A data‐informed approach using individualised dispensing patterns to estimate medicine exposure periods and dose from pharmaceutical claims data

Abstract

Pharmaceutical claims data are often used as the primary information source to define medicine exposure periods in pharmacoepidemiological studies. However, often critical information on directions for use and the intended duration of medicine supply are not available. In the absence of this information, alternative approaches are needed to support the assignment of exposure periods. This study summarises the key methods commonly used to estimate medicine exposure periods and dose from pharmaceutical claims data; and describes a method using individualised dispensing patterns to define time‐dependent estimates of medicine exposure and dose. This method extends on important features of existing methods and also accounts for recent changes in an individual's medicine use. Specifically, this method constructs medicine exposure periods and estimates the dose used by considering characteristics from an individual's prior dispensings, accounting for the time between prior dispensings and the amount supplied at prior dispensings. Guidance on the practical applications of this method is also provided. Although developed primarily for application to databases, which do not contain duration of supply or dose information, use of this method may also facilitate investigations when such information is available and there is a need to consider individualised and/or changing dosing regimens. By shifting the reliance on prescribed duration and dose to determine exposure and dose estimates, individualised dispensing information is used to estimate patterns of exposure and dose for an individual. Reflecting real‐world individualised use of medicines with complex and variable dosing regimens, this method offers a pragmatic approach that can be applied to all medicine classes.

Key Points.

• Pharmaceutical claims data often do not record information on exact indications and directions for use, as well as the intended duration of medicine supply.

• Methods for translating available dispensing data into exposure periods, including specifying duration of use and estimates of daily dose, need to be robust and fit‐for‐purpose.

• This paper summarises the key methods commonly used to estimate medicine exposure periods and dose from pharmaceutical claims data when information on prescribed dose and duration of supply are not available.

• A data‐informed approach based on individualised dispensings patterns is proposed for estimating medicine exposure periods and dose that account for individual‐level variation in use over time.

• Reflecting real‐world individualised medicine use, this method offers a pragmatic approach that can be applied to all medicine classes.

Plain Language Summary.

Dispensing records often contain information on the dates of dispensings, item codes, and the quantity dispensed, but not the duration or directions for use. In medicines research, we often want to be able to translate basic dispensing information to continuous exposure periods; that is, determine the periods of time when someone was exposed to a particular medicine. While we can make reasonable assumptions around exposure for medicines that are prescribed and used relatively consistently across the population, it becomes harder for medicines where the patterns of use vary, not only between people, but also within an individual over time. This study summarises the key methods commonly used to estimate medicine exposure periods and dose from dispensing records; and describes a method using individualised dispensing patterns (IDP) to define time‐varying estimates of medicine exposure and dose. The IDP method looks at an individual's recent dispensings for a particular item—specifically the time between dispensings and the quantity supplied—and uses these details to estimate how long a new dispensing will last. Reflecting real‐world individualised medicine use as indicated by the pattern of recent dispensings, the IDP method offers an approach that can be applied to all medicine classes.

1. INTRODUCTION

Determining medicine exposure periods is critical in pharmacoepidemiological research for exposure‐outcome studies and in quantifying persistence and adherence to treatment. Pharmaceutical claims (i.e., dispensing) data are often used as a key information source for defining medicine exposure periods. ¹ In many jurisdictions, including across Europe ² and Australia, ³ however, pharmaceutical claims data do not record information on exact indications, directions for use, or the intended duration of medicine supply. As such, medicine use patterns are often indirectly measured using dispensing dates. Specific methods are therefore needed to translate available dispensing data into exposure periods, including specifying duration of use and estimates of daily dose.

This paper summarises the key methods commonly used to estimate medicine exposure periods and quantify the level of exposure from pharmaceutical claims data that are suitable when information on prescribed dose and duration of supply are not available. It also outlines a novel data‐informed approach using individualised dispensing patterns (IDP) to estimate medicine exposure periods and dose that accounts for individual‐level variation in use over time. Finally, to evaluate the impact of accounting for between‐ and within‐individual variability in assigning exposure periods using the IDP method, summaries of episode duration and person‐time exposed are compared with a conventional, population‐based approach to evaluating exposure periods for three medicine types.

2. COMMON METHODS USED TO ESTIMATE MEDICINE EXPOSURE PERIODS

The methods used to estimate medicine exposure periods using pharmaceutical claims data could be broadly grouped into fixed and data‐driven approaches. In most cases, the date of index dispensing is usually assigned as the start of the exposure period, with different approaches used to determine the duration of this period, accounting for potential gaps and overlapping dispensings. Where there are overlapping exposure periods assigned to different dispensings, these are deemed to represent a continuing period of exposure.

2.1. Pre‐processing considerations

Prior to creating continuous exposure periods, it is important to consider a range of practical considerations, such as potential lag effects, grace periods and carry‐overs (see Table 1). ⁴ ‘Lag time’ reflects the processes associated with the absorption phase of a medicine following its administration; in this way, a ‘lag effect’ accounts for the potential for an outcome to occur after an exposure period ends but during which pharmacological activity may still be plausible. ‘Grace periods’ reclassify brief interruptions in exposure between the end of the calculated duration from one dispensing to the next dispensing as ‘exposed’ to minimise misclassification. While a ‘carry‐over’ effect accounts for the potential extension to an exposure interval resulting from the use of medicine supplies remaining from previous dispensings.

TABLE 1.

