PharmacoEconomics (2018) 36:369–380

https://doi.org/10.1007/s40273-017-0597-y

ORIGINAL RESEARCH ARTICLE

Sensitivity of the Medication Possession Ratio to Modelling

Decisions in Large Claims Databases
Margret V. Bjarnadottir1 • David Czerwinski2 • Eberechukwu Onukwugha3

Published online: 11 December 2017

Ó Springer International Publishing AG, part of Springer Nature 2017

Abstract percentage points from the base case: the decision of

Objectives When preparing administrative medical and whether to use interval- or prescription-based study peri-
pharmacy claims data for analysis, decisions about data ods, and the decision of how to handle overlapping pre-
clean up and analytical approach need to be made. How- scription claims. The effect of other decisions was smaller,
ever, information about the effects of various modelling with a difference of 1–9 percentage points from the base
decisions on adherence measures such as the medication case.
possession ratio (MPR) is limited. We address this gap with Conclusions Some of the decisions considered had a large
this study. impact on the MPR. Therefore, it is important for
Methods We utilized cross-sectional administrative claims researchers to standardize approaches for study period
data for commercially insured members filling at least two length and overlapping prescription claims. We also con-
prescriptions for drugs within five classes of hypertension clude that transparent reporting of modelling decisions will
medication between 2008 and 2010. We divided nine facilitate the interpretation of results and comparisons
modelling decisions into three categories: data scrubbing, across studies.
study design, and MPR definition/calculations. We defined
the base-case settings with commonly used values, varied
each modelling decision singly and in combination, and
Key Points for Decision Makers
measured the effects on the MPR.
Results Claims data for 358,418 individuals were available
Information that documents the effects of various
for analysis. Two modelling decisions were found to be
modelling decisions on commonly used adherence
highly influential, each yielding a difference of over 25
measures such as the medication possession ratio
(MPR) is limited.
Electronic supplementary material The online version of this
article (https://doi.org/10.1007/s40273-017-0597-y) contains supple- Our analysis may explain why different studies of
mentary material, which is available to authorized users. compliance report different observed MPR values,
and it points out the key factors that can be
& Eberechukwu Onukwugha
eonukwug@rx.umaryland.edu standardized to better compare studies.
1 Understanding the consequences of critical
Department of Decision, Operations, and Information
Technologies, Robert H. Smith School of Business, modelling decisions and the sensitivity of the results
University of Maryland College Park, College Park, MD, to those decisions will help guide both researchers in
USA their analysis of large claims data and practitioners in
2
Department of Marketing and Decision Sciences, College of applying the findings.
Business, San José State University, San José, CA, USA
3
Department of Pharmaceutical Health Services Research,
University of Maryland School of Pharmacy, 220 Arch
Street, Baltimore, MD 21201, USA
370
1 Introduction
For population health researchers, administrative claims
databases present many benefits, including sample size,
real-world utilization information and generalizability, but
they also pose well-known challenges such as the lack of
validated measures, non-specificity of billing diagnosis
codes, and the inability to confirm actual drug utilization.
Notwithstanding these limitations, many studies on medi-
cation adherence utilize administrative claims data to
develop adherence measures.
A multitude of adherence measures exist [1–4]. Based
on publications indexed in MEDLINE, the medication
possession ratio (MPR) was first utilized by Sclar et al. [5]
in 1991 to investigate outcomes associated with participa-
tion in a health education programme as well as costs
associated with increased adherence [6, 7]. Schulz and
Gagnon [8] defined the term ‘drug possession ratio’ in a
1982 paper, but it was the MPR that became better known.
While studies do not uniformly recommend one adherence
measure over another, there is some consensus that the
MPR and proportion of days covered (PDC) estimates are
comparable in the case of monotherapy [9, 10], whereas, in
the case of polypharmacy, the PDC provides more con-
servative estimates than the MPR and may be more
appropriate [9, 10]. The MPR remains a very popular
measure of adherence [9, 11] and has been used in multiple
studies using administrative claims data. Many of these
studies are focused on economic and health outcomes
related to adherence. For example, the MPR has been
analysed as an outcome in studies that do not involve an
intervention, as an outcome in comparative-effectiveness
research studies, and as a predictor of clinical and cost
outcomes.
Studies have noted the variety of approaches to mea-
suring adherence [12, 13] and the differences (and incon-
sistencies) in terminology and approach [1, 14, 15] and
compared the various measures of medication adherence
[1, 9, 11]. A few studies have also focused on the accuracy
of claims-based measures of adherence compared with
other sources such as self-reporting [16–19]. Some
researchers point to a need for standardization among the
available measures [14, 19–21]. Published guidance [2] and
checklists [22] have been cited as important tools for
standardizing the terminology and calculation of adherence
measures.
A subset of the literature has addressed the method-
ological aspects of adherence measures. However, rela-
tively little attention has been paid to the impact of
modelling decisions on the value of the adherence measure.
These modelling decisions include decisions about mem-
bers’ inclusion and exclusion criteria, data scrubbing, and
M. V. Bjarnadottir et al.
parameter settings, all of which need to be automated (as
manual review is not possible with these large datasets).
Understanding the effect of these decisions can help guide
the application of modelling decisions, the interpretation of
study findings, and the translation of study findings to other
settings. We build on the existing literature by considering
a broad range of modelling decisions in order to build a
framework for comparing their relative impacts. Findings
from available literature indicate that (1) prescription-
based estimates are more conservative than interval-based
estimates [23], (2) the fixed period length should be
reserved for studies with a shorter time interval for each
individual [15], and (3) capping the MPR is useful [3],
particularly for studies with shorter overall intervals [24].
One study [25] focused on MPR during the initial coverage
period and found that modelling decisions (e.g. handling
