Health Risk Assessment and Underwriting

Introduction and context

Health insurance underwriting is the disciplined process of evaluating expected health-related costs, uncertainty, eligibility, and policy terms before accepting a risk or renewing coverage. It connects clinical information with actuarial analysis, product design, claims management, regulation, and ethics. The underwriter does not simply ask whether an applicant is ill. The central questions are: What is the probability of a health event? What cost and duration might follow? How uncertain is the estimate? Which controls are legitimate and proportionate? Can coverage be offered without unlawful discrimination or unacceptable exclusion?

Underwriting is particularly important in health insurance because medical expenditure is highly skewed: a minority of members may account for a large share of claims. Costs are also influenced by age, morbidity, provider prices, treatment patterns, benefit design, geography, technology, and social determinants of health. Consequently, underwriting must combine individual assessment with portfolio management. Excessively narrow classification may be unfair and inaccessible; insufficient classification may create adverse selection and threaten solvency.

The legal rules governing underwriting differ substantially across jurisdictions. Some markets permit individual medical underwriting; others require community rating, guaranteed issue, risk adjustment, or restrictions on pre-existing-condition exclusions. This chapter therefore presents general principles and requires the reader to apply the relevant insurance, privacy, equality, consumer-protection, and health-information laws of the applicable jurisdiction.

Learning objectives

By the end of this chapter, the learner should be able to:

1. Explain the purpose, stages, and terminology of health underwriting.
2. Apply medical, demographic, occupational, lifestyle, and group-underwriting methods.
3. Assess pre-existing conditions using evidence-based and proportionate methods.
4. Distinguish individual risk selection from portfolio risk management.
5. Interpret basic frequency, severity, expected-cost, and loss-ratio measures.
6. Evaluate underwriting decisions against fairness, accessibility, privacy, and anti-discrimination requirements.
7. Critically assess the uses and limitations of predictive analytics and artificial intelligence.

1. Foundations of health risk assessment

1.1 Risk, uncertainty, and insurability

Risk refers to uncertainty about future outcomes that can be described probabilistically or financially. In health insurance, outcomes include hospitalization, surgery, outpatient treatment, prescription use, disability-related care, and death where relevant to the product. A risk is generally more insurable when losses are accidental or uncertain, measurable, sufficiently numerous for pooling, and not so correlated that one event overwhelms the insurer.

Health insurance also contains behavioural and informational problems:

• Adverse selection occurs when people with higher expected costs are more likely to seek or retain coverage than lower-risk people, especially when premiums do not reflect available risk information.
• Moral hazard occurs when insurance changes behaviour after enrollment, such as increased utilization because the marginal price paid by the member is lower.
• Information asymmetry exists when applicants or providers know more about health needs than the insurer.
• Risk pooling combines members so that individual uncertainty is replaced by a more predictable aggregate outcome.

Underwriting cannot eliminate these problems. It seeks to measure and manage them while preserving the social and commercial purpose of insurance.

1.2 The underwriting cycle

A robust underwriting process normally includes:

1. Product and market definition: identify benefits, target population, rating rules, exclusions, limits, and regulatory constraints.
2. Information collection: obtain only information that is relevant, necessary, lawful, and proportionate.
3. Verification and clinical interpretation: assess the reliability, recency, and significance of medical and non-medical information.
4. Risk classification: place the applicant or group into a permitted risk category.
5. Decision and terms: accept, accept with modified terms where lawful, defer for information, or decline only where justified and permitted.
6. Pricing and capital assessment: estimate expected claims and uncertainty, including expenses, risk margin, and capital requirements.
7. Communication and review: provide understandable reasons, appeal routes, documentation, and monitoring for consistency.

The underwriter should separate clinical facts from judgments. For example, “diagnosed hypertension, controlled on one medicine, no complications” is evidence; “high risk” is an interpretation that must be justified in relation to the product and data.