Common features of medicines use incorporating medicine exposure periods using pharmaceutical claims data

Feature	Brief description	Strengths	Limitations
Lag effect	Accounts for the time period after a drug is discontinued during which an outcome could still be reasonably attributed to the exposure.	Takes into account the drug's pharmacology and theorized mechanism of action. Captures periods of potential lower level exposure.	Using different methods may lead to different estimates of drug effects.
Grace periods	Permissible time intervals from the end of the calculated exposure duration from one dispensing to the start of the subsequent dispensing (or ‘gaps’) during which an individual who appears to be unexposed is considered exposed. Variants include multiplying estimated exposure duration with some factor greater than one.	May minimise misclassification during apparent brief interruptions in exposure or allow for less than perfect adherence.	When the assumptions on gaps are distant from the real drug utilisation, the treatment episodes estimates may be biased. Using different methods may lead to different estimates of drug effects.
Carry‐overs	The practice of accounting for overlapping prescriptions, for example, overlapping days may be added to the episode duration or ignored.	Smooths random variation introduced by irregular dispensing patterns and assumes that early dispensings (relative to calculated exposure duration from previous dispensing) are an intentional result of planned stockpiling of drugs.	When the assumptions on carry‐overs are distant from the real drug utilisation, the treatment episodes estimates may be biased. Using different methods may lead to different estimates of drug effects.

Similarly, if additional patient healthcare data is available, it may be possible to identify intervals of time where an individual may remain exposed, but there are gaps in dispensing information, such as during a hospital admission. For these intervals, an approach for defining exposure and daily dose is required. The simplest approach is to assume medicine exposure during hospital admissions follow the same pattern as when not in hospital and thus, the exposure method is applied as normal. Assuming uninterrupted exposure may not be appropriate for all studies, and sensitivity analyses (e.g., excluding these intervals of time) should be conducted to assess robustness of the study findings in relation to how exposure during these periods are specified. Clear explanations of how potential lag effects, grace periods, carry‐overs and intervals where no dispensings could occur are accounted for when constructing exposure periods and in subsequent analyses are important, as there is wide variation in their application (see Table 2). ⁴

TABLE 2.

Commonly methods used to create medicine exposure periods from pharmaceutical claims data

Method	Main assumptions	Method of accounting for grace periods	Method of accounting for carry‐overs	Method of accounting for lag effect	Uses retrospective data	Factors in quantity dispensed	Accounts for individual patterns of dispensings	Main strengths	Main limitations	Methods paper(s) or example application
Fixed‐time windows	Assumes medicine use following each dispensing lasts for a fixed time, for example, 60 or 120 days, with overlapping dispensings belonging to a single treatment episode. Time window is sometimes based on the maximum time a single dispensing can last according to the dispensing regulations of the setting/system	Varies between studies; methods include predefining a certain number of days (from 0 to 180 days) or percentage of exposure duration from previous dispensing	Varies from none to all remaining supply	Varies between studies, sometimes informed by a medicine's pharmacology/mechanism of action	Yes	No	No	Simplicity of application	Does not take into account the dispensed amount, any dosage or any other features of the dispensing except for the dispensing date	Bedson et al. (2018) 30
Estimation of medicine coverage, COV	An averaged fraction of prescribed dose units is estimated comparing the accumulated dose units and the elapsed time tk−t1 from k consecutive prescriptions (for k ≥ 2 redemptions resembling regular use within a period comprising prescriptions up to as many days as 2.5 times the latest prescribed quantity).	User‐specified: applied example assumes people have 15 days of medicine supply, which is used to fill apparent gaps between prescriptions using prospective filling.	User‐specified: applied example allowed a maximal overlap duration corresponding to 25% of the quantity of the last overlapping prescription to be added.	Not specified	Yes	Yes	Yes	Takes into account regularity of previous purchasing behaviour and quantity dispensed in calculating exposure periods.	Taking into account all previous dispensings in calculating the exposure period weighs the history considerably and thus, is less reactive to current dose changes.	Meid et al. (2016) 12
PRE2DUP	Uses temporal dosages, package information with refill time distribution and peoples' personal dispensing patterns to construct exposure time periods and estimates the dose used during the period.	None	When exposure duration does not reach the next dispensing and the local dose has temporarily declined, a ‘stockpiling test’ calculates if the current and previous dispensing joined together would reach the next dispensing if they were both dispensed at the time of the previous dispensing. If they do, the medicine use period continues; otherwise, the medicine use period ends.	None	No	Yes	Yes	Takes into account local estimate of dose, regularity of purchasing behaviour and dose limits set by experts in calculating exposure periods.	Dependent on future dispensings. Implementation available in limited number of programs and computational complex (iteration process is run until results are stable). Multiple expert‐defined parameters are required for each drug package, which can equate to thousands of parameters for studies with large numbers of drug types.	Tanskanen et al. (2015) 13
Reverse waiting time distribution (WTD)	Estimates the probability of medicine use at any time after dispensing from a maximum likelihood‐based estimation algorithm for the reverse WTD, which is the distribution of time from the last dispensed prescription of each patient before an index time point.	None	None	None	Yes	Yes, if included as a covariate in the estimation algorithm	No	May reduce misclassification in that it avoids the need for binary classification of exposure. Based on an explicit and consistent mathematical model.	Requires the covariates underlying the variation in patterns of medicine use be available in the data and time‐invariant,	Stovring et al. (2017) 10 ; Stovring et al. (2017) 11

2.2. Methods for defining exposure periods

Among the simplest of existing methods for creating medicine exposure periods is the ‘fixed‐time window,’ which assumes medicine exposure from each dispensing lasts for a fixed‐time period, for example, 14, 30, or 60 days. ⁵ , ⁶ The time interval is often based on the maximum time that medicines are supplied according to jurisdictional systems and regulations. An alternative fixed‐time method defines exposure based on whether at least one dispensing occurs within a fixed interval of time, for example, each 30‐day period following the date of the index dispensing. Such approaches are termed ‘fixed’ in that they do not account for any individual characteristics of the dispensing, such as the quantity dispensed, total amount supplied (in milligrams or other unit of measure), or any characteristic other than the dispensing date.