overlap and gap length for consolidating claims) did not
have large effects on MPR. Compared with other studies,
this study expands the number of modelling decisions
considered, with the goal of analysing the relative sensi-
tivity of the observed MPR to changes in common mod-
elling decisions.
2 Methods
We utilized cross-sectional administrative medical and
pharmacy claims data for commercially insured members
from multiple health plans across the country who filled at
least two prescriptions for a hypertension medication
between 2008 and 2010. The following five drug classes
were analysed: angiotensin-converting enzyme inhibitors
(ACEIs), angiotensin II receptor blockers (ARBs), beta
blockers (BBs), calcium channel blockers (CCBs) and
diuretics.
We calculate adherence using the MPR [2, 5, 22]. The
MPR is a ratio of two quantities: the sum of the days
supplied and the length of the observation period, or:
days supply
MPR ¼ :
length of observation period
This formula is consistent with the definition of MPR
proposed by Cramer et al. [2] as ‘‘the number of doses
dispensed in relation to the dispensing period.’’ Intuitively,
the MPR describes the fraction of time during the
observation period that a member has access to
medication. For each member, for the base case
considered, the observation period was defined as the
time from the member’s first hypertension prescription
until whichever came first: the end of the supply of the
medication from the member’s most recent prescription or
the end of a fixed study period. The days’ supply is the sum
of the days’ supply of all prescriptions during the
Sensitivity of the MPR in Claims Analysis
observation period. It is often adjusted for prescription
overlaps, possible duplications and errors.
We divided the modelling decisions into three cate-
gories: (1) data scrubbing, (2) study design, and (3) MPR
calculations. We define these concepts in the following
sections, and Table 1 summarizes the categories, decisions
and values investigated. For each decision, we defined a
base case, the value or the scenario against which all others
were compared. For each decision, we varied the decision
as described in the table and kept the values for all other
decisions at their base value. Finally, to study the com-
pounding of the modelling decisions, we selected the most
extreme values of each decision to study the maximum
variation in MPR associated with these decisions when
considering multiple decisions at a time.
2.1 Data Scrubbing
Data scrubbing encompasses multiple actions that are
needed to take the data from a raw form to a form usable
for analysis (by, for example, addressing data errors or
omissions). The data used in this study included all ‘un-
scrubbed’ claims to highlight some of the issues that may
arise and to illustrate how different decisions made during
this critical analytical phase can ultimately affect the end
results. In this study, we specifically analysed the effects of
decisions made regarding claims with unusual values (large
or negative) and multiple claims per day (duplicates), as
detailed in Table 1.
2.2 Study Design
The category ‘study design’ included all decisions about
inclusion criteria and other study parameters such as the
length of the observation period. The inclusion criteria in
MPR studies usually specify a minimum number of pre-
scriptions, commonly two [26–28]; however, we also
observed different definitions, such as a prescription and a
refill for the same medication category within a year [29]
and a single prescription [30]. Additionally, inclusion cri-
teria commonly require a prescription-free period, usually
defined as preceding the first prescription by 6 months
[28, 29, 31] or 12 months [26, 30, 32], especially if the goal
is for the start of the study period to align with the
beginning of a treatment episode. We called this period the
‘clear period’ and used 180 days as our base value.
Other study design decisions include the length of the
study and the treatment of the end of the study period. In
the literature, a 1-year study period is common [27, 31, 33],
although some studies evaluate multiple endpoints between
6 months and 2 years [26, 29, 32] or multiple events as
endpoints [28, 30]. We defined the base case for the length
of the study period as 365 days. A related but independent
371
choice involves the treatment of the end of the study period
for each member. We compared four alternatives. The first
approach was to use a fixed study period, that is, an
interval-based approach. For example, if the study period is
1 year but a member has his or her last prescription 8
months into the study, the MPR is calculated over the
entire 1-year period. Alternatively, the study period can be
prescription based; therefore, a second approach was to
define the end of the study period as the minimum of either
the nominal study period or the date of the last prescription
within the study period plus its days’ supply, that is, the last
dose. We used this approach as our base case. A third
approach was to define the end of the study period as the
date of the last prescription before the end of the nominal
study period (which excluded the entirety of the member’s
last prescription within the study window). Finally, a fourth
approach defined the end of the study period as the date of
the last dose from the last prescription within the nominal
study period.
2.3 Medication Possession Ratio (MPR)
Calculations
When analysing MPR calculations, we included two dif-
ferent modelling decisions, one regarding overlapping
prescriptions and the second regarding MPR values greater
than 1.
Overlapping prescriptions can be treated in multiple
ways. The simplest approach is to sum up the days’ supply
of all prescriptions within a study period, regardless of any
overlap (this is our base case). An alternative approach is to
merge the overlapping prescriptions. The third option is to
extend the coverage period beyond the second prescription
by appending the overlapping supply to the end of the
second prescription. These different treatments of overlap
are demonstrated in Fig. 1. When the overlap is appended,
the underlying assumption is that the overlap was caused
by an early refill. When the two prescriptions are merged,
the underlying assumption is that the early refill was a
replacement.
The choice between merging or appending overlapping
prescriptions may be informed by the disease setting and
study population, but it may also be a function of the length
of the overlap—a smaller overlap may be due to early
refill, but a longer overlap may be due to replacement.
Researchers often choose a threshold k such that overlaps
of durations shorter than the threshold are appended but
longer overlaps are merged. For example, Laliberte et al.
[28] choose an overlap threshold of 7 days. Researchers
have also developed specific decision rules based on the
prescribed drug classes [26] or the source of the prescrip-
tion [31].
372 M. V. Bjarnadottir et al.

