Time dependent relative risks in life insurance medical underwriting
-
Upload
robert-kneepkens -
Category
Documents
-
view
26 -
download
1
Transcript of Time dependent relative risks in life insurance medical underwriting
Title:
Time dependent relative risks in life insurance medical underwriting
Author:
Robert F. Kneepkens, MD MScEBP
Affiliation:
Chief medical officer Achmea division Pension & Life
Financial support:
Achmea division Pension & Life
Address for correspondence:
Achmea division Pension & Life; Postbus 90106, 5000 LA, Tilburg, The Netherlands,
TSA06/054; phone: 0031-6-5145-1528; e-mail: [email protected].
Correspondent:
Robert Kneepkens, MD MScEBP
Time dependent relative risks in life
insurance medical underwriting
Abstract
Introduction: Life insurance medicine focuses on mortality hazards in specified periods.
People are free to insure their lifes for shorter or longer terms. Because the chosen terms
range from one year to a life time, life insurers have to take into account the fact that the
predictive value of risk indicators can and will change over time. The time a risk indicator
keeps its predictive value, will be dependent on its biological effects, volatility, and
treatability. For a given applicant this implies that the RH calculated for his medical condition
should be dependent on the term of the insurance.
The main objective of this study is to determine if some commonly used risk indicators –
previously used to study age dependency of relative risks – have a predictive value that
increases with the observation period (1)
.
Methods: This population-based cohort study uses NHANES-datafiles from the Third
National Health and Nutrition Examination Survey (NHANES III) and the NHANES Linked
Mortality Files 2010. Only participants aged 20 to 69 that were examined in mobile
examination centers, without a history of some prevalent high risk diseases were included.
The observed mortality was compared to the expected mortality in a Generalized Linear
Model (GLM) with Poisson error structure with two reference populations, which both can
serve as preferred reference for life insurers: The United States Life Tables 2008 (USLT) and
the 2008 Valuation Basic Tables (VBT) based on the insured population of 35 US Life
insurers. The Time dependency of the RH’s of the systolic blood pressure (SBP), aspartate
aminotranseferase (ASAT), lactate dehydrogenase (LDH), serum albumen and albuminuria,
was assessed, with correction for ethnicity, household income, history of diabetes mellitus,
BMI and serum cholesterol. To be able to compare the results with the results of the Age
Dependency Study (ADS), we used the same data, the same risk indicators, and the same
statistical analysis method, the Generalized Linear Model (GLM).
Results: Contrary to the results in the univariate analyses, the multivariate analyses show
no difference between the USLT and VBT models. In both models the RH’s of SBP and
albuminuria increase over time, while the RH of LDH decreases over time. Only the slopes
are different, reflecting the rate of mortality increase between USLT and VBT. The RH’s of
ASAT and serum albumen are independent of time in both models.
Discussion: Time dependency of RH’s can be assumed to exist with many risk indicators.
Time dependency may take two forms: decreasing the RH and increasing the RH of a variable
as time passes. Medical underwriting guidelines should therefor differentiate between short
term and long term life insurances. Medical directors should realize that the presence of time
dependency diminishes the value of short term clinical studies for life insurance medical
underwriting. Life insurance mortality studies should always contain a discussion of the time
dependency of the predictive value of the variables under study.
Key Words: Time dependency, generalized linear model, life insurance, mortality,
relative hazard.
Time dependent relative risks in life
insurance medical underwriting
Introduction
Life insurance medicine focuses on mortality hazards in specified periods. People are free
to insure their lifes for shorter or longer terms. Because the chosen terms range from one year
to a life time, life insurers have to take into account the fact that the predictive value of risk
indicators can and will change over time. Some risk indicators are indicative of imminent
death, while others do only predict a raised relative risk in a more distant future.
Medical underwriting guidelines seldom include time-dependency of risk indicators. Time
is primarily used in two ways. The first years after assessment the risk may be deemed too
high and the insurance is postponed or only possible with temporary flat extra premiums. The
risk may also be deemed too high at the end of the requested insurance term, in which case a
shorter term will be offered. In most cases, these attenuations of the insurance proposal are
not based on specific risk indicators, but on the presence of certain diseases, like a
malignancy.
