts as different events.

Enrollment should use a well-defined population denominator, preferably the original assigned eligible participants. Policy counts give weight to people with multiple contracts and may increase when one contract is replaced by two smaller policies. Persons covered, households protected, property exposures and sum insured measure different features of the population. Premium volume is influenced by prices, policy limits, terms and savings components. An increase in gross written premium may occur with unchanged or declining numbers of insured people. It cannot be converted to insurance coverage without information about contract and risk exposures.

A suitable longitudinal measure is restricted covered time during the observation period, measured per eligible household, worker or property. For each eligible unit, count months or days with effective insurance meeting the predefined protection threshold. An automatic renewal may continue nominal coverage while a new exclusion or inadequate sum insured reduces meaningful protection. Measure continuity, coverage level and adequacy separately. For group policies, the relevant insured-person denominator requires dependent rosters and effective eligibility, not simply a single master contract count.

Costs and the distribution of economic value

Acquisition costs include channel remuneration, staff and technology resources, marketing expenditure and costs of underwriting or enrollment attributable to acquisition. Marginal and fully allocated expense differ. A digital implementation can require large fixed investment and low marginal expense; a broker may have a high commission and reduce insurer underwriting effort. Record price year, currency, allocation convention, horizon and perspective. A commission is a transfer between parties from a social welfare perspective to the extent it merely reallocates money; the resources used to perform intermediation and any behavioral distortions remain economically consequential.

For an experimental deployment, incremental cost per additional protected unit equals the difference in mean costs per assigned eligible unit divided by the difference in probability of the specified protection outcome, where the denominator is positive. A near-zero incremental effect creates unstable ratios; report numerator and denominator independently and quantify joint uncertainty. Cost per acquired contract measures an operational quantity and should not be substituted for a causal incremental cost-effectiveness ratio. Compare net cost, customer value and financing sustainability from clarified perspectives.

In an invented study allocating 1,000 eligible households to each arm, self-service costs 20,000 currency units and produces 300 policies; assistance costs 35,000 and produces 400. Simple acquisition costs per policy are 66.67 and 87.50. The incremental cost per additional initial policy is 15,000/100 = 150. Suppose 225 self-service customers and 240 assisted customers have effective continuous cover at month 12. Conditional retention is 75% versus 60%, whereas durable protection per assigned eligible household is 22.5% versus 24%. The incremental cost per additional continuously covered household is 15,000/15 = 1,000, assuming the stated expenses capture the relevant differential costs for the horizon. The assisted arm still reaches more people, despite weaker conditional retention; equality of quality and protection scope would require separate evidence.

This example exposes two denominator changes. The marginal value of the intervention is defined over assignment, while the operational retention ratio is defined over purchasers. Changing the horizon also changes the outcome purchased with additional distribution resources. A reasonable operational decision may prefer a less expensive short-term sales route or a costlier long-term protection strategy depending on the purpose; the comparison must make the tradeoff explicit.

Persistency, lapse and competing risks

Persistency measures continuation of a contract or coverage status under specified rules. Lapse due to nonpayment, cancellation by the insured, rescission, expiration, replacement, death and employer termination have distinct meanings. Replacement can represent better coverage rather than failure, and a life policy ending after an insured death is not equivalent to abandonment. Define whether the estimand is original-contract persistency, insurer retention or coverage at any provider. Administrative coverage periods may require a grace-period rule and a consistent approach to reinstatement.

Comparing time to lapse only among purchasers conditions on a postassignment event. For deployment evaluation, measure coverage and lawful continuation for everyone originally eligible. Survival analysis can characterize durations among observed policies while acknowledging the selected cohort. Competing risks should remain separate where substantively important; treating death or loss of eligibility as ordinary independent censoring can distort the probability of lapse. If an intervention affects employment, it may affect both group eligibility and coverage duration; an analysis aimed at the full institutional effect should retain that consequence.

Advice, complaints and claims outcomes

Advice quality can be evaluated through standardized scenarios, chart review, customer comprehension tests and realized outcomes. Identify the coverage need before recommendation, define permissible product comparisons and record fees, exclusions, limits and expected costs. A mystery shopper may elicit advice under controlled conditions without binding insurance. It can test response to standardized needs or information but cannot observe long-term financial protection for an actual policyholder. The Indian agent evidence in Anagol and colleagues' advice experiments illustrates why observed sales recommendations deserve distinct analysis from consumer welfare.

