Principles of alternative data for underwriting

This part will lay out what needs to be taken into account from a pricing point of view when considering such a strategy. A second article (Part 2: Getting practical: using alternative data for underwriting) will focus on the real example of physical activity data in the context of replacement or augmentation of traditional underwriting, and also on the known regulatory challenges that need to be taken into account.

Underwriting has always been “data-driven”. From the first applications asking the age of the applicant, to health questions asked for over 100 years, to modern lab and imaging testing required for today’s higher sum applicants, data has been the foundation of risk assessment. However, hopes of ubiquitous wearable and smartphone data, measuring individual lifestyle information and other digital biomarkers, are some of the aspects of new data that life and health insurance underwriting is considering using to risk assess applicants. How to effectively implement this remains a challenge to the industry. This newer and innovatively sourced data is often termed “alternative” data, which may imply replacement of one or more traditionally used risk factors. Electronic health records, prescription data, and proxy information like banking data are also being considered to enhance and individualise the insurance customer’s journey.

When contemplating any new risk data source like physical activity or steps, and how to use it effectively without impacting risk selection, a careful consideration is required as to what information will be forfeited, and how the new data source can or cannot determine the risks associated with that forfeited information. Simply considering the predictive value of a new data source is not considered to be sufficient.

Understanding the current information gathered through questionnaires and underwriting requirements is a key first step.

As opposed to simplified issue policies, fully underwritten products gather more information for the purposes of risk assessment, and take longer to finalise. It is in this fully underwritten space where expedience and a more frictionless journey is most desired, and how to use newer digital data, typically through some automated process, and foregoing some of the more traditional data sources (questions, labs, insurance medicals, or medical reports) is becoming ever more important.

Reducing the number of questions or information can lead to a faster issue of a policy, but at a price. Any of the above sections that are dropped lead to worse experience of portfolio – firstly as higher risk applicants are not identified and rated sufficiently (and hence the whole portfolio’s base price needs to increase), and secondly as those with higher risks may prefer to take out these policies as they know they will not be asked about their higher risk avocation, travel patterns, occupation, or health history. This anti-selection pull increases the proportion of poor risks in the applicant pool compared to a fully underwritten product.

Where risk selection may be impacted, an assessment needs to be made based on the prevalence of the risks related to the missed information, the mortality/morbidity impact of those risks, and the possibility of behavioral anti-selection of applicants. The pricing differentiation required when shifting from fully underwritten to simplified or accelerated products underlines the value of having more information to use when risk assessing.

This impact will usually result in an increase to mortality rates. These increases could typically range from a 30% to 70% in non-preferred markets. In preferred markets, it is not unusual to see a doubling of mortality rates or a 100% increase, although it could be even higher. This increase will differ significantly based upon competitiveness of the market, disclosure rates, type of product, age and demographic profile, distribution channel, as well as the difference in the actual questions/tests asked.

Underwriting questionnaires ask information about various aspects of individual risk. Ignoring financial information, the questions can broadly be divided into non health questions which involve travel, occupation and avocation, and health questions which ask about current and past medical information, including a question relating to family medical history. This latter health section also includes current clinical cardiovascular risk information like height and weight, and questions on hypertension or blood pressure, high lipids, diabetes etc. Depending on the market, availability of medical records and the sum being applied for, additional insurance medical exams or lab testing could also be performed, giving more objective values, especially the clinical cardiovascular risk factors (i.e. current health).

1 for 1 risk factor replacement: This is where one risk factor, test or data type is used as a 1-for-1 replacement for another. Using NT proBNP instead of resting ECG is a good example. Are the risks that outgoing and incoming tests may pick up the same, or similar? If they are not, is the protective value of the two tests similar? In other words, the number of declined or rated cases will be similar for each test. It follows that doing both tests in such a scenario (where there is little overlap) would be more protective than doing either one, but the market is typically reluctant to add additional testing, preferring to reduce testing or replacing onerous ones with ones that are more consumer friendly.

One newer (alternative) health risk factor used in place of many health risk factors: There is a perception that the correlation between certain health factors is very strong, particularly cardiovascular risk factors. This leads to a view that using one cardiovascular factor (or proxy) can reasonably assume the others are similarly good/average/poor. This belief regarding the correlations of risk factors or behaviours is not supported by the data. While less active obese applicants are more likely to have poor cardiovascular risk factors, there will be many less active obese applicants with reasonable cardiovascular risk markers, and similarly many active normal BMI applicants with rate-able cardiovascular risk factors like high blood pressure or poor lipids.

The data below from NHANES¹ considers BMI and the recording of “serious conditions”. What can be seen is how more than half of those with a serious conditions have a BMI under 30.

If the correlation between BMI and serious disease (which does exist but is not all that robust) was relied on too heavily (e.g., using just BMI and foregoing questions about serious diseases), 50% of serious conditions would be missed if using obese BMI only.

The lack of strong correlation, and the impact that missing the protective value of dropped (weakly) correlated risk factors could lead to missing of significant risk, as well as to anti-selective drift where those with poor “replaced” risk factors preferentially apply as they have “okay” values of the “new” risk factor. In the example above applicants with a good BMI but a history of serious medical conditions could disproportionally be drawn to that product.

