Report:: "MIS-REPORTING OF PRESCRIPTION DRUG UTILIZATION AND Expenditures in the MCBS"

MIS-REPORTING OF PRESCRIPTION DRUG UTILIZATION AND
EXPENDITURES IN THE
MEDICARE CURRENT BENEFICIARY SURVEY

By John A. Poisal, MBA
Office of Research, Development, and Information
Centers for Medicare and Medicaid Services

1

Table of Contents
Abstract ................................................................................................................... 3
Background ............................................................................................................. 4
Data ......................................................................................................................... 6
COLLECTING PHARMACY DATA................................................................................................................................. 6
Table 1. Follow-back study status of sample persons in 1999 Cost and Use file.........8
INCLUDING THE PROPER MEDICATIONS FROM THE BENEFICIARY AND THE PHARMACY................................... 8
M ATCHING BENEFICIARY-REPORTED MEDICATIONS WITH PHARMACY-REPORTED MEDICATIONS ............... 10

Methods................................................................................................................. 11
M ODELS OF MIS-REPORTING OF PRESCRIPTION DRUG UTILIZATION (DETERMINING THE NUMBER OF
PRESCRIPTIONS THAT SHOULD HAVE BEEN REPORTED)....................................................................................... 11

Results -- Utilization............................................................................................. 14
Table 2: Distribution of Net Adjusted Under-reporting Rates ..............................................14
Table 3: Model summaries and their associated formulas and net adjusted underreporting rates ................................................................................................................................................15
DETERMINING WHAT FACT ORS ARE PREDICTIVE OF UTILIZATION MIS-REPORTING ........................................ 15
Table 4: Variables Included in Original Regression Model...................................................16
GENERALIZING THE RESULTS TO THE FULL POPULATION .................................................................................. 19
Table 5: Imputation of reporting status variable .........................................................................19

Table 6: Detailed weighted statistics of reporting status variable, pre and post
imputation.........................................................................................................................................................20
Table 7: Example of calculations to estimate annualized under and over-reported
medications ......................................................................................................................................................20
CALCULATING THE FINAL UTILIZATION MIS-REPORTING RATE.......................................................................... 22
Table 8: Example of transforming annualized estimates into actual estimates ............22
Table 9: Example of EST_PURCH variable calculation .........................................................23

Results – AWP Expenditures ................................................................................ 23
Table 10: Results of Preliminary AWP Imputation by Reporting Status..........................24
Table 11: Results of Final AWP Imputation by Reporting Status........................................24
Table 12: Unweighted and Weighted Results of AWP Imputation by Category...........26
DETERMINING WHAT FACT ORS ARE PREDICTIVE OF EXPENDITURE MIS-REPORTING ...................................... 27
Table 13: Variables Included in Original Regression Model ................................................28
GENERALIZING THE RESULTS TO THE FULL POPULATION .................................................................................. 31
Table 14: Imputation of AWP expenditure reporting status variable ................................31

Table 15: Detailed weighted statistics of AWP expenditure reporting status variable,
pre and post imputation.............................................................................................................................31
Table 16: Example of calculations to estimate annualized under and over-reported
AWP expenditures ........................................................................................................................................32
CALCULATING THE FINAL AWP EXPENDITURE MIS-REPORTING RATE............................................................. 34

2

Table 17: Example of transforming annualized AWP estimates into actual AWP
estimates ............................................................................................................................................................34
Table 18: Example of EST_PURCH$ variable calculation ....................................................35
Discussion............................................................................................................. 35
Table 19: Percentile Distributions of Reported and Estimated Drug Use ......................35
Table 20: Mis-reporting Rates by Estimated Prescriptions per Beneficiary Category
................................................................................................................................................................................36

Table 21: Percentile Distributions of Reported and Estimated AWP Expenditures ..37
Table 22: Mis-reporting Rates by Estimated AWP Expenditure Level Category ........37
APPENDIX A....................................................................................................... 39
APPENDIX B....................................................................................................... 42
APPENDIX C ...................................................................................................... 46
APPENDIX D...................................................................................................... 47
APPENDIX E ...................................................................................................... 48
APPENDIX F....................................................................................................... 49
APPENDIX G...................................................................................................... 50
APPENDIX H...................................................................................................... 51

Abstract

One shortcoming of using household surveys to estimate healthcare utilization and
expenditures is that respondents tend to misreport their usage of medical services. 1 Using
data from both the 1999 Medicare Current Beneficiary Survey (MCBS) and drug
utilization data supplied by the survey respondents’ pharmacies, we examine the level of
mis-reporting of drug use and expenditures in the MCBS. 2 Findings suggest that
prescription drug expenditures are under-reported by 17% in aggregate, and that the
number of prescriptions used is under-reported by 17.7%. The data show that there are
various demographic factors that are predictive of a beneficiary’s likelihood to either
1

Groves RM, Survey Errors and Survey Costs, New York: John Wiley and Sons, 1989

3

over-report or under-report their medications, as well as the extent to which they
misreport their drug use and spending.

Background

Interest in prescription drug expenditures, as they relate to high out-of-pocket costs and
possible drug coverage expansion, remains high. Prescription drug spending rose almost
16% in 2001 and is projected to rise an average of 11.1 percent per year between 2002
and 2012. 3 Senior citizens are particularly vulnerable to these rising costs partially due to
higher incidence rates of chronic disease, many of which can be effectively treated with
prescription medication. Even seniors with employer sponsored insurance (ESI) drug
coverage, thought to be the most reliable source of coverage, are becoming increasingly
subject to high prescription drug expenses as the number of large employers (500+
employees) offering ESI coverage to Medicare-eligible retirees declined from 57 percent
in 1987 to 23 percent in 2001. 4

Adding a prescription drug benefit to Medicare has been the focus of debate on Capitol
Hill and elsewhere for the last several years. Recent findings suggest that Medicare
beneficiaries without drug coverage fill fewer prescriptions than their covered
counterparts, after controlling for factors like age, supplementary insurance status, and
income. 5 Moreover, many beneficiaries skip dosages or avoid filling prescriptions

2

Not all people under-report their drug utilization and spending, however, in the aggregate, MCBS
respondents report fewer medication purchases and lower expenses than they actually make and incur. As
a result, this analysis often refers to the mis -reporting of drugs as under-reporting.
3
Heffler, Steve. “Health Spending Projections for 2002-2012.” 2003. Health Affairs 7 Feb 2003
<>
4
Mercer, US Mercer/Foster Higgins survey of employer-sponsored health plans - key findings, . 2002,
Mercer Human Resource Consulting LLC and Mercer Investment Consulting Inc.: Washington, DC; Bos,
B. 15th Annual Mercer/Foster Higgins National Survey of Employer-Sponsored Health Plans. 2001.
Chicago, IL; DeWitt, D.L., Emerging facets of retiree benefits. Business & Health. 1988. 5(10): p. 8-12.
5
Poisal JA, Murray L, “Growing differences between Medicare beneficiaries with and without drug
coverage,” Health Affairs, Jan-Feb 2001, 74-85

4

entirely due to prohibitively high drug costs. 6 These findings emphasize the importance
of prescription drug coverage within the Medicare population.

In response to the legislative proposals to add a drug benefit to Medicare, CMS’s Office
of the Actuary and the Congressional Budget Office are regularly asked to make cost
projections for such proposals, many of which rely on survey prescription drug cost and
utilization data. When using survey data for this purpose, several assumptions must be
made to accurately project these costs including adjustments for survey mis-reporting,
institutional drug usage, and the degree to which demand would increase with the passing
of a new benefit.

This paper reports on an attempt to quantify the extent to which prescription drug use and
expenditures are mis-reported in one such survey—The Medicare Current Beneficiary
Survey (MCBS). The MCBS is an ongoing household panel survey of about 13,000
Medicare beneficiaries, funded by the Centers for Medicare and Medicaid Services
(CMS).7 Annually, CMS produces two file series: The Access to Care series and the
Cost and Use series. The Cost and Use series contains data on health care utilization and
expenditures for beneficiaries “ever-enrolled” in Medicare, including persons who
enrolled in the program or died during the year. This series also includes data on nonMedicare covered services such as prescription drugs, as well as data on Medicare
covered services.

Household surveys of health and health expenditures, such as the MCBS, are subject to
non-response and mis-reporting of medical events. 8 As a general rule, health events that
are farther removed in time and those that are less prominent are less likely to be recalled
at the time of interview. 9 Prescription drug purchases are no exception. During each
6

Steinman M, Sands L, Covinsky K, “Self-restriction of medications due to cost in seniors without
prescription coverage,” Journal of General Internal Medicine, Vol. 16, No. 12, Dec. 2001, 793-799
7
Adler G, “A Profile of the Medicare Current Beneficiary Survey,” Health Care Financing Review,
Summer 1994: 153-163
8
Chulis GS, Eppig FJ, “Matching MCBS and Medicare Data: The Best of Both Worlds,” Health Care
Financing Review, Vol. 18, No. 3, 211-229
9
Cohen SB, Burt VL, “Data Collection Frequency Effect in the National Medical Care Expenditure
Survey,” Journal of Economic and Social Measurement, 1985, Vol. 13: 125-151

5

interview, respondents are asked about all of their medication use since their last
interview. Using The MCBS takes several steps in an attempt to minimize recall error by
beneficiaries. For example, respondents are asked to retain and bring to their interview
any prescription bottles, packages, or receipts associated with their medication use. They
are also encouraged to make notes on calendars provided by the survey to record all of
their health care events. Finally, utilizing CAPI (computer-assisted-personal- interview),
MCBS interviewers are furnished with a list of all prescription drugs reported in previous
interviews so they can ask whether the respondent has taken any of those drugs during
most recent reporting period.

However, to date there have been no efforts to assess what mis-reporting occurs in the
wake of these efforts.

Our work provides an answer to that question via a multi-step process. First, we
collected and compared data from a survey of MCBS beneficiaries and their pharmacies.
We then determined the mis-reporting rates for MCBS sample persons for whom we had
complete survey and pharmacy data. Finally, we generalized our results to the entire
non- institutionalized MCBS population through a series of micro-simulation models.
This effort culminated in an estimate of the direction and magnitude of reporting errors as
well as the identification of the social, economic, and demographic correlates of those
errors.

