The All‐Payer Claims Database After the Gobeille Supreme Court Decision: Assessing the Impact on Fatal Opioid‐Related Overdose Research

ABSTRACT

Objective

To determine the consequences of the Gobeille v. Liberty Mutual Supreme Court decision on the representativeness of the Massachusetts all‐payer claims database (APCD). The loss of individuals captured in the APCD may vary demographically, geographically, and in the capture of fatal opioid‐related overdose (OOD).

Data Sources and Study Setting

We used 2013–2021 data from the Massachusetts Public Health Data Warehouse (PHD), which links person‐level APCD records to other datasets. The APCD includes commercially‐insured health claims mandatorily reported pre‐Gobeille. Post‐Gobeille, reporting from self‐insured plans, a subset of commercially‐insured plans, became voluntary.

Study Design

In a repeated cross‐sectional design, we compared the APCD population characteristics in 2015 to each subsequent year 2016–2021.

Data Collection/Extraction Methods

We compared pre‐post Gobeille statewide APCD demographic distributions using standardized mean differences and assessed geographic distribution changes by ZIP Code. We identified fatal OOD using death certificates. We used annual and monthly interrupted time‐series models with publicly‐available state records as the control to quantify the pre‐ (2013–2015) and post‐ (2016–2021) Gobeille changes in the total APCD population and in fatal OOD when linked to the APCD within the PHD.

Principal Findings

Within APCD, the commercially‐insured population decreased by 38% post‐Gobeille. State‐level age increased slightly and sex distributions remained stable, while proportions of White non‐Hispanic individuals decreased. Suburban ZIP Codes had the highest losses of individuals. In 2021, under 80% of fatal OODs could be linked to the APCD, compared to 95% linkage pre‐Gobeille. The change in monthly fatal OOD rates when linked to the APCD was 0.55 persons higher per 100,000 people post‐Gobeille (95% CI: 0.05, 1.05) than the change observed in official statistics.

Conclusions

The Gobeille decision negatively impacted APCD geographic and racial representativeness in Massachusetts, which should be addressed to improve external validity in Massachusetts and other states using APCDs to assess health services.

Callout Box

What is known on this topic

• ○The 2015 Gobeille decision resulted in substantial loss of individuals with self‐insured plans from all‐payer claims databases, which, as a result, no longer provide comprehensive coverage of state insured populations.
• ○After Gobeille, the Massachusetts state all‐payer claims database was found to have the same statewide sex distribution but had an older age distribution compared to those in years prior.

What this study found

• ○Post‐Gobeille, White non‐Hispanic and suburban individuals were disproportionately underrepresented in the Massachusetts all‐payer claims database.
• ○In 2021, we were unable to link over 20% of fatal opioid‐related overdoses to records in the all‐payer claims data, endangering generalizability.
• ○The Gobeille decision had little impact on the state‐level age and sex distribution of the all‐payer claims database over time and on the estimation of linked fatal opioid‐related overdose rates.

1. Introduction

The ongoing opioid‐related overdose epidemic in the US remains one of the most pressing public health challenges of our time, with more than 80,000 estimated deaths reported in 2023 [1]. In Massachusetts, fatal opioid‐related overdose rates are among the highest in the US, and there have been more than 2000 fatal opioid‐related overdoses per year in the state between 2016 and 2023 [2]. To understand the level of severity of the opioid‐related overdose epidemic, as well as factors contributing to opioid‐related overdose mortality, complex data systems, such as all‐payer claims databases (APCDs), have been used to identify factors contributing to risk, ultimately helping to inform health services policy, prevention, and treatment interventions.

APCDs are data systems maintained at the state level that capture all health, dental, and pharmacy insurance claims across public and private insurers in a state as well as insurance enrollment information [3]. Currently 29 states have an APCD in some form and 10 have expressed a strong interest in creating one [4]. The specific reporting requirements vary across states in terms of the types of payers required to submit data and voluntary (4 states) versus mandatory reporting (25 states) [4]. For example, the Massachusetts APCD requires mandatory reporting for all public and private payers except Medicare Fee‐for‐Service, which submits data only to the Center for Medicare and Medicaid Services [3]. Thus, the APCD historically captured health care service delivery in the state for all other insured individuals [5]. Due to their comprehensive and longitudinal capture of individual persons, APCDs have been leveraged by states and external researchers to understand a wide variety of health care issues including costs, access to health care services, disease burdens, and equity using race and ethnicity and locational information included in APCDs, with direct policy relevance for state public health agencies and clinical providers [3, 6, 7, 8]. In particular, APCDs have been used to study a wide array of aspects related to the opioid‐related overdose crisis, and in Massachusetts, the APCD has been linked with other data sources and used to estimate the prevalence of opioid use disorder, identify potentially inappropriate opioid prescribing patterns, evaluate post‐overdose outreach program effectiveness, and access to medications for opioid use disorder [9, 10, 11, 12, 13].