One newer (alternative) risk factor where non-medical questions are forfeited: While simplified products already skip many of these questions related to avocation, travel and occupation, the sums assured are limited, and the base price is adjusted for accepting these risks. Market wide practice also reduces company-specific or productspecific anti-selection.

If, in a fully underwritten market, these typical non-medical questions are not asked due to an alternative data source being used as a proxy, careful consideration must be made as to the prevalence of these risks that may be missed (and their associated mortality/morbidity risk), and more importantly the risk of anti-selection. For example, while someone who exercises regularly, has a healthy diet, and goes for regular physical/ medical checkups can be thought of as being very risk aware or risk averse, it may be unwise to assume they are unlikely to travel to dangerous destinations for work or leisure regularly, or unlikely to partake in hazardous avocations like sky diving. Ideally these non-medical questions should remain.

The graphics below indicates the smaller calculable risk of missing, for example hazardous travel and avocation risk, but then highlights the more significant and unknown risk of anti-selection in a new product that forfeits these non-medical questions. Anti-selection can include increased prevalence of a hazardous risk in a portfolio, and also the risk of increased sums being taken out when the risk question is not asked.

The graph below assumes simple additional premium of 5 per mille on average for an average portfolio, and disclosure of 1 in 2000 high risk applicants.

On the graphic below are shown scenarios with increasing anti-selection leading to 2x, 4x and 10x the number of high-risk applicants. The final bar shows premium impact when with increased sum assured.

A better use case in this category may be knowing via some data source if someone has a valid passport, and assuming they do not, that they currently do not participate in hazardous travel (and hence do not need to be asked); or using credit card or smart phone data sources and determining from that only travel for safer destinations has been purchased, and that no hazardous sports or activities have been paid for or recorded.

We have created a framework that one could consider using when assessing the value of new alternative data sources. This framework will help you address the different ways in which both underwriting and pricing are impacted.

Context: Who will you target?

Prevalence of rate-able conditions are very different by age. Younger applicants are less likely to have current medical conditions, or less significant family history, and not knowing those may have far less portfolio impact compared to not knowing them in a middle-aged applicant group. Older applicants may have outlived their family history risk, and travel and even risky avocations may imply high levels of health and resilience that could even counterbalance the increased risks these bring. Knowing your context, who the target market will be, and what their health or behaviour profile is, can help better understand and thereby price the portfolio more accurately which results in a better mitigation of the risks when considering alternative data.

Correlations: What are the impacts of the risks that are you missing?

Combining target market knowledge above, with known prevalence of risk – based on previously underwritten applicants or research in a market – will help understand what risks may be missed, particularly where the replacement is not 1-for-1 i.e. 1-formany replacement, or dropping medical or non-medical questions), or the protective value is less than the risk factor being replaced. Being able to monitor those risks closely, as well as assess the correlations of the new risk factor to any missed risk factors is important when reducing the risk factor information collected and used.

Credibility: The flaw of relying solely on “Predictive” studies

Predictive studies often compare the predictive value of one risk factor, or multiple risk factors, to another. While this is useful for considering the overlap of protective value, it is important to realise that there is often no consideration of the non-medical avocation, travel, or occupation risks in these studies. It is simply the comparison of a new risk factor compared to traditional clinical ones. It is important to assess the overall credibility of the risk factors for underwriting, which includes the additional value from the new risk factor as well as the impacts of any missing risk factors. Using NT-proBNP as an alternative to resting ECG because it’s as predictive did not mean that medical or non-medical questions did not need to be asked.

Accuracy: Accuracy of matching an underwriting decision

Often “confusing” information is a marketing message for certain new models using alternative data – i.e., that a “new” data’s underwriting decision model has a high accuracy (e.g., 99%) matching traditional underwriting outcomes. This needs to be understood in the context of an underwritten portfolio. Depending on age and sex, a single additional death per 1000 or even per 2000 in a year can double the expected claims and lead to significant losses. Matching 99% of traditional underwriting is not sufficient if e.g., 1 in 20 of the 1% that are missed by the new underwriting method or model (i.e., 1 per 2000) have a high mortality risk that would lead to an additional short-term death.

Anti-selection: How will you be anti-selected against?

Anti-selection will always exist, but it increases in a market where one insurer is considering omitting a risk factor question or test which is generally included in the market. The knowledge that for one risk factor even bad values are not looked at or assessed, means that an anti-selective drift of those with worse risks occurs at that one insurer, leading to worsening experience. It is not dissimilar to an aggregate smoker product with aggregate pricing in a market with mostly smoker/non-smoker rates. Smokers would preferentially choose the aggregate (cheaper price) than pay smoker rates, leading to a progressively higher and higher proportion of smokers in the aggregate pool, and eventually leading nonsmokers to abandon the product, further worsening the experience.

In part 2 we will make things more practical, focusing on physical activity data and how this stacks up against the considerations raised in this paper. What could be feasible, what to watch out for, how to mitigate risk, and navigate the regulatory landscape will be discussed. Most importantly we will focus on how to embrace this exciting new world of alternative data through experimenting and learning.

¹ NHANES - National Health and Nutrition Examination Survey Homepage (cdc.gov)