Data

Collecting pharmacy data

To test the extent of misreported prescription drug use and spending in the MCBS, a
pharmacy follow-back study was designed, and conducted in the first four months of
2000,

6

Four types of MCBS respondents were omitted from the study. Respondents who were
institutionalized for all of calendar year 1999 were not asked to participate. Similarly,
persons who lived in the community during 1999 but were institutionalized at the time of
their spring interview were excluded. 10 Respondents for whom a proxy answered and
beneficiaries who were not enrolled in Medicare for all twelve months of 1999 (including
deaths) were also excluded. 11

The remaining survey participants (n=9,384) were asked if they would request patient
profiles from all the pharmacies where they obtained their drugs in 1999. 12 Sample
persons who had not reported any medication use in 1999 were still asked to participate
in the study. In these cases, the beneficiaries were asked to identify the pharmacies that
they would normally use to fill a prescription.

Sample persons who agreed to participate were asked to supply the names and addresses
of every pharmacy they used during 1999. As a means to help beneficiaries recall their
pharmacies, interviewers suggested the use of medicine labels, receipts, phone books, and
pharmacy directories. Beneficiaries who reported no prescription drug use during 1999
were asked to supply the names of pharmacies they normally used. Each respondent was
asked to sign a pre-printed letter requesting a profile of their 1999 drug utilization from
each pharmacy on their list. The letters contained return envelopes addressed to Westat,
Inc, the contractor that administers the MCBS for the Centers for Medicare and Medicaid
Services.

Of those asked to participate in the pharmacy follow-back study (Table 1), more than half
were “complete responses” (meaning not only did they participate, but all of their
reported pharmacies submitted prescription profiles on their behalf). A small percentage
(6%) of those asked refused to participate, and 4% reported no pharmacies. The 8,406
10

Spring (Round 26) interviews are those interviews conducted between the months of January and April
of 2000.
11
There are times when a sample person is unable to participate in the MCBS interview. Where possible,
someone familiar with the beneficiary’s health care utilization and expenditures serves as a proxy and
answers on their behalf.

7

respondents who supplied pharmacy names reported 11,102 pharmacies. Westat, Inc.
received about three-quarters of the requested profiles (8,126), which were entered into
machine-readable format using a computer-assisted data entry system (CADE). About
one- fourth of the beneficiaries had unusable data (missing or invalid dates) turned in
from one or more of their pharmacies. Only respondents for whom all pharmacies
returned usable profiles were examined in this analysis. Thus, the effective response rate
was 57 percent (Table 1).

Table 1. Follow-back study status of sample persons in 1999 Cost and Use file.
Sample
Persons

% of All

13,106

100%

Excluded from the follow back:
A. No 1999 event level drug data collected:
1. In facility for all of 1999
2. New enrollee in 1998 or 1999
B. Didn't receive Round 26 interview
1. Proxy interviews
2. Spring interview was facility interview
3. Deaths and refusals

3,722

28%

Asked to participate in follow-back study

9,384

Total Sample, 1999 Cost & Use

A. Refused
B. No pharmacies reported by beneficiary
C. Reported one or more pharmacies
1. Partial complete (at least one, but not all
pharmacies responded)
2. Pharmacy non-response or unusable data
3. All pharmacies reported usable data

% of Study
participants or
non-participants

100%

946
638

25%
17%

1,162
227
749

31%
6%
20%
72%

100%

570
408

6%
4%

813

9%

2,291
5,302

24%
57%

Including the proper medications from the beneficiary and the pharmacy

12

The total number of respondents in the 1999 MCBS Cost and Use file is 13,106; not all were selected to
participate in the study.

8

A number of editing steps were necessary prior to analysis of the data. From the MCBS,
all beneficiary-reported drug names were standardized, correcting any misspelled words
as well as reformatting drug names. Over-the-counter medications that were reported by
the respondent were dropped. 13

In preparing the pharmacy profile data, profiles were excluded if they contained
prescription drug events with either a missing month or a missing day. Just over 1% of
the profiles received were rejected for unusable or missing dates resulting in 84 sample
persons (found within the 2,291 persons categorized as, “Pharmacy non-response or
unusable data”) being dropped from the analysis. As with the beneficiary-reported data,
all drug names were standardized and any over-the-counter medications were deleted.
All pharmacy-reported data for a given respondent were then concatenated into a single
file.

The next step was to ensure that beneficiary-reported drugs and pharmacy-reported drugs
were from exactly the same time periods. Unlike the pharmacy profile data, beneficiaryreported data do not have recorded dates of purchase: In the standard MCBS interview,
respondents are not asked for this exact date because such a practice would significantly
increase the respondent’s recall burden, particularly when a medication is refilled several
times. Although the survey does not capture dates of drug purchases, it does establish a
recall reference period with specific beginning and ending dates. Drug purchases for
calendar year 1999 were recorded in four rounds of interviews, numbered 23, 24, 25, and
26. Round 23 took place between the months of January and April, 1999. Because the
reference period for any interview is the previous four months, drug purchases recalled in
round 23 could have occurred during the end of 1998 or the beginning of 1999.
Likewise, the round 26 interview took place between January and April of 2000, meaning
recalled drugs could have been purchased in either late 1999 or early 2000. All reported
drugs for rounds 24 and 25 (June-December) were purchased in 1999, therefore, the
survey data analyzed was limited to those rounds.

13

MCBS interviewers are instructed to not collect over-the-counter medications.

9

Including the proper drugs from the pharmacy reports involved a simple process of date
comparisons. For each person, all beneficiary-reported drugs collected in rounds 24 and
25 were included and all pharmacy-reported drugs that fell between the beginning date of
the round 24 reference period and the ending date of the round 25 interview were
included. The results were a total of 101,144 pharmacy-reported drug events and 96,878
survey-reported drug events.

Matching beneficiary-reported medications with pharmacy-reported
medications

An initial attempt was made to match beneficiary-reported drugs to pharmacy-reported
drugs electronically. For each event in the survey-reported file, a variable,
(MATCH_KEY) was created that contained the sample person’s personal identification
code, the drug name, and a sequence number: there was one record per beneficiary, per
drug, per purchase (including refills). For example, for the fictitious beneficiary whose
BASEID was 00001234, the records would read in the following way:
BASEID

DRUGNAME

SEQUENCE

MATCH_KEY

00001234

AMOXIL

001

00001234AMOXIL

001

00001234

CIPRO

001

00001234CIPRO

001

00001234

CIPRO

002

00001234CIPRO

002

00001234

FUROSEMIDE

001

00001234FUROSEMIDE001

00001234

FUROSEMIDE

002

00001234FUROSEMIDE002

00001234

FUROSEMIDE

003

00001234FUROSEMIDE003

00001234

FUROSEMIDE

004

00001234FUROSEMIDE004

The same process was carried out on the pharmacy-report file, and the records from the
two files were matched on the variable MATCH_KEY. The automated merge produced
64,273 matches, 36,871 events that appeared only in the pharmacy file, and 32,605 events
that appeared only in the survey file.

10

Examination of the ‘pharmacy-only’ and ‘survey-only’ records revealed many missed
matches. There were many events in which a generic name was reported by one party
while the brand name was reported by the other. There were other events in which the
drug name was converted into different standardized names. For instance, if a
beneficiary reported the name, ‘Cardizem’, and the pharmacy reported the drug,
‘Cardizem SR’, then the prescriptions would fail to match during the electronic merge.

This manual review of the electronic match improved the agreement between pharmacyreported and survey-reported events in the aggregate. The matched figure increased by
9,246 to 73,519. The pharmacy-only figure fell to 27,625 and the survey-only figure
dropped to 23,359.

Unmatched survey-reported drugs were further classified into one of two categories:
survey-over-reports or omitted-pharmacy-under-reports. A prescription was assigned
survey-over-report status if there were any mentions of that drug in the pharmacy file.
We assumed that survey-over-reports occurred when the beneficiary “telescoped” a refill,
that is, they reported a refill that occurred in an earlier round, or that never occurred at all.
A drug was assigned omitted-pharmacy- under-report status if that drug name did not
appear in the pharmacy data. We assumed that these events occurred because the
beneficiary failed to report all of their pharmacies and that the prescription was filled in
one of these ‘omitted’ pharmacies. It is possible that a fraction of the drugs categorized
as survey-over-reports were, in fact, purchases made at omitted pharmacies.

Methods
Models of mis-reporting of prescription drug utilization (Determining the
number of prescriptions that should have been reported)
We explored three mis-reporting models. For each scenario, the following definitions
apply:
P=Sum of all prescriptions reported by beneficiary’s pharmacies 14
14

We assumed that all pharmacy-reported drugs were reported accurately.

11

M=Number of matched prescriptions
O1=Number of non- matched survey-only prescriptions that were deemed a result of
survey over-reports
O2=Number of non- matched survey-only prescriptions that were deemed a result of
omitted-pharmacy- under-reports
S=All survey reported prescriptions, or M+O1+O2
R=Net adjusted under-reporting rate

The models vary in their assumptions regarding the source and nature of survey-only
events. In the first mis-reporting rate model, all survey-reported drugs are divided by all
pharmacy-reported drugs. Here, sample persons are assumed to have reported
pharmacies completely, and O2 reflects errors of recall on the part of the beneficiary.
MODEL 1: R=1-(S/P)
See under-reporting and over-reporting examples in Appendix A

In the second model, all unmatched sur vey-reported drugs were assumed to have been the
result of an under-reporting of pharmacies. This model was further divided into two
possibilities. In model 2A we assumed that all unmatched drugs were perfectly reported,
that is, that there was no mis-reporting of drugs obtained from the omitted pharmacies.

