Self-reported information and pharmacy claims were comparable
for lipid-lowering medication exposure
David W. Browna,*, Robert F. Andaa, Vincent J. Felittib
aCenters for Disease Control and Prevention, Atlanta, Georgia, USA
bDepartment of Preventive Medicine, Southern California Permanente Medical Group, San Diego, California, USA
Accepted 22 August 2006
Objective: To examine agreement between self-reported exposure to lipid-lowering medications and objective evidence of filling pre-
scribed lipid-lowering medications.
Study Design and Setting: Using data from 7,918 adults from the Adverse Childhood Experiences (ACE) Study, we calculated the
sensitivity, specificity, and positive (PVþ) and negative (PV?) predictive values, and likelihood ratios for self-reported exposure to
lipid-lowering medications compared to exposure obtained from pharmacy claims (gold standard) both overall and by age, sex, race/eth-
nicity, education, and ACE Score.
Results: Eight percent (n5 655) of adults self-reported lipid-lowering medication exposure, and 379 adults filled at least one lipid-low-
ering prescription within 60 days of the baseline exam during 1997. The sensitivity of self-reported exposure was nearly 94%; the spec-
ificity was 96%; the PVþwas 54%; and the PV?was nearly 100%. Values for sensitivity, specificity, PVþ, and PV?were similar across
Conclusion: A self-reported measure of lipid-lowering medication exposure was accurate with high sensitivity and specificity while the
PVþof self-reported lipid-lowering medication exposure was relatively low. These findings suggest that self-reported exposure to lipid-low-
ering medications may be useful in surveys that examine the prevalence of hyperlipidemia, but may overestimate actual exposure in studies
monitoring trends in use of lipid-lowering medications. ? 2007 Elsevier Inc. All rights reserved.
Keywords: Sensitivity; Specificity; Validity (epidemiology); Data collection; Cholesterol; Epidemiology
High blood cholesterol (HBC) is an important heart-dis-
ease risk factor. Self-reported exposure to lipid-lowering
medication(s) is often included in definitions of HBC in
observational studies. For example, HBC may be defined
in the presence of total serum cholesterol concentrations
>240 mg/dL, self-reported physician-diagnosed HBC, or
self-reported exposure to lipid-lowering medication. Self-
reported exposure to lipid-lowering medications, therefore,
may influence prevalence estimates of HBC and surveil-
lance of trends in HBC if exposure is misclassified.
The validity of self-reported physician-diagnosed HBC
has been examined ; however, little is known about
how self-reported exposure to lipid-lowering medication
compares to exposure data obtained from administrative
pharmacy claims, which are not subject to recall or report-
ing bias. To address the issue of agreement between self-re-
ported exposure to lipid-lowering medication and evidence
for their use obtained from administrative claims, we used
data from the Adverse Childhood Experiences (ACE)
Study, a study designed to examine the long-term relation-
ship between ACEs (e.g., childhood emotional, physical, or
sexual abuse and household dysfunction during childhood)
and a variety of health behaviors and outcomes in adult-
hood. In addition, we examined possible sociodemographic
characteristics (age, sex, race/ethnicity, education) and the
ACE Score (an integer count of eight different categories
of ACEs) that may influence agreement between the self-
report and objective measures of lipid-lowering medication
exposure. In doing so, we address a concern  regarding
* Corresponding author. Centers for Disease Control and Prevention,
4770 Buford Hwy NE (MS K67), Atlanta, GA 30341, USA. Tel.: 770-
488-2567; fax: 770-488-5965.
E-mail address: firstname.lastname@example.org (D.W. Brown).
0895-4356/07/$ e see front matter ? 2007 Elsevier Inc. All rights reserved.
Journal of Clinical Epidemiology 60 (2007) 525e529
potential recall bias as it relates to studies of child maltreat-
ment and related traumatic stressors.
The ACE Study methods have been previously described
[3e5]. The study, a collaborative effort between Kaiser Per-
manente (San Diego, CA) and the Centers for Disease Con-
trol and Prevention (Atlanta, GA), was approved by the
institutional review boards of Kaiser Permanente and the
Office for Protection from Research Risks at the National
Institutes of Health. Participants were drawn from the more
than 50,000 adult members of the Kaiser Health Plan in San
Diego who are evaluated annually at Kaiser’s Health Ap-
praisal Center. The annual exam includes a standardized
health history, psychosocial evaluations, and physical
This analysis is based on data from 7,918 adults aged
>45 years with a baseline appointment between April 1
and October 31, 1997 and complete information on age,
sex, race/ethnicity, and education for whom self-reported
exposure and objective evidence of filling prescribed
lipid-lowering medication was available. As part of the
baseline questionnaire, persons were asked, ‘‘Have you
been told to take medicine to lower your cholesterol?’’
