Risk Factors for Ill Health Insurance Claims - The American Academy ...

JOURNAL OF INSURANCE MEDICINE Copyright Q 2003 Journal of Insurance Medicine J Insur Med 2003;35:17–25

ORIGINAL RESEARCH

Risk Factors for Ill Health Insurance Claims William T. Hamilton, FRCP, FRCGP; G. Harry Hall, MD, FRCP Objectives.—This study examined the information available at application for income protection insurance, to determine if any factors were predictive of a claim. The strength and significance of such factors were assessed and a predictive model was developed. Background.—The factors underlying life assurance risks are well known, but this is not the case for income protection insurance. For accurate underwriting of income protection insurance, it is important to know what information available at application has power to predict a claim. Improving the scientific accuracy of underwriting is good business practice, as well as answering the demands of disability legislation. Methods.—We studied all data available at application for 959 current claimants and 1417 non-claimants, using a case-control study design. Information included applicants’ description of their occupation, marital status, build and habits, plus a questionnaire asking about their personal health. For some applicants medical reports were available as well. Information was transcribed onto a database, and univariate and multivariate analyses were performed. A predictive scoring system was established and its performance measured by receiver operating characteristic curves. Results.—Significant associations with claiming were found for many variables, including age (odds-ratio 1.04, p , 0.001), height (0.11, p 5 0.03), smoking (2.10, p , 0.001), abstinence from alcohol (1.56, p 5 0.01), recent medical advice (1.34, p 5 0.06), and having had a lower gastrointestinal disorder (1.51, p 5 0.04). Using all the information from the application, a predictive model was constructed. This model had good predictive power with an area under the receiver operating characteristic curve of 72%. Conclusions.—Classical underwriting factors were generally shown to have predictive power for income protection insurance. The predictive scoring strengthens the scientific basis for underwriting and could be developed to simplify and expedite the underwriting process.

M

Address: GH Hall, 87, Polsloe Road, Exeter EXI 2HW, UK; e-mail: [email protected]; WT Hamilton, Division of Primary Care, University of Bristol, Cotham House, Bristol BS6 6JL, UK; e-mail: w.t.hamilton@ btopenworld.com. Correspondents: GH Hall, MD; WT Hamilton, FRCP. Key words: Insurance, disability; income protection; insurance claim review; risk factors; case-control studies.

insurance—which pays benefits if a policyholder is unable to work as a result of illness—there is much less evidence about which factors are important. Among the reasons for this is that the same condition may disable a manual worker, but have little or no effect on an office worker.

any variables have been identified which affect lifespan, such as smoking1 and obesity.2 The identification of these factors and the quantification of their effects have played an important part in the development of accurate underwriting of life insurance. However, for income protection (IP) 17

JOURNAL OF INSURANCE MEDICINE

Secondly, individuals vary greatly in their ability to manage with a disability, with the apparently same condition disabling one person but not another. These considerations reduce the value of classical risk factors, as these predict disease rather than functioning. Moreover, few research reports in the medical literature use the ability to work as an outcome measure. Although it is clear that some factors have an effect on morbidity as well as mortality (diabetes for example), the level of research evidence supporting this is much weaker. Accurate underwriting of income protection insurance requires studies designed with inability to work as an endpoint. Insurers need to be fair as well as accurate. This fairness means that each applicant is assessed fully and offered a premium rate that truly represents his or her risk to the pool of insured people. This has always been good business practice. In recent years most countries have brought in legislation covering this area, such as the Americans with Disabilities Act (United States) and the Disability Discrimination Act (United Kingdom). Broadly speaking, these laws require insurers to give services to disabled people on terms equivalent to able people. Underwriting has not been forbidden, but any adverse terms an insurance company offers must be justified. In practice, this means support by actuarial or medical evidence. Experienced underwriters can develop considerable skill in assessing income protection applications. In reality, this is an art not a science. With this skill, most underwriting decisions are ‘‘correct,’’ but it can be difficult to prove that the decision is fair because of the shortage of research literature. Proof requires some scientific backing. We investigated 3 main issues. First, can any information available at the time of application be linked with the likelihood of claiming? Second, if this is possible, what are the relevant characteristics taken singly or together that determine this? Third, can a measure of ‘‘claiming tendency’’ be derived that can usefully guide an insurance rating? This would provide scientific support for premi-

