RESEARCH DOSSIER
Insurance Distributor Sales Data Quality and Attribution Bias
Abstract. Insurance distribution generates heterogeneous records of inquiries, referrals, quotations, advice, binding, issuance, renewal and cancellation. Their reliability determines the validity of channel comparisons, acquisition economics, producer incentives and conduct indicators. This monograph develops an insurance-specific synthesis of transaction semantics, data lineage, identity resolution, missingness, descriptive allocation and causal inference. It distinguishes contractual sales assignment, conserved marketing credit and the causal increment in eligible insurance outcomes. Comparative documentary research identifies relevant international supervisory principles, European Union instruments and selected United Kingdom, United States, Australian and Ontario frameworks. Synthetic audits demonstrate overlapping attribution, missing-outcome identification bounds, case-mix reversals and uncertainty in incremental acquisition costs. The study establishes methodological requirements for credible measurement across insurance institutions, while distinguishing policy acquisition from increased protection and fair customer outcomes. Its contribution consists of original analytical synthesis and reproducible illustrations of established methods. It estimates no actual insurer's performance and introduces no new named theory.
Keywords: Insurance distribution; insurance intermediation; sales data quality; attribution bias; record linkage; data provenance; missing data; multi-touch attribution; causal inference; acquisition cost; insurance market conduct; consumer protection.
Introduction
Insurance distribution connects prospective policyholders, intermediaries, insurers and supporting organizations in a transaction process that can extend for minutes, months or years. An individual may discover a policy through a comparison service, discuss requirements with a broker, obtain a quote from an insurer, and complete payment through a separate interface. A commercial placement may involve a retail broker, a wholesale intermediary, a delegated underwriting authority and several participating carriers. Each participant observes different stages of the same economic activity and records them according to its own operational obligations.
Data quality is the suitability of records for an explicitly specified measurement or analytical purpose. Attribution bias is systematic error in the assignment or inference of contribution arising from an incomplete or distorted observation of events, incompatible classifications, allocation conventions or invalid causal identification. Accurate data can still support biased attribution when the measured quantity is confused with another quantity. An intermediary's entitlement to commission may be valid even when its causal contribution to new demand is unknown. A correctly registered sale may be credited to several acquisition channels, each of which genuinely participated.
The independent subject of this dossier is cross-platform measurement integrity. The institutional economics of referrals and the mechanisms of funnel abandonment belong to neighboring research objects; they enter the present inquiry insofar as they affect the production of data and the interpretation of channel indicators. The two governing domains are Insurance Distribution and Intermediation and Insurance Governance and Market Conduct.
Three questions organize the inquiry. Which events and identities permit reliable cross-institutional measurement? How do overlapping records, missing observations and causal ambiguity distort sales comparisons? Which statistical, institutional and operational audits provide a defensible basis for reporting and evaluation?
The method combines comparative documentary analysis, established statistical literature and synthetic demonstrations. The inspected regulatory sources comprise international supervisory principles, original EU instruments and selected national or provincial frameworks. Source access varies between original texts or excerpts and publisher or indexed abstracts. The latter support only statements established within their abstracts. No proprietary insurance microdata, participant-level experiment or independent replication of external datasets is claimed. The calculations are methodological illustrations, not estimated industry prevalence. The cutoff identifies the research session rather than a first publication date.
Distribution Sales Data Model and Event Taxonomy
Observational units and institutional roles
Insurance records describe entities, relationships, events and states. An entity may be a person, organization, insurer, intermediary, contract, exposure or payment transaction. A relationship defines the role of one entity relative to another, such as policyholder, insured, beneficiary, payer, broker or authorized representative. An event changes or documents an institutional state. A policy state is a derived representation of the contract at a specified time.
These distinctions are necessary because the counting units differ. A group policy can cover hundreds of insured individuals. A household may hold multiple policies. A single commercial risk can produce several carrier participations, and one insurance application can generate alternative quotations. A sales record keyed to a customer cannot therefore be assumed to identify a contract or exposure. Conversely, a policy number may be insufficient to distinguish policy versions, coverage terms and reinstatements.
A distributor's activities may include origination, referral, advice, placement, delegated binding, payment facilitation and continuing service. The same organization can perform several roles, and administrative responsibility can change over time. An intermediary's organizational identity is distinct from an individual producer's identity, appointed role or authority. The source system should preserve both when the research question depends on them.
The appropriate analytical unit follows from the intended outcome. The number of new policies measures contract acquisition; unique purchasers measure the reach among parties to contracts; exposed people or properties concern insured exposure; premium measures monetary activity; claims and complaints measure subsequent outcomes with their own observation windows and denominators. Their relationships are empirical and institutional, not arithmetic identities.
