Aaron Yelowitz · Teaching Applications Interactive teaching lesson · fictional data
← All teaching applications
Synthetic teaching exercise · fictional data

Adverse selection and underwriting in life insurance

Why might people holding individual coverage look no riskier — or even less risky — than nonholders, even if higher-risk people apply more often?

This exercise asks what a researcher can infer from coverage and later deaths, then reveals one possible mechanism in a fictional population. It is not a replication or estimate from either paper below.

The Yelowitz lens

Research: Harris and Yelowitz (2014) find no compelling adverse-selection evidence in a broad age cohort using life-insurance holdings linked to death records. Their study also examines new purchases and lapses; it does not observe applications or underwriting. Underwriting is one possible explanation, not a finding identified by that paper.

Harris, Yelowitz, Talbert, and Davis (2023) use university payroll and health-claims data and find worse-health employees are more likely to elect supplemental group coverage. Automatic basic employer coverage is a different margin.

Public communication: a February 2021 coauthored IZA commentary and June 2021 journalist coverage in The Wall Street Journal discuss COVID-era insurer offerings, a different question. Teaching focus here: the stage and sample you observe change the comparison you can make.

Main exercise · individual term

Where does the sample change?

Researcher view

The idea: application, underwriting, acceptance, and holding are different stages. A comparison among holders and nonholders does not expose the risk mix of applicants.

Loading the fictional lesson data…

Selection ladder in a fictional population

The exact numbers are in the tables immediately below.

Researcher-visible synthetic comparison: individual term holding and later death
Coverage statusPeopleDeaths in windowDeath share
Same comparison within fictional age bands (holders differ from nonholders in age composition)
Age bandHolders: deaths / peopleNonholders: deaths / people

Fictional outcome window: as defined in the data dictionary. The 2014 SIPP study used 1990–1991 panels linked to deaths through 1996, with at most about six years of follow-up from panel start. This table does not reproduce its observation window or regressions.

Interpret the difference

The synthetic choice rule uses latent risk as a shortcut for risk-related information or behavior available to the person; it does not mean people know their exact mortality type.

Ask whether higher-risk fictional people applied more often, where the composition changed, and what the researcher could have learned without seeing applications. The screen in this simulation is a candidate mechanism; the 2014 paper did not identify it as the cause of its finding.

These rows are designed to make the mechanism visible. Any small death count is unstable; do not use it for significance claims, real mortality rates, or policy forecasts. A positive coverage–death association would not automatically prove private information; a zero or negative association among holders would not prove applicants were low risk.

Inspect the synthetic data

Version and seed loading…

The observed extracts intentionally omit hidden risk, private news, applications, and underwriting fields not established in the source studies. In particular, the group researcher-style CSV includes issued supplemental coverage but omits attempted requests and EOI pass/fail outcomes. The oracle file is explicitly fictional and retains those generated fields for model inspection. The in-page reveal is an instructional sequence, not a security boundary: the full fictional truth is also available in its oracle download. The 50-row audit file is a worked checksum, not a statistically powered sample.

Published evidence and related work

Privacy details

Privacy and analytics

You can read this website and use its publications, teaching materials, recordings, and contact links without accepting analytics.

Your choice

Optional Google Analytics 4 helps Aaron Yelowitz understand how readers use the website. It starts only after you select Allow analytics. Keep analytics off is equally available. On a fresh visit, no Google analytics requests are made before a choice or when analytics is declined.

Use Privacy & analytics in the footer to change your choice. Withdrawing consent stops new measurement, removes accessible Google Analytics cookies, and reloads the page to stop the loaded analytics code. Events already queued or in flight from a previously consented visit may finish during that transition. Consent or refusal is remembered in this browser for 180 days. Clearing browser storage also clears that choice. A Global Privacy Control signal keeps optional analytics off.

The same settings include Exclude this browser for Aaron's own visits and anyone else who wants a persistent local exclusion. That exclusion has no automatic expiry and blocks this site's Google analytics requests. It applies only to this browser and device; it does not exclude another browser or device. Re-enabling analytics after exclusion requires a fresh consent choice.

What is measured after consent

Analytics may record the public page viewed, its publication or course identifier, broad browser/device information, and approximate geographic region derived by Google. Events describe use of HTML reading links, public file and citation downloads, successful citation copying, outbound links, email links, and native recording playback milestones. Search events include only query length, result count, and whether the search index was complete; the words you search for are not sent.

Page URLs sent by this integration exclude query strings and fragments, including campaign tags. Referrals are reduced to the referring site's origin. Unknown or missing-page URLs are reported as the generic error page. Link events use a generated registry of public resources and identifiers, rather than arbitrary link text or query values. The integration does not send email-address text, entered form content, full search queries, or private source files.

A link click does not prove a completed download or an actual professional inquiry. Playback events describe browser playback, not a verified listener. Visits without analytics consent are not counted, so these reports do not represent every reader.

Google and browser storage

Google Analytics uses cookies after consent to distinguish browsers and measure visits. This integration sets a 180-day cookie lifetime without extending it on each visit. The data is not described as anonymous: Google receives network information, such as an IP address, when the browser makes an analytics request, and can derive approximate region and device information. Advertising features, Google Signals, user-provided data, and advertising personalization are not enabled by this integration.

The consent and local-exclusion choices use this browser's local storage. Google provides more information in its privacy policy (https://policies.google.com/privacy) and explanation of how partner sites use its services (https://policies.google.com/technologies/partner-sites).

Website delivery and external links

Cloudflare serves the website and public files. Delivering pages involves technical network information needed for connectivity, security, and service operation, independently of optional audience analytics. This notice does not promise that infrastructure providers have no technical logs.

External publishers, media providers, institutions, and profile sites apply their own privacy practices when you follow their links. Email links open your email application; they do not submit an inquiry through this website.

For questions about this website, use the existing Contact page (https://yelowitz.com/contact/).

Recommended citation

Yelowitz, Aaron. 2026. Adverse Selection and Underwriting in Life Insurance (life-selection-synthetic-v3.1). Interactive teaching lesson. https://yelowitz.com/teaching/applications/adverse-selection-life-insurance/