Extending on the ‘fixed‐time window method’, the ‘fixed‐dose method’ assumes that a fixed‐dose or number of units is used across a given time period (e.g., one tablet per day). The dose amount is often selected on the basis of the average dose per day when the medicine is used for its main indication. ⁷ , ⁸ This method is most suitable for medicines with fixed dosing regimens, or when used relatively consistently across the population. However, for many medicines (e.g., opioid analgesics) it is expected that patterns of use will vary not only between individuals, but also within an individual over time.

To better characterise individual‐level patterns of use for such medicines, data‐driven methods can use the patterns of each individual's dispensings to determine exposure periods. These methods rely on data for the frequency of an individual's dispensings for each medicine as a representation of individual medicine use (and hence exposure).

The waiting time distribution (WTD) method, for example, takes a data‐driven approach to evaluating a population estimate of exposure duration for a single dispensing of the medicine of interest. The time between a specified start day (e.g., 1st of January 2004) and when users are first dispensed a specific medicine inside a subsequent time window (e.g., the remaining calendar year) is evaluated. From the resulting distribution of time intervals, the number of days corresponding to a user‐specified threshold (e.g., 80th percentile) is identified. This value is then considered the maximum interval of time between two dispensings that would represent a continuous treatment episode, ⁹ and is subsequently applied to all dispensings for the population. An extension of this approach, the reverse WTD model, evaluates the time from the last dispensing to the end of a specific time window. Mathematical properties underlying the reverse WTD allows the exposure duration estimate to vary depending on observed individual and dispensing characteristics (e.g., age, number of tablets). ¹⁰ , ¹¹ By capturing variation in the patterns of use for different sub‐populations, the reverse WTD method takes a more individualised approach to determining exposure periods compared to the original WTD method.

In the estimation of drug coverage (COV) method, the prior dispensing history for an individual is used to approximate the exposure duration for a given dispensing based on the accumulated dose units and elapsed time from previous dispensings. ¹² This method accounts for the frequency and quantity of previous dispensings to calculate exposure durations. However, by incorporating all previous dispensings since the index dispensing, recent changes in dose (i.e., potential dose tapering or dose escalation) may be diluted.

Finally, the prescription 2 drug use period (PRE2DUP) method uses data on refill time distributions and individual‐level dispensing patterns to establish exposure periods and estimate dose during those periods. ¹³ The decision procedure is multi‐step in that it involves specifying a set of expert‐defined global parameters; a pre‐processing step in which a sliding average uses information from previous and subsequent dispensings to generate individualised dose estimates; and then a core process that is applied to dispensings in chronological order to calculate expected refill length of each dispensing, the dose during that interval, and continuous periods of exposure. Depending on the purpose of its application, a potential weakness of PRE2DUP is that the sliding average in the pre‐processing step uses information after the current dispensing. This may violate the assumption of certain statistical methods, such as time‐to‐event analyses, which require time‐varying covariates be defined without looking into the future. ¹⁴

With respect to using future information, there are several time‐related biases that are important to consider when undertaking exposure‐outcome studies. ¹⁵ Immortal time bias, for example, can occur in a cohort study when the follow‐up period includes time periods during which no outcome events can occur (i.e., individuals are ‘immortal’). In the case of quantifying time‐varying exposure, this may occur when future dispensing information is used to identify when a new medicine is initiated to classify an individual as being exposed at baseline. Although immortal time and other time‐related biases can significantly impact exposure‐outcome relationships, they can be avoided with appropriate study design and analysis strategies. ¹⁵

3. COMMON METHODS USED TO ESTIMATE DOSE OR LEVEL OF EXPOSURE

Determining dose–response relationships are often central to understanding the relationship between exposure to a particular medicine and a given outcome. This requires exposure estimates from dispensing data to shift from evaluating duration to quantifying the level of exposure as measured by either duration or dose (see Table 3 for some common methods). The simplest approach for quantifying the level of exposure extends on the methods for defining exposure periods described above to measure, on any given day, the cumulative duration of use calculated either as the total duration of exposure since the start of the exposure period or the summed amount of exposed time over a recent time‐fixed interval (e.g., exposed person‐days in past 12 months).

TABLE 3.