um rates and help to rebut allegations of unfair discrimination. METHODS We performed a case-control study, examining data available at application to the insurance scheme for a random selection of claimants (cases) and non-claimants (controls). The study population was drawn from the 55,838 policyholders with an IP policy active in the year 2000 from the Permanent Insurance company (now Liverpool Victoria), based in Exeter, Devon, United Kingdom. Policies from 1963 to 2000 to individuals aged between 18 and 65 years were available for study. This included those who had been given substandard terms. IP policies offer sub-standard terms in 2 separate ways. For applicants with an impairment potentially leading to one of several illnesses (such as hypertension, which increases the risk of stroke, myocardial infarction and renal failure), an extra premium or loading is applied. Other applicants declare illnesses that may recur (such as backache); these may be dealt with by an exclusion clause, excluding benefits for the specified condition. Cases were policyholders claiming benefit for any period during the year 2000 (n 5 1246). This included new claims beginning in 2000, as well as claims running into that year from an earlier start. Controls were any policyholder who had not claimed in 2000, even if they had done so previously. From these 2 populations, 959 cases and 1417 controls were selected using a computerized random numbers table. The size of the study was constrained by UK legislation, in that research on pre-existing records had to be completed by October 2001 under the terms of the Data Protection Act. The samples chosen for study were representative of the whole insured population in terms of age and sex (data not shown). The information available for all cases and controls included: 1. Policy details, including name, age, sex of 18

HAMILTON ET AL—ILL HEALTH CLAIMS

the applicant, date of the inception of the policy, terms of the contract, and claims details. All these were available electronically and could be transferred directly to our database. Names were replaced by study numbers. 2. The application form, which included a short medical questionnaire asking about habits such as smoking and alcohol intake and previous illnesses. 3. Claims data, including dates, conditions claimed for, and the number of previous claims.

cused enough for classification yet broad enough to include significant numbers. After initial data collection, all tables were searched for missing or degraded data, and errors corrected by reference to the original files. STATISTICAL ANALYSIS The primary endpoint of the study was to determine whether, at the time of application for a policy, there were any differences in the characteristics of the cases and the controls. Means (and t-tests) were used when the data was normally distributed and medians (and Mann-Whitney U test) when the distribution was skewed. Pearson chi-square test was used for proportions. Odds ratios (OR) were calculated by univariate and multivariate logistic regression. Two further multivariate techniques were utilized: principal components analysis and recursive partitioning. Stata Statistical Software, release 7.0 (College Station, Tex: Stata Corporation) was used for the analyses. Items with less than 25 mentions in total were not analyzed. A scoring system for selected studied variables was derived from the multivariate results. The ability of these scores to predict case or control status was assessed by constructing receiver operating characteristic (ROC) curves.

Some applicants had further medical information available from the time of application, such as a report from their usual doctor or the findings from a medical examination requested at application. Both of these types of reports are requested for applicants applying for either larger sums of insurance cover or revealing adverse medical history on their application form. Batches of 50 were collated for coding with each batch containing a random mixture of cases and controls. A clerk extracted the microfilms containing the application details. By this means, the coder was blinded to casecontrol status when she was coding. A database was designed with entry forms for each of the information sources described above. This was based on our previous work in this field.3,4 Some of the questions on the application form were of multiple-choice type. Answers to these were potentially ambiguous. Ambiguities were resolved by using the applicant’s expanded answer in the open box on the form. Occasionally, applicants had entered information in the open box alone, which was also used. Conditions described in response to the specific questions, or in the open box—or more usually both—were grouped together for analysis. Minor changes had occurred in the design of application forms over the 37 years, but it was possible to accommodate the differences in 1 classification. We derived 21 disease groups for analysis from our previous work.3,4 These were fo-