Risk indicators can fluctuate fast, specially if treated. The time a risk indicator keeps its
predictive value, will be dependent on its biological effects, volatility, and treatability.
Arythmias might be short lived, but so may be their bearers, despite the excellent possibilities
for treatment. On the other hand, smoking habits tend to lead to an accumulation of corporal
damage, while smokers sometimes seem to be incurable. In the former case the predicitve
value will be lost soon, while in the latter the predictive value will be increasing year after
year.
In life insurance medicine the RH’s are calculated by comparing the observed mortality to
an expected mortality that is based on sex, smoking status, age at start, and time since medical
underwriting. This means that time dependency is integrated in the expected mortality and in
the standard premiums for life insurances, and therefor should be integrated in the
underwriting guidelines. For a given applicant this implies that the RH calculated for his
medical condition should be more dependent on the term of the insurance than is currently the
case. At least, the risk assessor should be more consciencious of the time dependency of a the
risk indicators present.
If time dependencies are used, acurate risk assessment will possible for every period of the
insurance term, making lower premiums possible without an accompanying higher risk of
profit loss due to excess policy loss or retention. Incorporating known time dependency in
underwriting guidelines could be profitable for both insurer and applicant.
Objectives
It stands without doubt that some commonly used risk indicators have a predictive value
for mortality that decreases with the observation time. However, life insurers need to be able
to predict long term mortality as acurate as short term mortality. The main objective of this
study is to determine if some commonly used risk indicators – previously used to study age
dependency of relative risks – have a predictive value that increases with the observation
period (1)
. To be able to compare the results with the results of the Age Dependency Study
(ADS), we used the same data, the same risk indicators, and the same statistical analysis
method, the Generalized Linear Model (GLM). GLM combines the advantages of life table
analysis – taking expected mortality into account – and Cox’ proportional hazards models –
allowing for many risk determinants in a proportional equation.
We expect that in a direct comparison of observed versus expected mortality the RH’s for
variables that signal any form of senescence, detoriation of organ damage – like serum and
urinary albumen – will be dependent on the time elapsed since the start of the insurance.
Other, more “volatile” variables are more apt to signal acute risk and will show decreasing
RH’s when time progresses. For the SBP, both outcomes are possible, as the result will
depend on the access to good health care and therapy adherence, and therefor on the
population under study.
Because time dependency of RH’s can differ according to the life tables used, this study
also looks at the effect of the life tables: the one-dimensional general population tables or the
two-dimensional insured lifes population. We expect the increase of RH’s to become more
prominent when one-dimensional tables are used, and the decrease to become more prominent
when two-dimensional tables are used.
Study design
The NHANES-datafiles from the Third National Health and Nutrition Examination
Survey (NHANES III) and the NHANES Linked Mortality Files 2010 were used for this This
population-based cohort study. The NHANES program from the National Center for Health
statistics (NCHS) has been designed to assess the health and nutritional status of non-
institutionalized adults and children in the United States. The anonymized datafiles are
publicly available for researchers (2)
.
To compare the observed mortality to the expected mortality we used a Generalized
Linear Model with Poisson error structure with two reference populations, which both can
serve as preferred reference for life insurers. The United States Life Tables 2008 (USLT)
served as unselected age and sex matched reference population (3)
. We used the insured
population of 35 US Life insurers as selected – medically underwritten – reference
population, which has been published as the 2008 Valuation Basic Tables (VBT) of the
American Academy of Actuaries and the Society of Actuaries (4)
.
Study and base population
Because insured populations will differ from state to state, company to company, and
product to product, we used the population-based study from the NCHS and the Centers for
Disease Control and Prevention (CDC), NHANES III, conducted from 1988 through 1994.
Population-based studies like NHANES tend to have results that are generalizible and contain
more and publicly accessible data. The NCHS searched death certificates, the National Death
Index (NDI), the Social Security Administration (SSA), and the Centers for Medicare and
Medicaid Services (CMMS) to retrieve the mortality status in 2010.
Inclusion and exclusion criteria
The study is restricted to participants aged 20 to 69 that were examined in a mobile
examination center (MEC), to imitate as much as possible the underwriting practice of life
insurers.