In 2015, the US Supreme Court issued a decision in Gobeille v. Liberty Mutual Insurance Co. that prohibited states from mandating reporting to APCDs for individuals on self‐insured plans [14]. Self‐insured health plans, a subset of commercially‐insured plans offered by beneficiaries' employers, are often used by large organizations that are able to assume all liability for medical debt for those covered [15, 16]. With the passage of the Employee Retirement Income Security Act (ERISA) in 1974, companies were allowed to forgo purchasing insurance from a third party and instead pay employee beneficiary claims directly using a designated pool of money. Such self‐insured plans were almost universally adopted by large corporations with over 5000 employees (≥ 90% adopted) and moderately adopted by mid‐size corporations of 500 to 4999 employees (50%–80% adopted) [17]. After the Gobeille decision, managers of self‐insured plans largely stopped reporting to state APCDs and for those who kept reporting, it may be due to infrastructure already being in place and inertia. For example, an analysis by the Center for Health Information and Analysis found that the Massachusetts APCD lost data for approximately 1.75 million individuals on self‐insured plans in 2016, the first year following the Gobeille decision, which represented 75% of the self‐insured population [5]. This was a substantial loss for a state comprised of nearly 7 million people, thereby significantly limiting the utility of the APCD as a virtual census of insurance claims, and thus people, in the state.

Since individuals covered by self‐insured plans may not be captured in an APCD, and their population characteristics may be different from the rest of the population, there is concern about the representativeness of APCDs for population‐based studies post‐Gobeille. Initial analyses of this impact in the first several years after the decision have shown little difference in the distribution of age and sex in the full Massachusetts population captured by the APCD; however, among private payers, the loss of the pediatric population post‐Gobeille was over 40% [5].

While these analyses point to continued representativeness of the Massachusetts APCD post‐Gobeille, it is unclear whether this generalizability is maintained for other characteristics such as race and Hispanic ethnicity, geography, and individuals experiencing health outcomes such as fatal opioid‐related overdoses. Given that APCD‐based linked data systems are widely used for state‐level health‐related analyses, it is critical for policy makers and researchers to understand the impact of the Gobeille decision when using APCDs to drive policy change and evaluate effectiveness. APCDs are invaluable public health analytic tools that can be used to address urgent public health crises, such as the ongoing opioid‐related overdose epidemic, but their limitations must be understood and taken into account. Since the Massachusetts APCD historically had served as the “spine” for linking data within the state's Public Health Data Warehouse (PHD), and the PHD also includes death certificate data to capture fatal opioid‐related overdoses, we have the opportunity to gain an understanding of the magnitude of the impact of Gobeille on public health service analyses relying on APCD‐based linked data systems.

The overarching goal of the current analyses was to determine the impact of the Gobeille decision on how best to use the APCD within the PHD to determine the extent to which statistical weighting approaches would be needed to ensure the generalizability of other opioid‐related overdose prediction modeling efforts. Specifically, we aimed to extend prior APCD analyses by comparing changes in demographics using time trends and by exploring any changes in the spatial distribution of individuals captured in the APCD. Further, we aimed to assess the impact of Gobeille on the analysis of fatal opioid‐related overdose within the PHD to ensure our results are generalizable when used to inform opioid‐related overdose prevention policy.

2. Methods

We used data from the Massachusetts Department of Public Health's Public Health Data Warehouse (PHD) for the years 2013 to 2021 [18]. At the time of the study, the PHD consisted of over 35 administrative datasets linked at the person level via a unique identifier derived from the APCD. Thus, the APCD functioned as the spine of the PHD. Individuals who appeared in the APCD pre‐Gobeille, even those on self‐insured health plans, could be linked to their records in administrative datasets post‐Gobeille because they had been assigned a random identification number in the spine; however, post‐Gobeille, self‐insured individuals new to the PHD have no records in the APCD, and therefore did not appear in the APCD spine, and had no associated random identification number within the PHD to be used to link to other administrative datasets. Therefore, data about these people from any source is not available within the PHD. The population in this study consisted of all Massachusetts residents who, in a given year, appeared at least once in the APCD monthly membership eligibility file, which catalogues all individuals with health insurance assessed on the 15th of each month. We created a monthly time‐series of the count of individuals captured in the APCD monthly membership eligibility file stratified by insurance type.

