Impact of parameter settings on the estimation of drug exposure with the individualized dispensing patterns method

Abstract

Aim

To assess the impact of parameter settings when using the individualized dispensing patterns (IDP) method to estimate exposure duration from dispensing data.

Methods

We conducted a cohort study using a 2% representative sample of the French healthcare insurance database from 2013 to 2022. The IDP method derives drug exposure distribution in a population sample, applies a selected percentile of this distribution to each individual's initial dispensing and then adjusts subsequent exposures at the individual level, incorporating a possible grace period to link consecutive dispensings into continuous exposure episodes. We studied two drugs with contrasting prescribing patterns: tramadol and rivaroxaban. We varied four parameters independently, compared median episode duration estimates and assessed consistency of exposed days estimates using kappa coefficients.

Results

With the original IDP parameters, median continuous exposure episodes were 47 days for tramadol and 79 days for rivaroxaban. The impact of parameter choice was drug‐specific: Percentile selection affected both drugs (kappa 0.58–0.65 for tramadol; 0.89–0.93 for rivaroxaban), sample population definition affected tramadol only (kappa 0.76), and grace period affected rivaroxaban (kappa 0.74–0.87). Heterogeneity across individual‐level kappa values was always high (I ² > 80%). These results informed the selection of drug‐specific parameters, yielding clinically more plausible median episode durations: 15 days for tramadol and 96 days for rivaroxaban.

Conclusion

Consumption patterns in the study population should guide percentile and grace period selection. Tailored accordingly, the IDP method provides an adaptable framework for estimating drug exposure from databases. Transparent reporting of parameter selection is essential to ensure reproducibility and interpretability.

What is already known about this subject

• The individualized dispensing patterns method estimates drug exposure from reimbursement data and can be applied across diverse drug use patterns.

• This method requires parameter settings, but their influence on exposure estimates has not been quantified.

What this study adds

• Some parameter settings influence drug exposure estimates and should be guided by expected use and confirmed in the study population.

• When tailored accordingly, the individualized dispensing patterns method provides an adaptable framework for estimating drug exposure from databases.

• Transparent reporting of parameter selection is essential to ensure reproducibility.

1. INTRODUCTION

Pharmaco‐epidemiological studies using real‐world data are essential for evaluating drug safety and identifying potential adverse effects in unselected populations.

Although healthcare database structures vary across countries, many, such as the French national health data system, ¹ include detailed data on drug dispensing but lack clinical information including treatment indication, prescribed dose or planned treatment duration. As a result, daily dose and treatment duration must be inferred from dispensing data. ²

Accurately estimating drug exposure is particularly important when studying dose–response relationships or drug–drug interactions, the latter also requiring precise definition of concomitant drug use.

Several methods have been developed to estimate drug exposure from dispensing data. These include covariate‐based methods, relying on population‐level assumptions or predefined rules such as defined daily dose ³ , ⁴ or medication coverage, ⁵ and cumulative sequence‐based methods, which reconstruct individual exposure over time from dispensing sequences, as in PRE2DUP ⁶ or reverse waiting time distribution. ⁷

However, studies have identified several limitations, including the influence of the decision rules required by these methods on exposure estimates, ⁸ disregard for inter‐individual dose variations ⁹ and reliance on assumptions that may not adequately reflect patient consumption behaviour. ¹⁰

The individualized dispensing patterns (IDP) method was developed to capture individual‐level patterns of drug use and dose changes over time for a wide range of drug types. ¹¹

However, the application of the IDP method requires setting several methodological parameters. When creating the IDP method and in previous studies, authors selected specific parameter settings and occasionally varied some choices, but comprehensive sensitivity analyses to evaluate the variation associated with each choice were not performed. ¹² , ¹³ , ¹⁴ , ¹⁵

Therefore, the objective of this study was to assess how parameter choices influence the estimation of drug exposure duration using the IDP method, based on data from the French health data system. Specifically, we evaluated the impact of the definition of the reference population used to assess exposure distribution; the percentile threshold selected from this distribution; the weighting applied to previous dispensings when estimating episode durations; and the duration of the grace period allowed to link consecutive dispensings into continuous exposure episodes.