RESULTS There were significant differences between cases and controls in their age, sex and build. These are summarized in Table 1. Marital status was also requested on some application forms, with answers available for 1397 (58.8%) applications. Single people at application were least likely to claim, with separated being next (OR for separated vs. single 1.17 [95% confidence intervals 0.65–2.08]), and married having the highest claim rate (OR for married vs. single 1.38 [1.06–1.79]). However, single people were significantly younger than married or separated people (p , 0.001 ANOVA), and in the multivariate analyses marriage status lost its significance. We analyzed several aspects of the insur19


Table 1. Demographic Variables for Cases and Controls

Variable Age Sex Build

At time of study: Median (IQR*) At application: Median (IQR*) Female: Number (Percent) Height (m): Median (IQR*) Weight (kg): Median (IQR*)

Claimant (cases) N 5 959 53.5 37.6 198 1.75 73.0

(47.8–57.9) (32.0–42.5) (20.7%) (1.70–1.80) (66.7–81.6)

Policyholders (controls) N 5 1,417 46.0 34.2 217 1.78 74.8

(39.1–52.5) (29.5–39.1) (15.5%) (1.73–1.83) (67.4–82.5)

Significance p p p p p

, , , , 5

0.0001** 0.0001** 0.001† 0.0001** 0.03**

* IQR Interquartile Range; ** Mann-Whitney; † Pearson chi-square.

ance decision. This was to check for inconsistencies between our results and previous findings. Such inconsistencies (there were none) would have cast doubt on the representativeness of our sample and reduced the generalization of our results. There was the expected association between deferment period and likelihood of claim. The deferment period is the length of time between the onset of incapacity and the first payment of benefit. A higher percentage of claimants had deferment periods of 4 weeks or less (38%) than policyholders as a whole (21%). However, short deferment policies should (all things being equal) lead to more claims, because short illnesses may trigger a claim on these policies but not on a longer deferment policy. The mean duration of claims was significantly longer as the deferment period lengthened (4.93 months for 4 weeks, 4.86 for 13 weeks, 5.54 for 26 weeks, 5.81 for 52 weeks: p , 0.01, ANOVA). Workers in manual jobs have more periods of incapacity when compared with professionals. This is allowed for by charging them more for each unit of insurance. The company operated 4 rate classes, ranging from rate class 1 (professional) to rate class 4 (manual). As expected, there were more claims in rate classes 2 through 4 when compared with rate class 1 [for rate class 2; OR 5 1.69 (CI 1.35– 2.12); for rate class 3, OR 5 3.42 (CI 2.52– 4.65); for rate class 4, OR 5 3.14 (CI 2.22– 4.44)]. We analyzed policies that were given sub-

standard terms. Policies with a loading made more claims [OR 1.88 (CI 1.40–2.54)]. However, for policies with an exclusion, the OR was 1.08 (CI 0.83–1.38). Exclusions were correlated with 3 conditions on application: backache (correlation coefficient 0.27, p , 0.001), joint disorders (0.23, p , 0.001) and psychological or psychiatric disorders (0.17, p , 0.001). The number of affirmative answers to the questionnaire on the form was significantly higher in cases than controls (p 5 0.05, t-test with unequal variances). The mean number for cases was 1.73 [CI 1.63–1.82], and for controls 1.63 [1.55–1.70]. The mean number of previous illnesses declared on the form by claimants was 1.63 [CI 1.54–1.72], and for other policyholders 1.57 [1.50–1.65]. This difference was not significant (p 5 0.17, t-test with unequal variances). The results of the univariate logistic regression are shown in Table 2. Categories with less than 25 mentions in total were: specialist investigation awaited (4 mentions), specialist advice planned (7), chronic fatigue and its synonyms (24), cancer (3), blood or lymphatic disorders (21), and small bowel disease (13). All variables from Tables 1 and 2 with a p value below 0.2, along with marital status, and the number of past illnesses and positive answers on the application form were entered into a multiple logistic regression model. The results are shown in Table 3. A ROC curve was derived from the logistic 20