We excluded participants with a history of a potentially life-threathening condition: heart
attack, congestive heart failure, stroke, emphysema, cancer, because the hazard of mortality
of these participants would be dictated too much by their medical history and could lead to
insurability. This exclusion should also ensure that the population resembles the insured
population more and will enable testing some often used risk determinants. We also excluded
MEC-examined participants that did not undergo the full blood and urine tests under study,
because life insurers would decline anyone who would refuse to undergo all tests that were
deemed necessary for the risk assessment. For a flow chart of participants, please see the ADS
(1).
Methods
The hazard of death relative to a life table of a large reference population, like the general
population or the insured population, can be analysed using GLM for RH’s, with fixed time-
periods of one year, the log of the observed total mortality as link, the log of the expected
total mortality according to the life tables as offset, and a Poisson error structure (5)
. The
model is discussed in more detail in the ADS (1)
.
Data preparation
We linked the anonymized datafiles of the NHANES III Adult Household Interview,
Examination, Laboratory, and Mortality Linkage 2010 dataset, which are publicly available
for researchers at the website of the NCHS (2)
.
For each participant a separate case was created for each time period of one year he was in
observation, with only the last year being broken because of death or censoring. Expected
mortality was drawn from two sets of life tables, USLT and VBT, depending on sex, age at
the time of interview, time elapsed since this interview and his smoking status at the time of
interview. We used the life tables for non-smokers of the VBT for never-smokers and former
smokers who stopped smoking more than two years before the household interview. we used
the life tables for smokers for all others, including participants with unknown smoking habits.
For this study we used the SBP, ASAT, LDH, serum albumen and albuminuria as main
variables, and ethnicity, household income, history of diabetes mellitus or high blood sugar,
BMI and total cholesterol as additional variables. Detailed information on the data collection
in the NHANES III survey can be found in the ADS and at the website of the NCHS (1,2)
.
Parameters and endpoints
This study focuses on five continuous variables that are widely used in life insurance
underwriting: SBP, ASAT, LDH, serum albumen and albuminuria. These variables were
measured in most participants and have predictive value for all-cause mortality in a relatively
healthy population (6,7,8,9)
. They also account for many undocumented chronic diseases that
could otherwise render the model unstable. To enhance the stability, the model was corrected
for ethnicity (Non-Hispanic white, Non-Hispanic black, Mexican-American, Other),
household income (> or < 20,000 US$), history of diabetes mellitus, BMI (8 categories) and
serum cholesterol (5 categories). Age, sex, smoking status and time since the interview
determine the expected mortality and were therefore not added to the linear predictor with
explanatory variables.
Main outcomes
The main goal of this study is the assessment of the time dependency of the RH’s of the
SBP, ASAT, LDH, serum albumen and albuminuria for the ULST and VBT 2008 reference
populations. The variables are tested on time dependency in the generalized linear model, by
adding interaction terms between the variable and the year since the household interview, the
first year being year 1.
Other outcomes
Because it is probable that the strength and the direction of the time dependency depends
on the life tables used, all models were run with both sets of life tables (USLT, VBT). In VBT
the mortality increases faster during the observation period, which might lead to a negative
time dependency, while in USLT the slower increasing mortality may lead to a positive time
dependency, i.e. the RH of the variable increases over time.
Test procedures
In GLM the choice between models is made using the deviance of the log-likelihood from
the maximal possible log-likelihood, where there is an explanatory variable for every case.
The significance of explanatory variables was tested with a Type III test. The decision if
explanatory variables could stay in the model was made using the likelihood ratio test. As cut-
off point we used a P-value of 0.1.
We used standardized deviance residuals, leverages and Cook’s distances to verify that no
case contributed too much to the model compared to other cases.
All tests and analyses were performed with SPSS 22 (IBM SPSS Statistics for Windows,
Version 22.0. Armonk, NY: IBM Corp. Released 2013).
Results
20,050 adults participated in the household interview in the NHANES III Survey, but
6,470 participants were excluded because of age or not being examined in the mobile
examination center. From 10 participants the mortality status could not be retrieved in 2010.