2. METHODS

The reporting of this study follows the RECORD‐PE guidelines, and the checklist is provided in Table S1. ¹⁶

2.1. Study design and database

This population‐based cohort study used the Échantillon du Système National des Données de Santé (ESND), a 2% pseudo‐anonymized sample of the French healthcare insurance database Système national des données de santé (SNDS) containing demographical and medical information. ¹⁷ , ¹⁸ It contains information on prescription drugs dispensed in retail pharmacies: dispensing date, pharmaceutical form, dosage and quantity dispensed. Indication, prescribed daily dose and treatment duration are not available in the database. The study period lasted from 1 January 2013 to 31 December 2022.

2.2. IDP method

The IDP method has been developed to estimate the daily dose and duration of exposure to a drug using pharmacy dispensing data. ¹¹

The IDP method consists of four steps:

1. Among individuals meeting a baseline criterion, a per‐unit (i.e., one pill and one tablet) exposure duration estimate is calculated by dividing the number of days between consecutive dispensings by the quantity dispensed at the earlier dispensing. From the resulting distribution, a percentile is chosen to represent the population‐level medication use.

2. For each individual in the cohort, exposure duration from the first dispensing is estimated by multiplying the dispensed quantity by the population‐level per‐unit estimate chosen in Step 1. If the next dispensing occurs within this estimated period plus a predefined grace and lag period, it is considered part of the same continuous exposure episode (Figure 1).

3. For subsequent dispensings within the same exposure episode, the per‐unit exposure duration is updated using a weighted average of the last three estimates to account for evolving consumption patterns and refill behaviour.

4. If the time to the next dispensing exceeds the estimated exposure duration plus the grace and lag period, the current exposure episode is closed. The next dispensing starts a new exposure episode. This process is repeated until the last dispensing over the whole study period (Figure 1).

FIGURE 1.

Definition of continuous exposure episodes from dispensing data in the individualized dispensing patterns method.

This approach was designed for medications with irregular or patient‐driven dosing regimens, to reflect actual medication use patterns. ¹¹

As the objective of this study was to assess how estimates of drug exposure vary according to parameter choices, we focused on the grace period rather than the lag period. The lag period refers to the time window during which an outcome may still be attributed to exposure after the end of an exposure episode and is therefore not relevant to exposure estimation alone. In contrast, the grace period defines the allowable delay between consecutive dispensings for which exposure is considered continuous and directly affects the construction of exposure episodes.

In developing the IDP method, the authors chose specific settings for four parameters: the baseline criterion to estimate the population‐level time per unit (individuals with at least two dispensings within 365 days in the original IDP article), the percentile used from the population distribution (80th percentile), the weights given to the first, second and third previous dispensing when computing the moving average (respectively 3/6, 2/6 and 1/6) and the allowed grace period (7 days).

Two articles, including one co‐authored by the method's creator, applied the IDP method with different percentile choices (50th and 75th), without providing the rationale for these selections. ¹² , ¹³ In other articles, the IDP method was used, but parameter settings were not reported. ¹⁴ , ¹⁵

2.3. Modification of methodological parameters

To explore the potential impact of methodological choices, we modified IDP parameters as follows: The inclusion criteria for the analysis at the population level were set at two dispensings in 183 days, or five dispensings in 183 days with a maximum of 60 days between two dispensings. The percentile of time per unit was set at the 50th or 90th percentiles. The weight given to the three previous dispensings was set equally (1/3 each). Finally, the maximum grace period for two dispensings to be part of the same episode was extended to 14 or 28 days.

Daily exposure to the drug was first assessed in the population using the choices described in the index IDP article. ¹¹ Then, the parameters were modified one by one, and daily exposure was reassessed.

The IDP method is based on units of dispensing (e.g., one pill and one syrup dose). However, this is questionable for drugs with a wide range of dosage per unit and for which patients may use different dosages per unit of the same drug. To account for this variability, we applied the IDP method using milligrams dispensed rather than number of units.

2.4. Studied drugs

To assess whether the impact of methodological choices differs for drugs with different prescription patterns, two drugs were studied: tramadol and rivaroxaban.

Tramadol is an opioid analgesic drug that can be used for short‐term and long‐term treatments, with daily doses depending on pain intensity. It is available in France as immediate‐ and prolonged‐release formulations, with single‐pill dosages ranging from 50 to 300 mg.

Rivaroxaban is an oral anticoagulant prescribed for the treatment of deep vein thrombosis or pulmonary embolism, as well as for arterial embolism and thrombosis prevention. Single‐pill dosages range from 2.5 to 20 mg. The prescribed dosage depends on the indication but is generally fixed for an individual.