MODEL 2A: R=1 – (S/ (P+O1+O2))
See under-reporting and over-reporting examples in Appendix A

In model 2B, we modify the assumption about mis-reporting in omitted pharmacies. We
assume that the same reporting percentage observed from the known pharmacy(ies)
occurred in the omitted pharmacies.
MODEL 2B: R=1-(S / (P+((O1+O2) * P/M)))15
See under-reporting and over-reporting examples in Appendix A

15

When the beneficiary’s pharmacy(ies) reported no drugs, the denominator was set to O2, which in this
case is equal to S.

12

The third model combined aspects of the preceding two. In this model, survey-only
events involving drugs encountered in the pharmacy data were considered over-reports
and prescriptions not seen in the pharmacy data were classified as omitted-pharmacyunder-reports. As with model 2, two alternative specifications are possible depending
upon the assumption of mis-reporting in omitted pharmacies. Model 3A assumed perfect
reporting of drugs:

MODEL 3A: R=1-(S / (P+O2))
See under-reporting and over-reporting examples in Appendix A

Model 3B, similar to model 2B, assumed the same rate of over or under-reporting in
omitted pharmacies as in reported pharmacies.
MODEL 3B: R=1-(S / (P+(O2 * P / (M+O1))))16
See under-reporting and over-reporting examples in Appendix A

We adopted model 3B for both our utilization and expenditure analyses, with one
modification. Analysis of the imputations for the constant reporting rate assumption in
this model uncovered some cases where the imputed number of estimated additional
scripts purchased for under-reporters seemed unrealistic. Consequently, we established a
cap on that number of any given beneficiary. First, an unconstrained number of
estimated additional scripts was calculated for each beneficiary using the methodology
described above. Next, the mean (2.54) and standard deviation (8.01) of these additional
prescriptions were tabulated. Finally, for each beneficiary, the number of additional
scripts purchased for an under-reporter was truncated at O2 plus two standard deviations
above the mean of the estimated additional prescriptions purchased.

For instance, assume a respondent reported 32 prescriptions and their pharmacy reported
just 3, of which 1 drug event matched. Further assume that each of the 31 remaining
survey-reported prescriptions are classified as omitted-pharmacy-under-reports.

13

Unconstrained, the model would have resulted in an additional 93 prescriptions from an
omitted pharmacy. Under the constraint described, the number of additional scripts is
capped at (31+20.5) = 52.
Results -- Utilization

The distribution of net adjusted utilization under-reporting rates using model 3B is shown
in Table 2 below:
Table 2: Distribution of Net Adjusted Under-reporting Rates
Percentile of the population Net adjusted mis-reporting
percentage
Maximum

100% under-reported

99th Percentile

100% under-reported

95th Percentile

100% under-reported

90th Percentile

67% under-reported

75th Percentile

36% under-reported

50th Percentile (Median)

10% under-reported

25th Percentile

0

10th Percentile

33% over-reported

5th Percentile

67% over-reported

1st Percentile

200% over-reported

Minimum

799% over-reported

MODE

100% under-reported

The net adjusted utilization under-reporting rates of all models are displayed in Table 3
below.

16

Identical to scenario 2B, when the beneficiary’s pharmacy(ies) reported no drugs, the denominator was
set to O2, which in this case is equal to S.

14

Table 3: Model summaries and their associated formulas and net adjusted underreporting rates

1
2a
2b
3a
3b

Under-reporting Rate Under-reporting Rate
(unweighted)
(weighted)
S/P
4.2%
4.4%
S/ (P+O1+O2)
22.2%
22.2%
S / (P+((O1+O2) * P/M))
27.3%
27.4%
S / (P+O2)
13.3%
13.5%
S / (P+(O2 * P / (M+O1)))
14.7%
14.9%

Note: The net adjusted under-reporting rate for Scenario 3B incorporates the capping procedure described.

Determining what factors are predictive of utilization mis-reporting
A multi-step process was used to determine the overall net adjusted utilization underreporting rate of prescription drugs in the MCBS. First, we analyzed the demographic
data of the follow-back participants to determine those factors that were predictive of a
person’s reporting status (over-, under-, or perfect-reporter). Second, we determined
those factors that were predictive of the degree to which a person under or over reports
their prescription use. After these models were developed, they were applied to those
beneficiaries not in the follow-back study so that an aggregate estimate of underreporting could be made.

We analyzed the demographic characteristics of the 5,302 full pharmacy follow-back
participants to determine which factors were predictive of utilization under-reporting.
The results from the pharmacy follow-back were merged to the 1999 Cost and Use Public
Use File. Several variables were analyzed via multinomial logistic regression to test their
predictive power of a person’s reporting status. Table 4, below, lists the independent
variables and our null hypotheses associated with each.

All of these variables are collected in the survey itself, with the exception of Total
Average Wholesale Price (AWP) and Drug Coverage. Total AWP was estimated by
merging beneficiary-reported drug data with First DataBank’s drug pricing compendium,
Bluebook. Dependent on each drug’s name, form, strength, and prescription size, an

15

average wholesale price is imputed. This imputation scheme is discussed more
thoroughly later in the paper.

Drug coverage is also a derived field. For the purposes of this analysis, drug coverage is
assigned if a beneficiary has at least one month of drug coverage in 1999. Beneficiaries
are categorized as ‘covered’ if one or more of the following occur:
•

Medicare+Choice beneficiaries: They belong to a plan that offers prescription
drug coverage as part of its basic benefit package or they purchase such coverage
via an added premium

•

Medicaid beneficiaries: They are fully entitled, as determined by CMS
administrative data, or they self- identify Medicaid drug coverage

•

Privately insured beneficiaries: They report a private plan (employer-sponsored or
individually purchased) that covers their prescription drugs

•

‘Other public’ insured beneficiaries: They report drug coverage from state-based
pharmaceutical assistance programs, the Veteran’s Administration, the
Department of Defense, or any other public source

•

All beneficiaries: They report any third party reimbursements 17

Table 4: Variables Included in Original Regression Model
Concept

Null Hypothesis

Total Prescriptions

As the number of prescriptions increases,
so does the likelihood of mis-reporting

Total AWP

As the level of expenditures increase, so
does the likelihood of mis-reporting

Age

The disabled and the oldest old would misreport drug use to a greater degree relative
to the youngest old

Race/Ethnicity (White, African-American,

There would be greater mis-reporting of

Other)

drugs among non-whites

17

Very rarely, survey respondents will report that they do not have prescription drug coverage via any
private or public plans, yet they report that a private or public plan made a drug payment on their behalf

16

Health Measures (Self- reported health

Healthier beneficiaries would report their

status, presence of chronic conditions)

drug use more accurately

Income

Higher income enrollees would report their
drug use more accurately

Utilization of Medicare Covered Services

Beneficiaries with increased utilization of
Medicare covered services would misreport more often

Gender

Females would be more likely to misreport their drug use

Drug Coverage

Covered enrollees would mis-report more
often relative to non-covered enrollees

Supplementary Health Insurance

Likelihood is high for differences between

(Medicare+Choice, Medicaid, Employer-

certain types of supplemental insurance

sponsored, Individually- purchased, Other
public, Fee- for-Service Medicare Only

Reporting status was one of three types. Under-reporters were those beneficiaries whose
reported medications totaled less than the estimated number of prescriptions purchased,
where the estimated number of prescriptions purchased equals the sum of pharmacyreported drugs plus imputed pharmacy drugs. Over-reporters were those enrollees who
reported more prescriptions than were estimated purchased. Persons were labeled as
perfect-reporters if their reported drug use matched the total estimated purchased drugs.
The unweighted frequencies of each category among the pharmacy follow-back
participants were 3,037 (57.4%), 1,269 (24%), and 990 (18.7%), respectively.

Study members averaged 365 days in the community during 1999 while non- members
spent just 332 days in the community. 18 In order to adjust for this experience,
prescriptions per beneficiary were standardized to annual figures (ANN_TOTSCRIP) for

17

modeling development purposes by dividing 365 days by each beneficiary’s community
days (C_DAYS). The final equation was as follows:
ANN_TOTSCRIP=((365/C_DAYS)*TOTSCRIP
Note: Variable definitions can be found in Appendix B

Relative to perfect-reporters, ten factors were statistically significantly predictive of
reporting status as determined by the following multinomial logistic regression
equations:19
LogP(O)/P(U)=

Exp(Xiβ m) / Exp(Xiβ n )

LogP(U)/P(P)=

Exp(Xiβ m) / Exp(Xiβ n )

LogP(P)/P(O)=

Exp(Xiβ m) / Exp(Xiβ n )

Where P(U)=Probability of being an under-reporter
P(O)=Probability of being an over-reporter
P(P)=Probability of being a perfect reporter
Subscript u=under-reporter
Subscript o=over-reporter
Subscript p=perfect reporter
T=Annualized total prescriptions
C=Number of chronic conditions
V=Number of Dr. visits
Y=(age 65-79=1, All Others=0)
A=African-American=1, All Others=0
O=Other race=1, All Others=0
R=Drug coverage=1, No coverage=0
18

This variation is explained by the rules for inclusion in the study: Non-participants included
beneficiaries who began receiving Medicare benefits during the year, beneficiaries who died during the
year, and those who moved between facilities and the community.
19
The original model tested the following variables: Number of beneficiary-reported prescriptions, total
AWP, age category, race and ethnicity, health status, number of chronic conditions, income, number of

18

M=Medicare risk plan=1, No risk=0
D=Medicaid=1, No Medicaid=0
I=Income

Not all variables were significant for both over-reporters and under-reporters. The
relative risk ratios are shown in Appendix C.

Generalizing the Results to the Full Population
To generalize these results to the population as a whole, the multinomial regression
model described above was used to impute reporting status. Three new variables were
created (Prob_Under, Prob_Over, and Prob_Perfect) dependent on the values of every
sample person’s demographics. These fields represented probabilities and combined,
they summed to 1. Second, a random number between 0 and 1 was assigned to each
beneficiary. A person’s reporting status was determined based on a comparison of that
number to the three probability variables. None of the beneficiaries in the pharmacy
follow-back had their reporting status changed. Tables 5 and 6 illustrate the mechanics
of the imputation and their corresponding results:20

Table 5: Imputation of reporting status variable
BASEID
00000001 00000002
RANDOM
0.75223 0.421134
PROB_UNDER
0.5509
0.5023
PROB_OVER
0.2203
0.2115
PROB_PERFECT
0.2288
0.2862
REPORT_STATUS
OVER
UNDER

inpatient hospitalizations, number of doctor visits, number of home health visits, number of outpatient
procedures, gender, prescription drug coverage status, supplementary health insurance status
20
Application of the model of reporting status to the total population resulted in trivial changes in the
relative share of over and under-reporters.