Table 1 Modelling decisions and parameters studied

Category Decision/parameter Alternatives investigated

Data Large claims: Claims with days’ supply [30 or 1. Left as is

scrubbing 90 for retail/mail order, respectively 2. Rounded down to 30 (retail) or 90 (mail)
3. Rounded down to 90 (both retail and mail)
4. Large claims excluded (as defined in 2 and 3)
5. Members with large claims excluded (as defined in 2 and 3)
Data Claims with negative numbers 1. Left as is (assumed to be corrections)
scrubbing 2. Made positive
3. Excluded all negative claims
4. Excluded members with any negative claims
Data Multiple claims on same date 1. Left as is (used all claims)
scrubbing 2. Keep only the claim with largest days’ supply
3. Keep only the claim with smallest days’ supply
Study design Inclusion criteria: Minimum number of 1. At least two prescriptions
prescriptions 2. At least three prescriptions
3. At least four prescriptions
Study design Inclusion criteria: drug-free time until start of 1. 0 days
therapy 2. 90 days
3. 180 days
4. 270 days
5. 360 days
Study design Study length Time from first prescription:
1. 180 days
2. 270 days
3. 360 days
4. 365 days
5. 450 days
6. 540 days
Study design Handling end-of-study 1. Fixed study window: Study ends after exactly 1 year, even if last
prescription concluded before 1 year
2. Study ends after whichever comes first: Conclusion of last prescription
or 1 year. If the 1-year mark occurs in the middle of a prescription, the
prescription is truncated at the 1-year mark
3. Same as (2) except if the 1-year mark occurs in the middle of a prescription
the whole prescription is included
4. Study ends on date of last fill prior to 1-year mark
MPR Overlap in prescription coverage 1. Leave as is
calculations 2. Truncate first prescription at point overlapping prescription begins
(merging)
3. Append overlapping prescription to the end of the first prescription
4. Append overlapping prescription if overlap \7 days, otherwise merge the
overlap
MPR Treatment of members with MPR [100% 1. Leave as is
calculations 2. Cap at 1
3. Discard if greater than 1.5. Cap remaining at 1
Bold formatting indicates the base case for each decision
MPR medication possession ratio
Sensitivity of the MPR in Claims Analysis
Fig. 1 Treatment of overlapping prescriptions. The figure depicts a
member with two prescriptions within a study period that overlap and
demonstrates the different treatment of overlapping prescriptions. For
Depending on the treatment of overlapping prescriptions
and on the decisions made during the data scrubbing per-
iod, members with multiple claims may have an MPR
value that exceeds 1. As a result, another study design
decision is whether to truncate such high MPRs down to 1,
which is a common approach in the literature [27, 32, 34]
and which we select as our base case. Aguilar et al. [29]
chose to truncate MPRs above 1 and smaller than 1.5 down
to 1 but to eliminate members with MPR greater than 1.5,
since MPRs of that magnitude may indicate data problems.
2.4 Comparison of MPR and Proportion of Days
Covered (PDC)
As mentioned in the introduction, another commonly used
measure is the PDC, which was developed by the Phar-
macy Quality Alliance [35] and endorsed by the Centers
for Medicare and Medicaid Services [36]. The choice
between PDC and MPR depends on the goals and setting of
the research study [9, 14, 21, 37], among other consider-
ations. PDC is calculated as follows:
days ‘‘covered’’
PDC ¼
length of study period
MPR and PDC are nearly identical when measuring
adherence to a single drug class. The difference lies in the
treatment of overlapping prescriptions. In the base case of
373
simplicity, we implicitly assumed that the second prescription ended
well before the end of the study period. MPR medication possession
ratio
MPR, the days’ supply is simply summed over the
observation period. In contrast, PDC appends overlapping
prescriptions. Therefore, by changing the base case in the
treatment of overlapping prescriptions for MPR, as we
discuss in Sect. 2.3, we also analyse PDC. We contrast the
results for PDC and MPR below and in the Electronic
Supplementary Material (ESM).
3 Results
The study period extended from 1 October 2008 to 30
September 2010. The study population consisted of
449,589 commercially insured members aged between 20
and 65 years. We excluded members receiving multidrug
hypertensive therapies, including those taking combination
drugs during the study period (see Fig. 2). The study cohort
consisted of 358,418 members who filled at least two
prescriptions for a single hypertension medication. Mem-
bers of the cohort filled 3.1 million prescriptions for anti-
hypertensive drugs, of which 87% were filled at retail
pharmacies and 13% were filled by mail. On average, a
member filled 8.7 ± standard deviation (SD) 7.2 prescrip-
tions, with an average total days’ supply of 370 ± 561. This
large SD reflects the number of duplicate and/or abnor-
mally large claims in raw claims data. The cohort had a
mean age of 50 ± 10 years, and 55% of the study cohort
374 M. V. Bjarnadottir et al.