Event definitions
| Event family | Unit | Essential distinction |
|---|---|---|
| Discovery | Session or contact | Anonymous, identified or shared |
| Referral | Referral occurrence | Referrer, recipient and acknowledgment |
| Application | Submission or version | Draft, submitted, withdrawn |
| Quotation | Offer version and risk | Indicative, bindable, expired |
| Advice | Recommendation or decision record | Product selection versus advice |
| Binding | Accepted coverage transaction | Authority, effective date |
| Issuance | Policy or policy version | Contract documentation |
| Payment | Financial transaction | Authorization, settlement, reversal |
| Renewal | New term or continuation | Acquisition versus persistence |
| Cancellation | Termination event | Effective date and reinstatement |
| Claim or complaint | Subsequent case | Exposure and reporting lag |
A quote can mean an initial price indication, a carrier-specific offer, a fully underwritten offer or an alternative offered within a single customer request. The research event dictionary retains the source meaning and the harmonized definition. A sale can mean payment, acceptance of binding instructions, policy issuance or an in-force contract at a specified maturity point. The denominator should correspond to the defined sales outcome and its population.
A useful event model includes a stable event identifier, source-specific message identifier, event version, record source, actor roles, application and contract keys, event and ingestion timestamps, legal or coverage effective dates, product, jurisdiction, status and transformation lineage. The original message ID distinguishes transport events and retries; the business event ID distinguishes underlying occurrences. Deduplicating messages alone does not establish that underlying business events are unique.
Time requires separate fields. A quote can be generated Monday, accepted Wednesday and received by a reporting warehouse Thursday, while coverage begins next month. Sales series based on processing date describe operational activity; series based on binding or inception describe different contract cohorts. Backdated cancellation and reinstatement create further differences between observed history and the reconstructed state as of a given date.
Event precedence may be established through causal business references, such as a reversal identifying an original payment, when source clocks are inconsistent. A system should retain unresolved states where relevant events are unavailable. Unknown is a valid data-quality category. Treating absent observations as evidence of absent real-world events is a substantive error.
Denominators and financial consistency
Define B as the number of unique eligible quote requests with at least one qualifying binding and Q as the number of eligible quote requests with adequate outcome ascertainment in the same cohort. The ratio B/Q is then a request conversion measure. Counting all contracts in the numerator when some requests produce several policies instead yields a contract-per-request ratio that may exceed one. Both measures can be useful when labeled accurately.
Acquisition cost requires a coherent expenditure scope and outcome definition. Total media, commission, platform and operational costs may include transfers between participants. Adding commission expense in an insurer's ledger and commission revenue in an intermediary's ledger as separate costs can double count the same transfer in an aggregate analysis. Acquisition expenditure, written premium and earned premium refer to different accounting concepts and periods. Currency conversion, taxes, refunds, reversals and net/gross distinctions must be explicit.
The OECD insurance market report (2026) illustrates the use of aggregate premium-to-GDP measures. These support macroeconomic comparison, not a measure of individual protection or distributor productivity. The World Bank's Global Findex Database 2025 supplies demand-side context about financial inclusion and digital connectivity. Such context does not establish broker activity or the causal effects of a distribution channel.
Claims counts and premium measures require appropriately matured cohorts. A low claims count in recently sold policies may reflect short exposure, delayed reporting, exclusions or selected risks. Evaluating distribution through claims outcomes requires linkage to insured exposure, coverage terms, time at risk and the relevant claims definition. Fair treatment, protection and acquisition are consequently separate dependent variables.
Channel and Referral Event Provenance
The classification problem
Marketing media, actors and transaction interfaces are different dimensions. An affiliate website is an acquisition medium or commercial partner, a broker performs an institutional activity, and an insurer's portal is a transaction interface. A customer can encounter all three during a single purchase. A database that forces these dimensions into one mutually exclusive channel field must specify the rule that makes them exclusive.
Provenance is evidence about the origin and transformation of a record. Moreau and Missier's (2013) PROV data model describes entities, activities and agents. Its established vocabulary can identify a source record, an extraction or transformation activity, the responsible organization and a derived dataset. The analytical value is the ability to trace an indicator to the evidence and processing choices that produced it.
A referral event may retain a referrer key, recipient key, original lead identifier, timestamp, acknowledgment status and source namespace. Referral IDs alone generally do not establish personal identity or causal contribution. Some partners may recycle identifiers, transmit a lead to multiple organizations or assign a referral code after a conversion. The evidence necessary to interpret the identifier includes its namespace, validity period, assignment procedure and link to a verified transaction.
A sequence containing a referral, an application and a binding supports a documented history conditional on the quality of the event keys. It cannot identify the outcome that would have occurred without the referral. Likewise, an application without a referral key may represent direct origination, offline assistance or loss of a tracking field. The category unknown should be retained unless separate evidence supports classification.
Reconciliation and reporting incentives
A distributor's placement ledger, insurer's policy system, platform callback and commission statement are evidence of different activities. The insurer's accepted policy transaction can be authoritative for binding status, subject to delegated authority and later corrections. The distributor may hold the better evidence of customer advice and referral history. A financial statement may establish an entitlement or payment, rather than the event's acquisition origin. Field-specific and time-specific authority is more defensible than designating one institution as authoritative for every attribute.
| Reconciliation exception | Possible explanation | Evidence |
|---|---|---|
| Lead without recipient acknowledgment | Delivery failure or duplicate | Sender and recipient logs |
| Application without referral | Offline contact or missing key | Contact records and source coverage |
| Binding without quote | Delegated authority or missing linkage | Policy transaction and authority |
| Several sales claims for one policy | Shared contribution or duplication | Contract key and role history |
| Commission without current active contract | Lag, transfer or adjustment | Financial and contract histories |
Reconciliation should preserve legitimate many-to-many relationships rather than remove intermediate contacts to obtain an attractive total. A co-brokerage arrangement can create several legitimate participation records. A single exclusive sale count, however, must correspond to one defined sale unit.
