Introduction
Medication use is an essential element of studies of drug safety or
effectiveness[1]. There have been several successful efforts to
define medication use, however, these efforts have focused on specific
applications of using medications within particular analyses[2–5].
Although real-world data (RWD) is being increasingly used to generate
real-world evidence (RWE) to guide clinical, policy, and regulatory
decision making, there are limitations associated with conducting
studies using these data. Medication data are not captured for research
purposes and often need to be transformed from unstructured to
structured data when incorporated into studies. The lack of formal
conceptual and operational medication use definitions may result in
failing to capture the varied dimensions of patients’ medication use.
Therefore, it is important to establish conceptual and operational
definitions of medication use to aid investigators and the scientific
community in designing, conducting, reporting on, and comparing across
studies.
To our knowledge, there is no uniform structure to help researchers
conceptualize or operationalize medication use in their studies. To
address this gap, we developed a structure that includes considerations
to address when developing medication use definitions. The structure can
be used by researchers and policy makers to understand and contextualize
findings from pharmacoepidemiology studies. We categorized our
considerations in two major categories: considerations for a conceptual
definition, or how medication use would ideally be defined; and
considerations for an operational definition, or how the medication usewas defined.
Considerations for the Conceptual
Definition of Medication Use
As illustrated in Figure 1, we suggest five key considerations for
developing the conceptual definition of the medication use.
Consideration 1: The context under which the medication
is being studied (or what about medication use the researcher trying to
capture). Consider whether the medication use is the central
exposure variable or a covariate in a study assessing effectiveness or
safety of a treatment, or an outcome in a study assessing prescribing
practices.
Consideration 2: The research concept of interest. That
is, whether the research focuses on a specific ingredient that may be
found in more than one medication, on a specific medication, or in a
class of medications.
Consideration 3: The routes of administration that may be
of interest. Many medications have different indications and uses in
their different formulations and have correspondingly different routes
of administration. For example, corticosteroids may be administered via
many routes and for myriad indications. The conceptual definition should
also note whether any routes of administration are excluded.
Consideration 4: Medication dose if of interest. If of
interest, the conceptual definition should include a description of the
dose for each administration or daily dose, and an estimated cumulative
dose. It may also be important to describe how different dosage forms
(e.g., parenteral versus oral) will figure into the dose calculation if
multiple forms are available. Researchers should also consider that
specific dosages of some drugs are highly difficult to identify. They
might want to incorporate definitions that consider dose changes over
time.
Consideration 5: The ideal timing and duration of
medication use. Considerations about whether the date the patient
initiated the medication, the timing of subsequent uses of that
medication, relative timing of medication use and health events, whether
continuous or cumulative exposure (i.e., duration), and whether
discontinuation or changes to a different medication (as a proxy for
standard of clinical care, lack of drug tolerance, or success/failure of
a treatment) are of interest will depend on the study question. It is
also useful to consider that the drug effect may persist after treatment
discontinuation or that there may be gaps in treatment, which may have
implications for the timing of administration in the analysis. This will
usually require an understanding of the pharmacological characteristics
of the medication.
Considerations for the Operational
Definition of Medication Use
As illustrated in Figure 2, we suggest five key considerations for
developing the operational definition of the medication use.
Consideration 1: The underlying pattern of health seeking
behavior and its documentation within the health system that give rise
to the observable data. There are various types of encounters patients
may have as they move within the health care ecosystem that may
influence the operational definition. A strategy to screen for the
records or occurrences of the medication use, should be informed by the
format in which the medications data are captured and stored. For
example, during hospitalization, a patient’s medication use may be
collected in different ways; hospital clinical staff use patients’
charts information to enter patient medication use into electronic
health record (EHR) (e.g., clinician-generated prescription orders,
medication administrations, pharmacy dispensing); hospital clinical
staff ask patients about past and current medication use; or
prescription orders recorded as part of hospital discharge instructions.
This information may be transformed to a common data model, used for
billing and payment purposes, and potentially included in medical claims
(though detail related to medication use in claims may be limited).
Consideration 2: The characteristics of data source that
were used (e.g., location,) and approach used to identify the medication
used. As can be seen in Figure 3, identification of a medication
requires finding of the likely location (e.g., specific data
tables/views and fields) where relevant medications-related information
is stored, which is heterogenous. Therefore, we suggest that decisions
regarding appropriate sources of medication information first be guided
by knowledge of clinical practice and how the data are produced and
sorted (e.g., clinician decision to order a medication versus
administration of the medication to the patient). The format of the
sorted data should also be considered. For example, whereas
administrative and claims data are highly structured and standardized in
format by virtue of the specific requirements to which providers must
adhere to receive payment, other data sources such as EHRs are often
more heterogeneous. Medication information may be stored as coded data
within structured fields, text within semi-structured or discrete
fields, or within free-text clinical notes. Coded medications data may
take the form of organization-specific or local codes, or may include
standard terminology or classifiers (e.g., RxNorm, NDC, Multum,
FirstDataBank). Whether medication information is stored as codes or as
text, it is useful to submit supplementary materials include the
structured code lists or list of text strings used to identify
occurrences of the medication use of interest. Similarly, it is
essential to report operational definitions and methods when using
information from semi-structured or unstructured data. Regardless,
researchers might consider engaging individuals with expertise of the
specific data source being used.
Consideration 3: Actual timing elements related to
medication use. The ability to define various aspects of timing of
medication use is an integral component of pharmacoepidemiologic
research questions. The researcher should describe how they identified
the date of patient initiation of a medication and the timing of
subsequent uses of that medication, date of medication discontinuation,
and relative timing of medication use and health events of interest, as
they were defined in the conceptual definition. They should also specify
how they calculated the duration of medication use if it is of interest,
including any assumptions of medication use tied to the definition
(e.g., identifying a prescription within an EHR and assuming the
medication was filled and used for the full days supplied).
Consideration 4: Other features of medication use. The
features of medication use that are of interest should be defined. These
features might include whether prevalent or incident use is of interest
and how it was defined; whether patients can have multiple qualifying
episodes of medication use or only one; and how dose was calculated, if
of interest. Researchers should also describe how they addressed missing
data and discrepancies if they are using more than one data source.
Consideration 5: Algorithm validation. If an algorithm
was used to identify the medication, researchers should specify whether
the algorithm for the medication measure was validated.