Fig. 2 The study flow diagram

was female. Table 2 details the sample size by drug class,
sex, average age, average number of prescriptions and
mean MPR for the base case by drug class.
In the following sections, we first describe the sensitivity
of the MPR for each modelling decision, followed by the
compounded scenario, ending with a comparison of the
MPR results with the PDC results. For all decisions anal-
ysed, the scenario studied did not affect the relative order
of the drug classes in a statistically meaningful way; CCBs
consistently had the highest measured mean MPR, fol-
lowed by ACEIs and ARBs, then BBs and, finally,
diuretics. In all the following calculations, the mean MPR
was calculated across all members, and any percentage
point change was always calculated as change from the
base case. Results by drug class, as well as additional
details, are provided in the ESM.
3.1 Data Scrubbing
When analysing our data, we found that 10.3% of claims
exceeded the 30- and 90-day supply for retail and mail-in
prescriptions, respectively. Overall, 18.5% of members had
at least one claim exceeding these limits. Of the retail
claims that exceeded a 30-day supply, 83% were for
exactly 90 days. The mean MPR ranged from 75.6 to
80.9% across different modelling decisions when using a
30-day limit for retail claims (and a 90-day limit for mail
order). When a 90-day limit was used for both retail and
mail-in prescriptions, the effect was much smaller, with the
mean MPR ranging from 80.6 to 80.9%.
Significantly fewer claims were negative (1.3%), and
5.2% of the population had at least one negative claim. The
effect of addressing negative claims was small, with mean
MPR ranging from 80.9 to 81.8%. Finally, there were
several ways to address multiple claims per day. In the base

Table 2 The population size by drug class and sex, the average number of prescriptions and the mean medication possession ratio by drug class
Class Females Males Total Age No. of prescriptions Mean MPR

All classes 195,918 (54.70) 162,500 (45.30) 358,418 (100) 50.2 ± 10.0 8.7 ± 7.2 80.9 ± 21.3
ACEI 43,942 (22.4) 71,692 (44.1) 115,634 (32.3) 51.1 ± 9.2 8.7 ± 6.8 82.8 ± 21.2
ARB 15,191 (7.8) 19,708 (12.1) 34,899 (9.7) 52.7 ± 8.5 9.7 ± 7.5 82.5 ± 23.0
BB 55,952 (28.6) 41,057 (25.3) 97,009 (27.0) 49.5 ± 10.6 8.9 ± 7.9 81.6 ± 21.3
CCB 19,334 (9.9) 14,931 (9.2) 34,265 (9.6) 50.2 ± 10.4 8.9 ± 7.8 84.8 ± 20.9
Diuretics 61,499 (31.4) 15,112 (9.3) 76,611 (21.4) 48.6 ± 10.5 7.8 ± 6.9 76.3 ± 25.5
Data are presented as n (%) or average ± standard deviation unless otherwise indicated
ACEI angiotensin-converting enzyme inhibitor, ARB angiotensin receptor blocker, BB beta-blocker, CCB calcium channel blocker, MPR
medication possession ratio, SD standard deviation
Sensitivity of the MPR in Claims Analysis 375

case, all claims were used. The alternatives, selecting—at
most—one claim per day, resulted in a lower mean MPR,
and if the claim with the lowest days’ supply value was
selected, the mean MPR was 78.2%. Table 3 summarizes
the main results of the data scrubbing decisions.
3.2 Study Design
We studied two inclusion criteria: the minimum number of
prescriptions and the length of the clear period. The mean
MPR increased by 1% when we increased the required
prescriptions from two to three and by an additional 2%
when four prescriptions were required. The second inclu-
sion criteria was the length of the clear period. The base
case was defined as 180 days, which had the lowest average
mean MPR, whereas the largest value was for 0 days, 4%
larger than the base case. The population size decreased
significantly with the length of the clear period.
The study period measures the time from the day of the
member’s first prescription fill until the end of the obser-
vation. The mean MPR decreased as the length of the study
increased. When the study duration was 180 days, the mean
MPR was 86.5%, but it dropped to 78.1% when the study
duration was 540 days. This decrease was observed across
drug classes, but the effect was the largest for diuretics,
where the mean MPR dropped from 84.0 to 73.4%. As with
the length of the clear period, the population size decreased
significantly as a function of the study period.
In our analysis of the endpoint of the study period, we
found that enforcing a fixed study period reduced the mean
MPR significantly, dropping it from 80.9% for the base
case to 56.6%. This was largely due to the effect of
members who may have stopped treatment during the study
period. The highest mean MPR, 81.4%, was measured
when we defined the endpoint as the day of the member’s
last dose.
Table 4 summarizes the main results of study design
decisions.
3.3 MPR Calculations
It was quite common for a member to have two prescrip-
tions that overlapped in time. In our dataset, 53.9% of the
population had at least one overlap. If we simply summed
the days’ supply of any overlapping prescriptions (the base
case), the resulting mean MPR was 80.9% compared with
58.7% when overlapping prescriptions were merged, a 27%
difference. When overlaps were appended, the resulting
mean MPR was a little lower than the base case, as some
appended prescriptions exceeded the 365-day base-case
study period and therefore did not fully count towards the
MPR (please refer to the ESM for details).
Rounding down the MPR also had significant effects. If
large MPR values were not treated, the resulting mean
MPR was 87.1%. But when members with MPR above 1.5
were excluded from the study, and other large values ([1)
were rounded to 1, we found a mean MPR of 80.2%.
Table 5 summarizes the main results of the MPR calcula-
tion decisions.
3.4 Simultaneous Decisions
So far, we have reported the effects of each decision in
isolation. However, in any given study, all of these deci-
sions need to be made simultaneously. We therefore
analysed the possible resulting mean MPRs for 576 dif-
ferent combinations of the aforementioned modelling
decisions. In other words, we analysed all possible com-
binations of the scenarios listed in Table 6, selected based
on their impact on MPR and/or their use in the literature.
The lowest resulting mean MPR obtained in this study
was 32%. This was obtained when members with large

