The Effects of Trigger Events on Changes in Children's Health Insurance Coverage

F. EFFECTS OF TRIGGER EVENTS

We have seen that transitions in health insurance coverage among children are often preceded by changes in their parents' employment, AFDC participation, family income, or family composition, although the frequency of these events varies by the type of transition. While this gives us a measure of the potential role of these events in effecting the transitions that we observe, it is quite possible for transitions to be frequently preceded by particular events but for these same events to be followed only infrequently by transitions. It might be the case, for example, that there are important mediating factors that must be present if a transition in health insurance coverage is to be produced by a particular trigger event. If these factors are not captured in our survey data, we cannot identify them and measure their impact, but the nature of the relationship between possible trigger events and transitions may suggest their presence.

To provide a measure of the effects of possible trigger events on the occurrence of transitions we examined the frequency of transitions as a function of the prior occurrence of these events. To do so, we defined selected changes occurring between months m-1 and m as potential trigger events and then estimated the relationship between these events and the likelihood that a transition in health insurance coverage was recorded over the next four months. We did this separately for each of the four types of coverage, with the outcomes of interest in each case being transitions to any of the other three types of coverage versus no transition. We present our findings in two forms: first, as the results of a logistic regression of transition outcomes on the full set of possible trigger events and, second, as estimates of the frequency of each type of transition among the subset of children experiencing a given event. The regression results give us a measure of the relative importance of individual events in predicting transitions while the conditional frequencies tell us in a more intuitive form how often transitions actually occurred after events that the regression analysis identified as the strongest predictors.

1. Regression Results

In this section we present the findings from an application of logistic regression analysis to estimate the impact of a child's experiencing a possible trigger event on the likelihood that the child will make a transition from his current coverage. As in the preceding section, we present separate analyses of children whose initial coverage is ESI, uninsured, Medicaid, or other insurance.

Methodology. Analyses of transitions in health insurance coverage often focus on spell length and use proportionate hazard models to estimate the impact of fixed or time-varying characteristics on the exit rate from a particular coverage status. Typically, events have a limited role--if any--as predictors. ⁽²⁴⁾ Given our interest in trigger events, we have approached the problem differently. Trigger events can occur at any point in the history of a spell, and by definition their impact is relatively quick. Rather than asking how the occurrence of such events affects the length of a spell or how it affects the monthly exit probability, we want to know how the occurrence of such an event affects the probability that a child will exit one state or enter another in the next few months. This is fundamentally different than wanting to know the impact of personal characteristics on spell length (or exit rates), and it requires a different approach. Our basic model utilizes a four-category "multinomial" dependent variable identifying exits from one of our four types of coverage into each of the other three versus a fourth category indicating no exit during the four-month time span. We estimated a separate model for each of the four original sources of coverage. The predictors are potential trigger events. Except for one additional variable added to adjust for the SIPP seam bias, the models include no other predictors. We opted for this reduced form rather than estimating structural equations in which we attempted to include all of the characteristics that may affect exits from particular types of coverage and transitions into others because our research is exploratory and we wanted to focus on the role of events as predictors of change in coverage.

Children with ESI. Table 14 presents the results of a logistic regression analysis of children's loss of employer-sponsored insurance. The regression model was estimated with a multinomial dependent variable indicating whether the child made a transition into uninsured, Medicaid, or other insurance or remained covered by ESI. ⁽²⁵⁾ The predictors are the several trigger events, expressed as binary variables (coded 1 or 0 to indicate whether or not the event occurred in the reference month). ⁽²⁶⁾