2. Sources of underwriting information

Medical information

Medical underwriting may consider diagnoses, symptoms, treatment, medication, hospitalization, surgery, laboratory results, functional status, prognosis, and treatment adherence where lawfully obtained and relevant. Important dimensions include:

• Severity: current clinical impact and potential complications.
• Stability: controlled, improving, deteriorating, or recurrent condition.
• Recurrence: probability of repeated episodes.
• Treatment intensity: routine monitoring versus expensive specialist care.
• Comorbidity: interaction among conditions, not merely a count of diagnoses.
• Prognosis: expected course over the policy period.

A diagnosis alone is not a reliable cost estimate. Two people with the same diagnosis may have very different risks because of disease control, complications, access to care, and treatment pathways.

Demographic and geographic information

Age, family composition, location, and sometimes sex-related risk patterns may be relevant where legally permitted and actuarially supported. Demographic variables are ethically sensitive because they may act as proxies for protected characteristics or social disadvantage. Geographic variables can reflect provider prices, availability, epidemiology, and travel distance, but may also reproduce structural inequality. Classification should therefore be tested for necessity, proportionality, and disparate impact.

Occupational information

Occupation may indicate exposure to injury, infectious disease, physical strain, stress, or irregular working hours. It is more directly relevant to disability and occupational health products than to all forms of medical expense insurance. Job title alone is often inadequate; actual duties, protective equipment, working environment, and exposure frequency are more informative.

Lifestyle information

Smoking, alcohol use, physical activity, diet, and participation in hazardous activities may correlate with health costs. However, self-reported lifestyle data can be incomplete, culturally biased, or difficult to verify. Underwriting should avoid moral judgments and should not penalize legally protected disability or health status. Incentive programs should distinguish supportive prevention from coercive surveillance.

Financial and insurance information

Prior coverage, claims history, benefit utilization, requested limits, and continuity of insurance may assist underwriting. Claims history is not equivalent to underlying morbidity: utilization can reflect access, provider practice, benefit design, or a member’s prudent decision to seek care. It must therefore be interpreted cautiously.

3. Classification methods and decision frameworks

3.1 Individual medical underwriting

Individual underwriting evaluates the applicant’s expected risk relative to the product’s standard population. Typical outcomes are standard acceptance, acceptance with lawful exclusions or limits, a premium adjustment where permitted, postponement pending evidence, or declinature. Decisions should be based on documented underwriting rules rather than intuition.

A sound decision framework asks:

• Is the information relevant to the insured benefit?
• Is it sufficiently current and reliable?
• Is the expected additional risk material?
• Is the proposed term lawful, understandable, and proportionate?
• Are comparable applicants treated consistently?
• Would a less restrictive alternative manage the risk?

3.2 Community, modified community, and experience rating

In community rating, a broad population pays premiums based on pooled experience, subject to applicable rules. Modified community rating permits limited variation using specified factors. Experience rating uses the claims experience of a group, usually with credibility adjustments, to influence future premiums. Markets may combine these methods with risk adjustment or subsidies so that insurers do not have incentives to avoid high-need members.

3.3 Group underwriting

Group underwriting assesses the group rather than individually selecting every member. Relevant factors include group size, age and morbidity profile, industry, geographic spread, prior claims, contribution structure, enrollment rules, waiting periods, benefit design, and employee turnover. Large groups may produce more credible experience, but a claims-heavy year can still be random or caused by one unusual event.

Group risks include:

• Selection bias: voluntary enrollment may attract members expecting treatment.
• Group instability: mergers, layoffs, or rapid growth alter the risk pool.
• Credibility limitations: small groups have volatile results.
• Benefit-induced utilization: richer benefits may increase use.

A responsible group underwriter examines whether observed claims reflect true morbidity or changes in coding, provider prices, network access, benefit limits, or reporting practices.