Complaint incidence combines adverse experiences and willingness or ability to submit complaints. More accessible complaint support can increase reported complaints despite improved underlying service. Pair complaint frequency with harmed customers, independently assessed severity, response times and resolution. Claims frequency is determined by covered exposure, loss incidence, reporting, contract scope and potentially preventive behavior. Claim settlement measures should distinguish reported claims, accepted claims, payments, denials and time to decision. A payment-to-premium ratio is informative about product economics but can be low in a quiet year for sound catastrophe cover or high in a year with severe events. It is not a complete consumer value measure.

Consider eligible customers assigned to different distribution strategies. An outcome such as timely indemnification of a verified covered loss can be defined at the population level, with zero payment for people who remain uninsured; this combines risk occurrence, enrollment and payment. Claims acceptance among claimants answers a narrower service question and is subject to selection by both loss and insured status. If the research concerns service after a claim, a separate design that randomizes claimant assistance at claim notification may better isolate that effect. It cannot independently establish the acquisition effect of the original channel.

Profitability and consumer welfare

Insurer profit is a distinct outcome from consumer surplus or social welfare. A distribution arrangement may attract low-risk customers and improve the insurer's portfolio but make coverage less available to higher-risk people. Profit calculations should use earned premium, expected or developed claims costs, operating expenses, reinsurance where relevant and capital consequences consistently through time. A short horizon for long-tailed liabilities may misrepresent ultimate loss costs.

Consumer welfare under uncertainty depends on the distribution of possible losses and their marginal utility consequences, not solely on realized claims. Insurance transfers resources across states of the world. Appropriate protection can have economic value even when no claim is made. In evaluating distribution, use price and coverage information, adequacy of protection, customer preferences and realized service evidence. No unique welfare weight for financially vulnerable households, consumers with large exposures or insurer profitability is implied by the observed data. Report alternatives and assumptions where a social valuation is constructed.

The methodological conclusion is that policy sales, risk exposure, service, financial value and welfare should be measured in their own units. A credible joint evaluation shows whether an apparent gain survives the movement from acquisition to ongoing effective protection, and who bears its costs.

Heterogeneous Effects and Cross-Country Comparison

Effect modification and access

An institutional channel may benefit consumers with limited internet access more than digitally fluent consumers, people with complex risks more than those with standardized needs, or applicants unfamiliar with local product requirements more than experienced buyers. These are plausible effect-modification hypotheses; observed differences in subgroup outcomes do not establish them. Compare causal contrasts within each subgroup using preassignment membership. An effect significant in one subgroup and nonsignificant in another is not evidence by itself of a statistically significant interaction. Power, baseline event probabilities, measurement and multiple-testing adjustments matter.

Relative and absolute effects may differ in their rankings. An increase from 2% to 4% doubles enrollment but adds two covered people per 100 eligible units. An increase from 40% to 46% raises relative enrollment by 15% but adds six per 100. For prioritization under constrained capacity, absolute incremental coverage is often relevant; relative gains describe proportional change, and equity concerns may justify attention to small groups with initially low coverage. Report both and preserve denominators.

The California outbound telephone study includes subgroup analyses by income, age, language and other characteristics, with important sample-size limitations for some groups. It supplies an example of heterogeneity in an enrolled-applicant pathway; it cannot establish that the same pattern holds in all product markets or national systems. In a new study, a prespecified language-support comparison could distinguish an interpreter or multilingual adviser effect from personal telephone contact itself.

Transporting findings between institutions

External validity starts with mechanisms, not geography. In a voluntary rainfall index insurance program, season and crop exposure define timing and payout structure; trust and basis risk can shape willingness to purchase. In US health coverage, subsidies, enrollment windows, regulated plan benefits and alternate public programs shape the choice. In commercial property insurance, a broker may help assemble risk information and access capacity. A standardized causal design can compare analogous service mechanisms across settings without treating the product, risk or legal environment as interchangeable.

To transport an estimated effect, identify the baseline variables modifying it in the source and target populations. A standardization over the target population can be meaningful when conditional effects are applicable and the target population has support in source data. If a source trial excluded employer-covered people, one cannot transport its average result to employer participants without an additional substantive assumption and overlap. Differences in prices, underwriting, licensing and compulsory coverage rules may change the intervention itself rather than simply modifying its effect. Transportability should be distinguished from statistical extrapolation.