19

Table 6: Detailed weighted statistics of reporting status variable, pre and post
imputation
Status

Follow-back Imputed
Full Sample
Percent
Percent
Percent
Under-reporters
57.2%
55.7%
56.3%
Over-reporters
23.8%
23.1%
20.3%
Perfect reporters
19.1%
21.2%
23.4%

Following the assignment of reporting status, we estimated the degree to which
respondents either under- or over-reported their drug events. As not all variables were
predictive of a person’s reporting status, separate models were developed to estimate the
magnitude of under-reporting and over-reporting within their respective categories.
Using only follow-back participants, an inflation factor was computed at the person level
to annualize the number of prescriptions that had been identified as either under or overreported. The factor was computed by dividing the total number of prescriptions in the
public use file by the number of survey-reported prescriptions in the follow-back study.
Given that the public use file represents one full year, and the follow-back analysis was
limited to two rounds, on average, the inflation factor was about 1.5. The result was two
new variables, UNDER365 and OVER365, where the former represented the annualized
number of under-reported drugs for beneficiaries who were identified as under-reporters
and the later represented the annualized number of over-reported prescriptions for
beneficiaries identified as over-reporters. The table below demonstrates.
Table 7: Example of calculations to estimate annualized under and over-reported
medications
BASEID
00000100 00000200 00000300
REPORT_STATUS
OVER
UNDER PERFECT
TOTSCRIP
25
30
10
MTOT
20
20
6
FACTOR
1.25
1.50
1.67
U_EVNTS
0
10
0
O_EVNTS
4
0
0
UNDER365
0
15
0
OVER365
5
0
0

20

Limited to under-reporting follow-back participants, a multi- linear regression model was
developed to determine which factors were predictive of the number of annualized underreported prescriptions. The final equation follows:
UNDER365=α u Ti+χu Wi+δ uH i+ε uV i+φ uY i+ηu Ai+ϕuR i+κuE i+λuP i+ ß
Where Subscript u=under-reporter
T=Annualized total prescriptions
W=Annualized total AWP
H=Health Status
V=Number of Dr. visits
Y=(age 65-79=1, All Others=0)
A= African-American=1, All Others=0
R= Drug coverage=1, No coverage=0
E=Employer Sponsored Insurance=1, No Employer Sponsored
Insurance =0
P=Individually Purchased Insurance=1, No Individually Purchased
Insurance =0
Note: The output from the model can be found in Appendix D.

Next, the model was used to predict a new variable, UNDER_RX, for all non-study
members for whom an under-reporting figure had to be imputed. For the study’s
participants, UNDER_RX equaled UNDER365.

An identical process was followed for predicting the annualized number of over-reported
drugs among over-reporters in the follow-back. That model’s equation follows:
OVER365=α oTi+χo C+δ oI +ε oFi+φ o Ai+λoO i+ ß
Where Subscript o=over-reporter
T=Annualized total prescriptions
C=Number of chronic conditions
I=Income
F=Female=1, Male=0
A=African-American=1, All Others=0

21

O=Other Ethnicity=1, All Others=0
Note: The output from the model can be found in Appendix E.

The variable OVER_RX, was assigned to all those deemed to be over-reporters. Like the
under-reporters, the value for OVER_RX for follow-back participants was set equal to
OVER365.

Calculating the final utilization mis -reporting rate
The next step was to transform the annualized number of over- or under-reported drugs to
reflect the actual experience of each individual beneficiary. Reduction ratios were
developed by dividing the number of prescriptions reported by the beneficiaries
(TOTSCRIP) by their annualized number of prescriptions (ANN_TOTSCRIP). Table 8
illustrates how the annualized number of over- and under-reported medications were then
multiplied by these ratios resulting in the variables, REAL_UNDER and REAL_OVER.
Table 8: Example of transforming annualized estimates into actual estimates
BASEID
00000100 00000200 00000300
REPORT_STATUS
OVER
UNDER PERFECT
TOTSCRIP
25
30
10
ANN_TOTSCRIP
30
40
15
RATIO
0.83
0.75
0.67
UNDER365
0
15
0
OVER365
5
0
0
REAL_UNDER
0
11.25
0
REAL_OVER
4.17
0
0

The variable EST_PURCH (Table 9) was created to represent the number of actual
prescriptions believed to have been purchased by the beneficiary and was calculated
using the following equation:
EST_PURCH=TOTSCRIP+REAL_UNDER-REAL_OVER

22

Table 9: Example of EST_PURCH variable calculation
BASEID
00000100 00000200 00000300
REPORT_STATUS
OVER
UNDER PERFECT
TOTSCRIP
25
30
10
ANN_TOTSCRIP
30
40
15
RATIO
0.83
0.75
0.67
UNDER365
0
15
0
OVER365
5
0
0
REAL_UNDER
0
11
0
REAL_OVER
4
0
0
EST_PURCH
21
41
10

To determine the final net adjusted utilization under-reporting ratio, weighted
calculations of TOTSCRIP and EST_PURCH were taken resulting in a final underreporting estimate of 17.7%.

Results – AWP Expenditures
There was a significant amount of overlap between the methods used to determine the
utilization mis-reporting estimate for prescription drugs in the MCBS and those used to
determine the mis-reporting estimate for medication expenditures in the survey. The
methodologies are identical up to and including the selection of model 3B. From there,
several other steps were undertaken.

In order to estimate expenditure mis-reporting, all survey-reported and pharmacyreported events were electronically passed through a published industry source (First
DataBank’s Bluebook) that assigns a unit average wholesale price (AWP) to each
prescription. The imputation algorithm attempts to match on as many characteristics of
the drug as possible, including drug name, drug form, drug strength, and prescription
size. In cases where form or strength were not collected in the survey, values for those
fields are imputed using probabilities that are proportionate to their relative use among
the dual-eligible population. CMS has drug utilization data at the National Drug Code
(NDC) level for dually-eligible beneficiaries. 21 These data are merged by NDC to the

21

One of the limitations in this study is the use of drug utilization data by dual-eligible beneficiaries to
estimate the relative usage of forms and strengths within drug names for imputation purposes. IMS Health

23

First DataBank file to provide relative usage counts for the various forms, strengths, and
package sizes within each drug. Once a unit AWP is assigned, that value is multiplied by
the prescription size to estimate an “event price”, or total AWP for that prescription. If
we were unable to match on the name of the drug, or if we weren’t given a prescription
size on which to multiply the assigned unit AWP, we didn’t ascribe an event price during
this phase of the imputation.

Table 10: Results of Preliminary AWP Imputation by Reporting Status

Events
Translatable Events (name can be matched to First DataBank)
Event Prices Assigned (Unit AWP can be multiplied by prescription size)
Average Event Price

SurveyPharmacyReported Reported
96,882
101,149
93,051
98,475
84,926
83,405
$ 50.89 $ 50.00

In order to fill in missing event prices, a two-step imputation process was implemented
separately for both the survey-reported and pharmacy-reported data. First, for each of the
two groups, average event prices were calculated for all unique drug names. These
averages were then merged, by drug name, back onto their respective files. This left 588
drug events with missing event prices in the survey file and 3,042 drug events with
missing event prices in the pharmacy file. Those remaining were assigned an event price
equal to the average of all other drugs in their respective file (survey- or pharmacyreported).

Table 11: Results of Final AWP Imputation by Reporting Status

also offers recent drug utilization figures by National Drug Code. We chose to use CMS administrative
data for cost reasons.

24

Event Prices Assigned Via First DataBank
Event Price Imputed via Individual Drug Name Averages
Event Prices Imputed using Overall Event Price Average
Average Event Price Following all Imputation

SurveyPharmacyReported Reported
84,926
83,405
11,368
14,702
588
3,042
$ 50.31 $
49.62

In order to pass the results of our imputation through model 3B, all survey-reported and
pharmacy-reported drugs needed to be organized into the same categories as described in
the model (P, M, O1, O2). Unlike the utilization analysis, these distinctions needed to be
made electronically. This was accomplished in a multi-step process. First, all of the
survey-only and pharmacy-only drugs that didn’t match electronically (for the reasons
described in the section, “Including the proper medications from the beneficiary and the
pharmacy”) were downloaded to a flat file. For each beneficiary, every manually
matched survey-only and pharmacy-only drug was flagged and the data were uploaded
again. Next, an unduplicated list of all of the beneficiary’s electronically and manually
matched drug names was created. Then, all of the beneficiary’s non- matched surveyreported drugs were electronically compared to that list. Drug names that matched during
that comparison were classified as survey-over reports while those that didn’t were
flagged as omitted-pharmacy under-reports.

The next step was to tabulate the number of drug mentions that were categorized into the
matched (M), survey over-report (O1), and omitted-pharmacy under-report (O2)
categories. We then compared these figures to their corresponding estimates that resulted
from our manual classifications prepared during our utilization mis-reporting estimation
process. All sample persons with matching estimates were considered ready for
expenditure analysis and were output to a file. All of the survey-reported drug mentions
for the remaining beneficiaries were then downloaded to a flat file. Once more, through a
manual process, these events were compared to the electronically and manually matched
survey drugs to determine the category (O1, O2) in which the event belonged. All drugs
that were deemed to be omitted-pharmacy under-reports were flagged and the data were
uploaded again. At this point, we could accurately electronically identify for each sample

25

person all of their drugs that were matched, pharmacy-only, survey over-reports, and
omitted-pharmacy under-reports.