4. Quantitative assessment of health risk

4.1 Frequency, severity, and expected claims

For a defined period, claim frequency is the number of claims divided by the number of exposure units, such as member-years. Average severity is total incurred claims divided by claim count. Expected claims per exposure unit can be represented as:

Here, E[C] is expected claims cost per exposure unit; λ is expected claim frequency per exposure unit; and μ is expected cost per claim. The relationship assumes that frequency and average severity are defined consistently and that their product is an appropriate approximation or exact decomposition for the portfolio being studied. Units must agree: claims/member-year multiplied by currency/claim equals currency/member-year.

For heterogeneous members, a more general expression is:

where w_i is the proportion or exposure weight of risk class i, p_i is its probability of a claim during the period, and s_i is its conditional mean cost if a claim occurs. The formula assumes classes are mutually exclusive and collectively exhaustive, and that the probabilities and severities are estimated for the same coverage, period, and currency basis.

Illustrative example. Suppose a portfolio has 10,000 member-years. It records 2,000 claims costing £4,000,000. Frequency is 2,000 ÷ 10,000 = 0.20 claims per member-year. Average severity is £4,000,000 ÷ 2,000 = £2,000 per claim. Expected claims cost is therefore 0.20 × £2,000 = £400 per member-year, which also equals £4,000,000 ÷ 10,000.

4.2 Premium, expenses, and risk margin

Claims cost is not the full premium. A simplified prospective premium relationship is:

where P is premium per exposure unit; E[C] is expected claims cost; E[X] is expected administrative, acquisition, and servicing expense; RM is a risk margin or allowance for uncertainty and capital; and M is any permitted margin or contribution to profit. This is a conceptual pricing identity, not a universal regulatory formula. Taxes, reinsurance, commissions, trend, investment income, and statutory levies may require additional treatment.

The loss ratio is:

It is meaningful only when the numerator and denominator use compatible periods, reserving bases, and accounting definitions. A high loss ratio may indicate underpricing, adverse selection, benefit expansion, medical inflation, poor utilization management, or inadequate reserves; it does not identify the cause by itself.

4.3 Portfolio risk and concentration

Individual underwriting focuses on a person or group; portfolio underwriting examines aggregate exposure. Important measures include age and morbidity mix, geographic concentration, provider concentration, high-cost claimant exposure, reinsurance attachment points, claim development, persistency, and scenario sensitivity.

A simple aggregate model is:

where S is aggregate claims cost, N is the number of claims, and X_j is the cost of claim j. The model is useful for explaining aggregation but does not by itself specify the distribution of N or X_j, dependence between claims, inflation, reserve uncertainty, or catastrophe events. Portfolio models must address correlation, especially for epidemics, provider failures, or regional disasters.

5. Pre-existing conditions

A pre-existing condition is generally a disease, symptom, diagnosis, or treatment present before a defined coverage date. Its legal meaning, disclosure requirements, waiting periods, and exclusions vary by jurisdiction and policy wording. The key underwriting challenge is distinguishing:

• a condition genuinely present before coverage;
• a newly diagnosed condition with earlier undocumented symptoms;
• a recurring condition whose treatment is continuous;
• an unrelated later event; and
• a condition whose apparent cost is actually driven by another cause.

Assessment should consider the date of onset, objective evidence, treatment history, current status, expected future cost, and the wording of the contract. Broad exclusions based on vague symptoms can be unfair and difficult to administer. Where exclusions or waiting periods are lawful, they should be specific, time-limited where appropriate, transparent, and reviewed against current evidence.

Underwriters should avoid treating a pre-existing condition as a permanent and uniform risk category. Effective treatment may substantially reduce risk, while an apparently minor condition may signal serious complications. Independent clinical review may be appropriate for complex or high-value cases, subject to consent, confidentiality, and auditability.

6. Fairness, privacy, and anti-discrimination

Health underwriting operates within a social contract: people seek protection from costs they may not be able to bear, while insurers require sustainable pooling. Fairness therefore has at least four dimensions:

1. Procedural fairness: consistent rules, competent review, reasons, and appeal.
2. Distributive fairness: reasonable allocation of costs and access across groups.
3. Epistemic fairness: accurate, relevant, and non-stereotyped evidence.
4. Outcome fairness: monitoring for unjustified disparities in acceptance, price, exclusions, service, and claims outcomes.