The logistic regression model necessitated by the nature of the dependent variable is non-linear, so the effects of the trigger events cannot be expressed simply as net changes in the probability of observing a transition. ⁽²⁷⁾ We have elected to express the effects of the individual trigger events as odds ratios. An odds ratio indicates how much the likelihood or "odds" of a child losing employer-sponsored insurance is increased by the occurrence of a particular event. ⁽²⁸⁾ With the multinomial dependent variable the odds ratios express the effects of the trigger event in terms of the likelihood of a child making the indicated transition versus remaining covered by ESI. For example, in the first row of Table 14 the coefficient of 2.68 in the uninsured column implies that the odds of a child becoming uninsured in the next four months are increased nearly 3 times by the father's losing employment. ⁽²⁹⁾ The coefficient of 4.58 in the Medicaid column implies that the odds of a child enrolling in Medicaid in the next four months are increased between 4 and 5 times by the father's losing employment whereas the coefficient of .43 in the other insured column indicates that the odds of a child moving from ESI to other insurance are actually reduced by 57 percent (1 minus .43) by the father's loss of employment, although this particular effect is not statistically significant. ⁽³⁰⁾

All six of the variables that represent actual or potential reductions in employment have significant and positive effects on the likelihood of a child's leaving ESI to become uninsured. The strongest effects are associated with the father's reducing his hours of work below 30 or changing jobs. The effects of changes in the mother's employment are consistently weaker than the corresponding changes in the father's employment, but they are still relatively strong. The only other events with significant effects on the likelihood of a child losing ESI and becoming uninsured are a drop in family income and a reduction in family size--both of which increase the likelihood of a transition to uninsured. These effects are weaker than the effects of employment changes. That the reduction in income continues to increase the likelihood of a loss of insurance after controlling for employment changes underscores the importance of the parents' ability or willingness to pay for coverage when free or heavily subsidized coverage is not available.

We have no explanation for the significant effect of a reduction in family size. We observed the appearance of this variable earlier as a prior event in transitions from ESI to uninsured, but we counted it as a relevant event solely on the strength of its empirical association with these transitions --that is, without a clear theoretical justification.

Turning to the next two columns of Table 14 we find, first, that only the parents' loss of employment and a parent's leaving the family affect the likelihood of a child's leaving ESI and enrolling in Medicaid. The parent's leaving the family may not only take away employer-sponsored coverage but place the family in a position where the child, at least, can qualify for Medicaid. This same event has a significant but negative effect on the child's moving from ESI to other insurance, and our interpretation is that the father's departure and associated loss of income may eliminate other insurance as a potential source of coverage. It is consistent with this interpretation that a marked increase in family income should also have a positive and significant effect on the likelihood of a child's obtaining other insurance, but we are at a loss to explain why a reduction in family income should have the same effect. Finally, the father's reducing his hours below 30 has a very strong positive effect on the likelihood of a child's replacing ESI with other insurance. It is not clear why the reduction in hours should so often result in an exit from ESI rather than the parent's assumption of the full costs of maintaining coverage, which we would continue to count as ESI. Further research is needed to understand the rationale behind such choices.

It is intuitively understandable that all six employment variables should have independent effects on the likelihood of transitions from ESI to uninsured, because each of these changes in employment carries the potential to change the employee's access to employer-sponsored coverage or the cost of maintaining that coverage. At the same time, children who enroll immediately in Medicaid rather than becoming uninsured must not only lose their ESI but qualify for Medicaid. Our results suggest that children whose parents lose their employment have some increased likelihood of qualifying for Medicaid whereas children whose parents change jobs or reduce their hours do not.

The strength of the coefficients on parents' employment loss may help to explain why the same two variables do not have a stronger effect on transitions from ESI to uninsured: rather than becoming uninsured, the children of parents who lose their employment may become covered by Medicaid. The results for other insurance, on the other hand, seem to underscore the fact that obtaining other insurance requires an ability to pay. That is, the parents' loss of employment has no effect on transitions from ESI to other insurance because parents who lose employment are not able to pay for other insurance. At the same time, fathers who change jobs or reduce their hours of work may retain their ability to pay for other insurance if they lose ESI. Nevertheless, we are surprised that the father's reduction in hours should have such a strong, positive effect on the likelihood of a child moving from ESI to other insurance.