Their averages are listed in Table 12 below.
Table 12: Unweighted and Weighted Results of AWP Imputation by Category 22
SURVEY-REPORTED
All Matched
Electronically Matched
Manually Matched
Unmatched
Survey Over-reports
Pharmacy Under-reports
PHARMACY-REPORTED
All Matched
Electronically Matched
Manually Matched
Unmatched

Number of
Unweighted
Weighted
Scripts
Average AWP Average AWP
73,519 $
49.30 $
49.48
64,273 $
50.30 $
50.41
9,246 $
42.34 $
42.85
23,359 $
53.45 $
53.62
12,779 $
50.95 $
51.54
10,580 $
56.48 $
56.09
Number of
Unweighted
Weighted
Scripts
Average AWP Average AWP
73,519 $
50.21 $
50.39
64,273 $
50.64 $
50.79
9,246 $
47.20 $
47.56
27,625 $
48.04 $
48.41

After classifying and pricing each drug, the next step was to calculate the total surveyreported AWP expenditures and the total adjusted pharmacy-reported AWP expenditures
at the person- level. The formula for determining total adjusted pharmacy-reported AWP
was contingent on whether the beneficiary was deemed to have fully reported all of their
pharmacies. Where pharmacies were fully reported, the following equation was used:

PHARMAWP = M * PharmMat$ + (P-M) * PharmOnly$
Where:
PharmMat$=Average AWP of beneficiary’s matched pharmacy-reported drugs
PharmOnly$= Average AWP of beneficiary’s non-matched pharmacy-reported drugs

For beneficiaries for whom pharmacies were deemed under-reported, total adjusted
pharmacy-reported AWP expenditures were estimated with the following formula:
22

For each category of drugs, the Weighted Average AWP was calculated by multiplying the AWP of each
drug by that person’s MCBS sampling weight, summing those results, and then dividing by the sum of
those weights.

26

PHARMAWP = M * PharmMat$ + (P-M) * PharmOnly$ + ((O2 * P / (M+O1) * O2$
Where:
PharmMat$=Average AWP of beneficiary’s matched pharmacy-reported drugs
PharmOnly$= Average AWP of beneficiary’s non-matched pharmacy-reported drugs
O2$=Average AWP of beneficiary’s omitted-pharmacy under-reported drugs
Calculating the total survey-reported AWP expenditures for all beneficiaries required just
one formula:

SURVAWP = M * SurvMat$ + O1 * O1$ + O2 * O2$

Where:
SurvMat$=Average AWP of beneficiary’s matched survey-reported drugs
O1$= Average AWP of beneficiary’s survey over-reports
O2$=Average AWP of beneficiary’s omitted-pharmacy under-reported drugs

In preparation for the modeling phase of the estimate, total survey-reported AWP
expenditures had to be estimated for all of the non- followback participants. For this
cohort, event prices were derived by passing their reported prescriptions through First
DataBank’s Bluebook using the identical ‘event price’ algorithm described earlier. Drugs
with missing AWP values following the Bluebook imputation were imputed by
separately calculating average AWP by drug name and merging those values back onto
the file.

A SAS-transport dataset was created containing 11,141 records (one per person). Each
record contained BASEID, PHARMAWP, and SURVAWP. For non- followback
participants, PHARMAWP was missing. This file was downloaded and converted to a
Stata file.

Determining what factors are predictive of expenditure mis-reporting

27

Similar to the utilization mis-reporting method, a multi- step process was used to
determine the overall net adjusted under-reporting of prescription drug expenditures in
the MCBS. We began by analyzing the demographic data of the follow-back participants
to determine those factors that were predictive of a person’s AWP expenditure reporting
status. Next, we determined those factors that were predictive of the degree to which a
person under- or over-reports their prescription expenditures. After these models were
developed, they were applied to those beneficiaries not in the follow-back study so that
an aggregate estimate of expenditure mis-reporting could be made.

We re-analyzed the demographic characteristics of the pharmacy follow-back participants
to establish which factors, if any, were predictive of expenditure mis-reporting. Multiple
variables were examined using multinomial logistic regression to test their predictive
power of a person’s expenditure reporting status. Table 13, below, lists the independent
variables and our null hypotheses associated with each.

Table 13: Variables Included in Original Regression Model
Concept

Null Hypothesis

Total Prescriptions

As the number of prescriptions increases,
so does the likelihood of mis-reporting

Total AWP

As the level of expenditures increase, so
does the likelihood of mis-reporting

Age

The disabled and the oldest old would misreport drug expenditures to a greater degree
relative to the youngest old

Race/Ethnicity (White, African-American,

There would be greater mis-reporting of

Other)

drug expenditures among non-whites

Health Measures (Self- reported Health

Healthier beneficiaries would report their

Status, presence of chronic conditions)

drug expenditures more accurately

Income

Higher income enrollees would report their
drug expenditures more accurately

28

Utilization of Medicare Covered Services

Beneficiaries with increased utilization of
Medicare covered services would misreport more often

Gender

Females would be more likely to misreport their drug expenses

Drug Coverage

Covered enrollees would mis-report more
often relative to non-covered enrollees

Supplementary Health Insurance

Likelihood is high for differences between
certain types of supplemental insurance

Reporting status was one of three types. Under-reporters were defined as those
beneficiaries whose reported total AWP expenditures totaled less than the estimated total
AWP expenditures. Over-reporters were those enrollees who reported total AWP
expenditures greater than was estimated from pharmacies. Persons were labeled as
perfect-reporters if their reported AWP drug expenditures matched the total estimated
AWP expenses. The unweighted frequencies of each category among the pharmacy
follow-back participants were 3,221 (60.8%), 1,564 (29.5%), and 511 (9.7%),
respectively.

Not unlike use, there were significant expenditure differences between the follow-back
participants and non-participants. The weighted average total AWP for the follow-back
participant was $1,364.77, or 16% higher than the $1,173.22 estimated for nonparticipants. Resembling the method to annualize prescriptions described in the
utilization mis-reporting model, total AWP expenses were standardized to annual figures
(ANN_TOTAWP) for modeling purposes. The exact equation was as follows:
ANN_TOTAWP=((365/C_DAYS)*TOTAWP
Note: Variable definitions can be found in Appendix B

29

Relative to perfect-reporters, there were ten factors that were statistically significantly
predictive of reporting status as determined by the following multinomial logistic
regression equations:23
LogP(O)/P(U)=

Exp(Xiβ m) / Exp(Xiβ n )

LogP(U)/P(P)=

Exp(Xiβ m) / Exp(Xiβ n )

LogP(P)/P(O)=

Exp(Xiβ m) / Exp(Xiβ n )

Where :
P(U)=Probability of being an under-reporter
P(O)=Probability of being an over-reporter
P(P)=Probability of being a perfect reporter
Subscript u=under-reporter
Subscript o=over-reporter
Subscript p=perfect reporter
T=Annualized total prescriptions
W=Annualized total AWP
G=Level of self-reported health status
C=Number of chronic conditions
V=Number of Dr. visits
O=Other race=1, All Others=0
M=Medicare risk plan=1, No risk plan=0
D=Medicaid=1, No Medicaid=0
E=Employer-sponsored=1, No Employer-sponsored=0
F=Female=1, Male=0
23

The original model tested the following variables: Number of beneficiary-reported prescriptions, total
AWP, age category, ethnicity, health status, number of chronic conditions, income, number of inpatient
hospitalizations, number of doctor visits, number of home health visits, number of outpatient procedures,
gender, prescription drug coverage status, supplementary health insurance status

30

Not all variables were significant for both over-reporters and under-reporters. The
relative risk ratios are shown in Appendix F.

Generalizing the Results to the Full Population
To apply these results to the entire population, the multinomial regression model
described above was used to impute reporting status. Three new variables were created
(Prob_Under$, Prob_Over$, and Prob_Perfect$) dependent on the values of every sample
person’s demographics. These fields represented probabilities and combined, they
summed to 1. Second, a random number between 0 and 1 was assigned to each
beneficiary. A person’s reporting status was determined based on a comparison of that
number to the three probability variables. None of the beneficiaries in the pharmacy
follow-back had their reporting status changed. Tables 14 and 15 illustrate the mechanics
of the imputation and their corresponding results:24

Table 14: Imputation of AWP expenditure reporting status variable
BASEID
RANDOM
PROB_UNDER$
PROB_OVER$
PROB_PERFECT$
REPORT_STATUS

00000001 00000002
0.75223 0.421134
0.5509
0.5023
0.2203
0.2115
0.2288
0.2862
OVER
UNDER

Table 15: Detailed weighted statistics of AWP expenditure reporting status variable,
pre and post imputation
Status

Follow-back Imputed
Full Sample
Percent
Percent
Percent
Under-reporters
60.9%
59.0%
59.8%
Over-reporters
29.3%
28.0%
28.5%
Perfect reporters
9.8%
13.0%
11.7%

31

Following the assignment of reporting status, we estimated the degree to which
respondents either under- or over-reported their drug expenses using separate models for
each category.

Using only follow-back participants, an inflation factor was computed at

the person level to annualize the level of AWP expenditures that had been identified as
either under- or over-reported. The factor was computed by dividing the sum of AWP
expenses estimated for all of the beneficiary’s purchases by the sum of their AWP
expenditures from the follow-back study. Given that follow-back participants had
approximately a full year of community exposure, and the follow-back analysis was
limited to two rounds, on average, the inflation factor was about 1.5. The result was two
new variables, UNDER365$ and OVER365$, where the first represented the annualized
level of AWP expenditures for beneficiaries who were identified as under-reporters and
the second represented the annualized level of AWP expenditures for beneficiaries
identified as over-reporters. The following table illustrates the mechanic of this
operation:
Table 16: Example of calculations to estimate annualized under and over-reported
AWP expenditures
BASEID
REPORT_STATUS
Full Year AWP
SURVAWP
FACTOR
U_AWP
O_AWP
UNDER365$
OVER365$

00000100
OVER
$ 1,000.00
$
800.00
1.25
$
$
200.00
$
$
250.00

00000200
UNDER
$ 1,200.00
$
800.00
1.50
$
400.00
$
$
600.00
$
-

00000300
PERFECT
$ 400.00
$ 240.00
1.67
$
$
$
$
-

Restricted to under-reporting follow-back participants, a multi- linear regression model
was developed to determine which factors were predictive of the level of annualized
under-reported AWP. The final equation follows:
UNDER365$=α u Wi+δ uCi +χ uV i+ε uY i +λu Li+φ uR i+ηu Ai+ϕu Zi+ ß

24

Application of the model of reporting status to the total population resulted in trivial changes in the
relative share of over and under-reporters.

