Identification of individuals benefiting from the kakaritsuke-yakuzaishi (family pharmacist) system in Japan: a retrospective cohort study using an employment-based health insurance claims database

We sourced data from the JMDC database (JMDC Inc., Tokyo, Japan), which contains basic information (e.g., month and year of birth, sex, and enrollment period in the database) and administrative claims data collected from 11,065,484 individuals under 75 years old (as of January 2021). Claims data includes diagnoses based on International Classification of Disease (10th revision) codes and prescribed drugs based on the WHO’s Anatomical Therapeutic Chemical classification system codes. All data are anonymized and linked by unique identifiers. The details of this database are presented in previous reports [9, 10].

We obtained data from 1.7 million individuals randomly sampled from the database. To account for the biennial health care plan reviews, we set the observational period between April 2018 and March 2020, and the pre-observational period between April 2017 and March 2018.

To identify the eligible cohort, first, we selected individuals who used the same community pharmacy at least once every six months during the pre-observational period from the database. Then, we divided individuals into users and non-users of the family pharmacist system based on the family pharmacist consultation fee claims during the pre-observational period. Lastly, we excluded individuals who met the following exclusion criteria from the eligible cohort: (1) being under 18 years old at the start of the observational period; (2) not receiving any pharmaceutical management during the observational period; (3) switching groups (user/non-user) during the observational period; (4) being unable to claim the fee related to the endpoint; (5) receiving pharmaceutical management at home or at a nursing home; and (6) being unable to be observed until March 31, 2020.

We used “choufukutouyaku-sougosayoutou-boushi-kasan” (henceforth, additional TDDI prevention fee) claims as indicators of potential benefit from pharmaceutical management. This additional TDDI prevention fee is divided into two types of claims, one related to leftover drugs adjustment and the other to TDDI prevention. We evaluated each claim separately to identify the modifiers of benefits resulting from the family pharmacist system.

The variables used in the analyses and their definitions are shown in Table 1. We analyzed these variables as potential benefit modifiers, except for “number of medical examinations” and “use of a pharmacy that can claim the family pharmacist consultation fees,” which we used as confounding factors instead to deal with potential heterogeneity in the frequency and quality of pharmaceutical services. We categorized age, types of drugs, and the number of medical institutions used based on a previous study [7], and collected data (excluding data on age and sex) six months prior to the observational period.

We summarized continuous variables and categorical variables as means with standard deviations and frequencies (%), respectively, and calculated the proportion of the additional TDDI prevention fee claim.

We applied multiple generalized linear regression model (link: identity; distribution: binomial) including the variables listed in Table 1 to estimate the adjusted absolute risk differences (aARDs) and 95% confidence intervals (CIs) for the association between the use of the family pharmacist system and the additional TDDI prevention fee claim [11]. To identify the modifiers of benefit, we developed models that include the use of the family pharmacist system, one candidate variable, a product term between the use of the system and one candidate variable, and other variables listed in Table 1 [12]. If the product terms’ 95% CI lower limits did not reach 0, we extracted the variables as the modifiers. We implemented sensitivity analyses that included individuals who switched groups (user/non-user) during the observational period in the eligible cohort to account for any differences in the probability that they claimed the additional TDDI prevention fee.

Based on the identified modifiers, we developed two scoring systems to identify subpopulations for which the benefit of using the system was enhanced, using the method shown by Zhao et al. [13, 14]. First, we divided the eligible cohort randomly into the training dataset (75%) and the test dataset (25%). Second, we conducted the following two modeling procedures for the training dataset: Modeling Procedure I, wherein we applied two multiple logistic regression models separately to each group (users/non-users) and Modeling Procedure II, wherein we applied a single multiple logistic regression model. Then, we created two separate models, collectively named Model I, composed of the identified modifiers and other variables listed in Table 1, and created a single model, referred to as Model II, comprising the use of the system, identified modifiers, product terms between the use of the system and the identified modifiers, and other variables listed in Table 1. We then applied Modeling Procedures I and II with stepwise selection, wherein we selected and eliminated terms that reached the 0.20 level of significance. The results are presented in Additional file 1. Third, we estimated the scores using the differences in Model I’s coefficients and using Model II’s coefficients of the product terms. Fourth, we calculated the distributions of the scores for the training dataset and the aARDs in each subpopulation having equal or higher q-th percentiles. The higher the aARDs were, the greater was the potential benefit for the analyzed subpopulation. We plotted graphs with q in the x-axis and corresponding aARDs in the y-axis, and calculated the area under the aARD curves (AUCs). Fifth, we applied the methodology described in the fourth step to the test dataset using the scoring systems developed in the second and third steps. Lastly, we compared the plots and aARDs obtained from the training dataset and test dataset to evaluate overoptimism. We determined the best scoring system based on higher AUC from the training dataset.

We conducted all statistical analyses using SAS software, version 9.4 for Windows (SAS institute Inc., Cary, NC, USA).

Among the 1.7 million individuals in the JMDC database, 162,340 met this study’s eligibility criteria (users: 1,214; non-users: 161,126) (Fig. 1). Table 2 shows their characteristics. We observed claims related to the adjustment of leftover drugs in 6.9% (84) of the user group and 2.7% (4,318) of the non-user group (aARD: 1.9; 95% CI: 0.6 to 3.2). We observed TDDI prevention claims in 8.9% (108) of the user group and 3.2% (5,168) of the non-user group (aARD, 1.9; 95% CI, 0.3 to 3.4).

Regarding the leftover drugs adjustment claims, we identified the prescription of antidepressants as a modifier (Table 3). The sensitivity analysis indicates similar results (see eTable 1 in Additional file 1).

Regarding TDDI prevention claims, we identified being female, being prescribed ≥ 6 types of drugs, using ≥ 2 medical institutions, and being prescribed proton pump inhibitors, antibiotics, probiotics, or traditional Japanese herbal medicine as modifiers (Table 4). The sensitivity analysis indicates similar results (see eTable 2 in Additional file 1).

We divided the cohort into 121,755 individuals in the training dataset (users: 901; non-users: 120,854) and 40,585 individuals in the test dataset (users: 313; non-users: 40,272).

We did not develop a scoring system for leftover drugs adjustment claims, as, except for antidepressant prescription, there were no variables to be considered. Conversely, we developed two scoring systems for TDDI prevention claims from the training dataset (Table 5). The coefficients from the multiple logistic regression model are shown in eTable 3 (see Additional file 1). Table 6 indicates the estimated aARDs in subpopulations having an equal or higher q-th percentile scores from both the training and test datasets. Figure 2 depicts the correspondences between q and aARDs. Scoring system II indicated higher AUC (512.5).