We obtained demographic information, including age, sex, race, Hispanic ethnicity, and residential ZIP Code, by linking individuals in the APCD monthly membership eligibility file to other datasets in the PHD with information on these characteristics, such as emergency medical services and hospital discharge data. We collapsed sex, race, and Hispanic ethnicity across years to be time‐invariant, while we allowed age and ZIP Code to change year‐to‐year. We calculated age by subtracting an individual's birth year from the given year of analysis. In each year, we used the first Massachusetts ZIP Code found for each individual across the PHD datasets as their residential ZIP Code for that year. The vast majority of individuals have only one recorded ZIP Code in a given year (Table S2). We converted ZIP Codes to US Census Bureau ZIP Code Tabulation Areas using a crosswalk provided by the Massachusetts Department of Public Health. We use ZIP Code Tabulation Areas because they are explicitly defined geographic units, while ZIP Codes only refer to mail delivery routes [19]. Individuals with no recorded birth year had missing age for all years, and those without a ZIP Code in a given year were designated as missing residential locational information for that year. Individuals with missing age or missing ZIP Code were dropped from their respective analyses.

We compared the characteristics of individuals and their spatial distribution in 2015 pre‐Gobeille to each year post‐Gobeille 2016 through 2021. We assessed differences in age, sex, race, and Hispanic ethnicity using standardized mean differences [20]. We used changes in the count of individuals in each ZIP Code Tabulation Area and percent change compared to the APCD population in 2015 to describe differences in the spatial distribution over time.

The Massachusetts Department of Public Health identified fatal opioid‐related overdoses in the PHD using Death Certificates data (death records). To understand the ability of the APCD to link to fatal opioid‐related overdoses identified in Death Certificates in the PHD in each month, we calculated the proportion of fatal opioid‐related overdose records that could be linked to the set of all unique identification numbers in the APCD monthly membership eligibility file in the given month, over the past 6 months, over the past 12 months, and over the entire history of the APCD.

The Massachusetts Department of Public Health publishes semi‐annual reports of rates of fatal opioid‐related overdoses per 100,000 persons in Massachusetts that do not rely on the PHD, and thus the APCD [2]. These official statistics represent the best available data on trends in fatal opioid‐related overdoses in Massachusetts and can serve as a benchmark against which to compare results obtained using linked data systems, such as the Massachusetts PHD.

Using the PHD, we calculated annual opioid‐related overdose rates per 100,000 persons by dividing the number of fatal opioid‐related overdoses linked to the APCD by the APCD's monthly membership eligibility file population. We modeled the impact of Gobeille on the APCD‐linked estimated rates within the PHD using controlled interrupted time‐series analysis [21]. Our control time‐series relied upon the official rates of fatal opioid‐related overdoses extracted from the Massachusetts Department of Public Health semi‐annual reports. We modeled both the annual rates from 2013 to 2021 as well as the monthly rates from June 2014 (the furthest back monthly data was available) to December 2021. The interrupted time‐series model quantifies the immediate impact of Gobeille on the rate of fatal opioid‐related overdoses, as well as any changes in the trend post‐Gobeille (Equation 1) [21].

$$E\left[Y_t\right]={\beta }_0+{\beta }_1\left(1-k\right)t+{\beta }_2\mathit{kj}+{\beta }_{3}k+{\beta }_{4}x+{\beta }_{5}x\left(1-k\right)t+{\beta }_6\mathit{xkj}+{\beta }_7\mathit{xk}$$ (1)

where $Y_t$ is the fatal opioid‐related overdose rate in year (or month) $t+2013$, where $t=0,1,\dots ,8$ corresponds to the years (or months) 2013 through 2021. The variable $j$ counts the number of years (or months) post‐Gobeille (number of years after 2015). The variable $k$ is an indicator for whether the year (or month) is pre‐Gobeille ($k=0$) or post‐Gobeille ($k=1$). The variable $x$ is an indicator for whether the rate of fatal opioid‐related overdose is calculated from APCD‐linked data within the PHD ($x=1$) or extracted from the official Massachusetts Department of Public Health semi‐annual reports ($x=0$). In this model, the coefficient ${\beta }_0$ is the 2013 fatal opioid‐related overdose rate in the yearly model or the rate in June 2014 in the monthly model provided in the Massachusetts Department of Public Health semi‐annual report. ${\beta }_1$ is the estimated annual trend in fatal opioid‐related overdose rates from 2013 to 2015. ${\beta }_2$ is the estimated annual trend 2016 to 2021. The coefficient ${\beta }_3$ estimates the difference in fatal opioid‐related overdose rates between 2015 and 2016 (i.e., pre‐post Gobeille), also in the Massachusetts Department of Public Health semi‐annual report. Coefficient ${\beta }_4$ is the 2013 fatal opioid‐related overdose rate within the PHD using linked APCD records. Coefficients ${\beta }_5$ and ${\beta }_6$ represent the difference in the annual or monthly trends from 2013 to 2015 and from 2016 to 2021, respectively, of fatal opioid‐related overdose rates as calculated within the PHD using linked APCD records and the Massachusetts Department of Public Health semi‐annual report. ${\beta }_7$ represents the immediate impact of Gobeille (i.e., comparing 2015 to 2016) by calculating the difference in fatal opioid‐related overdose rates estimated within the PHD using linked APCD records and the semi‐annual Massachusetts Department of Public Health report. Using the same model, we also quantified the impact of Gobeille on the total population captured by the APCD within the PHD by comparing to 5‐year American Community Survey estimates of the total Massachusetts population as the control [22]. In each time series, we checked for autocorrelation using an autoregressive integrated moving average [23]. We conducted a numerical search across a range of possible lags and selected the parameters that produced the lowest AICc. These autocorrelation parameters were then included in our controlled interrupted time series models.

