The length of time that children spend without health insurance coverage is exceedingly important to policymakers. Brief spells of uninsurance carry very different policy implications than long spells. While transitional spells of even one month in duration carry risks to the temporarily uninsured, the policy responses that they invite are clearly different than those demanded to address chronic lack of insurance. The advent of the Children’s Health Insurance Program (CHIP) gives heightened importance to the duration of uninsurance as a policy variable. Federal law allows states considerable freedom in determining how much of the uninsured population to cover and how to define eligibility. A number of states are specifying minimum periods of uninsurance before a child can qualify for the new coverage.
How long do spells of uninsurance last? Here we present findings with respect to the duration of spells without insurance. We contrast two different ways of identifying spells: new spells that began within a particular 12-month period and active spells that were in progress at the end of that period. In addition to looking at spells of uninsurance, we examine spells in which the uninsured were eligible for Medicaid and, theoretically, could have availed themselves of such coverage.
1. New Spells
Table 1 describes the duration of spells of uninsurance that began during FY93. For all of these spells, the first set of columns in Table 1 gives the distribution of completed duration in single months, up to a maximum of 13 or more. For most of this last category we did not observe the spell completion; we know only that these spells lasted at least 13 months. We combine all spells of 13 months or greater because 13 is the longest duration that we can observe with the SIPP panel for a spell starting in the final month of FY93. To produce representative estimates of longer durations, we must limit our sample to spells starting earlier in the year. For spells starting in the first six months of FY93, therefore, the second set of columns in Table 1 reports completed durations up to 19 months or more.1
Perhaps the most striking feature of these distributions is the heaping of reported durations at four months. Of the 9.2 million spells that began in FY93, more than one-third were completed in exactly four months. This is not a true reflection of spell durations but a phenomenon of the measurement process in the SIPP. This heaping of responses at four months, and to a lesser degree at multiples of four months (there are local peaks at 8, 12, and 16 months), is the result of a “seam effect.” As we mentioned above, SIPP respondents are interviewed every four months and asked extensive questions about their circumstances in each of the four preceding months. For reasons that are not fully understood, transitions in many characteristics are reported as occurring dispro- portionately between the four-month reference periods of interviews--that is, at the seams--rather than within these reference periods. It happens that this seam effect is particularly strong for transitions into and out of insurance coverage. For example, while we would expect only one quarter of new spells of uninsurance to begin in the first month of a reference period or end in the last month, 75 to 90 percent of new spells are reported as beginning in the first month or ending in the fourth month. The transitions, in other words, are reported to have occurred between the reference periods of the interviews. Explanations for this phenomenon include genuine recall error as well as various kinds of behaviors that respondents or interviewers may engage in to speed up or simplify the interview. Clearly, treating the previous four months as a single reporting period--for at least some kinds of responses--simplifies the interview, but if that is what is happening in the reporting or recording of health insurance coverage, then we have to interpret the data accordingly.
Because of the staggered interviewing, described earlier, the seam effect does not show up in the estimates reported for calendar months. But this is not to say that the seam effect does not influence the overall magnitude of these calendar month or point-in-time estimates. Unless respondents (or interviewers) are as likely to “round up” or “round down” their reported months of participation in a reference period, the monthly estimates of aggregate participation will be affected. In a later report, we consider how the seam effect might affect estimates of the number of children who are uninsured.
The seam effect is not unique to SIPP, by any means, and the research that preceded the initiation of the first SIPP panel in 1984 included an examination of this phenomenon. Panel surveys with annual interviews find the same problem when they attempt to measure monthly behavior. To the extent that recall error contributes to the seam effect in SIPP, the impact of such error will be magnified in surveys that attempt to measure behavior over longer periods of time--such as an entire year, which is the reference period for much of the data collected in the March supplement to the Current Population Survey (CPS). Indeed, the fact that CPS estimates of the number of people who had no health insurance coverage during the previous calendar year resemble point-in-time estimates more closely than they resemble the calendar year estimates produced by panel surveys is indicative of a serious response problem that may share some underlying causes with the SIPP seam effect.
Nevertheless, it is clear that we have to take account of the seam effect when we interpret data on spell durations, such as those reported in Table 1. Estimates of the duration of completed spells reflect the maximum impact of the seam effect because each reported spell is subject to this influence at both ends. Of the nearly 3.2 million spells of uninsurance reported as being completed in exactly four months, how many were really completed in four months, and how many were completed in three or five months, two or six months, or even one month or seven months? And how many spells of one, two, or even three months duration were not reported at all?