Those who responded affirmatively were considered to
have self-reported exposure to lipid-lowering medicine.
For prescribed lipid-lowering medication exposure, we
identified all Kaiser pharmacy claims during the calendar
year 1997 for bile acid resins (colestipol, cholestyramine,
colesevalam), fibrates (gemfibrozil, clofibrate, fenofibrate),
nicotinic acid (niacin), and 3-hydroxy-3-methyglutaryl co-
enzyme A (HMG-CoA) reductase inhibitors (lovastatin,
simvastatin, pravastatin, fluvastatin, atorvastatin, cervista-
tin). We examined exposure to these medications within
60 days (before or after) of the baseline appointment. Sec-
ondary analyses were completed using intervals of 15, 30,
and 90 days to examine the influence of interval length
on validity of test characteristics. Persons with at least
one pharmacy claim for a lipid-lowering medication were
considered to have documented exposure.
We compared self-reported exposure with that of lipid-
lowering pharmacy claims. Using pharmacy claims data
as the gold standard, we calculated the sensitivity, specific-
ity, and predictive value positive (PVþ) and negative (PV?)
(Table 1) of self-reported exposure overall and by age, sex,
race/ethnicity, educational attainment, and ACE Score. The
ACE Score  is a simple summation (integer count) of ex-
posure to each of the eight categories of adverse experi-
ences during the respondents’ first 18 years of life. The
categories are emotional abuse, physical abuse, contact sex-
ual abuse, a battered mother, household substance abuse,
household mental illness, incarcerated household members,
parental separation, or divorce [7e9]. Values for likelihood
ratios (LR), another measure of performance, for overall
self-reported exposure are also presented. We measured
agreement between the two sources of information with
the kappa (k-) statistic (?1< k < þ 1), which equals
0 when observed agreement is what would be expected
by chance alone and equals 1 when there is perfect agree-
ment. Guidelines for interpreting the k-statistic suggest
values <0.40 represent fair or poor agreement, 0.41e0.60
moderate agreement, 0.61e0.80 substantial agreement,
and 0.81e1.00 near perfect agreement . Analyses were
completed using SAS v9.1 (SAS Institute, Cary, NC).
The analysis sample included 4,308 women (54%) and
3,610 men (46%). The mean age (standard deviation) was
57 (15) years. Three-quarters of participants were white,
37% were college graduates, 41% had some college educa-
tion, and 15% were high school graduates. Only 7% had not
graduated from high school. Thirty-five percent (n5 2,760)
of participants had an ACE score of 0 whereas 26%
(n5 2,091) had a score of 1, 16% (n5 1,259) had a score
of 2, 10% (n5 767) had a score of 3, and 13% (n5 1,041)
had a score >4.
Of the 460 adults whowere prescribed >1 lipid-lowering
medication during 1997, 82% (n5 379) filled the prescrip-
followed by fibrates (15%), bile acid resins (10%), and nico-
tinic acid (3%).
Agreement between self-reported and confirmed expo-
sure to prescribed lipid-lowering medications is summa-
rized in Table 2. The prevalence of self-reported exposure
was 8% (n5 655). The accuracy (true positives and true
negatives as a proportion of all results) of the self-reported
lipid-lowering medication exposure was 96% and the
Agreement between self-report and pharmacy claims for identifying
exposure to lipid-lowering medication
TotalYes (þ)No (?)
a þb þc þd
Sensitivity, Pr (TþjDþ)5a/(aþc).
Specificity, Pr (T?jD?) 5d/(bþd).
PVþ, Pr (DþjTþ) 5a/(aþb).
PV?, Pr (D?jT?) 5d/(cþd).
LRþ, Pr (TþjDþ)/Pr (TþjD?).
LR?, Pr (T?jDþ)/Pr (T?jD?).
526D.W. Brown et al. / Journal of Clinical Epidemiology 60 (2007) 525e529
overall k-statistic value was 0.67 (95% confidence interval
[CI], 0.63e0.70). The sensitivity of self-reported exposure
was nearly 94% and ranged from 90% (age 45e54 years) to
100% (ACE Score 5 3) across participant characteristics.