The interrupted time‐series results allowed us to quantify the impact of Gobeille on the APCD and by extension, the PHD. We used the official statistics for fatal opioid‐related overdose rates based on ICD‐10 codes for causes of death to represent the counterfactual scenario regarding what would have been observed in the APCD‐linked data within the PHD in the absence of Gobeille and we assumed that no other changes to the APCD occurred simultaneously with Gobeille [19].

We queried data from the PHD using SAS software, Version 3.81 (Enterprise Edition) [24]. We used R Version 4.2.2 for descriptive statistics and visualizations, and the package “nlme” to fit the controlled interrupted time‐series models [25, 26].

3. Results

In the years prior to Gobeille, the APCD monthly membership eligibility file captured approximately 6.5 million unique individuals in a given year (in 2015, n = 6,709,816). In 2016, following the Gobeille decision, this number dropped to 5,377,460. Through our time‐series analyses, we observed the impact of Gobeille on the commercially insured population, which dropped by 35% in 2016 (in 2015, n = 4,068,155; in 2016, n = 2,661,653) (Table 1) (Figure 1).

TABLE 1.

Annual distribution of age, sex, and race/ethnicity and standardized mean differences in all‐payer claims data within the Public Health Data Warehouse population 2016–2021 compared to 2015, Massachusetts.