Overview of measures used to evaluate duration‐response and dose–response relationships in pharmacoepidemiological exposure‐outcomes studies

Method	Definition	Main strength(s)	Methodological consideration(s)	Example of application
Duration of use within recent interval	Model the duration of exposure within a specified interval, for example, 90 days	Simplicity of application	Does not capture other aspects of exposure, such as dose	Sylvestre (2012) 31
Continuous use duration	Various approaches used in measuring length of use, including length of time since last unexposed or total amount of time since first use of the specific formulation	Captures duration of exposure over time	Does not capture lifetime exposure or other aspects of exposure, such as dose	Hicks (2016) 32
Average daily dose	Calculated by dividing the total dose dispensed by the number of exposure days (evaluated either using the amount of days supply as recorded in the dispensing records or estimated using medicine use period exposure methods)	Takes into account recent changes in dispensed amounts	In the case where number of days supplied is available, this assumes medicines are used in an “as prescribed” manner	Park (2015) 33
			Where number of days supplied is not available and exposure methods are used to evaluate duration, dose evaluation has similar limitations to the selected exposure method
Cumulative dose	Increment cumulative dose at each dispensing with total dose dispensed on that day	Takes into account the time since exposure and the magnitude of the exposure	The sum of all doses up until a particular point in time does not account for changes in dose or variation in dosage that may impact risk	Tuccori (2016) 34
		Indirectly models the effective concentrations that may persist in the body after the dose has been taken	Time‐varying confounders acting as intermediates are not accounted for
		Does not need to take into account current estimates of exposure or assumptions around carry‐overs/ stockpiling for future dispensings
Weighted cumulative dose/duration	Calculated by (i) multiplying each dose taken in the past by a weight that represents the relative importance of that dose for the current risk, and (ii) summing the weighted past doses up to the current time.	Attempt to factor into time‐varying analyses that (1) the effect of exposure can cumulate over time and (2) that the different time points of the exposure have different impacts on the outcome.	Time‐varying confounders acting as intermediates are not accounted for	Abrahamowicz (2006) 16

With respect to dose, information on the strength and quantity of the medicine dispensed are available in most pharmaceutical claims databases, but specific dosing instructions are often not available. Even among databases with dosing information available, it is unclear to what extent an individual's actual use is consistent with the dosing information prescribed or recorded. Various approaches to investigating dose–response relationships have been proposed depending on the expected pharmacological activity of a given medicine. For example, where deemed clinically appropriate, cumulative dose can be calculated at each dispensing day as the sum of all previous total amounts supplied (in milligrams or other unit of measure). Such an approach accounts for the magnitude of the exposure, can model the effective concentrations that may persist in the body after the dose has been taken, and avoids the need to estimate a daily dose. Weighted cumulative dose (or duration) is a further extension that accounts for the potential for exposure to a medicine to accumulate over time and for the timing of the exposure to have different impacts on the outcome. ¹⁶ Such an approach multiplies and sums each dose taken in the past by a weight that represents the relative importance of that dose for the current risk of an outcome.

4. A NUANCED DATA‐DRIVEN APPROACH USING INDIVIDUALISED DISPENSING PATTERNS TO ESTIMATE MEDICINE EXPOSURE PERIODS

Many of the methods for estimating medicine exposure periods described earlier present limitations in some analytical scenarios; for example, exclusively using population‐based parameters, diluting recent changes in use, and the use of future information. To address these limitations and extend on the important features of existing methods, a method specifically designed for use when data on duration is unavailable and for medicines with complex and variable dosing regimens was developed.

The individualised dispensing patterns (IDP) method can be used when a time‐dependent exposure is needed that captures individual‐level use and dose changes over time. It has been developed primarily for application to databases, which do not contain information on duration of medicine supply or dose, but also may facilitate investigations even where such information is available (i.e., where actual use may deviate from the intended use). Core variables required for implementation of this method include the dates of dispensings, medicine item codes and the quantity dispensed. Information on strength and administration route are not essential but can facilitate investigations of differences in patterns of dispensings across these characteristics and (for medicine strength) evaluations of dose.

The IDP method uses data on the frequency an individual is dispensed medicines to estimate medicine exposure periods. As individuals can have multiple dispensings for a given medicine on the same day, a dispensing parameter (d_n) is defined as the nth day the individual had at least one dispensing record for a given medicine and the quantity dispensed (q_n), equal to the sum of quantities of the given medicine from all dispensing records on that day. The method is applied collectively to all strengths of a specific formulation of a medicine (e.g., tablets/capsules, transdermal patches, oral solutions), provided they are used for the same broad indication (e.g., by Anatomical Therapeutic Chemical code ⁸ ) or follow similar dosing regimens. It is applied to all dispensings of that formulation over the duration of an individual's observation period, starting with their first index dispensing.

4.1. Step 1: Define population‐based exposure estimate

First, a population per‐quantity exposure estimate (e_pop) is defined and applied to all index dispensings for a specific medicine formulation. The e_pop parameter provides an initial estimated exposure duration for an index dispensing as there is not yet sufficient dispensing data to establish an individualised dispensing pattern. The e_pop is calculated using data from all individuals with at least two dispensings of the formulation of interest. The time interval (in days) between any two successive dispensings (d_n and d_n + 1) over the observation period is calculated and divided by the quantity dispensed at the first of the two successive dispensings (q_n). To avoid an extreme skewed distribution, time intervals should be excluded where the duration between two dispensings exceeds a reasonable amount of time for the dispensings to be considered as part of a single continuing episode (e.g., if there is a gap of 1 year or more between dispensings ¹⁷ ). The e_pop is subsequently derived from the distribution of per‐quantity exposure estimates from the population dispensed the formulation of interest and expressed as the maximum interval between dispensings. ⁹ An 80th percentile threshold may be appropriate in defining this maximum interval, in line with previous exposure methods. ¹⁰ However, this should be reviewed in line with the clinical context, sensitivity of the estimates, and an awareness of the impact on exposure evaluations. For example, a higher threshold value will increase the number of exposure intervals, which overlap with subsequent dispensings and result in longer episodes.