Table 3 Summary of the effects of different modelling decisions in the data-scrubbing category
Decision Smallest and largest Case description Mean MPR in Mean PDC in
measures percentages percentages

Treatment of large claims (the base case is leave as is) Smallest Round down 75.6 ± 25.7 73.8 ± 25.5
Largest Leave as is 80.9 ± 23.1 78.6 ± 23.3
Treatment of negative claims (the base case is leave as Smallest Leave as is 80.9 ± 23.1 78.6 ± 23.3
is) Largest Made positive 81.8 ± 22.8 79.1 ± 23.1
Treatment of multiple claims on a single day (the base Smallest Use only the 78.2 ± 24.1 77.1 ± 23.9
case is use all claims) smallest claim
Largest Use all claims 80.9 ± 23.1 78.6 ± 23.3
The results for the smallest and largest mean medication possession ratio and proportion of days covered across drug classes are reported for
alternative approaches
Data are presented as mean ± standard deviation unless otherwise indicated
MPR medication possession ratio, PDC proportion of days covered
376 M. V. Bjarnadottir et al.

Table 4 Summary of the effects of different modelling decisions in the study design category
Decision Smallest and Case description Mean MPR in Mean PDC in
largest percentages percentages
measures

Inclusion criteria: minimum number of prescriptions Smallest Two 80.9 ± 23.1 78.6 ± 23.3
(base case is two) Largest Four 83.9 ± 19.7 80.5 ± 20.5
Inclusion criteria: length of the clear period (base case Smallest 180 days 80.9 ± 23.1 78.6 ± 23.3
is 180 days) Largest 0 days 85.1 ± 19.9 81.8 ± 20.6
Length of the overall study period (base case is 365 Smallest 540 days 78.1 ± 24.6 75.7 ± 24.6
days) Largest 180 days 86.5 ± 18.3 84.3 ± 19.2
The study endpoint for each member (base case is the Smallest Fixed study window 56.6 ± 31.1 54.1 ± 29.5
minimum of the study window or the last dose) Largest Date of last dose from a 81.4 ± 22.6 80.9 ± 22.5
prescription within the study
window
The results for the smallest and largest mean medication possession ratio and proportion of days covered across drug classes are reported for
alternative approaches
Data are presented as mean ± standard deviation unless otherwise indicated
MPR medication possession ratio, PDC proportion of days covered

Table 5 Summary of the effects of different decisions when calculating the mean medication possession ratio (across therapeutic classes)
Decision Smallest and largest Case description Mean MPR in % change from
measure percentages (SD) base case

MPR rounding (the base case is to cap at Smallest MPR above 1.5 excluded, 80.17 - 0.9
1) others capped at 1
Largest Not rounded 87.07 7.6
Overlapping prescriptions (the base case Smallest Merge overlapping 58.71 - 27.4
is to leave as is) prescriptions
Largest Leave as is 80.89 NA
The results for the smallest and largest mean medication possession ratio across drug classes are reported for alternative approaches, as well as
the percentage change from the base case
MPR medication possession ratio, NA not applicable, PDC proportion of days covered, SD standard deviation

Table 6 Scenarios studied in the simultaneous decisions study

Decision Scenarios studied

Large claims Leave as is, drop members with large claims

Claims with negative numbers Leave as is, make positive
Multiple claims on same date Summing up, using the smallest days’ supply, using the largest days’ supply
Inclusion criteria, minimum number of prescriptions Two, four
Inclusion criteria: drug-free time until start of therapy 90, 180, 270 days
Study length 180, 450 days (we did not consider 540 days due to small population size)
Handling end-of-study Fixed study window, last dose (prescription within study window)
Overlap in prescription coverage Leave it as is, merge any overlap
Treatment of members with MPR [100%: Leave as is, cap at 1, and cap at 1 after discarding those above 1.5
MPR medication possession ratio, PDC proportion of days covered