Commercial targets can affect the production of records. Where a payout depends on an attributed conversion, an institution may gain from expansive definitions of a qualifying contact or treating transferred clients as new business. Such potential mechanisms are analytical hypotheses, not evidence of misconduct by a particular institution. A study should document changes in definitions, incentives, systems and reporting deadlines before attributing discontinuities to changes in demand.
For example, an increase in one intermediary's claimed conversions accompanied by unchanged unique insurer bindings may reflect reallocation, improved matching or a shift in the path composition. It cannot be interpreted as incremental acquisition without further evidence. The audit should compare source-level changes, independently reconciled outcomes and allocation-version histories.
EIOPA's (2026) EU distribution market overview identifies differences in registration categories and data collection across markets. Registry counts cannot be treated as the number of active selling entities or connected directly to transaction productivity without harmonization. International comparisons should define the jurisdiction, line, legal category and counted entity before constructing ratios from registration and premium totals.
Identity Linkage, Duplicate Transactions and Attribution Windows
Identifying entities across systems
Cross-system linkage connects leads to customers, customers to applications, offers to bindings and policies to later terms. A durable identifier needs a known namespace and assigner. Personal identifiers can change, contact details can be shared, and organizational names can be inconsistent. A household email address does not establish whether two records represent the same insured individual. The role of each identity is particularly important for group and commercial contracts.
Deterministic linkage uses exact or normalized agreement on chosen keys. Where insurers and distributors share reliable contract identifiers, this approach can be transparent and efficient. Missing keys, transcription errors and reused values can create both false matches and false nonmatches. Probabilistic linkage estimates the relative evidential strength of a comparison pattern.
Fellegi and Sunter's (1969) record linkage theory provides a decision framework based on the relative probability of observed comparison patterns among matches and nonmatches. Under conditional independence as an approximation, a comparison vector has log likelihood ratio equal to the sum over fields of log(m/u), where m and u are the conditional comparison probabilities among true matches and nonmatches. Agreement on a rare identifying value can provide more evidence than agreement on a common one. Dependence among fields, household sharing and evolving identities require calibration or alternative models.
A score is evidence under the model rather than a definitive statement of contract ownership. Thresholds may yield accepted matches, rejected matches and a review region. Independent evidence can resolve some ambiguous pairs. Entity-level validation must also consider temporal and cardinality constraints: pairwise matches may imply an inconsistent identity cluster when transitive links are taken without review.
Harron et al. (2017), in their linkage quality evaluation guide, discuss reference data, comparisons of linked and unlinked records, and sensitivity analysis. The ONS data linkage policy distinguishes matching rates from precision and recall. Define TP as accepted true links, FP as accepted false links and FN as missed true links. Precision = TP/(TP + FP); recall = TP/(TP + FN). If 950 of 1,000 accepted links are genuine, precision is 95%. Recall is also 95% only if the number of missed genuine links is 50. The equality in that example is contingent on the assumed missed-link count.
A validation sample of accepted records alone cannot establish recall. Review of rejected candidates and an independent reference population may be needed, including candidates excluded by blocking rules. Source-specific error rates matter because false links can assign policies to the wrong channel, while missed links can transfer a genuine referral to the unknown category. A pooled accuracy statistic can conceal differential error across intermediaries.
Overlapping participation and exclusive allocation
Suppose a synthetic cohort contains 150 unique bindings and three channels: affiliate A, broker B and direct digital D.
| Recorded path | Unique bindings |
|---|---|
| A only | 30 |
| B only | 20 |
| D only | 40 |
| A then B | 30 |
| A then D | 10 |
| B then D | 20 |
| Total | 150 |
Each channel records 70 participating bindings. Their aggregate is 210, exceeding the unique count by 60, or 40% of the unique count. Each participation figure is descriptively meaningful; the total is not an exclusive sales count.
Equal-share allocation among channels on each path produces 50, 45 and 55 sale-equivalents for A, B and D. First-touch allocation produces 70, 40 and 40; last-touch allocation produces 30, 50 and 70. Each allocation conserves the 150 outcomes while producing a different ranking. The results depend on the meaning assigned to path position and participation, not on proof of causal contribution.
Let w_ic be a nonnegative credit assigned to channel c for eligible binding i. For exclusive credit, the sum over channels of w_ic must equal one for every eligible outcome. An unknown category can receive unresolved credits. Then the sum of all channel credits equals the number of eligible unique outcomes. This is an accounting identity defining consistency of exclusive allocation, not an original theoretical model.