Cross-country research should maintain a jurisdiction-year data dictionary specifying channel definitions, whether direct and ancillary activity is regulated, who can offer advice, what insurance products are included, who qualifies for coverage and how premiums and claims are reported. EIOPA's third IDD application report (2026) publication description identifies 2024–2025 market developments and draws on national-authority surveys and stakeholder input. Such information is valuable for mapping institutional differences, not an experiment identifying channel effects. The EU context cannot be assumed to encompass non-European markets. Likewise the IAIS comparative framework supplies supervisory concepts, not a uniform law or a source of individual channel treatment assignments.

Aggregate market statistics and the ecological problem

The OECD's Global Insurance Market Trends 2026 reports total gross premiums relative to GDP, including averages of 6.6% for OECD countries and 4.2% for reporting non-OECD jurisdictions in 2025. These are market penetration statistics, not the share of people with effective coverage. The denominators and reporting populations matter; changes can arise from premium rates, life-sector savings products, economic activity and risk prices. A country-level relationship between bank distribution shares and premium penetration cannot identify an individual-level effect of bank contact on protection.

The World Bank's Global Findex 2025 report draws on nationally representative surveys of about 148,000 adults across 141 economies in 2024 and includes connectivity and financial-access measures. This can help characterize the opportunity to reach digital audiences. It does not establish that every connected person can obtain an appropriate insurance policy, or reveal the causal effect of a particular distributor. A weak national digital infrastructure measure can coexist with effective community distribution; national income and institutional maturity may confound aggregate comparisons.

Country-level variation can contribute identification under a suitably specified policy design, but simple cross-country regressions often reflect simultaneous development processes. Hierarchical models can partially pool effects across institutions, estimate between-site dispersion and reduce extreme small-sample estimates. Their causal interpretation derives from the identification within each study or from the assumptions of an integrated design. Hierarchy alone does not make selected countries exchangeable or harmonize measurement. Report country-specific effects and uncertainty alongside a pooled summary when the interventions and outcomes are sufficiently comparable. Where interventions differ in substance, present contextual comparisons rather than a misleading composite effect.

Market equilibrium, inclusion and sustainability

A distribution trial estimates outcomes at a particular scale. If the intervention attracts customers who were previously uninsured, it may enlarge the market; if it mainly diverts existing customers from another distributor, firm-level gains overstate incremental protection. Competition may induce repricing or exit; improvements in search could compress commissions or alter the contracts insurers offer. Conversely, cross-subsidizing vulnerable populations through a group arrangement may make the new route sustainable only while enrollment is sufficiently broad.

Let the overall eligible market be observed under two alternative distribution policies. A provider's incremental market share combines newly insured people, switches from competitors and possible losses to other institutions. The change in total market coverage is a different quantity. A policy intervention aimed at reducing a protection gap must track coverage across providers or transparently limit its claim to the measured provider. A treatment that merely increases the firm's attribution share cannot establish insurance inclusion.

For legal and ethical comparability, preserve valid consent and data protections. Access to assisted options should not be denied on impermissible grounds. Audit who is reached, who is excluded by design, who receives unsuitable recommendations and who loses coverage subsequently. Causal analysis of inclusion should examine risk-bearing populations and meaningful protection, not treat particular channels as presumptively inclusive based on technology or brand.

Replication, Sensitivity and Interpretation

Evidence hierarchy and triangulation

The literature reviewed in this dossier provides complementary evidence. The insurance enrollment studies in California and Nicaragua identify specified interventions through randomized allocation. Gaurav and colleagues test targeted education, while the rainfall insurance experiments reported by Cole and colleagues identify several adoption frictions. Indian life insurance mystery-shopping and disclosure evidence concerns advice and remuneration responses. The insurer efficiency study addresses product quality and expense at the firm level. EIOPA and FCA evidence describes institutional risks and market practices, while IAIS and EU texts clarify roles and applicable standards. These designs answer different questions.

Triangulation uses these different objects to test whether candidate mechanisms are plausible. Evidence that assistance overcomes an enrollment barrier supports studying assistance in another market; evidence of conflicted recommendations supports carefully defining advice quality; a study of cost efficiency supports measuring outputs as well as input costs. Their coexistence does not supply a pooled effect of switching between seven channels. The scientific value of the synthesis lies in aligning intervention, mechanism, identification and outcome.

A reproducible data specification

A prospective research dataset should include stable pseudonymous eligible-unit identifiers and distinguish household, person, policy, insured risk, intermediary, carrier and cluster identifiers. Record a baseline eligibility timestamp, source of contact, initial channel assignment, randomization strata where applicable, receipt dates and service version, quote and product identifiers, premium and subsidy terms, underwriting disposition, activation, payment and lapse events, coverage intervals, claims and complaint records. Preserve original timestamps and codebook versions; identify when a variable becomes available to the analyst.

