Supplemental Appendix for Insurance Transitions and Changes in Physician and Emergency

Supplemental Appendix for “Insurance Transitions and Changes in Physician and Emergency Department Utilization: An Observational Study”

Michael L. Barnett, MD, Zirui Song, MD, PhD, Sherri Rose, PhD, Asaf Bitton, MD, MPH, Michael E. Chernew, PhD and Bruce E. Landon, MD, MBA, MSc

Table of Contents

Methods

1. Master patient identifier derivation and confidence
2. Defining insurance enrollment periods
3. Codes used for defining new physician visits
4. Defining primary care vs. specialty physicians
5. Difference-in-differences model specification
6. Other sensitivity analyses

Appendix Figure 1: Cohort Flow Diagram

Appendix Figure 2: Standardized Mean Differences in Covariates in Full and Matched Cohorts

Appendix Figure 3: Changes in New PCP Visits Stratified by Presence of Chronic Illness

Appendix Figure 4: Changes in New Specialist Visit Rates with Insurance Switching for Individuals with Initial Private Insurance Coverage, Sensitivity Analyses

Appendix Table 1: Pre-Period Utilization Trends 12 to 6 Months Before Switch

Appendix Table 2: Difference-in-Difference Sensitivity Analyses

Appendix Table 3: Full Regression Output for Difference-in-Difference Models – Initial Private Insurance Cohort

Appendix Table 4: Full Regression Output for Difference-in-Difference Models – Initial Public/Medicaid Insurance Cohort

Appendix Methods

1. Master patient identifier derivation and confidence

Because each health plan has its own unique method for identifying subscribers, it is difficult to create a master patient identifier that allows for following subscribers when they switch insurers.1As an example of this challenge, the member eligibility (ME) file in the APCD from 2009-2012 has 37.2 million separate plan-specific identifiers even though Massachusetts has only 6.6 million residents.This problem existsbecause there is no widely used standard such as full social security numbers (SSN), and in the absence of SSNs, other identifiers can have significant variability in quality.

To overcome the obstacle of duplication of identifiers across the same individual, the APCD uses a master patient identifier labeled in the database as the “member enterprise ID”. The APCD contracted with IBM Corporation to create probabilistic matches between different insurance enrollment entries for individuals. This algorithm incorporates information such as member name, SSN (when available), date of birth, address, and gender. More details on the derivation of the master patient identifier are available on the APCD website.2 The master patient identifier was validated by the state agency maintaining the APCD through manual examination of hundreds of matches as well as analysis of the concordance between the US Census and master patient identifier-derived population at the ZIP code level.

Because the master patient identifier is created using a probabilistic model, not all matches are perfect. This is indicated in a “confidence level” for matching assigned in the APCD as follows: “high” confidence for a nearly perfect match or better, “low” confidence for anything besides a perfect match, and “singleton” match for a member with entries only in the member enrollment file without medical or pharmacy claims. We included matches in all categories and controlled for any bias potentially introduced by matches measured with error by incorporating a control variable for match type in our adjusted models below.

1. Defining insurance enrollment periods

Due to issues of missing data and the administrative complexity of handling insurance switches, many individuals in the APCD who switch insurance have enrollment periods with some overlap. We dealt with this by making a series of assumptions which we believed to be conservative and leaned strongly towards removing “false positive” switches, which are switches where the date of switching is unclear or possibly erroneous. We followed the following steps:

1) If an enrollment period was missing a start date, it was excluded.

2) If an enrollment period was missing an end date (very common for Medicaid plans), then either December 31st of the year of data submission was used, or the date of “last activity” of the insurer on the member’s file was used, whichever was later.

3) If two enrollment periods for an individual were entirely overlapping (e.g. the individual appeared to have two insurance products simultaneously for an extended period), the following enrollment periods were excluded:

- If one of the overlapping plans was Medicaid (this is because Medicaid often provides secondary coverage which was not the focus of this study)

- If no medical claims had ever been filed under one of the enrollment periods

4) We also excluded any insurance switches with more than a 60-day lag between them (also because we required coverage for 11 months in any given year).

After excluding the enrollment periods above, we recalculatedthe existence of remaining overlapping insurance periods. We assumed that any remaining overlaps were due to an administrative lag between an individual’s enrollment in a new insurance carrier and the previous carrier removing the individual from enrollment. This has been reported as an endemic problem in Massachusetts Medicaid, for example.3 Therefore, we assigned the beginning of any new insurance enrollment periods as the “switch date” for individuals with 2 or more insurers in the study period.