Legally protected characteristics vary, but commonly include disability, race or ethnicity, sex, age, religion, and genetic information. Even apparently neutral variables may be proxies. A model using postcode, occupation, or purchasing behaviour can reproduce historical exclusion. Compliance therefore requires more than removing explicit protected variables; it requires testing the effects of the whole decision process.

Privacy principles generally include purpose limitation, data minimization, accuracy, security, retention control, transparency, lawful processing, and rights of access or correction. Medical information should be separated from unnecessary commercial data, access should be role-based, and disclosures should be recorded. Applicants should understand what is collected, why it is needed, how long it is retained, and how decisions can be challenged.

A practical governance test is: necessity, proportionality, consistency, explainability, and contestability. If an item of information is not necessary for the covered risk, it should not be collected merely because technology makes collection possible.

7. Technology, predictive analytics, and artificial intelligence

Predictive models can support underwriting by identifying patterns in claims, estimating expected costs, detecting inconsistent information, and prioritizing manual review. Techniques may include generalized linear models, survival models, decision trees, gradient boosting, and neural networks. Their value depends on data quality and governance rather than sophistication alone.

Key risks include:

• historical bias embedded in training data;
• proxy discrimination;
• data drift as treatment and populations change;
• poor calibration for minority groups;
• automation bias by staff;
• opaque explanations;
• privacy leakage and cybersecurity threats; and
• feedback loops in which prior decisions shape future data.

A model should be validated for discrimination, calibration, stability, clinical plausibility, and business impact. Human oversight must be meaningful: staff need authority, time, and evidence to override a model. Automated decisions should have documented logic, version control, audit trails, adverse-impact monitoring, and an accessible appeal process. AI should generally assist—not replace—clinical and ethical judgment in complex cases.

8. Applied case: Meridian Health Plan

The following is an illustrative case, not a description of a real insurer.

Background

Meridian Health Plan offers employer-sponsored inpatient and outpatient cover. A 650-member technology company requests renewal after claims rise from £2.1 million to £3.0 million. Membership has also increased from 500 to 650, and the employer introduced richer outpatient benefits.

Problem

The insurer initially proposes a large premium increase based on the latest loss ratio. The employer argues that two cancer cases and a one-time neonatal intensive-care claim distorted the result. Employees are concerned that a proposed health questionnaire may disadvantage workers with disabilities.

Analysis

The underwriter first standardizes exposure and claims periods, separates incurred from paid claims, and identifies benefit changes. The latest claims amount per member-year is £3,000,000 ÷ 650 = approximately £4,615, compared with £2,100,000 ÷ 500 = £4,200. The apparent increase is therefore about £415 per member-year, not the full 43% increase in total claims. The team then examines age mix, new joiners, provider price inflation, outpatient utilization, large claims, and claim development.

The two cancer cases are genuine but may not recur at the same rate. The neonatal claim is high severity but not necessarily evidence of a permanent change in group morbidity. Because the group is moderately sized, its experience is only partially credible; broader portfolio data should supplement it. The richer benefits plausibly explain part of the outpatient increase.

Approach and solution

Meridian uses a credibility-adjusted projection rather than relying solely on the latest year. It proposes a transparent premium adjustment, a revised outpatient utilization-management program, improved maternity and oncology case management, and voluntary wellness support that does not require disclosure of diagnoses. It rejects a blanket medical questionnaire for all employees and limits any lawful information request to the contractually relevant enrollment circumstances. Claims and health data are separated, access is restricted, and the employer receives aggregate reports rather than identifiable medical details.

Outcome

The employer accepts a moderate renewal increase and benefit-management changes. Six months later, outpatient growth slows, while cancer and maternity claims remain monitored without adverse action against affected employees. An internal review finds that the original proposal would have overreacted to a small number of high-cost claims and risked unfair treatment.