The specificity of self-report was 96%. Corresponding
LR valuesfor self-reported
LRþ5 23.47 and LR?5 0.06. The PVþof self-reported
lipid-lowering medication exposure was 54% and ranged
from 42% (Nonwhite) to 60% (Less than high school edu-
cational attainment) across participant characteristics. The
PV?was nearly 100% both overall and across participant
characteristics. Values for sensitivity, specificity, PVþ, and
PV?were similar across participant characteristics.
Values for the probability of self-reporting lipid-lower-
ing medication exposure among those who were prescribed
a HMG-CoA reductase inhibitor within 60 days of baseline
were similar to those in the Table 2 [overall sensitivity 5
95.5%; 95% CI, 92.5e97.3]; the overall value for spec-
ificity was 95.3% (95% CI, 94.8e95.7) and the value
for the PVþwas slightly lower at 44.9% (95% CI,
41.1e48.7) (data not shown).
Finally, we completed secondary analyses wherein the
interval from the baseline appointment for pharmacy data
was changed to assess the influence on validity measures
(Table 3). Measures of sensitivity, specificity, and PV?of
self-reported exposure remained high across interval values
Agreement between self-report and pharmacy claims for identifying exposure to lipid-lowering medication by participant characteristics
99.7 (99.5e99.8) Total
97.5 (84.3e99.7)95.3 (93.1e96.8) 60.0 (47.7e71.1) 99.8 (98.7e99.9)
95% confidence intervals reported in parentheses.
Overall kappa (k-) statistic50.67 (95% CI, 0.63e0.70).
aData are not reported for age groups 18e34 or 35e44 years due to small numbers of prescribed medications in these age groups.
Agreement between self-report and pharmacy claims for identifying exposure to lipid-lowering medication for several different
time intervals from baseline appointment
Rxindicates prescription medication. 95% confidence intervals reported in parentheses.
527D.W. Brown et al. / Journal of Clinical Epidemiology 60 (2007) 525e529
while values of PVþwere more affected by interval
changes as prevalences increased.
This analysis indicates that the sensitivity, specificity,
and PV?of self-reported lipid-lowering medication expo-
sure in a sample of relatively healthy adults were high
while PVþwas relatively low. The low PVþaffects how
an affirmative self-report should be interpreted since a pos-
itive result may not necessarily imply that the individual is
actually exposed to lipid-lowering medications. However,
this latter finding partly reflects the low prevalence of
lipid-lowering medication exposure in our sample at the
time of the survey, and these results may differ from those
of groups with a greater prevalence of exposure to dyslipi-
demia and lipid-lowering agents. For instance, in 2002 the
prevalence of lipid-lowering medication use in the US was
roughly 50% among patients with reported hyperlipidemia
. In a hypothetical cohort of 10,000 U.S. adults, where
the prevalence of lipid-lowering medication exposure is
50% and the sensitivity (94%) and specificity (96%), values
for self-report exposure are similar to those observed
herein, the PVþwould approximately equal 96%, an
increase from the 54% observed in our study, with little
change in the PV?(94%).
Although not directly comparable due to differences in
the self-report measure, our results are similar to those of
Boudreau et al. , who examined the accuracy of self-re-
ported exposure to HMG-CoA reductase inhibitors among
a sample of women who participated in a population-based
study within a large health maintenance organization (6-
month period: sensitivity O80%, specificity O95%,
Agreement between self-reported and prescribed medi-
cation exposure was similar across participant characteris-
tics, including the ACE Score. This latter finding is
important as it relates to concerns of overreporting of child
maltreatment among persons with chronic health conditions
in retrospective studies of childhood maltreatment . The
observation that measures of sensitivity, specificity, PVþ,
and PV?for this measure of lipid-lowering medication ex-
posure do not vary across the ACE Score suggests that the
accuracy of reporting exposure to lipid-lowering medica-
tions is not a function of ACEs.
This analysis used pharmacy claims data, which repre-
sent filled rather than written prescriptions, as a proxy for
actual use against which we compared the self-report mea-
sure. Pharmacy records are often used to assess medication
use and are commonly considered a source of complete in-
formation [13e15]. If enrollees obtained prescriptions for
lipid-lowering medications outside of this system, then
we may have underestimated prevalence of actual exposure.