	2015 mean (sd) n (%)	2016 mean (sd) n (%)	2017 mean (sd) n (%)	2018 mean (sd) n (%)	2019 mean (sd) n (%)	2020 mean (sd) n (%)	2021 mean (sd) n (%)
Statewide APCD population
Population	6,709,816 (100)	5,377,460 (100)	5,390,819 (100)	5,391,845 (100)	5,284,388 (100)	5,172,940 (100)	5,280,317 (100)
Age	39.2 (23.0)	39.8 (23.5)	40.2 (23.6)	40.7 (23.8)	40.9 (24.1)	41.3 (24.2)	41.3 (24.2)
SMD	—	0.029	0.043	0.064	0.072	0.092	0.09
Male	3,145,678 (46.9)	2,505,641 (46.6)	2,505,745 (46.5)	2,501,380 (46.4)	2,444,209 (46.3)	2,395,968 (46.3)	2,448,375 (46.4)
Female	3,429,605 (51.1)	2,784,411 (51.8)	2,808,790 (52.1)	2,823,772 (52.4)	2,783,961 (52.7)	2,730,168 (52.8)	2,776,688 (52.6)
Missing sex	134,533 (2.0)	87,408 (1.6)	76,284 (1.4)	66,693 (1.2)	56,218 (1.1)	46,804 (0.9)	55,254 (1.1)
SMD	—	0.083	0.020	0.020	0.088	0.088	0.088
American Indian and other NH	463,504 (6.9)	351,293 (6.5)	360,199 (6.7)	371,122 (6.9)	381,744 (7.2)	388,407 (7.5)	411,372 (7.8)
Asian NH	275,713 (4.1)	217,250 (4.0)	224,335 (4.2)	230,702 (4.3)	234,894 (4.5)	236,510 (4.6)	252,789 (4.8)
Black NH	497,253 (7.4)	457,270 (8.5)	466,486 (8.7)	472,026 (8.8)	462,287 (8.8)	454,770 (8.8)	469,883 (8.9)
Hispanic	722,157 (10.8)	691,061 (12.9)	716,032 (13.3)	733,863 (13.6)	730,952 (13.8)	729,369 (14.1)	770,186 (14.6)
White NH	3,983,428 (59.4)	3,119,932 (58.0)	3,094,852 (57.4)	3,076,538 (57.1)	2,999,323 (56.8)	2,936,822 (56.8)	2,941,944 (55.7)
Missing race/ethnicity	767,760 (11.4)	540,654 (10.1)	528,915 (9.8)	507,594 (9.4)	475,188 (9.0)	427,062 (8.3)	434,143 (8.2)
SMD	—	0.131	0.113	0.147	0.147	0.208	0.222
APCD commercially‐insured subpopulation
Population	4,068,155 (100)	2,661,653 (100)	2,693,663 (100)	2,691,994 (100)	2,725,400 (100)	2,598,850 (100)	2,515,651 (100)
Age	36.9 (19.9)	36.4 (19.0)	36.4 (18.8)	37.1 (19.1)	37.1 (19.0)	37.3 (19.0)	37.5 (19.1)
SMD	—	−0.025	−0.023	0.012	0.012	0.024	0.032
Male	1,952,032 (48.0)	1,276,909 (48.0)	1,287,413 (47.8)	1,282,612 (47.7)	1,298,597 (47.7)	1,245,742 (47.9)	1,211,132 (48.1)
Female	2,036,558 (50.1)	1,345,070 (50.6)	1,368,272 (50.8)	1,376,985 (51.2)	1,398,153 (51.3)	1,329,075 (51.1)	1,278,040 (50.8)
Missing sex	79,565 (2.0)	39,674 (1.5)	37,978 (1.4)	32,397 (1.2)	28,650 (1.1)	24,033 (0.9)	26,479 (1.1)
SMD	—	0.083	0.083	0.083	0.083	0.083	0.083
American Indian and other NH	355,913 (8.8)	233,141 (8.8)	236,788 (8.8)	239,866 (8.9)	248,543 (9.1)	245,566 (9.5)	247,496 (9.8)
Asian NH	171,170 (4.2)	111,115 (4.2)	117,750 (4.4)	122,645 (4.6)	134,789 (5.0)	133,909 (5.2)	139,875 (5.6)
Black NH	195,700 (4.8)	147,462 (5.5)	155,582 (5.8)	160,110 (6.0)	170,393 (6.3)	163,059 (6.3)	152,236 (6.1)
Hispanic	212,330 (5.2)	165,421 (6.2)	177,576 (6.6)	181,107 (6.7)	198,169 (7.3)	191,950 (7.4)	180,567 (7.2)
White NH	2,570,615 (63.2)	1,663,130 (62.5)	1,654,169 (61.4)	1,650,002 (61.3)	1,648,919 (60.5)	1,576,064 (60.6)	1,525,012 (60.6)
Missing race/ethnicity	562,427 (13.8)	341,384 (12.8)	351,798 (13.1)	338,264 (12.6)	324,587 (11.9)	288,302 (11.1)	270,465 (10.8)
SMD	—	0.067	0.100	0.120	0.120	0.120	0.178

Note: Bold indicates standardized mean differences and do not have an associated 95% confidence interval.

Abbreviations: APCD: all‐payer claims database, NH: non‐Hispanic, sd: standard deviation, SMD: standardized mean difference.

FIGURE 1.

Count of unique individuals in the all‐payer claims data within the Public Health Data Warehouse stratified by insurance type, Massachusetts, 2013–2021.

Among the full statewide APCD population, as well as the APCD commercially‐insured subset, individuals captured became slightly older post‐Gobeille (full APCD 2015: mean = 39.2, 2021: mean = 41.3; commercially‐insured 2015: mean = 36.9, 2021: mean = 37.5); however, we found no differences in sex proportions (full APCD male 2015: 46.9%, 2021: 46.4%; commercially‐insured male 2015: 48.0%, 2021: 48.1%) (Table 1). Both the full APCD population and the commercially‐insured subpopulation showed substantial changes in the race and Hispanic ethnicity distributions post‐Gobeille, primarily due to a decrease in the proportion of White non‐Hispanic individuals (full APCD 2015: 59.4%, 2021: 55.7%; commercially‐insured 2015: 63.2%, 2021: 60.6%) and increases in the proportions of Black non‐Hispanic individuals (full APCD 2015: 7.4%, 2021: 8.9%; commercially‐insured 2015: 4.8%, 2021: 6.1%), American Indian or Other non‐Hispanic individuals (full APCD 2015: 6.9%, 2021: 7.8%; commercially‐insured 2015: 8.8%, 2021: 9.8%), and Hispanic individuals of any race (full APCD 2015: 10.8%, 2021: 14.6%; commercially‐insured 2015: 5.2%, 2021: 7.2%) (Table 1). The proportion of Asian non‐Hispanic individuals increased in the commercially‐insured subpopulation (2015: 4.2%; 2021: 5.6%), but not the full APCD population (2015: 4.1%; 2021: 4.8%) (Table 1). Approximately 3% of individuals in the APCD were missing ZIP Code Tabulation Area information and between 2% and 5% were missing age information in a given year (Table S3). The vast majority of ZIP Code Tabulation Areas lost at least 25% of their commercially‐insured residents post‐Gobeille, with the highest losses of individuals occurring in suburban Massachusetts communities (Figure 2).