Learning points

1. Total claims growth is not the same as per-member risk growth.
2. Large claims require severity and recurrence analysis, not automatic exclusion.
3. Benefit design and provider prices can alter claims independently of morbidity.
4. Small and medium groups require credibility adjustment.
5. Aggregate reporting can support management while protecting privacy.
6. Fair underwriting considers both sustainability and access.

9. Best-practice control framework

A professional underwriting function should maintain:

• written, evidence-based underwriting manuals;
• clinical and actuarial sign-off for material changes;
• documented authority limits and referral rules;
• standardized evidence requests;
• regular review of acceptance, pricing, exclusion, and appeal outcomes;
• privacy impact assessments and secure information controls;
• model validation, fairness testing, and change management;
• complaint and appeal analysis;
• training in medical interpretation, equality, privacy, and communication; and
• board-level oversight of portfolio concentration and emerging risks.

Underwriting quality should be measured not only by loss ratio. Useful indicators include quote-to-bind rates, retention, claim emergence, referral consistency, complaints, appeal reversals, turnaround time, model calibration, outcome disparities, and the financial effect of underwriting changes.

Challenges and limitations

Health risk assessment is constrained by incomplete records, coding variation, uncertain prognosis, changing medical technology, treatment inflation, evolving regulation, and the inability of historical data to capture emerging risks. Correlation is not causation: a variable may predict cost without being a legitimate basis for individual action. Excessive reliance on claims experience can penalize prudent care-seeking or areas with historically poor access. Conversely, ignoring material risk differences can create cross-subsidies, adverse selection, and unstable premiums.

The most defensible approach combines quantitative evidence with clinical review, transparent rules, proportionality, and continuous monitoring. It recognizes uncertainty rather than presenting estimates as facts.

Chapter Summary

Health underwriting evaluates expected medical risk and uncertainty while supporting sustainable pooling. It uses medical, demographic, occupational, lifestyle, claims, and group information, but each source must be relevant, reliable, lawful, and proportionate. Individual, community, modified community, experience, and group underwriting serve different market structures. Frequency, severity, expected-cost, loss-ratio, and aggregate-claims concepts help translate evidence into portfolio decisions, but formulas depend on consistent definitions and assumptions. Pre-existing conditions require careful analysis of timing, treatment, prognosis, and contract wording rather than automatic categorization. Fairness requires procedural consistency, non-discrimination, privacy protection, explainability, and appeal. Predictive analytics can improve consistency, yet it must be validated for bias, calibration, security, and drift. Effective underwriting balances solvency and access through evidence-based governance, clinical judgment, and continuous review.

Key Terms

Adverse selection: The tendency for higher-risk people to seek or retain insurance disproportionately.

Aggregate claims: The total cost of claims over a defined period.

Anti-discrimination: Legal and ethical controls preventing unjustified unequal treatment.

Credibility: The degree to which a group’s own experience is considered reliable for prediction.

Experience rating: Premium adjustment based partly on a group’s past claims experience.

Frequency: Number of claims per exposure unit.

Group underwriting: Assessment of a defined membership group rather than unrestricted individual selection.

Loss ratio: Incurred claims divided by earned premium.

Moral hazard: Change in behaviour caused by the presence of insurance.

Pre-existing condition: A condition present, or defined as present, before coverage begins.

Premium: Amount charged for insurance coverage.

Protected characteristic: A legally protected personal attribute, varying by jurisdiction.

Risk classification: Placing risks into categories with permitted underwriting or pricing treatment.

Risk pooling: Combining many exposures to reduce individual uncertainty.

Severity: Average cost per claim.

Underwriting: Evaluation and selection of risks, terms, and conditions of insurance.

Utilization management: Methods designed to support appropriate and cost-effective use of care.

Review Questions

5 Short Answer Questions

1. Define adverse selection and explain why it matters in health insurance.
2. Distinguish claim frequency from claim severity.
3. What is meant by a pre-existing condition?
4. State two privacy principles relevant to health underwriting.
5. Why is group credibility important?