4.2. Step 2: Assign exposure days from index dispensing

At each individual's index dispensing date, evaluate the estimated number of exposure days as the product of the population per‐quantity exposure estimate (e_pop; evaluated in Step 1) and the quantity dispensed (q_n), for example, est_e_n = e_pop x q_n. If the interval of time to the next dispensing (t₁) is equal to or less than the estimated length of exposure, define exposure as current until the next dispensing (d_n + 1). Otherwise, define exposure status over the time interval as: current for the estimated number of exposure days following the dispensing; recent from the end of the current interval to 7 days afterwards (or until the next dispensing, whichever occurs earlier); and former from the time the recent interval ends to the next dispensing (or end of follow‐up, whichever occurs earlier). Sensitivity analyses varying the length of the recent interval are recommended to assess the robustness of resulting findings to this feature.

4.3. Step 3: Assign exposure days for each subsequent dispensing

For all non‐index dispensings, the time between prior dispensings is incorporated into estimations of the number of exposure days by use of a sliding weighted temporal average of the per‐quantity exposure estimate from the previous three dispensings in the same exposure episode as follows:

$$\mathrm{est}\_{\exp }_n={\mathrm{q}}_n\times \left(\frac{1}{6}{\mathrm{e}}_{n-3}+\frac{2}{6}{\mathrm{e}}_{n-2}+\frac{3}{6}{\mathrm{e}}_{n-1}\right)$$ (1)

Where e_n−1 is the per‐quantity exposure estimate from the (n−1)^th dispensing, calculated as the number of days between d_n−1 and the next dispensing (d_n), divided by the quantity dispensed at d_n−1 (i.e., q_n−1). The previous dispensings are weighted with weights 3/6, 2/6, and 1/6 for the first, second, and third previous dispensings, respectively. Greater weights are assigned to more recent dispensings to capture recent changes in the frequency of dispensings, while smoothing across multiple prior dispensings minimises the impact of potential stockpiling of medicines. For the second and third dispensings following an index dispensing, the population per‐quantity exposure estimate is used in place of unavailable individual parameter(s)—that is, e_pop replaces e_n−3, and e_n−2 for the second dispensing, and e_n−3 for the third dispensing following an index dispensing.

If the interval of time to the next dispensing is equal to or less than the estimated exposure period, define exposure as current until the next dispensing (d_n + 1). Otherwise, define exposure status over the time interval using the same approach used to define current, recent, and former exposure described above for index dispensings (see Figure 1).

FIGURE 1.

Two common scenarios representing different ways in which the individualised dispensing pattern (IDP) method accommodates complex and variable dosing regimens compared to the fixed‐dose method

4.4. Step 4: Calculate exposure days from the last dispensing

A similar approach is used to define exposure following each individual's last dispensing during the observation period; if the interval of time to the end of follow‐up is equal to or less than the estimated exposure period, define exposure as current until the next dispensing. Otherwise, define exposure using the same approach described in Step 2.

Where the interval between two dispensings includes at least 1 day with former exposure, it is taken to mean that the medicine supply has been exhausted and the dispensings do not belong to the same exposure period. In this case, an exposure period ends on the last day of recent use and a new period is started at the next dispensing. As a break in exposure is assumed to have occurred, the first dispensing in the new exposure period is considered an index dispensing and exposure for the subsequent interval is evaluated using the index dispensing method for evaluating exposure, that is, using the population estimate (e_pop) rather than sliding weighted temporal average estimate of exposure (e_n).

With the use of a recent exposure period, this feature incorporates a combination lag and grace period—the length of which should be informed by the medicines, population, and settings of interest. Furthermore, while this method focuses on specific formulations of a medicine rather than assessing a unique item type, this decision should again be considered on a case‐by‐case basis. For example, where it is common for two medicines to be used in combination, exposure to each medicine would be evaluated separately by the IDP method, such that an individual switching between these medicines or transitioning to their combined use can be identified.

5. EXTENDING THE INDIVIDUALISED DISPENSING PATTERN EXPOSURE PERIOD METHOD TO ESTIMATE DOSE

Similar to defining periods of exposure, the method for defining dose uses individual‐level data on pharmaceutical claims. It extends on the method for defining exposure periods by replacing the quantity dispensed with total amount in milligrams supplied (or other unit of measure, calculated by multiplying the strength by quantity dispensed) in Steps 1–3.

5.1. Estimating dose for index dispensing

Specifically, a population‐level daily dose estimate can be generated by replacing q_n from Step 1 with the total amount supplied on d_n (A_n). The population daily dose estimate is then defined as the daily dose estimate over current exposure days in the interval following the index dispensing. Where the total amount over this interval is greater than the amount dispensed at the index dispensing, the daily dose is revised to equal the result of dividing the total amount supplied on d_n (A_n) by the number of current exposure days in the interval.