2.5. Population

A cohort was created for each studied drug. Patients were included if they had at least one dispensing of the drug during the study period. Follow‐up began at the first dispensing of the drug and continued until death, 365 days after the last dispensing or study end.

Given the study objectives, censoring 365 days after last dispensing was decided to avoid artificially increasing the apparent consistency of daily exposure in patients without subsequent dispensings.

2.6. Analysis of agreement

Drug exposure was assessed for each patient from cohort entry to follow‐up censoring using the IDP index method and subsequently for each single parameter variation. Based on dispensing data, exposure episodes were defined by linking consecutive dispensings into a continuous episode when the time between dispensings did not exceed the predefined grace period, resulting in exposed and unexposed periods during follow‐up.

The number of exposure episodes, their median duration and the proportion of follow‐up time exposed were calculated in the study population for each parameter choice and were compared with those obtained using the IDP index method.

To assess consistency of exposure estimates between methods, we computed a Cohen's kappa coefficient for each patient, comparing daily exposure classifications (on drug, off drug) between methods and pooled the values using a random‐effects meta‐analysis, treating each patient as a separate unit to account for inter‐patient variability. ¹⁹ This pooled kappa quantifies the average impact of parameter modifications on patient exposure classification, with inter‐individual heterogeneity measured by the I ² statistic.

2.7. Sensitivity analysis

A sensitivity analysis was conducted to assess the impact of expressing exposure in units of dispensing rather than in milligrams for each drug.

2.8. Post hoc analysis

Three post hoc analyses were conducted. The first analysis explored the impact of varying individual parameters one at a time, using the drug‐tailored method as the reference. The second analysis examined the impact of including different numbers of previous dispensings when estimating the time allocated to a subsequent dispensing; we compared the use of two and four prior dispensings instead of three. The third analysis assessed agreement between the IDP and drug‐tailored methods: Daily exposure classification was compared using a meta‐analysis of kappa coefficients, and agreement in mean exposure episode duration at the patient level was evaluated with Bland–Altman plots. ²⁰ Because the differences in episode duration in days were proportional to the mean duration, the Bland–Altman analysis was performed on a logarithmic scale.

2.9. Nomenclature of targets and ligands

Key protein targets and ligands in this article are hyperlinked to corresponding entries in http://www.guidetopharmacology.org/ and are permanently archived in the Concise Guide to PHARMACOLOGY 2025/26. ²¹

Data access and management were conducted using SAS Enterprise Guide version 9.4, and analyses were conducted using RStudio Pro version 2022.12.0.

3. RESULTS

3.1. Characteristics of the cohort

Between 2013 and 2022, 309 103 patients in the ESND received at least one dispensing of tramadol and 32 558 of rivaroxaban.

Tramadol was used intermittently, with a median of 1.0 dispensing per year and only 6.8% of patients on continuous treatment for at least 6 months. In contrast, rivaroxaban had a median of 5.3 dispensings per year and 52.2% of patients on continuous treatment (Table 1).

TABLE 1.

Characteristics of patients in tramadol or rivaroxaban cohort in the ESND, 2013–2022.

	Tramadol (N = 309 103)	Rivaroxaban (N = 32 558)
Age, median [P25–P75], years	50 [35–65]	69 [58–79]
Male, number (%)	145 338 (47.0%)	16 678 (51.2%)
Follow‐up period, median [P25–P75], days	365 [365–932]	472 [365–1203]
Dispensings per follow‐up year, median [P25–P75]	1.0 [1.0–1.6]	5.3 [1.4–12.2]
Patients with ≥5 dispensings in 183 days and maximum 60 days between 2 dispensings, number (%)	21 113 (6.8%)	17 007 (52.2%)
Hospitalization period/follow‐up period (%)	0.77%	0.96%

Abbreviations: ESND, Échantillon du Système National des Données de Santé; P25, 25th percentile; P75, 75th percentile.

In both cohorts, patients spent only a negligible share of their follow‐up time in hospital (Table 1).

3.2. Impact of parameters modifications

For tramadol, median duration of exposure episodes varied widely across reference population definitions. Using the original IDP parameter choice (two dispensings in 365 days), the median episode duration was 47 days. When restricting the reference population to individuals with at least two dispensings in 183 days, the median duration decreased to 39 days. The average agreement was high but heterogeneous across patients (Figure 2A). Further narrowing the reference population to patients with five dispensings in 183 days with a maximum of 60 days between two dispensings resulted in a median duration of 29 days, with lower average agreement and greater heterogeneity (Figure 2B).