While there may be ways to tease out more information from the SIPP data, we have not attempted to do so here.2 Instead we simply recognize the limitations of these data. Specifically, despite what the SIPP data appear to tell us about the length of short spells of uninsurance, we would suggest that these data should not be used to draw inferences about the relative frequency of spell lengths shorter than six months or about longer spell lengths in groupings of fewer than four months. In other words, it may be reasonable to infer from Table 1 that about 60 percent of new spells of uninsurance in FY93 were completed in six months or less, but the data will not support inferences about how many of these spells may have been completed in two months or less, or four months or less. Similarly, it may be correct to infer that about 8 percent of spells lasted 9 to 12 months or that 3.5 to 4 percent lasted 15 to 18 months but certainly not that only .1 percent were completed in exactly 13 months. In general, we need to be conscious that the numbers reported at 4 months, 8 months, 12 months, or 16 months almost certainly include more spells of shorter or longer lengths than they do spells of the nominal length.
Spell lengths of six months or longer or 12 months or longer are becoming important in defining eligibility for CHIP coverage in a number of states. What can we infer about the relative frequency of such spells from the data presented in Table 1? Read literally, the data in Table 1 tell us that about 43 percent of the spells that began in FY93 lasted six months or longer (100 minus the 57.3 percent with lengths of five months or less), and 26.4 percent of the new spells in that year lasted 12 months or longer. Recognizing that most of the spells reported as having durations of exactly four or 12 months were actually longer or shorter than that, we would suggest that the true proportion of new spells with durations of six months or longer could be as much as 5 percentage points higher--up to 48 percent--and that the true proportion of spells with durations of 12 months or longer could be 2 to 3 percentage points lower than the observed figure, or closer to 24 percent.
2. Active Spells
Where Table 1 presented distributions of new spells of uninsurance that began in FY93, Table 2 presents distributions of all spells that were active at the end of FY93. For all children who were uninsured in September 1993, Table 2 tells us how long they had been uninsured. The distribution of these “current” durations can be compared to the distribution of completed durations of new spells, which is reproduced from Table 1 in the right hand columns of Table 2.
It is important to understand how the durations of active spells reported in the left side of Table 2 differ from the durations of new spells reported in the right side (and in Table 1). First, the durations of active spells are incomplete. More specifically, they are “right censored.” If we think of a duration as spanning from left to right on a time line, in other words, what we do not observe for these active spells is how far their durations extend to the right. All of the spells terminate at the same point in time--September 1993. We do not know how long these durations will be when they are completed. The second respect in which the active spells portrayed in Table 2 differ from the completed spells presented earlier is that they share a common month. All of the active spells include September 1993. While there is a lot of overlap in the spells reported in Table 1, these spells do not share a common month.
These two differences have important implications for how the distributions of durations differ. First, let us consider what the durations reported on the left side of Table 2 represent. Durations of one month represent spells that started in September. Ultimately, very few of these spells will end at one month. Instead, if we could follow them through to completion, we would find that they have essentially the same distribution as the spells reported in Table 1. Durations of two months refer to all spells that started in August 1993, minus those that ended in that same month--that is, spells with completed durations of one month. If we were to follow these spells to completion, they too would have the same distribution as the spells reported in Table 1 except for the complete absence of spells of one month duration. Similarly, the 12-month durations include all spells that started in October 1992 minus those that ended in any of the months from October 1992 through August 1993.3
What we notice first in the distribution of active spells is the absence of a seam effect. There are no peaks whatsoever at multiples of four months. This absence of a seam effect is due entirely to the right censoring. Most of these spells are not complete, so the seam effect is hidden. One quarter of the spells that will eventually end in four months shows up as one month in length, one quarter shows up as two months in length, one quarter appears as three, and one quarter as four months in length. Far from producing peaks at four months, the seam effect actually contributes to the relative evenness of the distribution of spells across durations of one to four months.
The second thing we notice is that apart from the higher proportion of spells with durations of three months or less, the spells that are active at the end of the year (and this would be true of spells that were active during any given month of the year) are longer, on average, than the spells that started during the year--even though most of the active spells are not yet complete. The median length of the active spells lies between 9 and 10 months whereas the median completed length of the new spells lies close to 4 months. Nearly 46 percent of the active spells have durations in excess of 12 months compared to just 20 percent of the spells that started during the year.
If we followed the active spells to their completion, we would obtain a distribution that is heavily skewed toward long durations. Why is this? It turns out that the representation of spells among active spells is directly proportional to their completed duration. Spells of one month duration are represented solely by spells that began in a single month: September 1993. Spells of two months duration are represented solely by spells that began in either of two months: August or September. Spells of three months duration are represented by spells that began in any of three months, and so on. Thus spells of 12 months duration are represented by spells that began in any of 12 months--October 1992 through September 1993--while spells of 36 months duration are represented by spells that began in any of 36 months--from October 1990 through September 1993. Compared to a distribution of spells starting in the same month or the same year, therefore, spells of exactly 36 months in length are represented at 36 times their relative frequency.