FIGURE 2.

ZIP Code tabulation area changes in the count (A) and percent (B) of the total all‐payer claims data population within the Public Health Data Warehouse and the count (C) and percent (D) of the commercially‐insured subpopulation between 2015 and 2021, Massachusetts.

Post‐Gobeille, the number of fatal opioid‐related overdoses captured in the PHD that were able to link to identification numbers was highest when using the longest look‐back window and decreased with shorter look‐back windows (Figure 3). For the 1‐month, 6‐month, and 12‐month look‐back windows, fatal opioid‐related overdose linkage to the APCD decreased steadily over time post‐Gobeille (Figure 3). However, linkage remained high in the post‐Gobeille period when using all available historical APCD data stretching back to the pre‐Gobeille period.

FIGURE 3.

Count and percent of fatal opioid‐related overdoses in Massachusetts in a given month that fail to link to all‐payer claims data identification numbers within the Public Health Data Warehouse in the given month, prior 6 months, prior 12 months, and all available data. January 2016, July 2016, and January 2017 are the first months where the 1‐month, 6‐month, and 12‐month look‐back periods do not include the pre‐Gobeille period, respectively.

In the PHD using linked APCD records, fatal opioid‐related overdose rates were consistently highest among white non‐Hispanic individuals over the study period; however, fatal opioid‐related overdose rates also rose for Black non‐Hispanic individuals and Hispanic individuals between 2013 and 2017, and accelerated among Black non‐Hispanic individuals after 2018 (Figure S1). Fatal opioid‐related overdose rates were also highest within the population receiving Medicaid and were lowest among the commercially insured population (Figure S2).

In controlled interrupted time‐series analyses, only the monthly model had significant autocorrelation. We determined that single data point lags were best to account for this autocorrelation. We found evidence for baseline, pre‐Gobeille parallel trends for the total statewide population in the APCD compared to American Community Survey estimates, as well as fatal opioid‐related overdose rates calculated using linked APCD records within the PHD compared to those reported in official statistics (Figure 4). Immediately after Gobeille, the total statewide APCD population declined by 1,146,709 persons (95% CI: −1,498,560, −794,858) when compared to the American Community Survey. This difference remained stable over time post‐Gobeille (trend difference = 110,440 (−90,987, 311,867)) (Table 2). Immediately after Gobeille, the fatal opioid‐related overdose rate calculated using the APCD population within the PHD was 4.5 persons per 100,000 (−1.05, 10.14) population higher than expected based on official statistics, although this difference was not statistically significant (Table 2). This difference narrowed over the years (trend difference = −0.25 (−3.45, 2.95)), but there was no evidence of divergence in fatal opioid‐related overdose rate trends between these two data sources after Gobeille. The results from the monthly controlled interrupted time series model largely mirrored those from the yearly model with two exceptions. First, there was a statistically significant greater increase in the rate of fatal opioid‐related overdose shift from 2015 to 2016 among the APCD‐link rate estimates compared to the Massachusetts Department of Public Health statistics (level shift difference = 0.55 (0.05, 1.05)). Second, the difference in the trends post‐Gobeille between the APCD‐linked rates and those from the Massachusetts Department of Public Health were negative (trend difference = 0.02 (−0.03, 0.06)) instead of positive.

FIGURE 4.

Annual total statewide population (A), annual rate of fatal opioid‐related overdose per 100,000 (B), and monthly rate of fatal opioid‐related overdose per 100,000 (C) calculated using linked all‐payer claims data within the Public Health Data Warehouse compared to the American Community Survey 5‐year estimates (A) and statistics in Massachusetts Department of Public Health official reports (B, C). The pre‐Gobeille period is 2013 to 2015 and the post‐Gobeille period is 2016 to 2021, Massachusetts.

TABLE 2.

Controlled interrupted time‐series comparing the total statewide population and fatal opioid‐related overdose rate calculated in the linked all‐payer claims data within the Public Health Data Warehouse to statistics in the American Community Survey 5‐year population estimates and Massachusetts Department of Public Health official opioid‐related overdose death statistics reports, respectively, 2013–2021, Massachusetts.