Financial shares, commercial attribution and contractual entitlement remain distinct. A commission agreement can legitimately allocate compensation differently from descriptive sales credit. A multilayer placement can measure signed participation shares rather than count one full policy for each carrier. Each table should identify the quantity it allocates and its reconciliation total.
Time-window selection and recurrent events
A lookback window defines which observed contacts qualify for attribution. A short window may exclude early research and complex broker consultation. A long window may include unrelated interactions. The choice depends on the insurance product, usual purchasing cycle, date definitions and the research question. The selected window is a modeling convention that can be tested rather than presumed universally appropriate.
Contacts after the binding event cannot cause that event, although they may affect subsequent service or renewal. Late arrival of a pre-binding event is a reporting issue rather than a post-binding activity. Both business-event and ingestion timestamps are necessary to separate these cases.
Renewal, cancellation and reinstatement require special care. A renewed policy can be genuine continued protection without constituting a new customer. A broker-of-record change can transfer administrative responsibility without creating a new contract. A renewal intervention may affect persistence rather than origination. Recent cohorts may not yet have experienced cancellations observable in older cohorts, producing differential censoring.
A mature analysis defines cohort eligibility, observation windows and follow-up requirements before comparing channels. Sensitivity analyses vary window length and outcome maturity; their effects on the unknown share, total eligible outcomes and relative allocations should be reported separately.
Consent, Missingness and Measurement Error
Selection mechanisms
Missing observations include unrecorded events, missing fields and records excluded by identity or consent filters. Website analytics omit contacts occurring offline; pseudonymous browser sessions may not link to policy administration; customer choices can limit optional processing. Insurer transaction totals may therefore be complete for a specific contract population even when the observed marketing path is incomplete.
Let R indicate observation of a value Y. Missing completely at random is independence of the missingness mechanism from relevant data under the adopted formulation. Missing at random permits dependence on observed data but requires conditional independence from missing values given the observed information. Missing not at random refers to mechanisms for which that conditional independence does not hold. These are assumptions about observation, not inferences established by a low missingness rate.
Rubin's (1976) inference and missing data identifies conditions for ignoring missingness in likelihood inference, including missing at random and distinct parameters. Their plausibility depends on the available observed variables and the data-producing process. If individuals with unusually complex risks avoid online systems, both measurement coverage and conversion may depend on partially observed complexity. Comparing only linked observations may distort the full target population.
Suppose 800 eligible quote requests are known, 600 have a resolved binding outcome, and 120 of those bind. Complete-case conversion is 20%. Without assumptions about the unresolved 200, the actual full-cohort conversion lies from 15% (120/800) to 40% (320/800). The interval reflects identification uncertainty, not sampling variance. A precise confidence interval around the observed 20% does not remove the missing-outcome ambiguity.
Similarly, suppose 1,000 known bindings consist of 350 assigned to A, 350 assigned to B and 300 unassigned. If A and B are the only possible true source categories, each category's share lies between 35% and 65%. The observed tie does not establish a population tie, and neither strict ranking is identified.
Correction and sensitivity methods
Inverse probability weighting can adjust for differential observation when the observation probability is identifiable from suitable covariates, is positive for the relevant units and supports the required missingness assumptions. If p(X) is the chance of observation given X, observed units weighted by its reciprocal may reconstruct the target distribution under that design. Very small probabilities increase variability, and unobserved determinants can invalidate the adjustment.
Multiple imputation creates plausible versions of missing values under a specified model, propagating uncertainty through analysis. Insurance transaction data require structural constraints. An imputed quote should have an admissible product and jurisdiction; a certificate should map to an eligible group contract; canceled or withdrawn offers should not become terminated policies automatically. The model needs to preserve structural nonapplicability and deliberate suppression separately from unobserved values where possible.
Sensitivity analysis is especially valuable for nonignorable mechanisms. In the 800-request illustration, if q denotes conversion among unresolved requests, the overall conversion is (120 + 200q)/800. Setting q to 0.10, 0.20 or 0.40 gives 17.5%, 20% or 25%. Tracking the decision implications over transparent assumptions may be more informative than a single imputed estimate.
Two-list capture-recapture provides one established method of estimating population completeness. Suppose list A has 600 entries, list B has 500, and the verified overlap is 300. The observed union is 800. Under assumptions of closed population, reliable matching, appropriate capture homogeneity and independence, the elementary estimate is 600 × 500 / 300 = 1,000. The Chapman adjustment is 601 × 501 / 301 − 1 = 999.34. If the overlap is 400 or 200 with unchanged list margins, the respective elementary estimates are 750 and 1,500.
Brenner's (1995) capture-recapture dependence study examines the effects of source dependence. Positive dependence can lead to underestimation and negative dependence to overestimation under the described two-list approach. Insurance distributor records may rely on shared infrastructure, reporting rules or joint customer behavior that creates dependencies. Capture-recapture does not identify the missing individuals or prove their distribution pathways. Where assumptions remain unsupported, the observed union and a sensitivity analysis are appropriate evidence.