For each outcome, maintain a definition, numerator, denominator, time origin, follow-up horizon, handling of multiple policies, data source and missingness rules. A linkage log records which records match and why unmatched records occur. A provenance table identifies changes in product terms or channel workflows. Validation should reconcile aggregate counts to source systems without assuming that reconciliation proves causal attribution. Research extracts should be access-controlled and appropriately minimized, with reproducible anonymized or synthetic data and analysis code shared where lawful.

A preanalysis plan specifies the primary assignment contrast, estimand, statistical estimator, stratification, precision, attrition treatment, subgroups, multiple-testing strategy and sensitivity checks. Post hoc hypotheses remain valuable as exploratory research when clearly distinguished. Record dates and versions of protocol amendments. Independent reproduction ideally starts from a frozen data and code version, verifies eligibility and flow counts, reproduces descriptive figures and estimates, checks alternative implementation and documents discrepancies.

Sensitivity analysis for observational designs

If observational adjustment is the primary identification strategy, evaluate covariate balance, propensity overlap, dependence on weight trimming, model specification, functional form and alternative baseline windows. A negative control outcome established before treatment can reveal residual imbalance or data leakage. A negative control influenced by the same unmeasured causes but unaffected by treatment may be useful when its substantive assumptions are defensible. Neither a null negative control nor good fit proves exchangeability.

Suppose an observational adjusted difference in coverage is 0.08, and a hypothetical unmeasured binary financial constraint is twenty percentage points less prevalent among assisted customers after measured adjustment. If that constraint reduces coverage by twenty points under an additive model and its distribution is the sole residual difference, the omitted-variable contribution is (−0.20)(−0.20) = +0.04, yielding an illustrative residual contrast of 0.04. A larger effect or imbalance could eliminate the observed contrast. The calculation is a transparent sensitivity scenario, not an estimate of the prevalence or impact of actual financial constraints. It specifies what would need to be true for the observational interpretation to change.

For incomplete follow-up, simple worst-case bounds can reveal fragility. In an invented 1,000-per-arm experiment, 900 outcomes are observed in each arm; 360 treated and 315 controls are continuously insured among observed cases. The full-denominator treated probability lies between 0.36 and 0.46 and the control probability between 0.315 and 0.415. The possible assignment difference under unrestricted missing binary outcomes ranges from −0.055 to +0.145. A complete-case difference of 0.05 is compatible with a negative full-cohort effect. Investigate outcome missingness and narrow bounds only with stated assumptions, justified auxiliary data or follow-up.

For staggered interventions, inspect event-time coefficients before implementation, alternative comparison cohorts, anticipation windows and sensitivity to plausible differential trends. A statistically inconclusive pretrend test does not prove parallel trends. A short preperiod may be especially misleading in insurance markets with annual renewals or seasonal risk exposure. Check product reform, underwriting appetite, economic shocks, distribution capacity and concurrent advertising. Use timing that reflects actual exposure to the intervention, not a contract signature after the relevant announcement.

Statistical and substantive uncertainty

Statistical confidence intervals describe sampling or randomization uncertainty under assumptions. Identification uncertainty concerns whether those assumptions plausibly hold. Implementation uncertainty concerns whether the assigned strategy can be delivered with the same staff and product choice at scale. Measurement uncertainty concerns coverage quality, unobserved alternative insurance and incomplete claims development. A narrow interval on an outcome recorded incorrectly is not a sound conclusion about protection.

A substantive report should present means and sample sizes by arm, absolute and relative contrasts, intervals, receipt rates, missingness, estimated subgroup contrasts, deviations and the assumptions used for each estimator. Where only a study abstract is available, report only its supported claims and do not replicate tables imaginatively. Where regulatory findings are descriptive, retain that status. A systematic review or meta-analysis could be a subsequent project if a predeclared eligibility protocol finds comparable intervention constructs and outcomes. Pooling presently dissimilar studies would conceal crucial differences in populations, products and assignment.

Interpretation for insurers and supervisors

Evidence can inform decisions at three levels. At an insurer, randomized offers support allocation of limited assistance resources and estimates of marginal costs, conditional on safeguarding customers. At a market or supervisor, linked datasets and staged policy variation may identify effects on competition, suitability and continuing protection across providers. At a social-policy level, effective coverage of previously uninsured groups and distributional welfare require population-based measures and institutional understanding beyond individual insurer sales metrics. The evidence threshold rises with the scope of the claim.