Coefficient	Interpretation	Total population estimate (95% CI)	Annual fatal opioid‐related overdose rate estimate (95% CI)	Monthly fatal opioid‐related overdose rate estimate (95% CI)
${\beta }_0$	Control value 2013	6,605,714 (6,431,522, 6,779,906)	14.23 (11.46, 17)	1.75 (1.43, 2.07)
${\beta }_1$	Control trend 2013–2015	50,264 (−84,665, 185,193)	5.7 (3.55, 7.85)	0.03 (0.00, 0.06)
${\beta }_2$	Control trend 2016–2021	−6837 (−149,267, 135,593)	−5.34 (−7.61, −3.08)	−0.03 (−0.06, 0.00)
${\beta }_3$	Control change 2015–2016	−12,061 (−260,857, 236,735)	3.17 (−0.79, 7.12)	0.21 (−0.15, 0.57)
${\beta }_4$	Control‐PHD value difference 2013	274,252 (27,907, 520,597)	−0.37 (−4.29, 3.55)	−0.38 (−0.83, 0.08)
${\beta }_5$	Control‐PHD trend difference 2013–2015	−189,323 (−380,141, 1495)	−0.42 (−3.45, 2.62)	−0.02 (−0.06, 0.02)
${\beta }_6$	Control‐PHD trend difference 2016–2021	110,440 (−90,987, 311,867)	−0.25 (−3.45, 2.95)	0.02 (−0.03, 0.06)
${\beta }_7$	Control‐PHD change difference 2015–2016	−1,146,709 (−1,498,560, −794,858)	4.54 (−1.05, 10.14)	0.55 (0.05, 1.05)

Note: The time span for the monthly model is June 2014 to December 2021.

Abbreviations: CI: confidence interval, PHD: Public Health Data Warehouse.

4. Discussion

The Gobeille decision limited the ability of APCDs to collect data on individuals with self‐insured health plans. It reduced the representativeness of the Massachusetts APCD and its ability to help evaluate public health outcomes generalizable to the full state population. The drop in the count of individuals may have been due to both natural changes and reduced capture of individuals on self‐insured plans, two factors that are impossible to tease apart. However, the sudden and dramatic decrease in the number of self‐insured individuals in the APCD from hundreds of thousands pre‐Gobeille to only a few thousand post‐Gobeille (99% decrease) suggests at least some impact of Gobeille on demographic and geographic representation of the APCD, especially when considering that individuals with other insurance types that were unaffected by Gobeille did not exhibit such a drop. Prior work has illuminated this challenge and has shown that the Massachusetts APCD was representative of over 98% of the Massachusetts population pre‐Gobeille, but only 76% of the population post‐Gobeille [5]. Despite such changes in state APCDs, few studies have directly addressed the potential impacts of Gobeille on APCD representativeness. Many studies either ignore the issue [27, 28, 29, 30, 31] or only acknowledge it in their limitations [32, 33, 34, 35, 36]. However, one study using the Rhode Island APCD acknowledged this limitation in their methods and adjusted for year of capture in the APCD to account for changes in APCD membership over time [37].

Results from our study confirm earlier findings from the Massachusetts Center for Health Information and Analysis showing that the distribution of sex post‐Gobeille is largely similar to the pre‐Gobeille period, but younger individuals were disproportionately lost [5]. Our results suggest that the impacts of Gobeille were most profound on its sustained impact on racial and ethnic and geographic representativeness of the state population, which may have significant impacts on the estimation of substance use and opioid‐related overdose events as well as analyses of disparities. While almost all ZIP Code Tabulation Areas lost population captured in the APCD, the largest decreases occurred among suburban communities across Massachusetts. These changes in the composition of the population suggest notable selection biases with respect to key social and demographic profiles in Massachusetts APCD and thus the need for statistical and perhaps spatial weighting approaches when using Massachusetts APCD systems post‐Gobeille to ensure representativeness to the underlying population of interest. Such methods may be able to approximate the population pre‐Gobeille; however, weighting procedures will still be biased by unknown factors that occur in the missing population post‐Gobeille.

The impact of Gobeille hindering the ability to use APCD as the spine to link multiple datasets within the PHD and subsequently to capture fatal opioid‐related overdose records is clear. Despite a large proportion of fatal opioid‐related overdoses occurring among individuals with Medicaid (Table S1), by 2021, about one‐fifth of fatal opioid‐related overdoses in a given month could not be linked to APCD insurance records in the prior year. This was because post‐Gobeille fatal opioid‐related overdoses among self‐insured individuals were only able to link to the APCD if they were present in the APCD pre‐Gobeille, thus explaining the high linkage when using all available APCD history stretching back pre‐Gobeille This increasing failure to link over time shows how representation of the Massachusetts population diminishes within the APCD, and thus the PHD, as more self‐insured individuals who were not present in the APCD pre‐Gobeille were unable to be identified in the APCD post‐Gobeille. Accounting for this linkage failure in analyses using a look‐back window that is unable to take advantage of APCD information pre‐Gobeille would require either a longer look‐back window in the post‐Gobeille period to allow for more APCD insurance records available to link to fatal opioid‐related overdoses or the use of weighting techniques to restore the unlinked fatal opioid‐related overdoses.