Privacy and measurement access
The General Data Protection Regulation (European Parliament and Council, 2016) specifies principles of accuracy, purpose limitation and minimization, as well as lawful processing conditions and separate conditions for special-category data. Consent is one possible lawful basis. The fact that an insurer or distributor holds an operational record does not by itself establish that the record can be reused for a different marketing purpose.
Pseudonymous identifiers reduce direct exposure but can support reidentification when joined across datasets. Researchers need defined purposes, lawful access, retention boundaries, security safeguards and a process to address errors. Where person-level linkage is inappropriate, aggregated outcomes and authorized controlled analysis can sometimes support the intended statistical question.
A distributor that observes more touchpoints may obtain a larger share of credit even when the true pattern of contribution is unchanged. Equal visibility cannot be assumed across direct, intermediary, offline and digital channels. Preserving unknown categories and examining measurement selection can support accurate inference without attempting to reconstruct prohibited customer histories.
Multi-Touch and Incrementality Attribution
Distinct estimands
Contractual assignment states which entity holds a specified institutional role or commercial entitlement. Descriptive marketing attribution allocates observed outcomes among recorded contacts under an explicit convention. Incrementality estimates the difference in outcomes attributable to a defined intervention relative to a counterfactual. All three can be valid at the same time, yet their numerical results need not coincide.
An insurer can issue a contract through its direct portal after a broker advised the customer. The broker may hold contractual rights to commission. The direct channel may obtain last-touch marketing credit. A randomized study may nevertheless indicate that a particular advertising contact had little incremental effect. Each statement concerns different evidence.
Consider the 150-binding synthetic cohort. Suppose costs are $4,000 for A, $4,500 for B and $5,500 for D. Their participation-based ratios are $57.14, $64.29 and $78.57 per observed participation. The equal-share conversion-equivalent ratios are $80, $100 and $100. Portfolio expenditure per unique binding is $14,000/150 = $93.33. It is incorrect to represent all three participation ratios as exclusive acquisition costs. The chosen measure changes management interpretation even when the underlying expenditure and contract records are identical.
Descriptive path models and causal inference
First-touch and last-touch attribution are straightforward to reproduce, given adequate observation and stable timestamp rules. Equal-share allocation embeds a symmetry convention, while time decay favors recent contacts. Predictive path models may use sequences, interactions and estimated conversion probabilities. All such approaches should be evaluated for their intended use and explicit assumptions.
Li and Kannan's (2014) multichannel conversion attribution study addresses carryover, spillover and field experimentation. The methodological relevance is the treatment of interdependent contacts. The study's findings do not establish quantitative effects in insurance. Applying such methods to insurance requires attention to renewal cycles, eligibility, underwriting, coverage and multiple intermediary roles.
A causal estimand specifies a counterfactual. For eligible individual i, let Y_i(1) be the potential binding indicator under a defined intervention and Y_i(0) under its comparison condition. The population average effect is the mean of their difference. Observed path credit cannot supply both potential outcomes for one person. Causal identification depends on research design and assumptions.
Confounding is pervasive in sales data. A prospect actively searching for insurance can be more likely to encounter advertising and more likely to purchase irrespective of advertising. A vehicle renewal date, mortgage closing or business expansion can independently affect both exposure to contact and need for coverage. Conditioning on a quote completion caused by the intervention can introduce selection bias or alter the target estimand.
Gordon et al.'s (2018) advertising measurement field experiments compare observational methods against randomized evidence, demonstrating that extensive covariates need not recover experimental treatment effects. The insurance-specific implication is that increasingly rich customer and policy records do not by themselves overcome confounding. The original empirical estimates concern general advertising, not distributor-level insurance outcomes.
Randomized lift and uncertainty
Randomized assignment can estimate an intention-to-treat effect when the design delivers valid assignment, measures outcomes appropriately and respects interference and follow-up conditions. An insurance distribution trial could vary eligible marketing contact, a compliant interface or the timing of an optional notification. It must preserve applicable obligations and avoid denying contractual rights or legally required information.
Suppose a hypothetical experiment has 10,000 eligible individuals per arm. Treatment produces 500 unique bindings; comparison produces 450. The rates are 5% and 4.5%, yielding a difference of 0.5 percentage points and relative lift of 11.1%. The independent-sample unpooled standard error is the square root of (0.05 × 0.95/10,000 + 0.045 × 0.955/10,000), or approximately 0.301 percentage points. The approximate 95% normal interval is −0.09 to 1.09 percentage points.
If incremental campaign expenditure is $10,000, the point ratio is $200 per additional binding. Because the effect interval includes zero, the data do not establish a finite positive incremental acquisition cost. Lewis and Rao's (2015) advertising returns research shows the importance of precision limitations even in large experiments. Insurance value additionally depends on retention, expenses, claims, reinsurance and capital. A point ratio can thus be unstable both statistically and economically.
Assignment at a broker, employer or regional level may reduce contamination, but inference then requires accounting for clustered assignment. Geographic spillover, competitors' interventions and evolving markets can invalidate simple comparisons. Quasi-experimental designs may be useful where assumptions such as parallel trends or credible threshold assignment are defensible. Placebo tests and pre-intervention checks help diagnose violations; their success does not logically prove every identifying assumption.