5.2. Estimating dose for subsequent dispensings

Similar extensions are applied for defining dose for all non‐index dispensings. The time between prior dispensings is incorporated into daily dose estimates by using a sliding weighted temporal average of the daily dose estimate from the previous three dispensings in the same exposure period. The per‐quantity exposure estimate from formula (1) is replaced with the result from dividing the total amount supplied from the (n−1)^th dispensing and the number of current exposure days between d_n−1 and the next dispensing, and so on for the second and third previous dispensings. Conversion factors may be applied for further extensions of dose estimates; for example, tailoring dose of opioid use by multiplying by an oral morphine equivalent conversion factor. ¹⁸

6. COMPARATIVE EVALUATION OF THE INDIVIDUALISED DISPENSING PATTERNS AND FIXED‐DOSE METHODS

To demonstrate the key benefits of the IDP method in capturing exposure of complex and variable dosing regimens, a practical application is used to directly compare exposure estimates generated from an approach that does not allow for any individual‐level variability; specifically, the fixed‐dose approach. The fixed‐dose method represents a conventional, population‐based approach to evaluating exposure durations, which varies only by the quantity dispensed. Three medicine types—fentanyl patches, oxycodone controlled‐release, and oxycodone immediate‐release—were chosen for this evaluation as examples of medicines with potential variability in patterns of use (for information on cohort, codes, and methods, see the published protocol ¹⁹ and Appendix). Comparisons of exposure to these medicines as assessed by the IDP and fixed‐dose methods were evaluated in two ways: first, to investigate differences at the individual‐level, summary statistics of exposure episodes for each method were evaluated. Second, lacking a gold‐standard measure of exposure, cross‐tabulations of person‐time exposed for the two exposure measures allowed for the relative level and type of discordance between methods to be evaluated.

The evaluation revealed that at the individual level, the IDP method defined fewer episodes of longer length compared to the fixed‐dose method (see Table 4). Correspondingly, intervals of person‐time during which individuals were formerly exposed tended to be longer when using the IDP method. This has direct implications for dose in that the fixed‐dose method defines a fixed population‐level estimate of dose across an observation period. In contrast, the IDP method registers a continuing exposure episode and adjusts (either up or down) an individual's dose based on the intervals of time between previous dispensings in the same episode (see Figure 1).

TABLE 4.

Episode summaries and person‐time exposed as evaluated by the individual dispensing patterns (IDP) method and fixed‐dose method for select medicine types

Exposure definition	Fentanyl patches (N = 15 088)		Oxycodone controlled‐release (N = 39 632)		Oxycodone immediate‐release (N = 523 527)
	Fixed‐dose method	IDP method	Fixed‐dose method	IDP method	Fixed‐dose method	IDP method
Number of episodes a	29 708	28 944	69 609	60 920	1 103 065	774 296
Episode length (days)
Median (IQR)	29 (30)	38 (42)	14 (14)	28 (20)	5 (0)	48 (0)
Max	1824	1589	1787	1498	1777	1669
Formerly exposed intervals (days)
Median (IQR)	15 (30)	37 (563)	14 (10)	510 (1198)	5 (0)	528 (874)
Max	1560	1806	1638	1813	1456	1813
Proportion of person‐years (column %)
Currently exposed	74.4%	62.7%	53.5%	58.4%	33.3%	57.5%
Recently exposed b	—	23.5%	—	23.5%	—	22.2%
Formerly exposed	25.6%	13.9%	46.5%	18.0%	66.7%	20.4%

Note: N, number of individuals with at least one dispensing included in comparative evaluation.

^a For the IDP method, an episode was defined as a continuous ‘current’ exposure with no more than a 7‐day break (i.e., a complete interval of recently exposed).

^b Recently exposed only specified in the IDP method and defined for a maximum of 7 days following an interval of time classified as currently exposed during which the frequency of medicine use may be reducing and where pharmacological effects of a medicine may still be experienced. A recently exposed interval may be followed by an interval of time classified as currently exposed (where there was a new dispensing for the medicine) or formerly exposed (where there was no new dispensing).

With respect to concordance of exposure levels, cross‐tabulations of person‐time exposed show that for all three medicine types, more than half of all person‐time was classified as the same exposure status by both methods (see Table 5). Discordance in exposure status was smallest for fentanyl patches (22.1% of person‐time differently classified between methods), with the IDP classifying 16.9% of all person‐time as exposed when the fixed‐dose method did not. Discordance was greater for both oxycodone controlled‐release (34.6%) and oxycodone immediate‐release (47.8%), whereby the IDP method classified more person‐time as exposed compared to the fixed‐dose method.

TABLE 5.

Cross‐tabulations of the distribution of person‐time as evaluated by the individual dispensing patterns (IDP) method and fixed‐dose method for (A) fentanyl patches, (B) oxycodone‐controlled release and (C) oxycodone immediate‐release

		Fixed‐dose method		Total (column %)
		Currently or recently exposed	Formerly exposed
(A) Fentanyl patches
IDP method	Currently or recently exposed	69.3%	16.9%	86.1%
	Formerly exposed	5.16%	8.73%	13.9%
	Total (row %)	74.4%	25.6%	100%
(B) Oxycodone controlled‐ release
IDP method	Currently or recently exposed	50.4%	31.5%	82.0%
	Formerly exposed	3.1%	15.0%	18.0%
	Total (row %)	53.5%	46.5%	100%
(C) Oxycodone immediate‐release
IDP method	Currently or recently exposed	33.2%	47.7%	80.8%
	Formerly exposed	0.2%	19.0%	19.2%
	Total (row %)	33.3%	66.7%	100%

7. DISCUSSION

Given the importance of defining medicines exposure in pharmacoepidemiological research, there is great interest in advancing available methods and ensuring they are fit‐for‐purpose and are applicable to complex and variable dosing regimens. This paper summarised the main strengths, assumptions, and considerations of some of the most common methods used to define medicine exposure periods and dose from pharmaceutical claims data. Building on this, a nuanced data‐informed approach (the IDP method) was developed to establish time‐dependent medicine exposure periods for medicines with complex and variable dosing regimens for use in time‐to‐event analyses.