5 Long Answer Questions

1. Explain the stages of the health underwriting cycle and the controls required at each stage.
2. Compare individual medical underwriting, community rating, and experience rating.
3. Discuss how medical, demographic, occupational, and lifestyle information should be evaluated.
4. Explain why fairness and accessibility may conflict with narrow risk classification, and propose a balanced approach.
5. Evaluate the benefits and risks of artificial intelligence in health underwriting.

3 Analytical / Application Questions

1. A portfolio records 3,600 claims costing £7.2 million across 12,000 member-years. Calculate frequency, average severity, and expected claims cost per member-year. Interpret the result.
2. An employer group’s claims rise sharply after richer outpatient benefits are introduced. Design an analysis to distinguish morbidity change from benefit-induced utilization and provider-price inflation.
3. An underwriting model produces lower acceptance rates for one ethnic community while using only postcode, income, and occupation. Identify possible mechanisms of indirect discrimination and recommend a governance response.

Multiple Choice Questions with Answers

Question 1

Which situation best illustrates adverse selection?

A. An insurer improves its claims payment system

B. People expecting high medical costs are more likely to purchase optional cover

C. A provider negotiates a lower hospital tariff

D. A member changes doctors after moving

Correct Answer: B. People expecting high medical costs are more likely to purchase optional cover

Question 2

A portfolio has 2,000 claims costing £4,000,000 across 10,000 member-years. What is the average severity?

A. £400 per claim

B. £2,000 per claim

C. £4,000 per claim

D. £20,000 per claim

Correct Answer: B. £2,000 per claim

Question 3

Which formula represents expected claims cost when frequency and severity are used?

A. Frequency ÷ severity

B. Frequency + severity

C. Frequency × severity

D. Severity − frequency

Correct Answer: C. Frequency × severity

Question 4

What is the principal purpose of credibility adjustment in group underwriting?

A. To ignore all group claims

B. To ensure small-group experience is not treated as fully reliable

C. To guarantee that every group receives the same premium

D. To replace clinical review with an algorithm

Correct Answer: B. To ensure small-group experience is not treated as fully reliable

Question 5

Which is the most appropriate interpretation of a diagnosis in underwriting?

A. The diagnosis alone determines future cost

B. The diagnosis should be considered with severity, control, complications, and prognosis

C. The diagnosis should always lead to declinature

D. The diagnosis is irrelevant if treatment has started

Correct Answer: B. The diagnosis should be considered with severity, control, complications, and prognosis

Question 6

Which practice best supports data minimization?

A. Collecting every available health variable

B. Retaining medical records indefinitely

C. Collecting only information necessary for the underwriting purpose

D. Sharing identifiable claims data with the employer

Correct Answer: C. Collecting only information necessary for the underwriting purpose

Question 7

A loss ratio is generally calculated as:

A. Earned premium ÷ incurred claims

B. Incurred claims ÷ earned premium

C. Claim count ÷ member count

D. Expenses ÷ claim count

Correct Answer: B. Incurred claims ÷ earned premium

Question 8

Why can claims history be misleading if interpreted without context?

A. Claims never contain financial information

B. Claims reflect only age and never benefit design

C. Utilization, provider prices, benefit changes, and access can affect claims

D. Claims are always random errors

Correct Answer: C. Utilization, provider prices, benefit changes, and access can affect claims

Question 9

Which is a key risk of automated underwriting models?

A. They cannot process any numerical data

B. They may reproduce historical bias through proxy variables

C. They eliminate the need for regulation

D. They always produce identical outcomes for all groups

Correct Answer: B. They may reproduce historical bias through proxy variables

Question 10

Which action best reflects proportionate underwriting?

A. Applying a blanket exclusion to everyone with any prior symptom

B. Requesting relevant evidence and considering current control and prognosis

C. Using social-media activity as the primary medical evidence

D. Declining all applicants with chronic conditions

Correct Answer: B. Requesting relevant evidence and considering current control and prognosis