We found a statistically significant effect of Gobeille on reducing the statewide population captured by the Massachusetts APCD when compared to the American Community Survey estimates, but no evidence for statistically significant differences between the fatal opioid‐related overdose rate as calculated using the APCD‐linked Death Certificate data in the PHD compared to official Massachusetts statistics. While there was an initial increase in the APCD‐linked estimates of fatal opioid‐related overdose rate post‐Gobeille, this increase was only statistically significant in the monthly model and the difference between the rates calculated using the APCD‐linked records within the PHD and those reported in official statistics was not sustained. The reason for this may be twofold. First, Medicaid recipients are at higher risk of opioid‐related overdose [38, 39, 40, 41] and are overrepresented among fatal opioid‐related overdoses that link to the Massachusetts APCD (Table S1). Thus, since Gobeille had no effect on Medicaid data reported to the APCD, the aggregated figures for fatal opioid‐related overdose rates in the full APCD population may be relatively unaffected by Gobeille. Second, the decline in the population in the APCD within the PHD may be proportional to the decline in the capture of fatal opioid‐related overdoses in the PHD when linking to the APCD, and thus the rate is unaffected.

This study has several limitations. First, age and ZIP Code information were not available for all individuals in the APCD. While the missingness figures were very low, it was difficult to determine whether this group was meaningfully different from the rest of the population since very little information was available about them. Second, comparisons were made between the APCD population in a given year post‐Gobeille to the population as it was in 2015 and do not take into account any changes to the population that may have occurred otherwise. We assumed that the 2015 population served as an adequate representation of the population between 2016 and 2021. Third, the APCD does not directly identify individuals as being on self‐insured plans; instead, it only categorizes them as “commercially‐insured.” The Massachusetts Department of Public Health designates individuals suspected to be on commercial self‐insured plans if they could no longer be identified in the APCD post‐Gobeille, but it is likely that some individuals designated as “commercially‐insured other” may in fact be on self‐insured plans. Fourth, fatal opioid‐related overdose linkage to the APCD was not complete pre‐Gobeille. Thus, we cannot be certain of the exact proportion of fatal opioid‐related overdose linkage failure to the APCD post‐Gobeille that was due to the Gobeille decision. Fifth, we focused on fatal opioid‐related overdose rates as an example health outcome, but the results presented here should not be presumed to apply to other health outcomes, particularly ones where the commercially‐insured population is of particular interest.

The Massachusetts Department of Public Health, in collaboration with the Center for Health Information and Analysis, is working on improved linkage procedures within the PHD to recover the representativeness of the APCD by leveraging information about individuals captured in other state‐level datasets. In the meantime, and for states without a linked data system similar to the PHD, Gobeille represents a major challenge to analyses using APCD data. While the impact of Gobeille on the representativeness of the APCD for the Massachusetts population was minimal with respect to sex and age, as well as analyses of fatal opioid‐related overdoses, we found that post‐Gobeille, the APCD population had a lower proportion of White non‐Hispanic individuals than pre‐Gobeille, and that a substantial portion of fatal opioid‐related overdoses post‐Gobeille was unable to be linked to the APCD. Representation of individuals also decreased disproportionately in suburban areas of Massachusetts, suggesting that weighting techniques that are specific to race, Hispanic ethnicity, and geography are needed post‐Gobeille to ensure that the APCD mirrors the state population.

Funding

This work was supported by National Institute on Drug Abuse (R01DA054267).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Annual fatal opioid‐related overdose rate per 100,000 population as captured by the Massachusetts Public Health Data Warehouse stratified by race and ethnicity, 2013–2021.

Figure S2: Annual fatal opioid‐related overdose rate per 100,000 population as captured by the Massachusetts Public Health Data Warehouse stratified by health insurance type, 2013–2021.

Table S1: Annual percentage of fatal opioid‐related overdoses linked to all‐payer claims data (APCD) within the Public Health Data Warehouse (PHD) by insurance status, Massachusetts.

Table S2: Annual distribution of individuals in all‐payer claims data (APCD) within the Public Health Data Warehouse (PHD) by number of unique recorded ZIP Codes, Massachusetts.

Table S3: Annual distribution of missingness of age and ZIP Code in all‐payer claims data (APCD) within the Public Health Data Warehouse (PHD), Massachusetts.

Data Availability Statement

The data that support the findings of this study are maintained by the Massachusetts Department of Public Health. Restrictions apply to the availability of these individually‐linked data, which were used under license for this study.