| A. Introduction | Table of Contents | C. Frequency of Transitions |
1. Data
The data source for this analysis is the 1992 panel of the Survey of Income and Program Participation (SIPP), which interviewed a nationally representative sample of household residents every four months over a span of three years and collected monthly data on health insurance coverage, family composition, family and personal income by detailed source, and a variety of additional variables. SIPP is thus an excellent source with which to measure transitions in health insurance coverage and to identify potential trigger events.
The most recent SIPP panel was started in 1996 and ran through the end of 1999, but the first longitudinal data file from this panel, which will cover 1996 and 1997, is not scheduled for release until March 2001. The next most recent panels, which were started in 1992 and 1993, cover about three years each. (1) We selected the 1992 panel because we had worked with it previously and because comparisons with other data suggest that the 1993 panel overstates the number of families below poverty. We focus our analysis upon transitions occurring between July 1993 and June 1994 to give us a representative set of transitions occurring late in the life of the 1992 panel and to allow us to look forward several months past the last transitions (September 1994 is the final month for which all components of the health insurance measures are available for the full longitudinal sample). While these data are nearly six years old, they nevertheless provide a rich source of information on transitions in health insurance coverage and the events that may help to precipitate them. Undoubtedly, whatever these data can tell us about the events that trigger changes in health insurance coverage remains relevant as we enter the next decade.
A transition in health insurance coverage involves both an exit, from the first coverage or origin, and an entry, into the second coverage or destination. Each type of exit or entry may be associated with a different set of potential trigger events, which suggests that we examine different types of transitions separately. We elected to group the transitions by the coverage that precedes the transition--that is, the original coverage. We examined transitions among four distinct sources of coverage: employer-sponsored insurance (ESI), Medicaid, other insurance, and a lack of coverage. ESI includes all coverage obtained through a current or former employer, whether or not the employer pays any part of that coverage. (2) "Other insurance" may include both privately purchased insurance and public insurance other than Medicaid or Medicare, which respondents identify directly, but from the survey questions we know only that such coverage was obtained in some way other than through a current employer or union, former employer, or the CHAMPUS or CHAMPVA programs. (3) Children may also have coverage that is not described adequately enough to be assigned to one of the three general sources of coverage. This is particularly true of children whose coverage is provided by an adult who lives outside the household--a divorced parent in most cases. While most of this unknown coverage is ESI, we elected not to assign such coverage to ESI but to exclude it from our typology altogether. Thus, movements into or out of unknown coverage are not counted among the transitions that we examine. (4)
While SIPP captures health insurance coverage on a monthly basis, the reporting of changes in health insurance coverage--as well as other types of transitions--is characterized by a substantial "seam bias." That is, reported transitions of many kinds fall disproportionately between rather than within the four-month reference periods for which the interviews collect data. If the timing of transitions were reported correctly, only one in four transitions would occur at the seams between reference periods. Instead, for the types of transitions in health insurance coverage that we examine in this report, between 66 and 99 percent were reported to have occurred at the seams (see Appendix Table A.1). (5) The seam bias for potential trigger events was weaker, with 34 to 76 percent of these changes being reported between rather than within reference periods. The seam bias affects the reported data in several ways that are relevant to our research. Both the temporal proximity and the sequencing of events may be misstated. Short spells are almost certainly underreported, and spell durations show a substantial heaping at multiples of four months. To use these data to investigate the impact of trigger events on transitions in health insurance coverage requires a number of accommodations, which we will discuss as we review our methodology and findings.
