Assessing the Impact of Parity in the Large Group Employer-Sponsored Insurance Market: Final Report. APPENDIX B: Methods

Introduction

The purpose of this appendix is to provide additional details on the methodology and analytic approach used in the study to obtain the results presented in the main section of the report. This includes details on specific codes used in data from the MarketScan CCAE Database, such as International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes used to identify diagnosis groups, and additional detail on the theoretical framework and regressions tests.

In Appendix E, we provide details on our decision process that led to the report's focus on outpatient services. We initially performed a broad set of descriptive analyses on yearly outcomes for the full set of service categories. From our initial survey of trends across all service categories and from simple regression analyses, the greatest change in use and spending was occurring in the outpatient service category, in correspondence with parity's implementation time frame. In the simple regression analyses, we statistically tested the significance of parity's impact in the annual outcomes using a basic ITS regression model with a linear time trend, parity indicator, and parity*year interaction term. Using a summarized table of all outcomes, we found that the clearest indication of impacts from parity appeared to be occurring in the outpatient service category, leading us to focus on outpatient services for our monthly descriptive and regression analyses.

Diagnosis Groups

For outpatient BH services, we relied on ICD-9-CM diagnosis codes. In the outpatient claims data, there are four fields that a physician or service provider can use to indicate a diagnosis related to the service claim. None of the fields is designated as the primary diagnosis. Therefore, we are unable to know which diagnosis of the possible four listed is the primary diagnosis being treated; for this reason, we use "any listed" diagnosis to determine the diagnosis group. This means that if any of the outpatient diagnosis fields has a MH or SUD diagnosis code, we identify the outpatient service as a BH service. We based MH and SUD diagnosis codes on HHS Agency for Healthcare Quality and Research Clinical Classifications Software (CCS) categories, with minor modifications. For example, we did not include diagnosis codes from the CCS category 653, Delirium, dementia, amnestic and other cognitive disorders. Table B1 contains the list of CCS categories that form the basis of the MH and SUD diagnosis groups. A full list of ICD-9-CM codes used to define the MH and SUD diagnosis groups is available in Appendix C. The outpatient non-BH diagnosis group then was defined by the lack of a MH or SUD diagnosis code on the service claim. We used this strategy of identifying BH services and non-BH services in numerous past studies.[37, 38]

In addition to examining the MH, SUD, and non-BH diagnosis groups, we further classified the SUD diagnosis group into two groups: (1) SUD diagnoses that relate to OUD; and (2) SUD diagnoses that do NOT relate to OUD. Table B2 lists the ICD-9-CM diagnosis codes used to identify outpatient services that fall into the OUD diagnosis group. Outpatient service claims with any listed OUD ICD-9-CM diagnosis code were classified into the OUD diagnosis group, whereas all other outpatient services originally identified as being SUD were classified into the non-OUD SUD diagnosis group. By splitting the SUD diagnosis group into OUD and non-OUD diagnosis groups, we were able to assess the impacts of parity on use and spending specifically for opioid-related SUD treatments.

Subpopulation Definitions

In addition to looking at the full population's use and spending for MH and SUD outpatient services, we examined overall MH and SUD service use and spending for two subpopulations: individuals with a SMI diagnosis and individuals with an OUD diagnosis. Our subpopulation analysis is different from our analysis of the full population in several ways. First, in our subpopulation analysis, we limited our analysis to the services used by the specific subpopulation identified. Thus, for the subpopulation SMI, we first identified the individuals in the MarketScan data that met our criteria for having an SMI, and then we proceeded with our comparison of MH, SUD, and non-BH use and spending trends ONLY for that subpopulation. Second, in contrast to the analysis of the full population in which we focused on outpatient services only, we included the full set of service categories in the subpopulation analysis. This means that our analysis of spending and use trends as affected by parity included, in addition to outpatient services, emergency department visits, inpatient admissions, prescription drug fills, and laboratory and radiology services. The SMI and OUD subpopulations are two of the most vulnerable groups of individuals with either an MH or SUD diagnosis. In this subpopulation analysis, we therefore were able to assess the impact of parity on each group's full use of MH and SUD services.

To define the SMI subpopulation, we used a strategy of requiring either one inpatient admission or two outpatient visits with a diagnosis from the following list of ICD-9-CM codes (also see Table B3). The ICD-9-CM codes used for identifying the SMI subpopulation were 295.x, 296.24, 296.34, 296.4, 296.5, 296.6, 296.7, and 296.8.

For the OUD population, we used a more complex strategy that used diagnosis codes, prescription drug fills, and selected service administration codes. We used additional identifiers for the OUD subpopulation to capture individuals receiving treatment for OUD (e.g., buprenorphine prescription fill for OUD treatment) who may not have been given an associated diagnosis. In other unpublished work, we have found that relying only on a diagnosis code for OUD population identification will miss some individuals who clearly are receiving treatment for OUD. We can only speculate why individuals who are receiving OUD treatment are not always given an OUD diagnosis; however, it does appear that this issue is less prominent in more recent years. We believe our method for identifying individuals with OUD is robust and alleviates worries of missing individuals in this subpopulation.