A decision to retain a high-cost intermediary might be supported by improved placement quality or durable coverage. A decision to simplify a self-service path might be supported by similar or better consumer outcomes at lower resource cost. A decision to change loan-related distribution might prioritize informed consent and meaningful alternatives even where initial uptake falls. These are conditional interpretations grounded in the intended outcome, not recommendations for particular products or individualized insurance decisions.

The appropriate reporting endpoint is an explicit evidentiary statement: the intervention, comparison, eligible population, time, outcome, estimator, principal identification assumptions and contribution to knowledge. That statement allows a reader to decide whether the result informs a proposed channel change. A detached channel coefficient cannot carry the same scientific meaning.

Conclusion

The causal evaluation of insurance distribution is a scientific inquiry into institutional changes in the availability, selection, maintenance and realization of insurance protection. It begins with an actionable intervention and follows the eligible population through the sequence from contact to effective coverage and service. Its central methodological task is separating changes induced by distribution from differences in the people, risks, products and markets observed within each distribution channel.

Randomized assignment identifies the effect of a defined offer or pathway when implemented and measured appropriately. Randomized encouragement can additionally support an instrumental-variable effect for a particular subgroup subject to further assumptions. Matching, weighting, natural experiments and longitudinal designs support causal inference when their identification conditions are credible in the institutional setting. The estimand must remain connected to treatment versions, time, eligibility and observed outcomes. The methods are complementary and answer distinct questions.

The empirical literature supports several important interpretations. Specific enrollment interventions can overcome procedural and informational barriers; education and trust can influence demand; remuneration and product-specific disclosure can affect advice; access through an existing institution does not automatically ensure durable uptake. Differences between acquisition expenses and outputs require quality-sensitive interpretation. The studies reviewed establish intervention-specific evidence, not a universal ordering of channels. Evidence from randomized loan-linked offer timing and employer framed-choice experiments further clarifies that these interventions have different endpoints and intervention units; they cannot be pooled into a single effect of bank or employer distribution. Employer, bank, broker, digital and embedded pathways merit further carefully designed field comparisons covering continuing protection, consent, service and market-level effects.

A coherent evaluation distinguishes operational performance from insured value, and neither can be inferred reliably from sales totals alone. The use of eligible-population denominators, restricted covered time, defined quality and contract outcomes, explicit expense perspectives and transparent sensitivity analysis provides a basis for assessing these consequences without conflating portfolio composition with causation. Where evidence is unavoidably selected or incomplete, the corresponding limits belong in the result.

The original contribution of this monograph is the integration of these methodological distinctions with the institutional economics of insurance distribution and the measurement of durable protection, supported by worked hypothetical comparisons and critical assessment of actual evidence. No new theory or original empirical discovery is claimed. The scientific conclusion is that channel effectiveness is the causal consequence of a specified institutional strategy for a specified population and outcome, under identified implementation and market conditions. This formulation makes distribution a legitimate object of empirical insurance science rather than a label attached retrospectively to sales.

References

  • Hernán, Miguel A., and James M. Robins. (2020). Causal Inference: What If. Boca Raton: Chapman & Hall/CRC. Verified source

  • Angrist, Joshua D., Guido W. Imbens, and Donald B. Rubin. (1996). Identification of causal effects using instrumental variables. Journal of the American Statistical Association, 91(434), 444–455. DOI: 10.1080/01621459.1996.10476902. Verified source

  • Callaway, Brantly, and Pedro H. C. Sant'Anna. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200–230. DOI: 10.1016/j.jeconom.2020.12.001. Author preprint, arXiv:1803.09015 (2018, revised 2020). Verified source

  • Berger, Allen N., J. David Cummins, and Mary A. Weiss. (1997). The coexistence of multiple distribution systems for financial services: The case of property-liability insurance. Journal of Business, 70(4), 515–546. Verified source

  • Cole, Shawn, Xavier Giné, Jeremy Tobacman, Petia Topalova, Robert Townsend, and James Vickery. (2009). Barriers to Household Risk Management: Evidence from India. Federal Reserve Bank of New York Staff Report 373. Published version: American Economic Journal: Applied Economics, 5(1), 104–135 (2013), DOI: 10.1257/app.5.1.104. Verified source