claims were dropped, negative claims were made positive, minimum of two claims to be included, the clear period
and the minimum days’ supply was selected if there were was set to 270 days, the study length was 450 days and a
multiple claims on a single day. A member needed a fixed study window was used. Overlaps were merged. At
Sensitivity of the MPR in Claims Analysis
the opposite end of the spectrum, the largest observed
average mean MPR was 106%. This was achieved when
large claims were left ‘as is’, negative claims were made
positive, and multiple claims per day were summed.
Members needed four prescriptions to be included, the
clear period was 90 days, the study length was 180 days
and the study period for each member ended on the day of
the last dose. Finally, overlaps were ignored (all days’
supply simply summed) and MPR values above 1 were not
truncated. Figure 3 summarizes the distribution of the
observed mean MPRs for the 576 runs. The different
shades represent different treatments of overlapping
claims.
3.5 Comparison with PDC
As discussed, the differences between MPR and PDC in the
single drug class case resulted in a different base case for
the treatment of overlapping prescriptions. As a result, the
base-case value for PDC (78.6%) was lower than that for
MPR (80.9%). With one exception, we consistently
observed that, for any decision scenario, PDC values were
lower (generally by 1.5–3%) than MPR values. The single
exception occurred when we specified the end of the study
as the date of the last prescription, in which case the
resulting mean PDC exceeded the mean MPR (see the
Fig. 3 The distribution of observed mean medication possession ratio
(MPR) for the different modelling decisions. The two colours
represent the treatment of overlapping claims; pink represents
combinations where overlapping claims are merged and green
represents the case when overlapping claims are treated as an early
refill
377
ESM for details and Tables 3, 4, 5 for the smallest and
largest mean PDCs).
4 Discussion
In this study, we investigated the impact of modelling
decisions that are necessary in order to prepare claims data
and conduct analysis in adherence studies. We examined
the MPR and PDC as a proof of concept to investigate the
impact of modelling decisions on commonly used mea-
sures in health services research studies. We considered
MPR modelling decisions in isolation (i.e. one-way anal-
ysis) and in conjunction with other modelling choices. Of
the nine decisions considered, two were especially
impactful: the use of interval- or prescription-based study
periods and the handling of overlapping prescriptions.
Other influential choices included the decision to cap the
MPR at 1, the treatment of large claims, and the length of
the clear period. Our observation that a fixed clear period
lowered the MPR is consistent with published literature
[24, 38]. We note that the effect of the decision regarding
claims with a large days’ supply is reduced because, in our
base-case analysis, MPR was rounded down to 1, which
then masks the effects of the choices made in handling
large claims.
One notable implication for study design is our finding
that, as we increased the number of required prescriptions,
the MPR of all drug classes increased slightly. This was
mostly because requiring more prescriptions excluded
members with only two (or three) prescriptions over a long
period of time, resulting in a smaller study population with
a higher adherence rate. This result is consistent with a
study that documented an inherent upward bias associated
with requiring at least two prescriptions [13]. Sample size
may also be affected by the parameters set for the clear
period, and this may happen in two ways. First, members
will be excluded if they do not meet the minimum time
requirement that is defined for the clear period. Second, if
the defined clear period is very long, a member is less
likely to satisfy both the requirement for the clear period
and the study duration. For our data, introducing a clear
period of 90 days reduced the population size by 72%.
Our findings are consistent with other work document-
ing the variability of adherence measures and the impact
and importance of modelling decisions. To be effective,
studies should provide sufficient detail for their readers to
understand the modelling decisions that were made. As one
study noted, it is not ideal to note that underlying mod-
elling (e.g. measurement) decisions related to the adher-
ence measure are proprietary [22], as this does not further
scientific knowledge. We echo the call from Sattler et al.
[9] for transparency in the algorithm used to define the
378
adherence measure as well as standardization of terminol-
ogy and modelling decisions. We believe that researchers
studying adherence in working age populations receiving
monotherapy would benefit greatly from the development
of standardized modelling decisions for (1) the use of an
interval- or prescription-based measure, (2) handling
overlapping prescriptions, and—to a lesser extent—(3)
whether or not to cap the MPR at 1. Additionally, if MPR is
not capped, it is critical to reveal the data-scrubbing deci-
sions regarding the handling of large, negative and multiple
claims.
With the growing number of adherence measures
[14, 39] available to researchers, information regarding
their susceptibility to changes in the underlying measure-
ment decisions provides additional useful information for
evaluating various measures. To date, this dimension of
adherence measures has not been explored in the literature
and represents a useful line of inquiry for future research.
Further, exploring these decisions in the context of multi-
therapy is important. With regard to empirical studies on
adherence, there is a need [40, 41] to better define the role
of medication adherence in cost-effectiveness studies. If a
higher level of adherence impacts drug effectiveness, and if
adherence levels differ between intervention and com-
parator drugs, the value of the MPR can materially impact
the incremental cost-effectiveness ratio and associated
decisions regarding cost effectiveness. To the extent that