There were three possible ways that an individual could be identified as part of the OUD subpopulation. First, we identified individuals who had either a primary inpatient diagnosis or two any listed outpatient diagnoses using the same list of ICD-9-CM codes listed in Table B2. Second, we identified individuals who had a prescription drug fill either for certain buprenorphine prescriptions that are primarily used for OUD treatment or for naltrexone prescriptions. A full list of National Drug Codes (NDCs) for the buprenorphine and naltrexone prescriptions is included in Appendix D. Third, although buprenorphine is received most often through a physician prescription, samples also may be given by the physician. In the latter case, a service administration code can be used that will appear in the outpatient service claim. Individuals receiving methadone treatment also can be identified through outpatient service claims by a methadone service administration procedure code. Outpatient buprenorphine and methadone service administration codes are listed in Table B4.

It is important to note that when identifying subpopulations, all claims and prescription drug data available within the full study period were used for individuals who met the general population criteria.

Service Categories: Outpatient and Other Service Category Definitions

Outpatient Service Category

The outpatient service category that was the primary focus of the main results included a wide range of service settings and levels of treatment. In fact, the only services that were excluded from the outpatient service category were inpatient admissions, emergency department visits, and laboratory and radiology services. Thus, the outpatient service category included intensive outpatient services, partial hospitalizations, and residential treatment services. In MarketScan data, inpatient admissions are identified by the existence of a room and board revenue code. We identified services all related to an inpatient admission and placed them into a separate inpatient analytic file. In the outpatient analytic file, we identified emergency department visits using the last two digits of a MarketScan variable called SVCSCAT that identifies the detailed service type. Only emergency department visits that ended in a discharge from the emergency department rather than an admission to the hospital were included in the MarketScan outpatient analytic file. Emergency department visits that led to an inpatient admission were included in the inpatient analytic file. Other than treat-and-release emergency department visits, it was important to identify laboratory and radiology services, which often do not contain diagnosis codes and therefore do not differentiate which services are related to BH. We identified laboratory and radiology services using the MarketScan variable STDPLAC, which identifies the place of service. We also placed prescription drug claims in a separate analytic file, similar to inpatient admissions. To identify the set of services included in the outpatient service category, we used the MarketScan outpatient file and excluded emergency department visits and laboratory and radiology services (see Table B5).

Other Service Categories

Although we decided to focus on the outpatient service category for our main analysis, including monthly descriptive and regression analyses, we did examine the full set of service categories in a preliminary analysis on annual spending and use trends. Results from this preliminary analysis are discussed in Appendix E. The full set of service category definitions are detailed in Table B5.

When identifying which inpatient admissions were BH (MH or SUD), we used any listed diagnosis field out of 16 total possible fields in the MarketScan inpatient analytic file. Similarly, for emergency department visits we used any listed diagnosis out of the four possible fields in the outpatient analytic file. Because laboratory and radiology services often do not have populated diagnosis fields, we categorized all laboratory and radiology services as non-BH.

For prescription drug claims, we used a MarketScan variable indicating the therapeutic class to identify MH prescription drugs and a list of NDC codes to identify SUD prescription drugs. The following is a list of MH therapeutic classes: antidepressants, antipsychotics, stimulants, stimulants non-amphetamine, anxiolytic/sedative/hypnotics (barbituates, benzodiazepines, not elsewhere classified), antimanic agents not elsewhere classified, and certain central nervous system agents.

For SUD medications, the full list of NDC codes is detailed in Appendix F.

The methodology used to categorize service categories resulted in mutually exclusive service categories whereby each service or prescription drug fill fell into one service category only. Similarly, services and prescription drug fills were identified as either BH or non-BH.

Spending Decomposition Framework

We used a spending decomposition framework in this report as a theoretical structure to analyze the spending and use outcomes. A spending decomposition framework uses the following spending decomposition equation to break health spending per enrollee into the relevant components that drive spending changes. We expressed the spending decomposition equation in terms of BH spending per enrollee:

BH spending per enrollee [spending] =
(% enrollees using BH services) [access] × (average units used per user) [intensity of service utilization] × (average cost per unit) [reimbursement]

In addition to the three components that make up spending in this equation, another factor that could be incorporated into the spending decomposition framework is whether services used are in-network or out-of-network. For insurers, prices of in-network services are contractually negotiated and generally are lower than those charged by out-of-network providers. For enrollees, out-of-network services often impose higher out-of-pocket cost sharing incentivizing enrollees to use in-network providers.

Interrupted Time Series Regression First Order Serial Correlation Tests

First Order Serial Correlation

One important statistical issue with ITS regression is serial correlation. We tested for first order serial correlation using a Durbin-Watson test statistic for all regressions. Test statistics across various ITS regressions showed some evidence of first order serial correlation, leading us to estimate the full set of ITS regressions correcting for first order series correlation. Comparing results of coefficients, we found that coefficient signs and magnitudes were very similar for all regression results. We therefore opted to present results that do not correct for first order serial correlation.