The IDP method uses a simple and pragmatic data‐driven approach that is based on individual dispensing patterns. It reduces the need to rely on pre‐specified duration data or dosing instructions to determine an estimated exposure period and dose of a given medicine dispensing. Even when duration and/or dosing information are available, real‐world use may vary from what was prescribed; this method provides an alternate approach to creating individualised medicine exposure periods that correspond to an individual's pattern of medicine use. Duration estimates rely on information from past dispensings only, and this way avoid the issue of using future information, which can violate the assumptions of most time‐dependent statistical analyses. Further, by accounting for only a select number of previous dispensings from the same treatment episode, random variation in the per‐quantity exposure and daily dose estimates is reduced, while accounting for recent changes in use. This method endeavours to move exposure and dose estimates away from solely relying on prescribed dose, while also addressing the issue of pro re nata (prn) dosing; it estimates exposure and dose based on dispensing patterns aligned with real‐world medicine use. Indeed, a comparison of the IDP and fixed‐dose methods for medicines with known variable dosing regimens (opioids) revealed that while most person‐time exposed was classified consistently by the two methods, the IDP method tended to define fewer episodes of longer length. Finally, the inclusion of a recent exposure period provides a soft transition between current and former use during which the frequency of use may be reducing and where pharmacological effects of a medicine may still be experienced. ⁶

A worked empirical example using opioid medicines was used to demonstrate the practical application of the IDP method to a widely used medicine class with often complex and individualised dosing regimens. Future work validating this method against measures of actual medicines use will be beneficial. One approach is to assess the level of agreement between IDP exposure periods derived from pharmaceutical claims data and self‐reported medicine use data collected from medication diaries or interviews. ²⁰ , ²¹ Although self‐report medicines use data can be affected by recall and social desirability biases, there are various methods for minimising these. ²² Another potential validation approach is to assess the level of agreement between IDP exposure periods and those generated from dispensing records, which document duration of medicine supply. ²³ It is important to note, however, that this approach is unlikely to capture dispensing patterns for individuals whose real‐world medicine use deviates from that prescribed.

There are several methodological considerations to take into account when implementing the IDP method. First, consistent with all methods for estimating exposure and dose in pharmacoepidemiological research, the assumption that, within each interval, medicines are used regularly at a consistent dose between dispensings does not account for intentional titration or tapering. Further, any variation in exposure resulting from non‐adherence would only be known if dispensing data were supplemented with information on the individual's actual medicine‐taking behaviour. Second, the IDP method assumes that supply of a medicine is exhausted at each dispensing; while this may not always be the case, the weighting approach used in evaluating the duration estimates minimises the impact of some irregularity of dispensings. Third, by leveraging individual‐level dispensing pattern information, the IDP method relies on the availability of longitudinal data with multiple data points for an individual to generate individualised exposure patterns. Generated exposure periods, however, may not represent how medicines are used among people with one (or few) dispensings. Finally, with respect to prn medicines, the approach used capitalises on the frequency of dispensings to estimate medicine exposure periods. However, the nature in which prn medicines are used means that dose estimates may be unavoidably over‐inflated or under‐estimated in different intervals during an exposure period as they may not be used consistently over this period.

Despite several studies comparing medicine exposure methods, ¹² , ²⁴ , ²⁵ , ²⁶ , ²⁷ there are clear advantages and disadvantages for each approach and there is unlikely to be a single approach that is suited to all research scenarios. Considering various time‐varying exposure and dose metrics in a single study is likely to provide greater insights of potential associations compared to a priori selection of a single metric, especially when there is uncertainty about the potential mechanisms linking the exposure with the outcome. This highlights the importance of the complete and transparent manner of reporting on operational details of exposure ascertainment in compliance with relevant reporting guidelines ²⁸ , ²⁹ as different choices may lead to different effects being estimated. Understanding how exposure was defined will aid correct interpretation of results and facilitate the comparison of findings from studies with similar research questions in different settings. In conclusion, the IDP method represents a new approach to evaluating exposure periods and dose from pharmaceutical claims data for studies in which exposure patterns are expected to vary, not only between individuals, but also within an individual over time. This method overcomes analytical issues in defining time‐varying medicines exposure when data on duration is unavailable, and importantly, relies only on prior individual dispensing patterns and the quantity dispensed (i.e., avoiding time‐related biases using future dispensing information). It offers a pragmatic approach that can be applied to all medicine classes and is well suited for analyses of medicines with complex and variable dosing regimens, and when interest is in capturing time‐varying periods of exposure, such as in exposure‐outcome studies.

AUTHOR CONTRIBUTIONS

Chrianna Bharat and Natasa Gisev drafted the manuscript. Chrianna Bharat, Natasa Gisev, Louisa Degenhardt, Sallie‐Anne Pearson, Andrew Wilson, and Timothy Dobbins developed the methodology. Chrianna Bharat and Luke Buizen developed the code. Luke Buizen pre‐processed the data. Chrianna Bharat analysed the data. All authors made revisions and approved the final manuscript. Chrianna Bharat had full access to all the data in the study and all authors shared final responsibility for the decision to submit for publication.