FIGURE 2.

Distributions of individual kappa coefficients for drug exposure duration concordance depending on the reference population choice in the ESND, 2013–2022.

For rivaroxaban, agreement was high across reference population definitions with a stable median episode duration of 79 days and minimal heterogeneity across patients (Figure 2C).

Percentile choice had a major impact for tramadol; median duration ranged from 19 to 92 days with low agreement. In contrast, variations in the grace period had minimal impact on agreement, as drug dispensings were generally widely separated.

For rivaroxaban, marked differences were observed across percentile choices; median durations ranged from 52 to 100 days, compared with 79 days using the original IDP method. However, agreement across considered percentiles remained high, likely due to fixed doses in each individual and converging estimates over continuous exposure episodes. In contrast, changes in the allowed grace period led to lower agreement, potentially due to irregular prescription refills in some patients (Table 2).

TABLE 2.

Impact of parameter settings on population‐level drug exposure estimation with the IDP in the ESND, 2013–2022.

	Numbers of exposure episode	Median episode duration [P25–P75]	Proportion of time exposed (%)	Pooled kappa coefficient compared to the IDP method [95% CI]	I 2 (%)
Tramadol (N = 309 103)
IDP method Parameters = reference population: 2 dispensings in 365 days, 80th percentile, weighting of previous dispensings: 3/6, 2/6 and 1/6, grace period: 7 days
	615 464	47 [24–78]	17.19
Different reference population
2 dispensings in 183 days	639 687	39 [20–69]	15.66	0.909 [0.908–0.909]	84.3
5 dispensings in 183 days maximum 60 days between 2 dispensings	673 568	29 [15–57]	13.58	0.763 [0.762–0.763]	94.2
Different percentile
50th	765 437	19 [10–37]	10.96	0.579 [0.578–0.579]	96.8
90th	563 253	92 [46–129]	24.46	0.651 [0.650–0.652]	98.9
Different weighting of previous dispensings
Equal for last three dispensings	606 968	47 [24–81]	17.42	0.876 [0.876–0.876]	73.2
Different grace period
14 days	569 589	47 [24–74]	17.32	0.972 [0.972–0.972]	95.2
28 days	518 627	47 [24–70]	17.73	0.957 [0.957–0.957]	98.1
Rivaroxaban (N = 32 558)
IDP method Parameters = reference population: 2 dispensings in 365 days, 80th percentile, weighting of previous dispensings: 3/6, 2/6 and 1/6, grace period: 7 days
	141 851	79 [41–155]	62.76
Different reference population
2 dispensings in 183 days	141 851	79 [41–155]	62.76	0.995 [0.995–0.995]	15.1
5 dispensings in 183 days maximum 60 days between 2 dispensings	141 851	79 [41–155]	62.76	0.995 [0.995–0.995]	15.1
Different percentile
50th	181 877	52 [24–110]	59.95	0.888 [0.887–0.889]	96.7
90th	123 576	100 [54–182]	64.27	0.928 [0.927–0.928]	95.4
Different weighting of previous dispensings
Equal for last three dispensings	133 282	84 [41–165]	63.36	0.941 [0.940–0.942]	99.0
Different grace period
14 days	85 010	106 [45–255]	64.66	0.871 [0.869–0.873]	99.6
28 days	46 691	126 [41–437]	67.08	0.737 [0.734–0.739]	99.7

Abbreviations: CI, confidence interval; ESND, Échantillon du Système National des Données de Santé; IDP, individualized dispensing patterns; P25, 25th percentile; P75, 75th percentile.

Except for the reference population analysis for rivaroxaban, the I ² statistic consistently exceeded 80%, indicating substantial inter‐individual heterogeneity in kappa values.

The number of previous dispensings used for the weighted mean applied to the next dispensing was also evaluated in a post hoc analysis, comparing two or four previous dispensings with the three used in the IDP method. The impact on episode duration was modest for both drugs: For tramadol, the median episode duration remained 47 days, whereas for rivaroxaban, it varied slightly from 74 to 83 days. Based on these results, we retained the weighted mean of three previous dispensings (Table S2).