Table 2. Univariate Analysis of Illness Factors

Odds Ratio

95% Confidence Interval

Significance (p value)

2.13 0.56

1.75–2.60 0.45–0.70

0.0000 0.0000

Family history of inheritable disease Current disease Current treatment Previous specialist advice or treatment Medical advice in the past 5 years

General Illness History 140 153 52 43 85 99 350 591 520 701

1.41 1.83 1.29 0.80 1.21

1.10–1.82 1.19–2.83 0.95–1.77 0.68–0.95 1.03–1.43

0.006 0.004 0.09 0.01 0.02

Lower gastrointestinal disorders Liver or gallbladder disease Upper gastrointestinal disorders Gynecologic disorders* Psychological or psychiatric disorders Muscular or connective tissue disorder Joint disorders (excluding backache) Backache Endocrine disorders Disorders of hearing or vision Cardiovascular disorders Renal disorders Glandular fever Respiratory disorders Genitourinary disorders Skin disorders Disorders of bone

Specific Illness History 125 136 23 25 74 88 55 64 83 100 32 38 214 269 117 162 14 18 180 249 38 56 15 21 14 22 273 432 36 59 143 261 100 194

1.41 1.37 1.26 0.92 1.25 1.25 1.23 1.08 1.15 1.08 1.00 1.06 0.94 0.91 0.90 0.78 0.73

1.09–1.83 0.78–2.41 0.91–1.69 0.59–1.44 0.92–1.69 0.78–2.01 1.00–1.50 0.84–1.39 0.58–2.30 0.88–1.34 0.66–1.52 0.55–2.04 0.48–1.82 0.76–1.09 0.59–1.37 0.62–0.97 0.57–0.95

0.09 0.28 0.15 0.70 0.16 0.35 0.05 0.57 0.69 0.46 0.98 0.87 0.86 0.29 0.62 0.03 0.02

Variable

Mentions in Cases (N 5 959)

Mentions in Controls (N 5 1,417)

Habits Current smoking Current alcohol use

278 763

228 1,238

*Analyzed for females only (cases 5 198, controls 5 217)

regression model and is shown in Figure 1. The area under the ROC curve was 66%. The recursive partitioning tree is indicated in Figure 2. Those terminal nodes with the biggest difference in case/control incidence can be identified easily. Principal components (factor) analysis revealed predictors corresponding to the variables with the largest coefficients in the multiple logistic regression. From the principal components analysis, a scoring system was derived to predict claims from the application data. One subgroup analysis was performed by examining the 756 claims lasting over 1 year. As this subgroup formed the majority of the

dataset, differences between it and the whole dataset are minor, but in all variables the OR moved further away from 1.0. Data from the doctor’s report or medical examination were added to the model where available (n 5 483), and a backward stepwise logistic regression performed, with only items having a p value under 0.1 retained. The additional variables in this model with strong predictive power were: units of alcohol consumed per week (OR 1.01, p 5 0.07), number of cigarettes per day (OR 1.04, p 5 0.009), the number of conditions listed by the doctor (OR 1.41, p 5 0.001) and the diastolic blood pressure (OR 1.02, p 5 0.06). The ROC curve from 21


Table 3. Multiple Logistic Regression Analysis of Illness Risk Factors 95% Confidence Interval

Significance (p value)

Age at entry (years) Male sex Single2

Age, Sex and Marriage 1.04 0.01 0.79 0.16 1.09 0.14

1.02–1.05 0.54–1.17 0.85–1.39

,0.001 0.2 0.5

Height (meters)3 Weight (kilograms)4

Height and Weight 0.11 0.11 1.00 0.01

0.01–0.82 0.99–1.01

0.03 0.9

1.58–2.77 0.45–0.90

,0.001 0.01

Variable 1

Standard Error

Odds Ratio

Habits Current smoking Current alcohol use

2.10 0.64

0.30 0.11

Family history of inheritable disease Current disease Current treatment Previous specialist advice or treatment Medical advice in the past 5 years