32

Where:
Subscript u=under-reporter
W=Annualized AWP
C=Number of chronic conditions
V=Number of Dr. visits
Y=Ages 65-79=1, All Others=0
L= Age 80+=1, All Others=0
R=Drug coverage=1, No drug coverage=0
A=African-American=1, All Others=0
Z=Other public coverage=1, No other public coverage=0
Note: The output from the model can be found in Appendix G.

Next, the model was used to predict a new variable, UNDER_RX$, for all non-study
members designated to be under-reporters. For the study’s participants, UNDER_RX$
equaled UNDER365$.

An identical process was followed for predicting the annualized level of over-reported
AWP expenses among over-reporters in the follow-back study. That model’s equation
follows:
OVER365$=α oWi+χo C+ε oY i+φ oLi+λ oHH i+ ß
Where:
Subscript o=over-reporter
W=Annualized AWP
C=Number of chronic conditions
Y=Ages 65-79=1, All Others=0
L= Age 80+=1, All Others=0
HH=Number of home health visits
Note: The output from the model can be found in Appendix H.

33

The variable OVER_RX$, was assigned to all those deemed to be over-reporters. Like
the under-reporters, the value for OVER_RX$ for follow-back participants was set equal
to OVER365$.

Calculating the final AWP expenditure mis-reporting rate
The next step was to convert the annualized level of over or under-reported expenditures
to reveal the actual experience of each individual beneficiary. Reduction ratios were
developed by dividing the level of AWP expenditures reported by the sample persons
(TOTAWP) by their annualized AWP expenditures (ANN_TOTAWP). Table 17
illustrates how the annualized levels of under- and over-reported AWP were then
multiplied by these ratios resulting in the variables, REAL_UNDER$ and
REAL_OVER$.
Table 17: Example of transforming annualized AWP estimates into actual AWP
estimates
BASEID
REPORT_STATUS
TOTAWP
ANN_TOTAWP
RATIO
UNDER365$
OVER365$
REAL_UNDER$
REAL_OVER$

00000100
OVER
$ 1,000.00
$ 1,200.00
0.83
$
$
200.00
$
$
166.67

00000200
UNDER
$ 1,200.00
$ 1,600.00
0.75
$
600.00
$
$
450.00
$
-

00000300
PERFECT
$ 400.00
$ 600.00
0.67
$
$
$
$
-

The variable EST_PURCH$ (Table 18) was created to represent the level of AWP
expenditures believed to have been purchased by the beneficiary and was calculated
using the following equation:
EST_PURCH$=TOTAWP+REAL_UNDER$-REAL_OVER$

34

Table 18: Example of EST_PURCH$ variable calculation
BASEID
REPORT_STATUS
TOTAWP
ANN_TOTAWP
RATIO
UNDER365$
OVER365$
REAL_UNDER$
REAL_OVER$
EST_PURCH$

00000100
OVER
$ 1,000.00
$ 1,200.00
0.83
$
$
200.00
$
$
166.67
$
833.33

00000200
UNDER
$ 1,200.00
$ 1,600.00
0.75
$
600.00
$
$
450.00
$
$ 1,650.00

00000300
PERFECT
$ 400.00
$ 600.00
0.67
$
$
$
$
$ 400.00

To determine the final net adjusted expenditure under-reporting ratio, weighted
calculations of TOTAWP and EST_PURCH$ were taken resulting in a final underreporting estimate of 17%.

Discussion
This analysis addresses the important issue of adjusting survey-reported drug use and
expenditure data to account for survey under-reporting. It has demonstrated several
critical findings regarding prescription drug data collection among the Medicare elderly
and disabled populations, including:
Utilization
•

Medicare beneficiaries, on average, under-report 17.7% of their drug utilization,
as measured in number of prescriptions filled or refilled (23.3 – Reported, 28.3 –
Estimated).

•

Many beneficiaries actually over-report their drug utilization.

•

Adjusted for under-reporting (Table 19), the data show that approximately 25% of
Medicare beneficiaries filled more than 40 prescriptions in 1999.

Table 19: Percentile Distributions of Reported and Estimated Drug Use
Percentile Estimate Prescriptions /Beneficiary Reported Prescriptions/Beneficiary
5%
0.0
0.0
10%
0.0
0.0
25%
9.0
6.0
50%
21.5
17.0

35

75%
90%
95%

•

40.3
63.7
80.4

34.0
54.0
70.0

The probability of mis-reporting drug use increases with increased utilization, as
well as increases in the number of chronic conditions.

•

The most accurate utilization reporters (Table 20) filled between 5 and 10
prescriptions in 1999.

•

Although the number of mis-reported drugs increases with utilization, the rate at
which they are mis-reported (Table 20) is relatively consistent following the
passing of the 15-prescription threshold.

Table 20: Mis-reporting Rates by Estimated Prescriptions per Beneficiary Category
Estimated Utilization Category Estimated Prescriptions/Beneficiary Reported
<=5
1.1
5.1-10
7.8
10.1-15
12.6
15.1-20
17.7
20.1-25
22.5
25.1-30
27.5
30.1-35
32.5
35.1-40
37.5
40.1-45
42.5
45.1-50
47.6
50.1-55
52.5
55.1-60
57.4
>60
84.8

•

Prescriptions/Beneficiary Percent Under- / Over-Reported
1.4
27.32% Over-reported
6.8
12.08% Under-reported
10.1
20.22% Under-reported
14.5
18.08% Under-reported
18.3
18.83% Under-reported
22.9
16.64% Under-reported
26.9
17.27% Under-reported
31.2
16.68% Under-reported
34.7
18.31% Under-reported
38.4
19.36% Under-reported
42.7
18.66% Under-reported
48.5
15.46% Under-reported
68.9
18.71% Under-reported

Among utilization over-reporters, heavy drug users and minorities tend to overreport to a greater degree.

•

Among utilization under-reporters, the number of medication purchases that
beneficiaries under-report increases with an increasing number of physician visits,
but decreases for those who are privately insured.

Expenditures
•

Medicare beneficiaries, on average, under-report 17% of their drug expenses
($1,253.25 – Reported, $1,510.23 – Estimated).

•

Many beneficiaries over-report their prescription drug spending.

36

•

The probability of mis-reporting drug spending (as measured by total AWP)
increases with increases in expenditures, as well as being enrolled in a
Medicare+Choice plan or Medicaid.

•

Analysis of the percentile distributions from Table 21 shows that, when adjusted
for under-reporting, the median spending level exceeds $1,000, up from an
unadjusted figure of $809.

Table 21: Percentile Distributions of Reported and Estimated AWP Expenditures
Percentile Estimated Total AWP Reported Total AWP
5%
$
0$
0
10%
$
0$
0
25%
$
326.85 $
225.03
50%
$
1,028.20 $
809.62
75%
$
2,108.99 $
1,720.67
90%
$
3,468.01 $
2,891.03
95%
$
4,595.98 $
3,936.13

•

The most accurate expenditure reporters, as shown in Table 22, tend to be those
beneficiaries who purchased between $250 and $500 in drugs, as measured by
AWP expenses.

Table 22: Mis-reporting Rates by Estimated AWP Expenditure Level Category
Average
Estimated AWP Category Average Estimated AWP Reported AWP Percent Under- / Over-Reported
<=$250
$
35.09
$
96.61
175.3% Over-reported
$251-500
$
374.93
$
338.91
9.6% Under-reported
$501-750
$
621.85
$
535.95
13.8% Under-reported
$751-1000
$
873.00
$
715.22
18.1% Under-reported
$1,001-1,250
$
1,122.71
$
930.05
17.2% Under-reported
$1,251-1,500
$
1,369.40
$
1,103.79
19.4% Under-reported
$1,501-1,750
$
1,623.46
$
1,309.45
19.3% Under-reported
$1,751-2,000
$
1,868.18
$
1,530.59
18.1% Under-reported
$2,001-2,250
$
2,125.82
$
1,673.73
21.3% Under-reported
$2,251-2,500
$
2,369.07
$
1,934.03
18.4% Under-reported
>$2,501
$
4,258.02
$
3,488.21
18.1% Under-reported

•

Among expenditure over-reporters, being an aged beneficiary mitigates the degree
to which you over-report.

37

•

The amount of expenses that beneficiaries under-report increases with an
increasing number of physician visits, among expenditure under-reporters.

•

Beneficiaries frequently report incomplete drug names (eg. Cardizem instead of
Cardizem CR) leading to drug cost estimates that are below the actual total
expenditure level . This more than offsets the practice of inadvertently reporting
more expensive brand name drugs when beneficiaries, in fact, received a less
expensive generic drug.

•

With respect to average drug prices, Medicare beneficiaries have a propensity to,
a) report drug purchases that were, to some extent, higher in cost, and b) not
report drug purchases that were somewhat less expensive, marginally offsetting
their recall error rate.

Finally, prior to this analysis, the Information and Methods Group (IMG) recommended
using the under-reporting estimate for physician visits (30%) as a proxy for the underreporting level related to prescription drugs. 25 Adjusted for a net-expenditure underreporting rate of 17%, 1999 MCBS data indicate that outpatient prescription drug
spending among the non- institutionalized Medicare population totaled approximately
46.7 billion dollars. Given that level of expenditure, being precise with respect to
assumptions made regarding survey mis-reporting takes on added significance as just a
one percentage point difference in the under-reporting level estimate can change the total
projected annual outlays by nearly 570 million dollars.

We believe the analysis

described here will help inform policy- makers and contribute to improved accuracy of
cost estimates of various prescription drug legislative proposals.

25

IMG is located within the Office of Research, Development, and Information and is responsible for
maintaining and analyzing the MCBS.