From the 13,570 remaining participants 1,261 were excluded from this study because they had
an history of one or more potentially life threathening diseases that were documented in the
household interview: heart attack, congestive heart failure, stroke, emphysema, cancer. 1,140
participants were excluded because of incomplete laboratory testing. This study is performed
with the data of 11,169 participants aged 20 to 69 that completed the household interview, the
medical examination and the laboratory testing. 1,150 died during the observation period. The
participants were observed for a median of 14.8 years (IQR 13.3 – 16.5). The total number of
completed and incompleted observation years was 168,047. Detailed information on the
baseline characteristics of the participants can be found in table 1 of the Age Dependency
Study (1)
.
Univariate analysis
The continuous variables SBP, ASAT, LDH, serum albumen and albuminuria were all
significant with P<0.001 in univaratiate analyses in GLM with expected mortalities
calculated from USLT and VBT. This changed when their interaction term over time since the
interview was introduced. The results are shown in tables 2a–e.
Table 2a – GLM, univariate analysis, systolic blood pressure
US general population (US Life Tables 2008) US Insured Lifes (2008 Valuation Basic Tables)
Value Df Value Df
Deviance 12,222 168,044 v/df=0,073 12,054 168,044 v/df=0,072
Akaike’s Information criterion 14,528 14,360
Likelihood Ratio Chi-square 88.6 2 P<0.001 85.3 2 P<0.001
RH=exp(B) CI 95% LR Chi-sq. RH=exp(B) CI 95% LR Chi-sq.
Intercept 0.186 0.126 – 0. 275 P=0.004 0.386 0.261 – 0.572 P<0.001
Systolic blood pressure 1.012 1.009 – 1.015 P<0.001 1.014 1.011 – 1.017 P<0.001
Time * systolic blood pressure 1.00017 1.00007 – 1.00026 P=0.001 0.99980 0.99970 – 0.99990 P<0.001
Table 2b – GLM, univariate analysis, aspartate aminotransferase
US general population (US Life Tables 2008) US Insured Lifes (2008 Valuation Basic Tables)
Value Df Value Df
Deviance 12,250 168,044 v/df=0,073 12,060 168,044 v/df=0,072
Akaike’s Information criterion 14,556 14,366
Likelihood Ratio Chi-square 60.4 2 P<0.001 79.3 2 P<0.001
RH=exp(B) CI 95% LR Chi-sq. RH=exp(B) CI 95% LR Chi-sq.
Intercept 0.87 0.81 – 0.94 P=0.029 1.65 1.53 – 1.79 P<0.001
Aspartate aminotransferase 1.0079 1.0045 – 1.0112 P<0.001 1.014 1.012 – 1.016 P<0.001
Time * aspartate aminotransf. 1.00015 0.99983 – 1.00046 P=0.353 0.99939 0.99906 – 0.99971 P<0.001
Table 2c – GLM, univariate analysis, lactate dehydrogenase
US general population (US Life Tables 2008) US Insured Lifes (2008 Valuation Basic Tables)
Value Df Value Df
Deviance 12,236 168,044 v/df=0,073 12,033 168,044 v/df=0,072
Akaike’s Information criterion 14,542 14,339
Likelihood Ratio Chi-square 74.3 2 P<0.001 107.2 2 P<0.001
RH=exp(B) CI 95% LR Chi-sq. RH=exp(B) CI 95% LR Chi-sq.
Intercept 0.44 0.36 – 0.53 P=0.649 0.75 0.62 – 0.90 P<0.001
Lactate dehydrogenase 1.0044 1.0032 – 1.0057 P<0.001 1.0076 1.0065 – 1.0088 P<0.001
Time * lactate dehydrogenase 1.00011 1.00004 – 1.00019 P=0.003 0.99983 0.99976 – 0.99991 P<0.001
Table 2d – GLM, univariate analysis, serum albumen
US general population (US Life Tables 2008) US Insured Lifes (2008 Valuation Basic Tables)
Value Df Value Df
Deviance 12,179 168,044 v/df=0,072 12,027 168,044 v/df=0,072
Akaike’s Information criterion 14,485 14,333
Likelihood Ratio Chi-square 131.5 2 P<0.001 112.6 2 P<0.001
RH=exp(B) CI 95% LR Chi-sq. RH=exp(B) CI 95% LR Chi-sq.