We could estimate the distribution of completed durations among spells that were active in September 1993 by taking the completed distribution of new spells reported in Table 1 and multiplying the number of one-month spells by 1, the number of two-month spells by 2, number of three-month spells by 3, and so on. Obviously, we cannot complete this exercise with the data presented in Table 1 because we do not know the completed duration of spells that exceeded 18 months in length. But if we make an assumption about the average completed duration of spells that were completed in more than 18 months, then we can estimate the relative frequency of such spells among active spells.
Table 3 displays this projected distribution of the completed duration of spells that were active in September 1993. These calculations assume, of course, that the distribution of completed durations did not change over time, so that the relative frequency of, say, an 18-month spell is the same for spells starting at any time between October 1991 and September 1993. Whether or not this assumption is satisfied, the projected distribution illustrates our point. In Table 3 we see that only 22 percent of these spells are completed in six months or less compared to about 60 percent of the new spells in Table 1. At the other end of the distribution, 49 percent of the spells reported in Table 3 extend beyond 18 months compared to only 15.5 percent of the spells in Table 1.
If these different ways of measuring the duration of uninsurance yield such different distributions, can they all be useful? The answer is yes, but the different distributions address different questions. Table 1 reflects a representative sampling of spells. It tells us that most spells of uninsurance among children are relatively short in length, 60 percent being six months or less. The distribution has a long tail, however, with at least 20 percent of spells (we suggested somewhat higher) exceeding 12 months in length and about three-fourths of these surpassing 18 months. It is this long tail that makes possible the skewed distribution that we see when we look at the spells of a representative sample of people with spells active in a give month. People who experience long spells are more likely to be uninsured in a given month than people who experience short spells, so children who are uninsured at a point in time exhibit disproportionately many long spells. Table 3 suggests that about half of the 9 million children who were uninsured in any given month of FY93 will have had spells of more than 18 months in length by the time they become insured (again), and most of these spells will extend well beyond 18 months. Finally, Table 2 addresses a question that is becoming very pertinent with the introduction of CHIP--namely, if eligibility is limited to children who have been uninsured for a specified number of months, how many children will satisfy these criteria? Table 2 indicates that with a 12 month minimum, about 47 percent of all uninsured children would meet this requirement. The percentage among children who meet the income and other requirements is likely to be higher. Lowering the duration requirement to 6 months would add another 16 percent of children--a relatively modest increase given that the minimum duration would be cut in half.
3. Medicaid-Eligible Uninsurance
The large number of uninsured children who appear to be eligible for Medicaid has captured the attention of policymakers and advocacy groups alike. In Technical Appendix A we present estimates that 30 to 33 percent of the children who were uninsured at a point in time in FY93 or FY94 were eligible for Medicaid, based on an eligibility simulation that accounted for a substantial majority but certainly not all Medicaid-eligible children. Here we ask how long children who are uninsured and eligible for Medicaid tend to remain both uninsured and Medicaid-eligible.
Table 4 provides information on the duration of spells of Medicaid-eligible uninsurance. A spell of Medicaid-eligible uninsurance begins in the first month that a child is both uninsured and Medicaid-eligible. A child may have been uninsured for one or more months before becoming Medicaid-eligible or may have been Medicaid-eligible at the start of a spell of uninsurance. Likewise, a child may have lost Medicaid eligibility before becoming insured or may have remained eligible through the end of a spell of uninsurance.4 Because patterns may vary, we cannot infer the full length of a spell of uninsurance from the number of months that a child was both Medicaid eligible and uninsured.
We began our Medicaid-eligibility simulation in October 1992, so we cannot identify new spells of Medicaid-eligible uninsurance starting before November 1992. (We need to observe the previous month to determine if a new spell has started.) The spells reported in Table 4, therefore, began in November 1992 or any of the subsequent months through September 1993.
Spells of Medicaid-eligible uninsurance tend to be much shorter than spells of uninsurance generally. More than one-half of all spells started in FY93 were completed in three months or less, and only 6 percent extended beyond one year. Among all spells of uninsurance started during the same period, 20 percent reached beyond one year, and more than 15 percent extended at least another six months beyond that.
Finally, the SIPP seam effect is much less pronounced for spells of Medicaid-eligible uninsurance than for all spells of uninsurance. We attribute this to the fact that simulated Medicaid- eligibility is not a characteristic that is reported by survey respondents or recorded by survey interviewers but a complex construct of a great many variables that change at different rates and times. Nevertheless, there are local peaks at four, eight, and 12 months. Taking this into account suggests that the proportion of spells of Medicaid-eligible uninsurance completed in three months or less could be as high as 60 percent.