38

APPENDIX A
Description of Various Under- and Over-reporting Models

MODEL 1: UNDER-REPORTING EXAMPLE
Survey-reported drugs: 12
Pharmacy-reported drugs: 16
Matched drugs: 8
Net Adjusted Under-reporting rate = 1-(12 / 16) = .25 = 25%
MODEL 1: OVER-REPORTING EXAMPLE
Survey-reported drugs: 22
Pharmacy-reported drugs: 12
Matched drugs: 10
Net Adjusted Under-reporting rate = 1-(22 / 12) = -.833 = -83%

MODEL 2A: UNDER-REPORTING EXAMPLE
Survey-reported drugs: 12
Pharmacy-reported drugs: 16
Matched drugs: 8
Survey-over-reports: 1
Omitted-pharmacy-under-reports: 3
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 1 + 3 = 4
Net Adjusted Under-reporting rate = 1-(12 / (16+ 4)) = .40 = 40%
MODEL 2A: OVER-REPORTING EXAMPLE
Survey-reported drugs: 22
Pharmacy-reported drugs: 12
Matched drugs: 10
Survey-over-reports: 10
Omitted-pharmacy-under-reports: 2
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 10+2 = 12

39

Net Adjusted Under-reporting rate = 1-(22 / (12+12)) = .083 = 8.3%

MODEL 2B: UNDER-REPORTING EXAMPLE
Survey-reported drugs: 12
Pharmacy-reported drugs: 16
Matched drugs: 8
Survey-over-reports: 1
Omitted-pharmacy-under-reports: 3
Reporting ratio for known pharmacy: 16 / 8 =2
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 4 * 2 = 8
Net Adjusted Under-reporting rate = 1-(12 / (16+ 8)) = .50 = 50%
MODEL 2B: OVER-REPORTING EXAMPLE
Survey-reported drugs: 22
Pharmacy-reported drugs: 12
Matched drugs: 10
Survey-over-reports: 10
Omitted-pharmacy-under-reports: 2
Reporting ratio for known pharmacy: 12 / 10 =1.20
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 12 * 1.2 = 14.4 = 14
Net Adjusted Under-reporting rate = 1-(22 / (12+14)) = .154 = 15.4%

MODEL 3A: UNDER-REPORTING EXAMPLE
Survey-reported drugs: 12
Pharmacy-reported drugs: 16
Matched drugs: 8
Survey-over-reports: 1
Omitted-pharmacy-under-reports: 3
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 3

40

Net Adjusted Under-reporting rate = 1-(12 / (16+3)) = .368 = 36.8%
MODEL 3A: OVER-REPORTING EXAMPLE
Survey-reported drugs: 22
Pharmacy-reported drugs: 12
Matched drugs: 10
Survey-over-reports: 10
Omitted-pharmacy-under-reports: 2
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 2
Net Adjusted Under-reporting rate = 1-(22 / (12+2)) = -.571 = -57.1%

MODEL 3B: UNDER-REPORTING EXAMPLE
Survey-reported drugs: 12
Pharmacy-reported drugs: 16
Matched drugs: 8
Survey-over-reports: 1
Omitted-pharmacy-under-reports: 3
Reporting ratio for known pharmacy: 16 / (8+1) =1.78
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 3 * 1.78 = 5.34 = 5
Net Adjusted Under-reporting rate = 1-(12 / (16+5)) = .428 = 42.8%
MODEL 3B: OVER-REPORTING EXAMPLE
Survey-reported drugs: 22
Pharmacy-reported drugs: 12
Matched drugs: 10
Survey-over-reports: 10
Omitted-pharmacy-under-reports: 2
Reporting ratio for known pharmacy: 12 / (10+10) =.60
Omitted pharmacy’s expected number of reported drugs (estimated additional scripts
purchased): = 2 * .60 = 1.20 = 1
Net Adjusted Under-reporting rate = 1-(22 / (12+1)) = -.692 = -69.2%

41

APPENDIX B
Description of Variables na mes and their sources
Variable Name

Description

Source

ANN_TOTAWP

Annualized Average Wholesale

Generated

Price (AWP)
ANN_TOTSCRIP

Annualized TOTSCRIP

Generated

BASEID

Unique beneficiary identifier

Cost and Use, 1999 and
Pharmacy Follow-back Study

BI_AF_AM

Binary, African-American

Cost and Use, 1999

BI_CAID

Binary, Medicaid insurance

Cost and Use, 1999

BI_EMP

Binary, Employer-sponsored

Cost and Use, 1999

insurance
BI_OLD

Binary, ages 80 and up

Cost and Use, 1999

BI-OP

Binary, Other Public insurance

Cost and Use, 1999

BI_OTHER

Binary, Other race (not White or

Cost and Use, 1999

African-American)
BI_PHI

Binary, Individually-purchased

Cost and Use, 1999

insurance
BI_RISK

Binary, Medicare+Choice

Cost and Use, 1999

insurance
BI_RXCOV

Binary, prescription drug

Cost and Use, 1999

coverage
BI_YOUNG

Binary, ages 65-79

Cost and Use, 1999

C_DAYS

Number of days in community,

Cost and Use, 1999

1999
CHRONIC

Number of chronic conditions

Cost and Use, 1999

DRVISITS

Number of doctor visits

Cost and Use, 1999

EST_PURCH

Estimated number of

Generated

prescriptions believed to have
been purchased by the beneficiary
in 1999
EST_PURCH$

Estimated level of AWP

Generated

expenditures believed to have
been purchased by the beneficiary
in 1999
FACTOR

Inflation factor, defined as

42

Generated

(TOTSCRIP/MTOT)
HEALTH

Self-reported health of

Cost and Use, 1999

beneficiary (ordinal scale, 1-5)
MTOT

Total number of beneficiary

Pharmacy Follow-back Study

reported prescriptions in Rounds
24 and 25
O_EVENTS

Number of over-reported drugs

Pharmacy Follow-back Study

among over-reporters: MTOTTOTALCAP
O_AWP

Level of over-reported AWP

Pharmacy follow-back, generated

expenses among over-reporters:
SURVAWP-PHARMAWP
OVER_RX

Annualized imputed number of

Generated

over-reported prescriptions
among over-reporters
OVER_RX$

Annualized imputed level of

Generated

over-reported AWP expenses
among over-reporters
OVER365

Annualized number of over-

Pharmacy Follow-back Study

reported drugs among overreporters
OVER365$

Annualized level of over-reported

Pharmacy follow-back, generated

AWP expenses among overreporters
PHARMAWP

Capped level of AWP expenses

Pharmacy follow-back, generated

estimated to have been purchased
by beneficiary in 1999
PROB_OVER

Probability of being designated as

Generated

an over-reporter
PROB_OVER$

Probability of being designated as

Generated

an AWP expenditure overreporter
PROB_PERFECT

Probability of being designated as

Generated

a perfect-reporter
PROB_PERFECT$

Probability of being designated as
an AWP expenditure perfectreporter

43

Generated

PROB_UNDER

Probability of being designated as

Generated

an under-reporter
PROB_UNDER$

Probability of being designated as

Generated

an AWP expenditure underreporter
RANDOM

Randomly generated number

Generated

between 0 and 1
RATIO

Reporting rate (TOTSCRIP /

Generated

EST_PURCH)
REAL_OVER

Transformed number of over-

Generated

reported drugs among overreporters
REAL_OVER$

Transformed level of over-

Generated

reported AWP expenses among
over-reporters
REAL_UNDER

Transformed number of under-

Generated

reported drugs among underreporters
REAL_UNDER$

Transformed level of under-

Generated

reported AWP expenses among
under-reporters
REPORT_STATUS

Beneficiary reporting status

Generated

(‘Under-reporter’,’Overreporter’,’Perfect-reporter’)
SURVAWP

Total level of beneficiary

Pharmacy follow-back, generated

reported AWP expenses in
Rounds 24 and 25
TOTALCAP

Capped total number of

Pharmacy follow-back, generated

prescriptions estimated to have
been purchased by beneficiary in
1999
TOTAWP

Total AWP expenses for 1999

Cost & Use, 1999

TOTSCRIP

Total prescriptions for 1999

Cost & Use, 1999

U_EVENTS

Number of under-reported drugs

Pharmacy Follow-back Study

among under-reporters:
TOTALCAP-MTOT
U_AWP

Level of under-reported AWP

44

Pharmacy follow-back, generated

expenses amo ng under-reporters:
PHARMAWP-SURVAWP
UNDER_RX

Annualized imputed number of

Generated

under-reported prescriptions
among under-reporters
UNDER_RX$

Annualized imputed level of

Generated

under-reported AWP expenses
among under-reporters
UNDER365

Annualized number of under-

Pharmacy Follow-back Study

reported drugs among underreporters
UNDER365$

Annualized level of underreported AWP expenses among
under-reporters

45

Pharmacy follow-back, generated

APPENDIX C
Relative risk ratios for reporting status
-----------------------------------------------------------------------------REP |

RRR

STD. ERR.