However, in a prepaid pharmacy system, it is highly un-
likely that expensive prescriptions would be filled outside
the system. Our measure of self-reported exposure captured
whether participants were told to use lipid-lowering medi-
cation, which we use as a proxy for self-reported use of
such medications. Misclassification (nondifferential) of
self-reported exposure would result in a decrease in sensi-
tivity, an increase in specificity, and little or no change in
predictive values. Also, our estimates derive from a sample
of adults attending a wellness clinic at a Southern Califor-
nia health maintenance organization (81% of all adults
within the San Diego Kaiser Health Plan attend this clinic
at least once in any 4-year period [3e5]). Performance
may vary in other target populations where the prevalence
of dyslipidemia and lipid-lowering medication use differs.
In summary, self-reported lipid-lowering medication ex-
posure was accurate with high sensitivity and specificity but
had only moderate positive predictive value, correctly
identifying slightly more than half (54%) of those with
objective evidence of lipid-lowering medication exposure.
These findings suggest that self-reported exposure to
lipid-lowering medications may be a useful addition to
self-reported HBC in surveys that examine the prevalence
of hyperlipidemia, but may overestimate actual exposure
in studies monitoring trends in use of lipid-lowering medi-
cations. Researchers using self-reported exposure should
consider the underlying prevalence of lipid-lowering medi-
cation use in their population of interest.
 Natarajan S, Lipsitz SR, Nietert PJ. Self-report of high cholesterol:
determinants of validity in U.S. adults. Am J Prev Med 2002;23:
 Widom CS, Raphael KG, DuMont KA. The case for prospective lon-
gitudinal studies in child maltreatment research: commentary on
Dube, Williamson, Thompson, Felitti, and Anda (2004). Child Abuse
 Felitti VJ, Anda RF, Nordenberg D, Williamson DF, Spitz AM,
Edwards V, et al. Relationship of childhood abuse and household dys-
function to many of the leading causes of death in adults. The Ad-
verse Childhood Experiences (ACE) Study. JAMA 1998;14:245e58.
 Anda RF, Croft JB, Felitti VJ, Nordenberg D, Giles WH,
Williamson DF, et al. Adverse childhood experiences and smoking
during adolescence and adulthood. JAMA 1999;282:1652e8.
 Anda RF, Felitti VJ, Bremner JD, Walker JD, Whitfield C, Perry BD,
et al. The enduring effects of abuse and related adverse experiences in
childhood: a convergence of evidence from neurobiology and epide-
miology. Eur Arch Psychiatry Clin Neurosci 2005; [Epub ahead of
 Dong M, Anda RF, Felitti VJ, Dube SR, Williamson DF,
Thompson TJ, et al. The interrelatedness of multiple forms of child-
hood abuse, neglect, and household dysfunction. Child Abuse Negl
 Straus M, Gelles RJ. Physical violence in American families: risk
factors and adaptations to violence in 8,145 families. New Bruns-
wick, NJ: Transaction Press; 1990.
 Wyatt GE. The sexual abuse of Afro-American and white-American
women in childhood. Child Abuse Negl 1985;9:507e19.
 Schoenborn CA. Exposure to alcoholism in the family: United States,
1988. Adv Data 1991;205:1e13.
 Landis JR, Koch GG. The measurement of observer agreement for
categorical data. Biometrics 1977;159e74.
528 D.W. Brown et al. / Journal of Clinical Epidemiology 60 (2007) 525e529
 Ma J, Sehgal NL, Ayanian JZ, Stafford RS. National trends in statin
use by coronary heart disease risk category. PLoS Med 2005;2:e123.
 Boudreau DM, Daling JR, Malone KE, Gardner JS, Blough DK,
Heckbert SR. A validation study of patient interview data and phar-
macy records for antihypertensive, statin, and antidepressant medica-
tion use among older women. Am J Epidemiol 2004;159:308e17.
 Christensen DB, Williams B, Goldberg HI, Martin DP, Engelberg R,
LoGerfo JP. Comparison of prescription and medical records in
reflecting patient antihypertensive drug therapy. Ann Pharmacother
 Caskie GI, Willis SL. Congruence of self-reported medications with
pharmacy prescription records in low-income older adults. Gerontol-
 Lau HS, de Boer A, Beuning KS, Porsius A. Validation of pharmacy
records in drug exposure assessment. J Clin Epidemiol 1997;50:
529 D.W. Brown et al. / Journal of Clinical Epidemiology 60 (2007) 525e529