3.3. Drug‐tailored analysis

Based on these findings, we defined parameter settings better aligned with the underlying consumption patterns of each drug. For both drugs, the reference population was defined as patients with at least five dispensings within 183 days, and a maximum of 2 months between two dispensings, as this criterion appeared most appropriate to define continuous exposure episodes and compute realistic times per unit from there. The decreasing weighting to the three previous dispensings (3/6, 2/6 and 1/6) was kept.

The selection of percentile and time grace parameters was informed by the results presented in Tables 1 and 2. Tramadol can be used with a wide range of daily doses. The 50th percentile was chosen to reflect predominant use patterns rather than extremes. Choosing higher percentiles would disproportionately capture atypical uses. Because a low percentile was chosen for the per unit exposure duration, the grace period was set at 28 days to avoid misclassification of continuous episodes with low daily doses of the drug.

In contrast, rivaroxaban is rarely used for less than 3 months and is often prescribed for long‐term or lifelong use. Therefore, the 90th percentile was selected to avoid underestimating the duration of initial dispensings, relying on subsequent dispensings to adjust individual estimates. Consistent with this percentile choice and the expected continuous exposure, a short grace period of 7 days was applied.

Compared with estimates with the standard IDP method, the median duration of tramadol exposure episodes decreased to 15 days, which is consistent with its mainly intermittent use (Table 3). The choice of percentile remains influential at the population level.

TABLE 3.

Impact of parameter settings on population‐level drug exposure using the drug‐tailored method as reference, in the ESND, 2013–2022.

	Numbers of exposure episode	Median episode duration [P25–P75]	Proportion of time exposed (%)	Pooled kappa's coefficient compared with the DT method [95% CI]	I 2 (%)
Tramadol (N = 309 103)
Drug‐tailored method Parameters = reference population: 5 dispensings in 183 days maximum 60 days between 2 dispensings, 50th percentile, weighting of previous dispensings: 3/6, 2/6 and 1/6, grace period: 28 days
	562 404	15 [8–29]	11.69
Different reference population
2 dispensings in 365 days	553 559	19 [10–37]	12.56	0.915 [0.915–0.915]	84.4
2 dispensings in 183 days	556 078	18 [9–35]	12.29	0.946 [0.946–0.946]	76.9
Different percentile
80th	537 380	29 [15–52]	14.54	0.724 [0.724–0.724]	96.0
90th	527 862	37 [19–58]	16.01	0.619 [0.618–0.620]	97.7
Different weighting of previous dispensings
Equal for last three dispensings	562 714	15 [8–29]	11.65	0.962 [0.962–0.962]	88.8
Different grace period
7 days	816 546	15 [8–29]	9.64	0.895 [0.894–0.896]	99.2
14 days	687 543	15 [8–29]	10.36	0.912 [0.912–0.913]	98.1
Rivaroxaban (N = 32 558)
Drug‐tailored method Parameters = reference population: 5 dispensings in 183 days maximum 60 days between 2 dispensings, 90th percentile, weighting of previous dispensings: 3/6, 2/6 and 1/6, grace period: 7 days
	126 840	96 [51–177]	63.93
Different reference population
2 dispensings in 365 days	123 576	100 [54–182]	64.27	0.983 [0.983–0.984]	79.6
2 dispensings in 183 days	124 025	100 [54–181]	64.22	0.985 [0.985–0.986]	78.5
Different percentile
50th	190 558	49 [23–103]	59.47	0.811 [0.809–0.812]	98.7
80th	141 851	79 [41–155]	62.76	0.949 [0.948–0.949]	92.6
Different weighting of previous dispensings
Equal for last three dispensings	117 675	103 [55–192]	64.58	0.933 [0.899–0.901]	99.2
Different grace period
14 days	81 556	117 [52–268]	65.30	0.875 [0.873–0.877]	99.7
28 days	45 983	133 [47–446]	67.47	0.749 [0.746–0.751]	99.7

Abbreviations: CI, confidence interval; ESND, Échantillon du Système National des Données de Santé; P25, 25th percentile; P75, 75th percentile.

No substantial changes were observed for rivaroxaban. Episodes were slightly longer than before, aligning more closely with the standard treatment durations recommended for venous thromboembolic diseases. Among all parameters, the choice of percentiles and grace periods remained the ones with the greatest impact.

For both drugs and all parameter variations, heterogeneity between patients remained high (I ² > 75%).