FUNDING INFORMATION

This work is supported by a National Health and Medical Research Council Health (NHMRC) project grant (#1138442). Chrianna Bharat is supported by National Drug and Alcohol Research Centre and UNSW Scientia PhD Scholarships. Louisa Degenhardt is supported by an NHMRC research fellowship (#1135991). The National Drug and Alcohol Research Centre is supported by funding from the Australian Government Department of Health under the Drug and Alcohol Program.

CONFLICT OF INTEREST

The authors declare no direct competing interests relevant to this study protocol. Louisa Degenhardt has received untied educational grant funding from Indivior, Mundipharma, Seqirus and Reckitt Benckiser. Sallie‐Anne Pearson is an employee of the Centre for Big Data Research in Health, UNSW Sydney, which has received funding from AbbVie Australia to conduct research unrelated to this work. Sallie‐Anne Pearson is a member of the Drug Utilisation Sub‐Committee (DUSC) of the PBAC. Andrew Wilson is paid by the Australian Commonwealth government as the chair of the Pharmaceutical Benefits Advisory Committee (PBAC).

RESOURCES

To facilitate adoption, SAS code implementing this method is provided in a dedicated public GitHub repository (https://github.com/c-bharat/IDP_exposure_method.git).

ETHICS STATEMENT

Approval for this study was obtained from the Australian Institute of Health and Welfare (AIHW) Ethics Committee (EO2016/4/314), NSW Population and Health Services Research Committee (2017/HRE0208), the ACT Health Human Research Ethics Committee (ETHLR.18.094) and the ACT Calvary Public Hospital Bruce Ethics Committee (5–2019).

APPENDIX A.

A.1. METHODS FOR THE COMPARATIVE EVALUATION OF THE INDIVIDUALISED DISPENSING PATTERNS METHOD WITH FIXED‐DOSE

A.1.1. Setting

The POPPY II Study is a population‐based cohort of adult (≥18 years) NSW residents who initiated a new opioid analgesic dispensing episode (defined as a opioid analgesic dispensing with no opioid analgesic dispensing in the preceding 365 days). ¹⁹ Data on prescription medicine dispensings were obtained from the Pharmaceutical Benefits Scheme (PBS) database and linked with data on mortality from the National Death Index (NDI).

A.1.2. Cohort definition

Separate cohorts and analyses were conducted for each of the three medicines—fentanyl patches, oxycodone immediate‐release, and oxycodone controlled‐release (see Table A1 for included PBS item codes).

For each medicine, the cohort was defined as all individuals who were dispensed the specific medicine between 1st July 2013 and 1st July 2017 and who had no prior dispensing for that medicine in the preceding 365 days. This allowed a minimum of 12‐months of follow‐up to evaluate exposure. Cohort follow‐up included all subsequent dispensings during the study period.

Observation commenced on the date of the index dispensing for the specific medicine and ended on the study end date (1st July 2018) or date of death, whichever was earlier. For each individual, observed person‐time was classified as exposed or unexposed to the specific medicine according to the fixed‐dose method and IDP method.

A.1.3. Fixed‐dose method

For each dispensing, the estimate of exposure duration for each medicine was calculated as the quantity dispensed multiplied by the recommended dose per day when used for its main indication (see Table A1 for daily doses used).

Overlapping exposure durations were assumed to indicate continuing exposure to the medicine. No lag or grace periods were used; however, a full carry‐over effect was employed. That is, where there was a new dispensing before the supply from the previous dispensing(s) were estimated to have all been used, the exposure interval for the new dispensing was extended according to the amount of remaining medicine supply. The end date of an episode was defined as the date following issue of the last dispensing for the specific medicine at which individuals had no more medicines available.

TABLE A1.

Recommended daily dose guidelines used in comparative evaluation of fixed‐dose method.

Medicine	PBS item codes	Recommended daily dose guidelines a , b
Fentanyl patch	05265D 05277R 05278T 05279W 05280X 05437E 05438F 05439G 05440H 05441J 08337T 08338W 08339X 08340Y 08878G 08891Y 08892B 08893C 08894D	1/3 patch per day (i.e., 3 days per patch)
Oxycodone controlled‐release	05227D 05247E 05248F 05249G 08385H 08386J 08387K 08388L 08681X 09399Q 09400R 05250H 05016B	2 tablets per day
Oxycodone immediate‐release	02622B 05191F 05195K 05197M 08464L 08501K 08502L 05198N	4 tablets per day (note this is within guideline recommendations to take 4–6 h)

Abbreviations: PBS, pharmaceutical benefits scheme.

^a Australian Medicines Handbook. Australian Medicines Handbook Pty Ltd: Adelaide; 2022.

^b MIMS Australia. MIMS Online; 2022.

Bharat C, Degenhardt L, Pearson S‐A, et al. A data‐informed approach using individualised dispensing patterns to estimate medicine exposure periods and dose from pharmaceutical claims data. Pharmacoepidemiol Drug Saf. 2023;32(3):352‐365. doi: 10.1002/pds.5567

DATA AVAILABILITY STATEMENT

This study used deidentified person‐level data from the POPPY II study. The data are not publicly available due to ethical restrictions.