Acquisition, protection and conduct
An outcome must match the policy question. Additional binding by one insurer can reflect switching from another insurer, as opposed to newly insured risk. A distributor's increase in acquired contracts may coincide with lower retention or unsuitable coverage. The intended measure of incremental market access must identify previously uninsured or inadequately protected exposures, the relevant policy terms and the duration of protection.
Underwriting selection complicates inference. Campaigns may attract prospects with different risk characteristics and approval probabilities. The total effect among eligible prospects incorporates the underwriting process. Restricting analysis to accepted applicants conditions on an intermediate event and answers a different question.
Retention and renewal effects require a later outcome horizon than origination. A campaign that increases initial binding while increasing early cancellation can have a different effect on in-force coverage. A credible experimental or observational evaluation should specify the relevant follow-up and treatment of censoring before analyzing outcomes.
Regulatory Transparency and Confidentiality
International supervision
The IAIS Insurance Core Principles, especially ICP 18 and ICP 19, address supervision of intermediaries, conflicts of interest and conduct of business. They supply international supervisory context rather than a directly enforceable global statutory attribution algorithm. Their relevance is institutional accountability for activities and incentives affecting customers.
Sales metrics can be useful for monitoring, but high conversion does not itself establish fair treatment. A low complaint rate can reflect good service or inaccessible complaint procedures. Proper conduct indicators require compatible exposure denominators, reporting periods and a connection to the underlying behavior or product outcome.
European Union
The Insurance Distribution Directive (European Parliament and Council, 2016) defines activities within distribution and specified exclusions, including some mere information-provision activities under the stated conditions. Articles 17–20 concern customer interests, remuneration-related conflicts, disclosure and demands-and-needs obligations. The regulatory status of an affiliate or platform depends on its actual activity and applicable national transposition, not a marketing label.
A lead code is therefore insufficient evidence of regulatory authorization or fulfillment of conduct duties. A sales report may support investigation of the transaction chain, but the relevant legal facts include activities undertaken, authority, relationship to the customer and operative jurisdiction.
EIOPA's Third IDD Application Report (2026) draws on national supervisory and stakeholder information. Its evidence environment illustrates the challenges of comparing distribution structures and regulatory measures. It is not an individual customer or transaction dataset. This study uses the original directive and selected supervisory sources; it does not claim an exhaustive consolidated legal analysis of every member state.
Lawful reuse and linkage of personal data additionally depend on the applicable processing purpose and responsibilities. Separate communications or tracking rules may be relevant in individual settings. A customer outcome can be statistically important while direct person-level disclosure remains inappropriate.
Selected national and provincial approaches
The United Kingdom FCA's demands and needs rules in ICOBS 5.2 illustrate requirements concerning proposed insurance contracts and customer needs before conclusion. A record of sales origin does not substitute for the required substantive evidence. Data systems should retain the distinction between marketing source, customer needs and advice records.
The NAIC producer licensing overview describes the state-centered US producer licensing environment. Producer identity and authorization checks must be differentiated from a finding of causal sales contribution. State, product and transaction date matter when determining applicable obligations. The NAIC Market Conduct Annual Statement provides standardized conduct data collection; the resulting indicators are not alone sufficient for identifying marketing effects.
Australia's ASIC Regulatory Guide 274 (2024) concerns product design and distribution obligations. Depending on the product and relevant legal scope, sales, cancellation, complaint and claims information can contribute to product review and distribution monitoring. A marketing attribution measure cannot substitute for the product oversight evidence required in the particular case.
Ontario's FSRA fair treatment guidance incorporates the CCIR/CISRO fair-treatment approach. This provides a provincial supervisory example rather than a uniform description of law across all Canadian jurisdictions. Comparable customer outcomes may support comparative research even when regulatory instruments and operative obligations differ.
Controlled evidence and governance
A credible reporting framework identifies who owns each source, prepares metrics, validates transformations, determines incentives, approves changes and investigates anomalies. Where a commercial team can both define credited outcomes and benefit from them, independent review of definitions and exceptions becomes particularly valuable.
Confidentiality and auditability can coexist through controlled evidence access, versioned aggregate extracts, independent reconciliation and documented methodological choices. Public reporting need not disclose individual customer histories or commercial agreements to specify a denominator, calculation and uncertainty range. More detailed evidence can be made available to appropriately authorized reviewers under lawful conditions.
The compared authorities support the importance of accountable distribution activity and customer-related evidence. They do not establish a universal first-touch, last-touch or causal attribution rule. Different regulatory questions demand different data, and each actual obligation requires examination of the applicable instrument and jurisdiction.
Reproducible Data-Quality and Sensitivity Audits
Audit design
A study begins with a target question and defined population rather than available dashboard fields. It records source versions, inclusion criteria, cohort dates, follow-up horizon, outcome definition, source ownership, transformation code and treatment of exceptions. Where lawful, fingerprints of input extracts and code versions permit reproduction of the analysis.