the value of the MPR depends on modelling decisions,
these decisions will be relevant for cost-effectiveness
analysis.
A few limitations are worth noting. Our study utilized
claims data for a working age, commercially insured
population. The results regarding the utilization patterns
of the five drug classes, adherence, and rankings by drug
class are not necessarily generalizable to older and/or
differently insured (e.g. Medicare) individuals. Our
multi-way sensitivity analysis utilized minimum and
maximum values to provide the necessary structure, and
it is possible that a study in a different population would
yield different minimum and maximum values and that a
multi-way sensitivity analysis of that population may
lead to different conclusions. However, it is worth noting
that by selecting modelling decisions that lead to
extreme values of MPR, we arrived at a mean MPR
ranging between 32 and 106%. Lastly, we did not
directly address potential errors in the days’ supply field,
which is not an audited field and is typically populated
by the pharmacist. But we did study the effects of data-
scrubbing decisions that may reflect some of the result-
ing complexities. Studies that examined the accuracy of
the days’ supply found mixed results [23, 42, 43]. Errors
could either be inadvertent (incorrect entry) or pur-
poseful [23] (e.g. entering 30 days’ supply to coincide
M. V. Bjarnadottir et al.
with insurance policies rather than entering the exact
days’ supply). When there is documented limited con-
cordance between days’ supply and the original pre-
scription, one option is to apply a correction factor as
shown in Blais et al. [23]. Finally, we did not study the
effects of taking members’ eligibility into consideration
when defining study endpoints, which is worthy of future
investigation.
5 Conclusion
Any use of administrative claims data to investigate med-
ication adherence requires researchers to make several
modelling decisions. We utilized an administrative claims
dataset to investigate the robustness of the MPR to nine
modelling decisions focused on preparing the administra-
tive claims data for analysis. The following modelling
decisions had the largest impact on the MPR: study period
length, overlapping prescription claims, and handling MPR
values above 1. For the sake of meaningful comparison
between studies, it will be important to utilize standardized
approaches and provide transparent reporting for these
important modelling decisions.
Data Availability Statement This study used proprietary
data that are not available to other researchers.
Acknowledgements The authors thank Husam Albarmawi for pro-
viding research assistance during the planning stages of this project.
Author Contributions The interpretation and reporting of the results
are the sole responsibility of the authors. Dr. Bjarnadottir will act as
the guarantor of the work presented in this paper. Dr. Bjarnadottir
contributed to the study design, data collection, and interpretation of
the analysis and drafted and revised the manuscript with input from
all co-authors. Dr. Czerwinski contributed to the study design, ran the
analysis and contributed to the interpretation of the results and the
writing and revision of the manuscript. Dr. Onukwugha contributed to
the study design, interpretation of the analysis, and drafted and
revised the manuscript.
Compliance with Ethical Standards
Funding This is an unfunded study.
Conflict of interest Drs. Bjarnadottir and Czerwinski have no con-
flicts of interest to declare. Dr. Onukwugha has received grant
funding from Bayer Healthcare Pharmaceuticals and Pfizer, Inc. as
well as consulting fees from Novo Nordisk.
References
1. Andrade SE, Kahler KH, Frech F, Chan KA. Methods for eval-
uation of medication adherence and persistence using automated
databases. Pharmacoepidemiol Drug Saf. 2006;15:565–74.
Sensitivity of the MPR in Claims Analysis
2. Cramer JA, Roy A, Burrell A, Fairchild CJ, Fuldeore MJ,
Ollendorf DA, et al. Medication compliance and persistence:
terminology and definitions. Value Health. 2008;2(1):44–7.
3. Karve S, Cleves MA, Helm M, Hudson TJ, West DS, Martin BC.
Prospective validation of eight different adherence measures for
use with administrative claims data among patients with
schizophrenia. Value Health. 2009;12(6):989–95.
4. Karve S, Cleves MA, Helm M, Hudson TJ, West DS, Martin BC.
Good and poor adherence: optimal cut-point for adherence
measures using administrative claims data. Curr Med Res Opin.
2009;25(9):2303–10.
5. Sclar DA, Chin A, Skaer TL, Okamoto MP, Nakahiro RK, Gill
MA. Effect of health education in promoting prescription refill
compliance among patients with hypertension. Clin Ther.
1991;13(4):489–95.
6. Sclar DA, Skaer TL, Chin A, Okamoto MP, Gill MA. Utility of a
transdermal delivery system for antihypertensive therapy. Part 2.
Am J Med. 1991;91(1A):57S–60S.
7. Sclar DA, Skaer TL, Chin A, Okamoto MP, Gill MA. Utility of a
transdermal delivery system for antihypertensive therapy. Part 1.
Am J Med. 1991;91(1A):50S–6S.
8. Schulz RM, Gagnon JP. Patient behavior patterns regarding
prescription refills. Contemp Pharm Pract. 1982;5(3):150–5.
9. Sattler EL, Lee JS, Perri M 3rd. Medication (re)fill adherence
measures derived from pharmacy claims data in older Americans:
a review of the literature. Drugs Aging. 2013;30(6):383–99.
10. Tang KL, Quan H, Rabi DM. Measuring medication adherence in
patients with incident hypertension: a retrospective cohort study.
BMC Health Serv Res. 2017;17(1):135.
11. Clifford S, Perez-Nieves M, Skalicky AM, Reaney M, Coyne KS.
A systematic literature review of methodologies used to assess
medication adherence in patients with diabetes. Curr Med Res
Opin. 2014;30(6):1071–85.
12. Stolpe S, Kroes MA, Webb N, Wisniewski T. A systematic
review of insulin adherence measures in patients with diabetes.
J Manag Care Spec Pharm. 2016;22(11):1224–46.
13. Vollmer WM, Xu M, Feldstein A, Smith D, Waterbury A, Rand
C. Comparison of pharmacy-based measures of medication
adherence. BMC Health Serv Res. 2012;12(12):155.
14. Hess LM, Raebel MA, Conner DA, Malone DC. Measurement of
adherence in pharmacy administrative databases: a proposal for
standard definitions and preferred measures. Ann Pharmacother.