1. Codes used for defining new physician visits

We defined a new office visit with the following Current Procedural Terminology (CPT) codes:

New Patient Office Visit: 99201-99205

New Patient Office Consultation: 99241-99245

New Patient Preventive Care Visit: 99381-99387 (Only applied to primary care physicians)

1. Defining primary care vs. specialty physicians

Similar to the master patient identifier problem described in part A of this appendix, a master provider identifier is not universally available in the APCD. This is because each insurance carrier uses their own internal identifier system to keep track of billing providers, ranging from physicians to physical therapists and laboratories. The National Provider Identifier (NPI) used by Medicare is a standard identifier for physicians, but is used infrequently to identify providers in all claims submitted by the insurance carriers in the APCD. Therefore, to identify individual physicians in the APCD across insurers and merge in their specialty information from the APCD provider file, we created a crosswalk file between the sets of insurer-specific identifiers and NPI. To do this, we took the following steps:

1) Using the provided “carrier specific values” file provided with the APCD, we categorized all specialty codes from the top 5 largest private insurers in Massachusetts as primary care or specialist and merged these definitions to the APCD provider file.

2) We took the entire medical claims database from the APCD and extracted every unique combination of insurer-specific provider identifiers and NPI.

3) We merged the file from step #2 to the APCD provider file, limiting to only physician entries with a primary address in Massachusetts.

4) To define a physician NPI for any given medical claim, we merged the file from step #3 back into the medical claims file and defined an NPI for any identifier in the following fields: service provider, billing provider or rendering provider.

5) We assigned any given claim to the NPI merged in the following hierarchy: rendering provider NPI > service provider NPI > billing provider NPI. Some insurers only used one or two of the fields above.

Once NPI and specialty were merged into the medical claims file in step #5, we could assign each new physician office visit in the database to a physician specialty, used for our primary outcomes.

1. Difference-in-differences model specification

We used linear regression and a difference-in-differences analytic design to assess the association between insurance switching and outcomes of new physician visits and emergency department utilization. We used a linear model because we were most interested in average effects, since with large sample sizes linear models generally outperform other statistical models at estimating population averages.4–6For each outcome, we fitted the following model:

Yi,t = 0+ 1Post_Switcht + 2Switcheri + 3Post_Switcht*Switcheri +

4Covariatesi+ 5Montht

where Yi,t is the number of events for each outcome forpatientiinmonth t, “Post_Switch” is a binary indicator for the 12-month period after aninsurance switch, “Switcher” is an indicator for whether patientiswitched insurance or is a matched control patient, and “Covariates” denotes the patient characteristics in Table 1. “Month” denotes fixed effects for each individual month in the study period. We fitted separate models for patients initially with private insurance or Medicaid insurance. The covariates included in the model depended on whether or not an individual was initially privately insured or had Medicaid insurance. The patient characteristics shown in Table 1 list the characteristics included in each model for these two major subgroups of individuals. All models were estimated accounting for clustering of observations within individuals using generalized estimating equations using SAS (v. 9.4, Cary, NC).7,8

We included post-switchperiod and switcher indicators to compare the outcomes for all switchers in the post-switch period versus the change inoutcomes in the matched control group from the pre- to post-switch periods. Therefore, 3, our quantity of interest, represents the average adjusted change in each outcome in the post-switch period attributable to insurance switches controlling for trends in the matched control population.

1. Other Sensitivity Analyses

We examined trends in new PCP visits for patients with chronic illness who may be more vulnerable to the effects of changing physicians. We defined chronic illness as the presence of one or more chronic illnesses common in the Medicaid and privately insured populations, using the presence in the prior year of a diagnosis of diabetes mellitus, hypertension, coronary artery disease, congestive heart failure, asthma, or chronic obstructive pulmonary disease, columns.

We also replicated our analysis excluding the 6 month period prior to switching due to the possibility of measurement error in the dates of switching and possible changes in behavior in anticipation of insurance switching. Finally, we replicated our analysis limiting our cohort to matched switchers and non-switchers whose identifiers were linked across time with high confidence.