The Actuarial Standards Board's (2016) ASOP No. 23 on data quality addresses appropriate data, review, reliance and disclosure in actuarial services. It does not require an audit of every datum or create a universal marketing obligation. Its relevance to the present synthesis is the explicit connection between intended use, review and communication of limitations.
| Dimension | Reproducible test |
|---|---|
| Structural integrity | Event and message key uniqueness |
| Semantic validity | Correct roles, states and transition logic |
| Completeness | Missing fields and source coverage by cohort |
| Linkage | Precision and recall on reference samples |
| Reconciliation | Unique contracts and signed financial amounts |
| Allocation | Exclusive-credit totals and unresolved category |
| Time | Event versus ingestion dates, maturity, leakage |
| Inference | Assignment, balance, clustering, sensitivity |
A high technical pass rate is insufficient if a field has the wrong substantive meaning. A conserved allocation can still be biased relative to its assumed target, and a high predictive score can coexist with poor causal identification. Audit results should distinguish these dimensions rather than average them into a single score that conceals the cause of a failure.
Validation sampling can combine representative strata and deliberately oversampled high-impact exceptions. Population error rates require known sampling probabilities or another defensible design. A targeted anomaly sample supports diagnosis but cannot be interpreted as representative without adjustment. Reports should preserve the size, sampling frame, adjudication process and uncertainty of each estimate.
Case mix and ranking reversal
Suppose two distributors handle simple and complex requests. The counts are synthetic.
| Risk group | Distributor A | Distributor B |
|---|---|---|
| Simple | 90 / 100 | 800 / 1,000 |
| Complex | 20 / 100 | 10 / 100 |
| Aggregate | 110 / 200 | 810 / 1,100 |
A converts 90% versus B's 80% among simple risks and 20% versus B's 10% among complex risks. Their aggregate rates are 55% and 73.64%. Distributor B has a larger proportion of simple risks, and the aggregate comparison reverses both within-group rankings. Standardizing to an equal mix of simple and complex risks produces 55% and 45%.
This example illustrates the importance of the target population in evaluating channel performance. Standardized rates can improve descriptive comparison while remaining distinct from the causal effect of replacing one distributor with another. Product classification may itself depend on upstream channel decisions, and unobserved differences may persist within strata. Adjustment variables and reference weights require substantive justification.
Reproducible synthetic calculations
The following standard-library fragment reproduces the core illustrative calculations. It does not use insurer or customer data.
from math import sqrt
paths = {"A":30,"B":20,"D":40,"AB":30,"AD":10,"BD":20}
channels = "ABD"
n = sum(paths.values())
participation = {c: sum(k for p,k in paths.items() if c in p)
for c in channels}
first = {c: sum(k for p,k in paths.items() if p.startswith(c))
for c in channels}
last = {c: sum(k for p,k in paths.items() if p.endswith(c))
for c in channels}
equal = {c: sum(k/len(p) for p,k in paths.items() if c in p)
for c in channels}
assert n == 150 and sum(participation.values()) == 210
assert all(sum(a.values()) == n for a in (first,last,equal))
assert equal == {"A":50,"B":45,"D":55}
missing_bounds = (120/800,320/800)
chapman = 601*501/301-1
effect = .05-.045
se = sqrt(.05*.95/10000 + .045*.955/10000)
ci = (effect-1.96*se,effect+1.96*se)
aggregate = (110/200,810/1100)
standardized = ((.9+.2)/2,(.8+.1)/2)
print(participation,first,last,equal,missing_bounds,
chapman,ci,aggregate,standardized)
Production implementation would introduce controlled access, explicit event types, time and currency normalization, independent fixtures and tests for corrections, reversals and many-to-many matches. The fragment is a transparent reproducibility aid for these synthetic results, not a packaged production service.
Sensitivity, corrections and research priorities
A sensitivity audit varies consequential assumptions rather than simply computing another dashboard. In linkage it varies thresholds and unresolved treatments; in attribution it varies observation windows and allocation rules; in missing-data analysis it varies plausible outcomes among unobserved units; in cross-channel comparisons it varies documented case-mix weights; in causal studies it examines assignment, interference, clustering, follow-up and identification assumptions.
The examples establish that distinct forms of uncertainty have distinct meanings. The completed-case conversion interval reflects an identification problem. The trial interval describes sampling uncertainty under the adopted design. Alternative path allocations describe modeling conventions. The case-mix reversal concerns population composition. Reporting all four as one confidence score would obscure their scientific interpretation.
Correction procedures should identify the original result, affected population, source and metric version, reason for revision, corrected result and decision implications. A late historical correction can change financial or conduct reporting without indicating that the underlying business event itself changed. If commercial incentives were affected, the responsible institution can determine applicable remediation requirements.
Future insurance research should seek authorized event histories independently reconciled to policy and outcome records. Linkage can be tested against reference subsets, source coverage quantified by documented strata, and attribution evaluated under alternative windows and allocation conventions. Where suitable and lawful, randomized or credible quasi-experimental research can estimate effects on new binding, retention and first-time coverage separately.