Intercept 53.7 27.1 – 106.3 P=0.322 64.3 32.2 – 128.3 P<0.001
Serum albumen 0.36 0.30 – 0.53 P<0.001 0.45 0.38 – 0.54 P<0.001
Time * serum albumen 1.0073 1.0040 – 1.0106 P<0.001 0.9947 0.9914 – 0.9981 P=0.002
Table 2e – GLM, univariate analysis, albuminuria
US general population (US Life Tables 2008) US Insured Lifes (2008 Valuation Basic Tables)
Value Df Value Df
Deviance 12,253 168,044 v/df=0,073 12,080 168,044 v/df=0,072
Akaike’s Information criterion 14,559 14,386
Likelihood Ratio Chi-square 57.4 2 P<0.001 59.5 2 P<0.001
RH=exp(B) CI 95% LR Chi-sq. RH=exp(B) CI 95% LR Chi-sq.
Intercept 1.067 1.006 – 1.131 P=0.015 1.93 1.82 – 2.05 P<0.001
Albuminuria 1.00012 0.99993 – 1.00031 P=0.257 1.00026 1.00007 – 1.00044 P=0.022
Time * albuminuria 1.000024 1.000002 – 1.000045 P=0.016 1.000011 0.999990 – 1.000032 P=0.293
In univariate analyses most variables have a lower RH in the last year of the observation
period than in the first year of the observation period, when expected mortality is derived
from VBT. However, the RH’s of most variables are increasing during the observation period
when the USLT are used. The overall mortality is rising faster in VBT than in USLT and
apparently the RH’s of these variables show a rate of increase in univariate analysis that falls
in between the rates of the VBT and USLT.
There are two exceptions to this general observation. The RH of ASAT is not time
dependent compared with USLT, making ASAT a relatively short term predictor. The RH of
albuminuria has little predictive value in the first years when used with USLT and is not time
dependent compared with VBT, suggesting albuminuria is more suitable as a long term
predictor than as short term predictor.
Multivariate analysis
All variables were significant in mulitvariate analysis with GLM with expected mortalities
calculated from USLT and VBT (results shown in the ADS (1)
). If a variable predicts relative
mortality better in the short term than in the long term – or vice versa –, the RH of the
variable will be time dependent. This means that an added interaction term “time * variable” –
with “time” representing the year of observation – should be significant in the multivariate
analysis, as is the case for most variables in the univariate analysis.
The RH’s of the interaction terms with time for the SBP, LDH and albuminuria are
significant in mulitvariate analysis when a limit of P=0.1 is used, both in the USLT and the
VBT model. The RH’s of the continuous variables themselves stay significant after
introduction of the interaction terms, except for albuminuria. This indicates that albuminuria
has a relatively low predictive value for mortality in the first years of the insurance.
Contrary to the results in the univariate analyses, the multivariate analyses show no
difference between the USLT and VBT models. In both models the RH’s of SBP and
albuminuria increase over time, while the RH of LDH decreases over time. Only the slopes
are different, reflecting the rate of mortality increase between USLT and VBT. The RH’s of
ASAT and serum albumen are independent of time in both models.
The final models can be found in Table 3. The order in which the interaction terms had to
be removed differed depending on the life tables used.
Table 3 – Generalized Linear Model, with all explanatory variables, and interaction terms for
age dependency
All ages US general population (US Life Tables 2008) US Insured Lifes (2008 Valuation Basic Tables)
Value Df Value Df
Deviance 11,732 168,024 v/df=0,070 11,561 168,024 v/df=0,069
Akaike’s Information crit. 14,078 13,907
Likelihood Ratio Chi-square 578.8 22 P<0.001 598.4 22 P<0.001
RH=exp(B) CI 95% LR Chi-sq. RH=exp(B) CI 95% LR Chi-sq.