We observed no significant change in our difference-in-differences estimates after excluding the 6 months prior to switching as a “washout period,” with the exception of the post-switch decrease in ED visits observed in the Medicaid population, which became non-significant (Appendix Table 2). There was also no significant difference in our overall results when restricting to individuals with member identifiers linked with high confidence in the APCD (Appendix Table 2). We also observe similar increases in new PCP visits for patients with chronic illness (p<0.001 for the post-switch period for all, Appendix Figure 3)

Appendix Figure 1: Cohort Flow Diagram

Appendix Figure 2: Standardized Mean Differences in Covariates in Full and Matched Cohorts

A

B

Appendix Figure 2 shows the standardized mean differences between switchers and non-switchers across all of the categories of covariates shown in Table 1 in the “raw,” or unmatched cohort (solid black dots) and matched cohorts (open white dots). The gray dotted line shows a standardized mean difference of 0.05, a very conservative threshold of similarity for balance between covariates in two separate groups. The covariates are sorted by the magnitude of their standardized mean difference in each figure. The covariates for the initially privately insured are shown in panel A and for those with initial Medicaid coverage in panel B.

Appendix Figure 3: Changes in New PCP Visits Stratified by Presence of Chronic Illness

Figure 3 shows trends in outcome rates per 100 persons for new primary care physician (PCP) visits in 30-day intervals (“months”) relative to insurance switching for switchers and their matched control non-switchers, stratified by initial insurance (Medicaid or Private, rows) and presence of one or more chronic illnesses common in these populations (defined as the presence in the prior year of a diagnosis of diabetes mellitus, hypertension, coronary artery disease, congestive heart failure, asthma, or chronic obstructive pulmonary disease, columns).

Unadjusted rates for switchers are shown with the solid red line (“Switchers”). The same trend for non-switchers is shown with the blue solid line (“Non-switchers”). The solid black line indicated the time point of insurance switching. 95% confidence intervals are shown for all unadjusted estimates, assuming a normal distribution of rates given the large sample size of individuals.

Appendix Figure 4: Changes in New Specialist Visit Rates with Insurance Switching for Individuals with Initial Private Insurance Coverage, Sensitivity Analyses

Appendix Figure 4 shows sensitivity analyses for trends in outcome rates per 100 persons for new specialist visits in 30-day intervals (“months”) relative to insurance switching for switchers and their matched control non-switchers. The top panel shows outcomes for the subgroup of individuals with employer insurance switches or plan cancellations (n= 74,030 including controls), the middle panel shows outcomes for the subgroup switching insurance on January 1st or July 1st from 2011-2013 (n=273,628 including controls), and bottom panel shows outcomes for the subgroup of individuals living in the same ZIP code before and after insurance switching (n=509,038 including controls). These analyses are restricted to individuals with initial private insurance coverage. Unadjusted rates for switchers are shown with the solid red line (“Switchers”). The same trend for non-switchers is shown with the blue solid line (“Non-switchers”). The solid black line indicated the time point of insurance switching. 95% confidence intervals are shown for all unadjusted estimates, assuming a normal distribution of rates given the large sample size of individuals.

Appendix sTable 1: Pre-Period Utilization Trends 12 to 6 Months Before Switch

Abbreviations: Primary care physician (PCP)

* Estimated from a linear regression model clustering at the individual level using generalized estimating equations, adjusting for all patient characteristics shown in Table 1 (depending on intial insurance type).

Appendix Table 2: Difference-in-Difference Sensitivity Analyses

Abbreviations: Primary care physician (PCP)

*Replication of the main analysis from Table 2 (shown in this Table as well) excluding the 6 months prior to switching. Therefore, the pre-period encompasses the period from 12 to 6 months prior to switching.

**Replication of the main analysis from Table 2, with the cohort restricted to only individuals with high confidence member identifiers in the all-payer claims database. We also repeated the switcher/non-switcher algorithm described in the Methods section with this cohort for this analysis.

*** Estimated from a linear regression model clustering at the individual level using generalized estimating equations, adjusting for all patient characteristics shown in Table 1 (depending on intial insurance type).

Appendix Table 3: Full Regression Output for Difference-in-Difference Models – Initial Private Insurance Cohort*

* Estimated from a linear regression model clustering at the individual level using generalized estimating equations. All regression coefficients are shown, excluding month level fixed effects for presentation.