General Illness History 1.28 0.23 1.72 0.57 1.09 0.29 0.84 0.14 1.34 0.21

0.90–1.83 0.90–3.28 0.64–1.83 0.61–1.15 0.99–1.81

0.2 0.1 0.8 0.3 0.06

Lower gastrointestinal disorders Upper gastrointestinal disorders Psychological or psychiatric disorders Joint disorders (excluding backache) Skin disorders Disorders of bone

Specific Illness History 1.51 0.29 1.13 0.28 0.89 0.21 1.19 0.19 0.93 0.16 0.74 0.16

1.03–2.21 0.69–1.84 0.56–1.41 0.87–1.63 0.67–1.30 0.48–1.13

0.04 0.6 0.6 0.3 0.7 0.2

All variables are categorical (‘‘yes/no’’) except the following: 1 For every year increase in age. 2 Single, divorced or widowed vs. married. 3 For every meter increase in height. 4 For every kilogram increase in weight.

this data had an area below the line of 72%, showing that the information from these sources has additional predictive value. DISCUSSION Our findings demonstrated clear differences between claimants and policyholders in many of the variables in the recorded application data, both by univariate and multivariate analysis. Any study relying on random sampling for its participants runs a risk of being unrepresentative of the population as a whole. However, our sample was large, and the insurance related variables tested gave results in keep-

Figure 1. Receiver Operating Characteristic Curve From Multiple Logistic Regression of Data From the Application Form 22


Figure 2.

Recursive Partitioning Tree

ing with other studies. Conversely, one advantage of this study design is that it relies on information collected at application. This avoids recall bias,5 which occurs when participants in a study are asked about pertinent factors which have occurred in the past. Also, we initially included all information available at application, reducing the risk of investigator selection bias. The truthfulness of the answers given by applicants was not tested. Answers to questions on the application form were sometimes at variance with answers on the 2 types of medical reports. These discrepancies were not more common in cases than controls. Completeness of the data was excellent, with the exception of marital status. The analyses of the data, in particular the 3 different multivariate analyses, all produced very similar results. Therefore, we are confident that the findings are robust. The variables showing the largest differences occurred in what are regarded as classical predictive factors1 (namely, age, sex,

height and occupation [as judged by rate class]) and smoking status, together with the interesting negative correlation of claims with alcohol consumption. Alcohol consumption shows a J-shaped curve when graphed against mortality, with higher mortality rates in non-drinkers and heavy drinkers. The company may not have granted very heavy drinkers a policy, so our studied population probably contains mostly moderate drinkers and non-drinkers. There was a strong association between non-drinking and claiming, but an association between increased quantity of alcohol consumed and claiming, as well. This suggests that there is a J-curve for income protection claims, too. Body mass index showed no association with claiming. This contrasts with mortality studies, where increased body mass indices are linked with higher death rates.2,6 It is possible that the ill effects of obesity occur mostly after the policies cease. In contrast, ‘‘medical’’ factors in the past history showed smaller differences, though 23


still statistically significant. Of these, having had treatment in the previous 5 years or reporting a lower gastrointestinal tract disorder showed the biggest correlation with claiming. Claimants also reported more illnesses on the application form. It is no surprise that having a recent illness at the time of application increases the chance of claiming as many illnesses recur. One possible explanation for the strong association between lower gastrointestinal disorders and claiming is that many of these were mentions of irritable bowel syndrome. Opinions about the etiology of irritable bowel syndrome vary with some authorities including it in the group of functional somatic syndromes.7 If so, it is not unexpected that irritable bowel syndrome should be a marker of future ill health. Surprisingly, neither backache nor mental disorders proved significant predictors. This may well be accounted for by the use of exclusion clauses in the policies. Policies with exclusion clauses had a similar claim rate to policies at standard terms. This is an important result, suggesting that exclusion clauses are largely successful in eliminating the extra risk posed by the medical problem declared at application. This is contrary to received insurance wisdom that policies with an exclusion clause still have more claims than those at standard terms. However, policies with an exclusion do not have less claims. This is relevant, because some applicants argue when offered a policy with an exclusion that they should get preferential rates. Their superficially plausible argument is that because they cannot claim for their excluded condition their overall chance of claiming is less. Our results show this reasoning to be wrong. What usefulness might the effect of these observations have on underwriting procedures? Firstly, our results underpin most current underwriting and will be of value in showing that underwriting decisions based on the factors we have found to be significant are indeed fair. However, alcohol usage and obesity are not as adverse predictors as previously thought. Indeed, alcohol in modera-