  • Gaurav, Sarthak, Shawn Cole, and Jeremy Tobacman. (2011). Marketing complex financial products in emerging markets: Evidence from rainfall insurance in India. Journal of Marketing Research, 48(SPL), S150–S162. DOI: 10.1509/jmkr.48.SPL.S150. Verified source

  • Anagol, Santosh, Shawn Cole, and Shayak Sarkar. (2017). Understanding the advice of commissions-motivated agents: Evidence from the Indian life insurance market. Review of Economics and Statistics, 99(1), 1–15. DOI: 10.1162/REST_a_00625. Author working-paper version, Harvard Business School Working Paper 12-055, revised 2015. Verified source

  • Bauchet, Jonathan, Amy Damon, and Vance Larsen. (2017). Microfinance bundling and consumer protection: Experimental evidence from Colombia. Journal of Development Effectiveness, 9(4), 443–461. DOI: 10.1080/19439342.2017.1363802. Verified publisher record

  • Atanasov, Pavel, and Tom Baker. (2014). Putting health back into health insurance choice. Medical Care Research and Review, 71(4), 337–355. DOI: 10.1177/1077558714533821. Verified publisher record

  • Thornton, Rebecca L., Laurel E. Hatt, Erica M. Field, Mursaleena Islam, Freddy Solís Diaz, and Martha Azucena González. (2010). Social security health insurance for the informal sector in Nicaragua: A randomized evaluation. Health Economics, 19(S1), 181–206. DOI: 10.1002/hec.1635. Verified source

  • Myerson, Rebecca, Nicholas Tilipman, Andrew Feher, Honglin Li, Wesley Yin, and Isaac Menashe. (2022). Personalized telephone outreach increased health insurance take-up for hard-to-reach populations, but challenges remain. Health Affairs, 41(1), 129–137. DOI: 10.1377/hlthaff.2021.01000. Verified source

  • Ravel, K., J. Ahrary, K. Avakian, A. Feher, and I. Menashe. (2022). Effect of personalized outreach on Medicaid to Marketplace coverage transitions: A randomized clinical trial. JAMA Health Forum, 3(10), e223616. DOI: 10.1001/jamahealthforum.2022.3616. Verified source

  • European Parliament and Council of the European Union. (2016). Directive (EU) 2016/97 of 20 January 2016 on insurance distribution (recast). OJ L 26, 2 February 2016, 19–59. Consolidated version dated 9 January 2024. Verified source

  • International Association of Insurance Supervisors. (n.d.). Insurance Core Principles and ComFrame Online Tool: ICP 18, Intermediaries. Accessed November 19, 2026. Verified source

  • European Insurance and Occupational Pensions Authority. (2022). Credit Protection Insurance (CPI) Sold via Banks: Thematic Review. Verified source

  • Financial Conduct Authority. (2014). General Insurance Add-Ons Market Study: Final Findings. Official study summary, updated April 1, 2016. Verified source

  • Financial Conduct Authority. (2014). TR14/11: Price Comparison Websites in the General Insurance Sector. Official review summary, marked historical August 2025. Verified source

  • European Insurance and Occupational Pensions Authority. (2026). Third Report on the Application of the Insurance Distribution Directive. Publication page, March 30, 2026, EIOPA-26-077 and EIOPA-26-078. Verified source

  • OECD. (2026). Global Insurance Market Trends 2026. Paris: OECD Publishing. DOI: 10.1787/c8be6585-en. Verified source

  • World Bank. (2025). The Global Findex Database 2025: Report. Washington, DC: World Bank. Verified source

  • Karaca-Mandic, Pinar, Roger Feldman, and Peter Graven. (2013). The Role of Agents and Brokers in the Market for Health Insurance. U.S. Census Bureau, Center for Economic Studies Working Paper CES-13-58. Source record

  • Staiger, Becky, Anran Li, Diane Alexander, and Molly Schnell. (2022). Enrollment Brokers Did Not Increase Medicaid Enrollment, 2008–18. Health Affairs, 41(9), 1333–1341. DOI: 10.1377/hlthaff.2022.00182. Source record

  • Behavioural Economics Team of the Australian Government and Australian Securities and Investments Commission. (2021). Supporting consumer decision-making about add-on insurance. Australian Department of the Prime Minister and Cabinet, official project report summary and trial design. Source record

  • Larsen, Vance J. (2016). Responsible microfinance bundling: Experimental evidence on separating insurance and credit offers. M.S. thesis, Purdue University. AAI10146718. Source record