Intercept 4.55 1.92 – 10.81 P<0.001 4.59 1.91 – 10.99 P<0.001
Low or no income, <20,000
US$ 1.74 1.54 – 1.97
P<0.001
1.53 1.35 – 1.73
P<0.001
History of DM 1.63 1.39 – 1.91 P<0.001 1.71 1.46 – 2.01 P<0.001
SBP, mmHg 1.0050 1.0009 – 1.0092 P=0.016 1.0060 1.0019 – 1.0101 P=0.004
Time * SBP 1.00059 1.00028 – 1.00089 P<0.001 1.00034 1.00004 – 1.00064 P=0.033
ASAT, U/L 1.0076 1.0057 – 1.0094 P<0.001 1.0075 1.0056 – 1.0094 P<0.001
Time * ASAT removed (P=0.302) removed (P=0.296)
LDH, U/L 1.0050 1.0027 – 1.0074 P<0.001 1.0066 1.0044 – 1.0088 P=0.005
Time * LDH 0.99974 0.99951 – 0.99998 P=0.041 0.99964 0.99941 – 0.99988 P<0.001
Serum albumen, mg/dL 0.50 0.42 – 0.58 P<0.001 0.57 0.48 – 0.68 P<0.001
Time * serum albumen removed (P=0.613) removed (P=0.816)
Albuminuria, ug/mL removed (P=0.939) removed (P=0.798)
Time * albuminuria 1.000024 1.000016 – 1.000031 P<0.001 1.000023 1.000015 – 1.000032 P<0.001
Ethnicity: P<0.001 P=0.052
- Non-Hispanic white (ref.) 1 1
- Non-Hispanic black 1.16 1.00 – 1.35 1.08 0.93 – 1.26
- Mexican-American 0.87 0.74 – 1.02 0.91 0.77 – 1.06
- Other 0.71 0.49 – 1.02 0.75 0.52 – 1.09
BMI, kg/m2 P<0.001 P=0.002
- Lower than 20.0 (ref.) 1 1
- 20.0 – 22.4 0.64 0.48 – 0.86 0.74 0.55 – 0.98
- 22.5 – 24.9 0.50 0.38 – 0.66 0.60 0.46 – 0.79
- 25.0 – 27.4 0.48 0.37 – 0.63 0.60 0.46 – 0.79
- 27.5 – 29.9 0.46 0.35 – 0.60 0.61 0.47 – 0.80
- 30.0 – 32.4 0.42 0.32 – 0.56 0.56 0.42 – 0.75
- 32.5 – 34.9 0.51 0.37 – 0.69 0.67 0.49 – 0.91
- 35.0 and over 0.54 0.40 – 0.72 0.78 0.59 – 1.04
Cholesterol (mg/dl) P<0.001 P<0.001
- Lower than 175 (ref.) 1 1
- 175 – 199 0.75 0.62 – 0.89 0.74 0.62 – 0.89
- 200 – 224 0.65 0.54 – 0.78 0.65 0.54 – 0.78
- 225 – 249 0.69 0.57 – 0.84 0.67 0.55 – 0.81
- 250 and over 0.65 0.54 – 0.79 0.64 0.53 – 0.77
DM = diabetes mellitus, SBP = systolic blood pressure, ASAT = aspartate aminotransferase, LDH = lactate dehydrogenase, BMI = body
mass index. Decision about inclusion of explanatory variables were made on the basis of a Type III test, using the P-value of the Likelihood
Ratio Chi-square (LR Chi-sq.) given in this table. In case of removal of a variable or interaction term, the last recorded P-value is given
between brackets.
A few cases with exceptional high ASAT, LDH and albuminuria had larger Cook’s
Distances than the rest of the cases, although none of the Cook’s Distances were larger than 1.
Sensitivity analysis consisted of removing these cases. Removal did not change the results of
the base models, but in both models with time dependency the RH’s of ASAT and “time *
albuminuria” were increased, without affecting the models themselves.
Discussion
Life insurance underwriting guidelines help underwriters and medical directors to
calculate a best estimate of the mortality ratio. This study shows that the duration of the life
insurance should be taken into account when the applicant’s data are assessed.
Most risk indicators, like SBP, ASAT, LDH, serum albumen and albuminuria, predict
mortality over a longer period. In the multivariate analysis of this study, the RH’s of two of
them – ASAT and serum albumen – were independent of the time elapsed since the household
interview and medical examination. Only one of them – LDH – had RH’s that decreased over
time. The RH of LDH decreases to 1.00 within 20 years, making LDH less suitable for
prediciting mortality in the more distant future. The RH’s of the remaining two variables –
SBP and albuminuria – increased over time. The RH of the SBP squared in 9 to 18 years,
depending on the life tables used. The RH of albuminuria even started at virtually 1.00 at the
start of the observation, rendering the variable unsuitable for short term mortality predictions.