As reflected in the median episode duration, the distribution of kappa coefficients and the Bland–Altman plots indicate greater variability in exposure estimates across parameter choices for tramadol than for rivaroxaban (pooled kappa 0.52; mean relative difference in episode duration = −60.1% when comparing the drug‐tailored method to the IDP). These results confirm the stronger influence of percentile selection and reveal a pattern for tramadol: Compared with the IDP, the drug‐tailored method tended to estimate shorter durations for short exposure episodes and longer durations for long episodes (Figure S1). This is consistent with the dominant impact of percentile selection on short episodes (one or a few consecutive dispensings) and the dominant impact of the grace period on longer episodes (many consecutive dispensings).

3.4. Sensitivity analysis

The index and drug‐tailored methods based on milligrams of drugs were compared with their equivalents based on dispensed units, as in the original IDP method. For tramadol, pooled kappas were 0.81 for the index method and 0.83 for the drug‐tailored method. For rivaroxaban, pooled kappas were 0.91 and 0.92, respectively. Heterogeneity remained high (I ² > 90%) (Table S3).

4. DISCUSSION

4.1. Key results

We conducted a methodological study on two drugs with distinct consumption patterns, tramadol and rivaroxaban, to assess how parameter settings affect real‐world exposure estimates in the IDP method. The choice of the reference population to compute the population‐level estimate had a substantial impact for tramadol, whereas the choice of the time grace to ascertain continuous exposure was more influential for rivaroxaban. Percentile selection affected estimates for both drugs. Compared with the original parameter settings, drug‐tailored parameters resulted in exposure durations more consistent with expected prescribing.

4.2. Strengths and weaknesses

The main strength of this study is the systematic approach used to evaluate the impact of each parameter choice, aimed at ensuring reproducibility and transparency in future research. Another strength lies in its use of empirical data derived from routinely collected real‐world data, capturing actual patterns of drug use. By examining tramadol (short‐term, variable dosing) and rivaroxaban (long‐term, stable dosing), the study captures opposite ends of drug use patterns. We used meta‐analyses of kappa coefficients to account for the full range of consumption profiles and to quantify inter‐individual heterogeneity.

A first limitation of our work is that we did not assess the impact of parameter choices on the estimated association between exposure and outcomes. Because the effect of exposure definitions on measures such as relative risks depends on the specific outcome and study context, evaluating this within a single empirical example would have limited generalizability. Therefore, we focused on how parameter choices influence exposure classification itself, using metrics derived directly from dispensing data (e.g., treatment episode duration), in order to provide a framework applicable across different pharmacoepidemiologic studies. Another limitation inherent to most pharmacoepidemiologic research is the lack of a gold standard for measuring true drug exposure at the individual level, which prevents directly determining the most accurate method based on dispensing data. More specifically, like all methods developed to estimate drug exposure from dispensing data, the IDP method assumes that all dispensed medications are taken between two dispensing dates, with no leftover, and that drug use is regular and with a stable daily dose between two consecutive dispensings within the same episode. In this study, parameter selection was therefore guided by comparisons between estimated episode durations and what would be expected in clinical practice, according to typical patterns of drug use.

Although two drugs with different patterns were studied, our study does not capture the full spectrum of real‐world drug use. Researchers applying the IDP method should assess the specific usage patterns of their study drugs to guide parameter selection.

Another limitation is that agreement between parameter settings was assessed based on the number of days on or off the drug only. Further approaches could have been chosen, for example, to assess the agreement between estimated daily doses. The meta‐analysis of kappa coefficients, with random effects at the patient level, appeared as the most appropriate approach to estimate agreement between methods. However, it treats each day as independent and ignores the temporal correlation of consecutive days within patients. As a result, the kappa coefficient overestimates agreement. Finally, this study did not account for either inpatient drug use, as hospital dispensings are not available in the ESND. When hospital stays exceed the duration covered by the dispensing plus the grace period, the IDP method may misclassify post‐discharge dispensing as a new episode. Given this study's objectives and the minimal follow‐up time spent hospitalized (Table 1), hospitalization periods were not specifically considered, although they may need to be accounted for in other settings.