Different questions require different data. A reliable policy ledger may establish the unique contract count without reconstructing a path. A customer survey may capture offline contacts but contain recall errors. A treatment assignment record may establish randomization but still require complete outcome data. A producer register identifies institutional identities rather than each sale. The integration of these sources is scientifically valuable when their definitions, observation conditions and confidentiality constraints permit the linkage.
Conclusion
Insurance distributor sales data are reliable when their event semantics, identities, provenance, outcome definitions and observation boundaries support the intended measurement. The same contract can legitimately generate multiple participation records, while an exclusive attribution total must reconcile to unique eligible outcomes. Missing source assignment and missing binding outcomes are distinct problems. Record linkage requires assessment of false and missed matches, including observations excluded before matching.
The central analytical distinction is between contractual assignment, descriptive marketing credit and causal incrementality. The first concerns authority or entitlements, the second concerns allocation of observed outcomes, and the third concerns counterfactual effects. Accurate administrative records and internally conserved marketing credit are compatible with substantial causal uncertainty.
The comparative institutional evidence identifies the significance of accountable distribution activities, conflicts, appropriate customer outcomes and lawful processing. It does not prescribe a universal attribution convention. Regulatory, financial and research measures should retain their separate denominators, purposes and sources of authority.
Five numerical findings follow from the synthetic analyses: overlap can inflate aggregate participation by 40% relative to unique bindings; alternative conserved allocations can change rankings; unresolved observations can produce a wide identification interval; composition can reverse channel comparisons; and a favorable experimental point estimate can remain too imprecise for a stable incremental cost ratio. These findings illustrate general measurement properties without estimating industry-wide effects.
The study's original contribution is the integration of established statistical and data-governance methods with the institutionally specific meaning of insurance transactions. An independently reconciled binding count is important evidence of acquisition, yet expanded insurance protection, fair conduct and causal distribution effectiveness remain separate research outcomes. Their credible evaluation requires explicit populations, definitions, observation processes and identifying assumptions.
References
EIOPA (2026). Third Report on the Application of the Insurance Distribution Directive. EIOPA-26-077, March 30, 2026. Verified source.
EIOPA (2026). Structure of the EU Insurance Distribution Market. Online market overview. Verified source.
IAIS (n.d.). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups. ICP 18 and 19. Verified source.
European Parliament and Council (2016). Directive (EU) 2016/97 of 20 January 2016 on insurance distribution (recast). OJ L 26, 19–59. Verified source.
European Parliament and Council (2016). Regulation (EU) 2016/679 of 27 April 2016 (General Data Protection Regulation). OJ L 119, 1–88. Verified source.
FCA (n.d.). Insurance: Conduct of Business Sourcebook, ICOBS 5.2, Demands and needs. Verified source.
NAIC (n.d.). Market Conduct Annual Statement (MCAS). Verified source.
NAIC (n.d.). Producer Licensing. Verified source.
ASIC (2024). Regulatory Guide 274: Product Design and Distribution Obligations. September 10, 2024. Verified source.
FSRA Ontario (n.d.). Fair Treatment of Customers in Insurance. Verified source.
OECD (2026). Global Insurance Market Trends 2026. OECD Publishing, Paris. DOI: 10.1787/c8be6585-en. Verified source.
World Bank (2025). The Global Findex Database 2025. World Bank, Washington, DC. Verified source.
Moreau, L., & Missier, P. (eds.). (2013). PROV-DM: The PROV Data Model. W3C Recommendation, April 30, 2013. Verified source.
Fellegi, I. P., & Sunter, A. B. (1969). A Theory for Record Linkage. Journal of the American Statistical Association, 64(328), 1183–1210. DOI: 10.1080/01621459.1969.10501049. Verified source.
Harron, K. L., Doidge, J. C., Knight, H. E., Gilbert, R. E., Goldstein, H., Cromwell, D. A., & van der Meulen, J. H. (2017). A guide to evaluating linkage quality for the analysis of linked data. International Journal of Epidemiology, 46(5), 1699–1710. DOI: 10.1093/ije/dyx177. Verified source.
Office for National Statistics (n.d.). Data Linkage and Matching Policy. Verified source.
Brenner, H. (1995). Use and limitations of the capture-recapture method in disease monitoring with two dependent sources. Epidemiology, 6(1), 42–48. DOI: 10.1097/00001648-199501000-00009. Verified source.
Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. DOI: 10.1093/biomet/63.3.581. Verified source.
Li, H. (Alice), & Kannan, P. K. (2014). Attributing Conversions in a Multichannel Online Marketing Environment: An Empirical Model and a Field Experiment. Journal of Marketing Research, 51(1), 40–56. DOI: 10.1509/jmr.13.0050. Verified source.
Gordon, B. R., Zettelmeyer, F., Bhargava, N., & Chapsky, D. (2018). A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook. Working manuscript, April 12, 2018. Verified source.
Lewis, R. A., & Rao, J. M. (2015). The Unfavorable Economics of Measuring the Returns to Advertising. Quarterly Journal of Economics, 130(4), 1941–1973. DOI: 10.1093/qje/qjv023. Verified source.
Actuarial Standards Board (2016). Actuarial Standard of Practice No. 23: Data Quality. Revised edition, December 2016. Verified source.