2. Methodology
To perform the analyses reported herein, we constructed a dataset consisting of 11-month snapshots providing measures of health insurance coverage and a variety of parental and family characteristics. Each snapshot consisted of data from a focal month, m, plus the next four months and the preceding six months. (6) Month m was any of the 12 months from July 1993 through June 1994. We aggregated these snapshots into a single database. A sample child who was in the universe of children under 19 for the entire period and was covered by ESI, Medicaid, other insurance or was uninsured is represented 12 times--once for each 11-month sequence. A child who was born into or who aged out of the universe of children during the year or was covered by an unknown source of insurance at any time is represented for only those months m in which the child was a member of the universe. (7)
Footnotes:
1. The 1992 panel collected calendar month data from January 1992 through December 1994 while the 1993 panel collected calendar month data from January 1993 through September 1995.
2. Thus coverage obtained under COBRA would be included under ESI. While it might have been useful to separate out such coverage in order to show transitions between coverage provided by the employer and coverage that a former employee was maintaining at his or her own expense, only about a quarter million children at any one time were reported to be insured with coverage that we interpret as COBRA, and there were too few sample observations of transitions between this coverage and broader ESI to support detailed analysis.
3. For the population of children, coverage under Medicare is exceedingly rare. In our SIPP sample there are only two children who were reported to have had Medicare coverage at any time between June 1993 and June 1994, and for one of these children the coverage changed to Medicaid. Rather than group Medicare with much more common forms of "other" coverage, or with Medicaid, we excluded it from our typology. Therefore, movements into or out of Medicare are not included among the transitions that we examine.
4. Many such movements may not involve actual transitions at all, or relevant trigger events may be unobserved (as is often true when the source of coverage is someone outside the survey household). Arguably, transitions between unknown coverage and a lack of coverage should be treated differently and counted among the transitions that we examine. Such transitions were rare, however, and may reflect excessive reporting error. Furthermore, when the coverage that preceded or followed a spell of uninsurance was provided by persons outside the household, we lack measures of key trigger events.
5. The types of reporting error that produce this seam bias are not peculiar to SIPP nor limited to longitudinal surveys. While some aspects of the seam bias in SIPP are unique to the SIPP design and may be accentuated by features of the SIPP questionnaires, misreporting of the starting and ending dates of spells and whether or not a spell occurred at all during a reference period is distressingly common (Moore et al. 1999; Mathiowetz forthcoming).
6. The asymmetry grows out of the two distinct analytical approaches that the dataset was designed to support and the different ways in which the SIPP seam bias was deemed relevant to either. These analytical approaches are discussed below.
7. For simplicity and based on what we knew of the SIPP seam effect, we captured most of our data items for only a subset of months (specifically, m-6, m-1, m, and m+4). We identified events and transitions by comparing selected pairs of months, as we will explain in the discussion of our findings. In so doing, we may have lost some events and transitions, but only if they were quickly reversed. It is a feature of this dataset that transitions measured between months m-1 and m will sum to the annual total of transitions occurring between July 1993 and June 1994.
We approached the analysis of trigger events in two ways. The first involved looking backwards from transitions observed between months m-1 and m to identify events that occurred with disproportionately high frequency among children with transitions compared to children without transitions. The second involved looking forward from events observed between these same two months to determine how often the events were followed by specific transitions. This second approach lent itself to the application of regression analysis to determine the extent to which alternative events predicted transitions and to generate estimates of the net effects of the impact of these alternative events.
With both approaches we had to contend with two general problems affecting the relative positioning of potential trigger events and the transitions with which they were associated. First, because of the seam bias in the reporting of change in the SIPP, events and transitions that occurred as much as a few months apart could be reported as occurring in the same month. This made it important to look for potential trigger events not only in prior months but in the same month as the transition. Second, as we detail in the next section, an additional transition often preceded or followed any transition that we observed. Events that were associated with the second transition could occur at the same time as or even prior to the first transition, creating spurious associations. The design of the regression analysis was intended to limit the effects of paired transitions.
| A. Introduction | Table of Contents | C. Frequency of Transitions |
Home Pages:
Office of Health Policy
(HP)
Assistant Secretary for Planning and Evaluation
(ASPE)
U.S. Department of Health and Human Services
(HHS)
Last updated September 2, 2000