T

P>|T|

[95% CONF. INTERVAL]

-------------+---------------------------------------------------------------UNDER-REPORTER
ANN_TOTSCRIP |

1.02012

.0033735

6.02

0.000

1.013515

1.026769

CHRONIC |

1.117897

.0433334

2.88

0.004

1.03593

1.206349

INCOME_C |

1.002726

.0013524

2.02

0.044

1.000073

1.005386

MPAEVNTS |

1.013994

.0029756

4.74

0.000

1.008166

1.019856

BI_YNG |

.741002

.0654738

-3.39

0.001

.6229249

.8814608

BI_RXCOV |

1.203728

.0857882

2.60

0.010

1.046465

1.384624

BI_AF_AM |

1.044207

.1790865

0.25

0.801

.7455314

1.462538

BI_OTHER |

.6271836

.1355481

-2.16

0.031

.4102124

.9589162

BI_RISK |

1.56772

.1891474

3.73

0.000

1.236898

1.987023

BI_CAID |

1.519315

.2183676

2.91

0.004

1.14558

2.014977

-------------+---------------------------------------------------------------OVER-REPORTER
ANN_TOTSCRIP |

1.038283

.0035828

10.89

0.000

1.031269

1.045346

CHRONIC |

1.077724

.0563766

1.43

0.153

.9724742

1.194364

INCOME_C |

1.000153

.0017065

0.09

0.929

.9968062

1.003511

MPAEVNTS |

1.006567

.0034203

1.93

0.055

.9998699

1.013308

BI_YOUNG |

.8007253

.0858158

-2.07

0.039

.6487068

.9883679

BI_RXCOV |

.8356108

.0734555

-2.04

0.042

.7030821

.9931206

BI_AF_AM |

1.601988

.2821842

2.68

0.008

1.133389

2.26433

BI_OTHER |

1.207341

.3083749

0.74

0.461

.7310036

1.99407

BI_RISK |

1.55242

.2134558

3.20

0.001

1.184953

2.033842

BI_CAID |

1.275708

.2367956

1.31

0.190

.8859092

1.837018

-----------------------------------------------------------------------------(PERFECT REPORTERS ARE THE COMPARISON GROUP)

46

APPENDIX D
Results of multiple linear regression, R_UNDER365
SURVEY LINEAR REGRESSION

PWEIGHT:

C99WGT

NUMBER OF OBS

=

2904

STRATA:

SUDSTRAT

NUMBER OF STRATA =

PSU:

SUDUNIT

NUMBER OF PSUS

=

POPULATION SIZE

= 8769604.1

F(

=

15.44

PROB > F

=

0.0000

R-SQUARED

=

0.1271

12,

416)

64
491

-----------------------------------------------------------------------------R_UNDER365 |

COEF.

STD. ERR.

T

P>|T|

[95% CONF. INTERVAL]

-------------+---------------------------------------------------------------ANN_TOTSCRIP |

.0827889

.0265151

3.12

0.002

.0306724

.1349053

ANN_TOTAWP |

.0015831

.000506

3.13

0.002

.0005886

.0025776

_IGENHELTH_2 |

-.8999554

.9018077

-1.00

0.319

-2.67249

.8725793

_IGENHELTH_3 |

-.4527387

.9828935

-0.46

0.645

-2.38465

1.479173

_IGENHELTH_4 |

1.199128

1.155594

1.04

0.300

-1.072232

3.470489

_IGENHELTH_5 |

4.416433

2.080764

2.12

0.034

.3266185

8.506248

MPAEVNTS |

.0985806

.0192787

5.11

0.000

.0606877

.1364735

BI_YOUNG |

-1.874666

.7879259

-2.38

0.018

-3.423362

-.3259697

BI_RXCOV |

2.342071

.7941136

2.95

0.003

.7812123

3.902929

BI_AF_AM |

3.981971

1.839239

2.17

0.031

.3668815

7.59706

BI_EMP |

-3.015366

.9407723

-3.21

0.001

-4.864487

-1.166245

BI_PHI |

-2.804111

.8997573

-3.12

0.002

-4.572616

-1.035606

_CONS |

7.074387

1.29127

5.48

0.000

4.53635

9.612424

------------------------------------------------------------------------------

47

APPENDIX E
Results of multiple linear regression, R_OVER365

SURVEY LINEAR REGRESSION
PWEIGHT:

C99WGT

NUMBER OF OBS

STRATA:

SUDSTRAT

NUMBER OF STRATA =

PSU:

SUDUNIT

NUMBER OF PSUS

=

POPULATION SIZE

= 3815147.3

F(

6,

256)

=

1269
65
326

=

29.66

PROB > F

=

0.0000

R-SQUARED

=

0.2691

-----------------------------------------------------------------------------R_OVER365 |

COEF.

STD. ERR.

T

P>|T|

[95% CONF. INTERVAL]

-------------+---------------------------------------------------------------ANN_TOTSCRIP |

.1732639

.015702

11.03

0.000

.1423451

.2041826

CHRONIC |

.4518025

.1929737

2.34

0.020

.0718191

.8317859

INCOME_C |

-.0204454

.0063774

-3.21

0.002

-.0330032

-.0078877

BI_FEMALE |

-1.095882

.4450839

-2.46

0.014

-1.972295

-.2194699

BI_AF_AM |

BI_OTHER|
_CONS |

3.079474
2.825821
1.437244

.8844495
.8334886
.6971909

3.48
3.39
2.06

0.001
0.001
0.040

1.337909
1.184603
.0644093

4.821039
4.467039
2.810079

------------------------------------------------------------------------------

48

APPENDIX F
Relative risk ratios for expenditure reporting status
-----------------------------------------------------------------------------REPORT |

RRR

STD. ERR.

T

P>|T|

[95% CONF. INTERVAL]

-------------+---------------------------------------------------------------UNDER-REPORTER
ANN_TOTSCRIP |

1.04274

.0088527

4.93

0.000

1.025494

1.060277

ANN_TOTAWP |

1.000202

.0001152

1.76

0.080

.9999761

1.000429

_IGENHELTH_2 |

.878652

.13271

-0.86

0.392

.6530668

1.18216

_IGENHELTH_3 |

1.315806

.1888364

1.91

0.056

.9925492

1.744343

_IGENHELTH_4 |

1.085319

.2053783

0.43

0.665

.748362

1.573993

_IGENHELTH_5 |

1.183464

.3459059

0.58

0.565

.6664874

2.101445

CHRONIC |

1.140309

.0613496

2.44

0.015

1.02594

1.267428

MPAEVNTS |

1.018277

.0053191

3.47

0.001

1.007881

1.02878

BI_FEMALE |

1.2477

.1393492

1.98

0.048

1.001902

1.553799

BI_OTHER |

.4610349

.1238251

-2.88

0.004

.2720146

.7814035

BI_RISK |

1.946958

.2913845

4.45

0.000

1.451015

2.612409

BI_CAID |

1.810141

.3830344

2.80

0.005

1.194484

2.743118

BI_EMP |

1.309765

.1468208

2.41

0.016

1.05089

1.632412

-------------+---------------------------------------------------------------OVER-REPORTER
ANN_TOTSCRIP |

1.049681

.0090032

5.65

0.000

1.032143

1.067518

ANN_TOTAWP |

1.000351

.0001138

3.09

0.002

1.000128

1.000575

_IGENHELTH_2 |

1.02427

.1547527

0.16

0.874

.7612271

1.378206

_IGENHELTH_3 |

1.446507

.2433406

2.19

0.029

1.039435

2.013002

_IGENHELTH_4 |

1.307621

.283259

1.24

0.216

.8544167

2.001218

_IGENHELTH_5 |

.9599647

.3102434

-0.13

0.899

.5087802

1.811258

CHRONIC |

1.12764

.0660925

2.05

0.041

1.005

1.265247

MPAEVNTS |

1.012854

.0054205

2.39

0.017

1.002261

1.023558

BI_FEMALE |

1.202219

.1348332

1.64

0.101

.9644924

1.49854

BI_OTHER |

.7379603

.2141953

-1.05

0.296

.4172549

1.305162

BI_RISK |

1.727666

.2956669

3.19

0.001

1.234392

2.418055

BI_CAID |

1.525787

.3182734

2.03

0.043

1.012812

2.298578

BI_EMP |

1.152943

.1506402

1.09

0.277

.8919428

1.490318

-----------------------------------------------------------------------------(PERFECT REPORTERS are the comparison group)

49

APPENDIX G
Results of multiple linear regression, R_UNDER365$

SURVEY LINEAR REGRESSION

PWEIGHT:

C99WGT

NUMBER OF OBS

=

3091

STRATA:

SUDSTRAT

NUMBER OF STRATA =

64

PSU:

SUDUNIT

NUMBER OF PSUS

=

POPULATION SIZE

= 9371125.4

F(

=

23.37

PROB > F

=

0.0000

R-SQUARED

=

0.1383

8,

425)

496

-----------------------------------------------------------------------------R_UNDER365 |

COEF.

STD. ERR.

T

P>|T|

[95% CONF. INTERVAL]

-------------+---------------------------------------------------------------ANN_TOTAWP |

.2266565

.0423352

5.35

0.000

.1434478

.3098651

CHRONIC |

41.34224

15.97401

2.59

0.010

9.945794

72.73869

MPAEVNTS |

4.474903

1.079446

4.15

0.000

2.353283

6.596523

BI_OLD |

-230.3471

102.4492

-2.25

0.025

-431.7079

-28.98622

BI_YOUNG |

-298.833

93.76712

-3.19

0.002

-483.1295

-114.5365

BI_RXCOV |

192.5127

33.95269

5.67

0.000

125.7797

259.2457

BI_AF_AM |

200.2031

91.72257

2.18

0.030

19.92513

380.4811

BI_OP |

302.4975

147.848

2.05

0.041

11.9067

593.0884

_CONS |

251.7298

89.43593

2.81

0.005

75.94617

427.5135

------------------------------------------------------------------------------

50

APPENDIX H
Results of multiple linear regression, R_OVER365$
SURVEY LINEAR REGRESSION

PWEIGHT:

C99WGT

NUMBER OF OBS

=

1564

STRATA:

SUDSTRAT

NUMBER OF STRATA =

65

PSU:

SUDUNIT

NUMBER OF PSUS

=

POPULATION SIZE

= 4705472.8

F(

5,

297)

366

=

25.76

PROB > F

=

0.0000

R-SQUARED

=

0.1294

-----------------------------------------------------------------------------R_OVER365 |

COEF.

STD. ERR.

T

P>|T|

[95% CONF. INTERVAL]

-------------+---------------------------------------------------------------ANN_TOTAWP |

.0933267

.0323955

2.88

0.004

.0295764

.157077

CHRONIC |

55.63052

17.13858

3.25

0.001

21.90392

89.35712

HHAEVNTS |

-.2236337

.0909783

-2.46

0.015

-.4026678

-.0445996

BI_OLD |

-200.2182

52.65341

-3.80

0.000

-303.8336

-96.60277

BI_YOUNG |

-160.098

58.29061

-2.75

0.006

-274.8067

-45.38923

_CONS |

279.7384

66.18351

4.23

0.000

149.4975

409.9794

------------------------------------------------------------------------------

51