4.3. Interpretation

Compared with other methods used for estimating drug exposure from dispensing data, the IDP method offers greater flexibility. Fixed‐dose approaches, as defined daily dose or one tablet a day, apply uniform assumptions that often do not reflect actual drug use. ³ , ²² , ²³ , ²⁴ , ²⁵ In contrast, the IDP method reconstructs exposure episodes from patient‐specific dispensing intervals and quantities, enabling individual‐level adjustments. The IDP method also improves upon the reverse waiting time distribution method, which incorporates patient covariates but typically assumes them to be time‐invariant within the study period, limiting the ability to capture changes in patient behaviour over time, ⁷ , ²⁶ and on the estimation of drug coverage (COV) method, which introduces variability but relies on predefined thresholds. ⁵ The IDP method provides more data‐driven and patient‐centred estimates of drug exposure. Finally, although algorithmic methods such as PRE2DUP enable automated reconstruction of exposure, they fully rely on expert‐defined rules for each drug, which limit scalability when analysing multiple medications. ⁶ The IDP method can rely on real‐life usage patterns observed in the study population.

4.4. Perspectives

Our study proposes a reproducible framework for estimating drug exposure based on a two‐step approach. First, expected drug use patterns (intermittent vs. continuous) are confirmed within the study population. Second, these patterns inform the selection of parameters, validated by comparing the ensuing estimates against expected exposure durations.

Restricting the reference population to individuals with continuous exposure appears relevant for all drugs, including those like tramadol, which display substantial variability in both episode duration and daily dose. For these drugs, selecting an intermediate percentile more accurately reflects the consumption of a majority of users. Because population‐level assumptions are applied to each first dispensing, selecting a higher percentile could lead to overestimation of episode duration. This consideration is particularly important for discontinuous treatments, where individual estimates cannot be adjusted based on subsequent dispensings. As the first two tramadol parameter choices were conservative, a more permissive grace period was used to allow greater flexibility between dispensings.

Because the choice of the reference population has greater impact for discontinuous treatments, applying a single population‐level assumption may be limiting. Stratifying the population by relevant covariates (such as age, gender or comorbidity) could refine assumptions for the first dispensings, allowing more accurate estimates of episode durations within patient subgroups, accounting for systematic differences in dispensing patterns.

Such an approach would integrate covariate‐based subgroup information with the sequence‐based method and could improve exposure estimation across the full spectrum of treatment use, from single dispensings to continuous use.

Regarding the choice between milligram or unit, we observed no major difference overall, although the high heterogeneity indicates that exposure classification diverges for many patients. The decision should thus be guided by the number of available dosages for a given drug and their expected or observed use across and within individuals.

Consequently, a tailored approach should be applied to each drug to ensure that parameter settings are grounded in expected and real‐world use, enhancing the accuracy of exposure assessment. Sensitivity analyses should be performed on key parameters, such as percentiles and grace periods, to ensure robust and reliable results. ¹⁰ , ²⁷ Clear reporting of methodological choices should be standard practice when using the IDP method to ensure transparency and enable critical appraisal of the study validity and generalisability. ¹⁶

5. CONCLUSION

When using the IDP method, parameter settings, especially percentiles and grace periods, are influential and should be guided by expected use and confirmed in the study population. This tailored approach enables accurate estimation of drug exposure across diverse drug use patterns and promotes transparent reporting of parameter choices to ensure reproducibility.

AUTHOR CONTRIBUTIONS

Conception and design of the study: L.G., C.L., O.S. Data analysis: L.G., C.L. Drafting manuscript: L.G., O.S. All authors participated in data interpretation, reviewed and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Supporting information

Table S1 The RECORD statement for pharmacoepidemiology (RECORD‐PE) checklist of items.

Table S2 Impact of parameter settings on population‐level drug exposure for modification in the number of previous dispensings included, 2013–2022.

Table S3 Impact of parameter settings on population‐level drug exposure between unit and milligram methods, in the ESND, 2013–2022.

Figure S1 Agreement of drug exposure estimates between IDP and drug‐tailored method. Subfigures (a) and (b) show individual kappa coefficients on days of exposure, and (c) and (d) present Bland–Altman plots of individual mean duration of episodes, in the ESND, 2013–2022.

Gosselin L, Lévy C, Sève N, et al. Impact of parameter settings on the estimation of drug exposure with the individualized dispensing patterns method. Br J Clin Pharmacol. 2026;92(11):3982‐3991. doi: 10.1002/bcp.70701

DATA AVAILABILITY STATEMENT

Access to the database was possible only through a secure connection to a specific server.