tion protects from ill health, as many epidemiological studies suggest. The results also show the added value of obtaining additional medical information from a doctor’s report or examination. These gave amounts of tobacco and alcohol use that may be more reliable than any figure put on the application form. Additionally, the diastolic blood pressure was predictive of claims. This accords with many epidemiological studies.8,9 The number in this study with endocrine, genitourinary, cardiovascular, or muscular disorders was too small to allow meaningful analysis, which is only to be expected in a young population. Clearly, an applicant mentioning one of these conditions, such as diabetes, would usually lead to an insurer requesting a medical report. Our results do not change this. Our multivariate analyses established a scoring system allowing us to predict claims from applicants up to an accuracy of 72%. While this is an exciting prospect, some caution must be exercised in extrapolating our findings until it can be repeated on another dataset. Ideally, complete validation of these methods would require a longitudinal study of several years duration. It may be more practical to test the formulas on different datasets from different companies. If our scoring system is validated, then all applications could be processed automatically (using bar coding or other techniques). This processing would be very speedy, and would generate a single score summarizing all the application data. Companies could then choose how to use this score. It could be used alone for applications below a certain threshold. Alternatively, companies could use a high score as a trigger for assembling further medical evidence. We are not suggesting that underwriting can be bypassed by a scoring system derived from our data, but that it could be complementary. Ratings arrived at in this way can be defended as objective because they are based on past experience. The study was funded by the Assurance Medical Society, whom we thank. We also wish to thank the 24

HAMILTON ET AL—ILL HEALTH CLAIMS Permanent Insurance Company for access to their data, and also Barbara Laws for her accurate work in data entry. The opinions in this article are wholly those of the authors.

5. Schulz K, Grimes D. Case-control studies: research in reverse. Lancet. 2002;359:431–434. 6. Calle E, Thun M, Petrelli J, Rodriguez C, Heath CJ. Body-mass index and mortality in a prospective cohort of US adults. N Engl J Med. 1999;341:1097– 1105. 7. Wessely S, Nimnuan C, Sharpe M. Functional somatic syndromes: one or many? Lancet. 1999;354: 936–939. 8. Staessen JA, Thijis L, Fagard R, O’Brien E, Clement D, de Leeuw PW. Predicting cardiovascular risk using conventional vs. ambulatory blood pressure in older patients with systolic hypertension. JAMA. 1999;282:539–546. 9. Hansson L, Zanchetti A, Carruthers SG, et al. Effects of intensive blood-pressure lowering and lowdose aspirin in patients with hypertension: principal results of the Hypertension Optimal Treatment (HOT) randomized trial. Lancet. 1998;351:1755– 1762.

REFERENCES 1. Brackenridge R, Elder W. Medical Selection of Life Risks. 3rd ed. New York, NY: Stockton Press; 1992. 2. Niverthi M, Ivanovic B. Body mass index and mortality in an insured population. J Insur Med. 2001; 33:321–327. 3. Hall GH, Hamilton WT, Round AP. Increased illness experience preceding chronic fatigue syndrome: a case control study. J Royal Coll Phys London. 1998;32:44–48. 4. Hamilton W, Hall G, Round A. Frequency of attendance in general practice and symptoms before development of chronic fatigue syndrome: a case-control study. Br J Gen Pract. 2001;51:553–558.

25