2006;40(7–8):1280–8.
15. Steiner JF, Prochazka AV. The assessment of refill compliance
using pharmacy records: methods, validity, and applications.
J Clin Epidemiol. 1997;50(1):105–16.
16. Choo PW, Rand CS, Inui TS, Lee ML, Cain E, Cordeiro-Breault
M, et al. Validation of patient reports, automated pharmacy
records, and pill counts with electronic monitoring of adherence
to antihypertensive therapy. Med Care. 1999;37(9):846–57.
17. Guenette L, Moisan J, Preville M, Boyer R. Measures of adher-
ence based on self-report exhibited poor agreement with those
based on pharmacy records. J Clin Epidemiol.
2005;58(9):924–33.
18. Kelly K, Grau-Sepulveda MV, Goldstein BA, Spratt SE, Wolfley
A, Hatfield V, et al. The agreement of patient-reported versus
observed medication adherence in type 2 diabetes mellitus
(T2DM). BMJ Open Diabetes Res Care. 2016;4(1):e000182.
19. Lau HS, de Boer A, Beuning KS, Porsius A. Validation of
pharmacy records in drug exposure assessment. J Clin Epidemiol.
1997;50(5):619–25.
20. Caetano PA, Lam JM, Morgan SG. Toward a standard definition
and measurement of persistence with drug therapy: examples
from research on statin and antihypertensive utilization. Clin
Ther. 2006;28(9):1411–24 (discussion 0).
379
21. Karve S, Cleves MA, Helm M, Hudson TJ, West DS, Martin BC.
An empirical basis for standardizing adherence measures derived
from administrative claims data among diabetic patients. Med
Care. 2008;46(11):1125–33.
22. Peterson AM, Nau DP, Cramer JA, Benner J, Gwadry-Sridhar F,
Nichol M. A checklist for medication compliance and persistence
studies using retrospective databases. Value Health.
2007;10(1):3–12.
23. Blais L, Vilain A, Kettani FZ, Forget A, Lalonde G, Beauchesne
MF, et al. Accuracy of the days’ supply and the number of refills
allowed recorded in Quebec prescription claims databases for
inhaled corticosteroids. BMJ Open. 2014;4(11):e005903.
24. Sperber CM, Samarasinghe SR, Lomax GP. An upper and lower
bound of the medication possession ratio. Patient Prefer Adher.
2017;11:1469–78.
25. Bjarnadottir MV, Malik S, Onukwugha E, Gooden T, Plaisant C.
Understanding adherence and prescription patterns using large-
scale claims data. Pharmacoeconomics. 2016;34(2):169–79.
26. Baggarly SA, Kemp RJ, Wang X, Magoun AD. Factors associ-
ated with medication adherence and persistence of treatment for
hypertension in a Medicaid population. Res Soc Admin Pharm
RSAP. 2014;10(6):e99–112.
27. Krousel-Wood M, Holt E, Joyce C, Ruiz R, Dornelles A, Webber
LS, et al. Differences in cardiovascular disease risk when anti-
hypertensive medication adherence is assessed by pharmacy fill
versus self-report: the Cohort Study of Medication Adherence
among Older Adults (CoSMO). J Hypertens. 2015;33(2):412–20.
28. Laliberte F, Bookhart BK, Nelson WW, Lefebvre P, Schein JR,
Rondeau-Leclaire J, et al. Impact of once-daily versus twice-daily
dosing frequency on adherence to chronic medications among
patients with venous thromboembolism. Patient. 2013;6(3):213–24.
29. Aguilar KM, Hou Q, Miller RM. Impact of employer-sponsored
onsite pharmacy and condition management programs on medi-
cation adherence. J Manag Care Spec Pharm. 2015;21(8):670–7.
30. Shin S, Song H, Oh SK, Choi KE, Kim H, Jang S. Effect of
antihypertensive medication adherence on hospitalization for
cardiovascular disease and mortality in hypertensive patients.
Hypertens Res. 2013;36(11):1000–5.
31. Han E, Suh DC, Lee SM, Jang S. The impact of medication
adherence on health outcomes for chronic metabolic diseases: a
retrospective cohort study. Res Soc Admin Pharm RSAP.
2014;10(6):e87–98.
32. Chen S, Macaulay D, Swallow E, Diener M, Farooqui S, Xie J,
et al. Real-world adherence and persistence associated with
nebivolol or hydrochlorothiazide as add-on treatment for hyper-
tension. Curr Med Res Opin. 2014;30(4):637–43.
33. Rolnick SJ, Pawloski PA, Hedblom BD, Asche SE, Bruzek RJ.
Patient characteristics associated with medication adherence. Clin
Med Res. 2013;11(2):54–65.
34. Eakin MN, Brady T, Kandasamy V, Fivush B, Riekert KA.
Disparities in antihypertensive medication adherence in adoles-
cents. Pediatr Nephrol (Berlin, Germany). 2013;28(8):1267–73.
35. Kuhle J. Update on medication quality measures in medicare part
D plan star ratings-2017: pharmacy quality alliance. http://
pqaalliance.org/measures/cms.asp. Accessed 6 Dec 2017.
36. Medicare 2016 Part C and D Star Ratings Technical notes - First
plan preview draft, Updated 8/5/2015. Centers for Medicare and
Medicaid Services; 2016. https://www.cms.gov/Medicare/Prescri
ption-Drug-Coverage/PrescriptionDrugCovGenIn/Downloads/
2016-Technical-Notes-Preview-1-v2015_08_05.pdf. Accessed 6
Dec 2017.
37. Malo S, Aguilar-Palacio I, Feja C, Lallana MJ, Rabanaque MJ,
Armesto J, et al. Different approaches to the assessment of
adherence and persistence with cardiovascular-disease preventive
medications. Curr Med Res Opin. 2017;33(7):1329–36.
380
38. Kozma CM, Dickson M, Phillips AL, Meletiche DM. Medication
possession ratio: implications of using fixed and variable obser-
vation periods in assessing adherence with disease-modifying
drugs in patients with multiple sclerosis. Patient Prefer Adher.
2013;7:509–16.
39. Vink NM, Klungel OH, Stolk RP, Denig P. Comparison of var-
ious measures for assessing medication refill adherence using
prescription data. Pharmacoepidemiol Drug Saf.
2009;18(2):159–65.
40. Campbell NKJ, Saadeldin K, De Vera MA. The duality of eco-
nomic issues with medication non-adherence in patients with
inflammatory arthritis. Curr Rheumatol Rep. 2017;19(10):66.
M. V. Bjarnadottir et al.
41. Hiligsmann M, Boonen A, Rabenda V, Reginster JY. The
importance of integrating medication adherence into pharma-
coeconomic analyses: the example of osteoporosis. Expert Rev
Pharmacoecon Outcomes Res. 2012;12(2):159–66.
42. Gross R, Bilker WB, Strom BL, Hennessy S. Validity and
comparison of two measures of days supply in Medicaid claims
data. Pharmacoepidemiol Drug Saf. 2008;17(10):1029–32.
43. Jackevicius CA, Paterson JM, Naglie G. Concordance between
discharge prescriptions and insurance claims in post-myocardial
infarction patients. Pharmacoepidemiol Drug Saf.
2007;16(2):207–15.