Important implications of these results are, that life insurance underwriting is rightfully
claimed to focus more on long term prognosis than curative medicine does, and that most
clinical studies have a too short observation period to be of much use to life insurance medical
underwriting.
This study failed to show important differences between the models with the USLT as
reference and the models with the VBT as reference. In both sets of models the results of
mulitvariate analyses concerning the time dependency were comparable. Only in univariate
analyses differences could be noted.
The major weakness of this study is the fact that the population under study is not a
population of insured lifes or life insurance applicants. As explained in the study design
section the use of insurance populations lowers the generalizibility, suffers from the fact that
most data are only collected for specific subgroups, and may contain competition sensitive
information. The choice for NHANES III data implies that the reported RH’s cannot be
copied to medical underwriting guidelines for life insurances. Further research in subgroups
of insurance populations is needed when the goal is to update underwriting guidelines.
A second weakness of this study is that not all possible forms of time dependency are
studied. Some variables used in medical underwriting may be indicative of more than one
disease processes, with as a result a biphasic mortality curve. These forms of time dependency
are not studied.
The major strength of this study is that the observed mortality has been compared directly
to the expected mortality for every year of observation separately. Life insurers do not only
want to know who dies first, but also what an applicant’s mortality will be compared to his
peers. The GLM combines this direct comparison with a linear predictor containing many risk
determinants, which leads to more stable results.
By using two sets of life tables to provide the expected mortality, the effect of the choice
of the life tables can be studied. The fact that the results in both sets of models are comparable
lends power to the findings of this study.
Conclusion
Time dependency of RH’s can be assumed to exist with many risk indicators. Time
dependency may take two forms: decreasing the RH and increasing the RH of a variable as
time passes. Medical underwriting guidelines should therefor differentiate between short term
and long term life insurances.
Medical directors should realize that the presence of time dependency diminishes the
value of short term clinical studies for life insurance medical underwriting. Life insurance
mortality studies should always contain a discussion of the time dependency of the predictive
value of the variables under study.
References
1. Kneepkens RF, Lindeboom R. Age dependent decline of relative risks in life insurance
medical underwriting. J Insur Med 2014; 44: ... - ...
2. National Center for Health Statistics and Centers for Disease Control and Prevention:
Analytic and reporting guidelines: The third National Health and Nutrition Examination
Survey, NHANES III (1988–1994). http://www.cdc.gov/nchs/nhanes/nh3rrm.htm#ag.
Accessed May 5, 2013.
3. Arias E. United States life tables, 2008. National vital statistics reports; vol 61 no 3.
Hyattsville, MD: National Center for Health Statistics. 2012.
4. American Academy of Actuaries / Society of Actuaries, Joint Preferred Mortality Project
Oversight Group. 2008 Valuation Basic Table. Paper presented at the March 2008
meeting of the National Association of Insurance Commissioners’ Life and Health
Actuarial Task Force, Orlando, Florida, March, 2008: 49-50, 53-54, 57-58, 61-62.
Accessed: September, 2013 from http://www.actuary.org/pdf/life/tables_march08.pdf.
5. Jong P van, Heller GZ. Generalized Linear Models for insurance data. Melbourne.
Cambridge University Press 2008, 2012.
6. Perloff D, Grim C, Flack J, Frohlich ED, Hill M, McDonald M, Morgenstern BZ. Human
blood pressure determination by sphygmomanometers. Circulation 1993; 88: 2460-2470.
7. Fulks M, Stout RL, Dolan VF. Using liver enzymes as screening tests to predict mortality
risk. J Insur Med 2008; 40: 191-203.
8. Fulks M, Stout RL, Dolan VF. Albumin and all-cause mortality risk in insurance
applicants. J Insur Med 2010; 42: 11-17.
9. Ohsawa M, Fujioka T, Ogasawara K, Tanno K, Okamura T, Turin TC, Itai K, Ogawa A,
Yoshida Y, Omama S, Onoda T, Nakamura M, Makita S, Ishibashi Y, Tanaka F,
Kuribayashi T, Ohta M, Sakata K, Okayama A. High risks of all-cause and cardiovascular
deaths in apparently healthy middle-aged people with preserved glomerular filtration rate
and albuminuria: A prospective cohort study. Int J Cardiol 2013.
http://dx.doi.org/10.1016/j.ijcard.2013.10.076.