Data Source and Study Population
The Massachusetts APCD contains detailed data on health care utilization, insurance eligibility, and provider credentialing across all commercial payers and public health insurance programs in Massachusetts, representing approximately 90% of the state's non-elderly population.14
15 The APCD uses individual identifiers from insurers to create a probability-based “master patient identifier” that can be used to follow enrollees when they change insurers (see online Appendix).16
Our study cohort included all Massachusetts residents aged 21–64 years as of December 31, 2013, who had at least 3 calendar years of continuous enrollment with any APCD-participating insurance provider during the 4-year period from 2010 to 2013, based on their master patient identifier. We chose a 3-year period of continuous enrollment to ensure 12 months before and after an insurance switch for our study design, and additionally used the prior-calendar-year claims for risk adjustment. We defined continuous enrollment as ≥11 months of insurance coverage in a year. We excluded any individual covered by Medicare during this period, as Medicare claims are not included in the APCD.
This project was approved by the Committee on Human Subjects at Harvard Medical School.
Identifying Insurance Carrier Switching
Our primary exposure was switching insurance carriers, which we defined when an individual (“switcher”) began enrollment with a new health insurance carrier (i.e., not changing plans within an insurance carrier). Because identifiers for individual plans in the APCD were not consistent within carriers, we were unable to examine switches between plans within a carrier. For each person, we defined an index insurance carrier (e.g., Aetna or United) based on their health insurance enrollment in January 2011. We assigned the date of a switch to the month of the first instance when a person began enrollment with a new carrier, even if their prior enrollment extended beyond this month, because we found that plans did not always accurately record when coverage ended (see online Appendix for details). We defined “Medicaid” coverage as enrollment in MassHealth or in a privately contracted MassHealth health plan. We studied switchers who changed insurance during calendar years 2011–2012, including January 1, 2013, because January 1 is a common date for insurance switching.
We defined new visits to primary care and specialist office-based physicians based on evaluation and management (E&M) Current Procedural Terminology (CPT) codes (see online Appendix for list of codes used) for new visits. A new visit CPT code can only be used if a patient has not been seen by any physician within that practice within the previous 3 years.17 Because payers in the APCD do not consistently use a single physician identifier such as the national provider identifier (NPI), we created a master physician crosswalk between NPI and insurer-specific identifiers and used this to define specialty (see online Appendix for details). We defined emergency department (ED) visits by any claim with an emergency care-associated E&M code (99281–99285, 99291, 99292) or an ED-associated revenue center code (0450–0459, 0981) in hospital claims.18
To measure comorbidities, we calculated hierarchical condition category (HCC) scores using software available from the Centers for Medicare and Medicaid Services.20 We assigned an HCC score from 2010 or 2011 for individuals who switched insurance in 2011 or 2012 plus January 1, 2013, respectively. To derive a proxy measure of socioeconomic status, we linked individuals’ five-digit ZIP Code to the publicly available “Area Deprivation Index,” a measure derived using a weighted combination of 17 census-level indicators of socioeconomic disadvantage.21
22 We also assigned members to an initial insurance market category (private vs. Medicaid insurance). Other variables included member age as of December 31, 2010, sex, and patient identifier linkage confidence (an indicator of whether a member’s identifier match was perfect or had some inference in linkage, see online Appendix). We also collected several variables related to insurance coverage for those with private insurance: plan type (e.g., preferred provider organization [PPO] vs. health maintenance organization [HMO]), employer size (ranging from individual to 500+ employees, see Table 1 for categories), and whether the employer was self-insured.
Because individuals switching insurance may differ from those with stable enrollment, we used propensity score methods to match switchers with non-switchers without replacement in our sample using all available patient characteristics (see online Appendix). We assigned non-switchers the same insurance switch date as the matched switcher to align the timeline of all individuals relative to this date.
To compare the characteristics of switchers versus non-switchers before and after matching, we used t-tests or χ2 tests as appropriate. We assessed the change in utilization of physician and ED visits using a difference-in-differences design. We defined the pre-switch period as the 12 months preceding a switch, and the post-switch period as the 12 months after a switch, including the month of the insurance change.
For the ED utilization analyses, we divided the indicator for the post-switch period by month to estimate changes in utilization over time versus the pre-switch period from 4 to 12 months before switching. We estimated linear models that included an indicator for period (pre/post or month fixed effects), an indicator for insurance switching, and interactions between the period and insurance switching indicators, which was our main quantity of interest. Even though the dynamics of insurance switching were more complex than simple pre/post changes, we used a standard difference-in-differences framework to estimate average overall effects in the post-switch period. The model further adjusted for all available patient factors noted above (see full model specification in online Appendix). We tested for the assumption of parallel trends between switchers and non-switchers in the pre-switch period by examining whether utilization trends were equivalent between the two groups, using the same model specification as described above. All models were estimated accounting for clustering of repeated observations within individuals using generalized estimating equations. Analyses were performed with SAS software (version 9.4, SAS Institute Inc., Cary, NC).
Sensitivity and Subgroup Analyses
Because some insurance switching might be motivated by a change in health status or a desire to switch healthcare providers, we performed sensitivity analyses focusing on subgroups more likely to be experiencing exogenous insurance switches. First, we repeated the analyses above after restricting to individuals who were most likely to have switched due to an employer switching plans or the cancellation of an entire plan contract (“plan cancellation”). We identified these by flagging all plans or employers with 20 or more members, with at least 75% of members switching insurance on the same day, using plan identifier codes and employer tax identifiers (available for 12% of the matched sample). Second, because those who move, whether due to a job change or not, may be more likely to establish care with new providers, we restricted our analyses to those whose ZIP Code of residence did not change (80% of the matched sample).
In addition, because new health plans might also entail higher enrollee cost sharing that might have independent effects on post-switch utilization, we repeated our analyses on the subset of switchers for whom there was available information on insurance deductibles (35% of the matched sample). We classified all insurance switches as those switching to a plan with a higher, equal, or lower deductible and repeated the analyses stratified by these populations.
Finally, we performed additional sensitivity analyses to replicate our analyses among patients with chronic illness, among patients with high confidence identifiers and excluding a 6-month “washout” period prior to insurance switching (methods and results in online Appendix).