Children without Insurance. Logistic regression analysis of the effects of potential trigger events on children who are without health insurance indicates that very few events were significantly associated with transitions out of uninsurance after controlling for other events. In Table 15 we see that increases in the hours worked by either parent had significant effects on transitions into ESI, as did a marked rise in family income and a parent joining or rejoining the family. This last event also had a very strong positive effect on the likelihood of a child enrolling in Medicaid. A parent's leaving the family also had a positive but much weaker effect on this same transition while the mother's changing jobs or losing employment had positive effects as well. The mother's losing employment presumably helps to qualify the child for Medicaid, but the fact that the child was previously uninsured suggests that the mothers that account for this effect held jobs that provided no insurance coverage but produced enough income to make the child ineligible for Medicaid. Income changes in both directions had significant, positive effects on transitions to other insurance, and the same was true of the mother's changing jobs. With respect to the income changes, recall that we saw the same result for transitions from ESI to other insurance. Here, too, it is difficult to explain why changes in both directions should affect transitions in the same way, but we saw this same phenomenon with respect to other insurance in our earlier analysis of events preceding transitions.

Children with Medicaid. Regression results for children who were initially covered by Medicaid are presented in Table 16. The odds ratios are strikingly similar for transitions into uninsured and transitions into Medicaid. The loss of AFDC is the single strongest predictor of transitions from Medicaid to uninsured, but it is also one of the strongest predictors of transitions from Medicaid to ESI. Trigger events that had significant effects on the transitions from Medicaid to uninsurance tended to have similar if not significant effects on transitions from Medicaid to ESI, and vice versa. The chief exceptions to this pattern are the mother's changing jobs, which had a significant if modest effect on the child's moving from Medicaid to ESI but no measured effect on the child's moving from Medicaid to uninsured, and the family's income falling markedly, for which

the reverse was true. An interpretation of the overall pattern is that the principal effect of these events is to move children out of Medicaid rather than pull them into ESI or uninsurance.

Our estimates of the effects of trigger events on transitions to other insurance are affected by the very small sample size of these particular transitions. We included these transitions in our regressions only to obtain a complete accounting of transitions. Nevertheless, there are some similarities with the findings for the other two transitions--in particular, the estimated effects for either parent's increase in hours worked, the father's loss of employment or reduction in hours below 30, and the rise or fall in family income (where the resemblance is to transitions from Medicaid to uninsured but not Medicaid to ESI).

Children with Other Insurance. Regression results for children whose initial coverage was other insurance are presented in Table 17. Because of the relatively small sample size of children with other insurance, odds ratios that would be significant in the regression results that we have already reviewed are not significant here, and some of the odds ratios are quite large. Rather than viewing these as evidence of very powerful effects on transitions, we are more inclined to see them as the result of large standard errors. The father's increasing his hours of work or the mother changing jobs had significant effects on the likelihood of a child leaving other insurance for ESI. Both of these make intuitive sense, but we cannot explain the significant positive effects of either parent's gaining employment on the likelihood of a child leaving other insurance to become uninsured. On the other hand, the significant positive effects of the father's losing employment or family income falling markedly do fit our priors here, and they suggest that with a major employment loss or reduction in family income the family's ability or willingness to continue paying for other insurance tends to decline. Finally, as in the previous table the sample size for transitions between other insurance and Medicaid is very small. We included these transitions, again, so that we could fully account for transitions out of other insurance, but we find these odds ratios difficult to interpret.

2. Conditional Frequencies of Transitions

The number of transitions attributable to a particular trigger event is a function of both the net effect of that event and its frequency in the population. ⁽³¹⁾ If two trigger events have similar effects on the likelihood of a particular transition, but one event occurs much more often than the other, then the more commonly occurring event will induce a larger number of transitions. Because of their nonlinearity, the net effects that we reported in the preceding section do not translate directly into probabilities that children will experience transitions, but the gross or unadjusted effects do, and estimates of the frequencies of individual types of events are readily obtainable. In this section we examine the frequencies of the 12 types of transitions while conditioning on each of the events that was included in our final regression models. While these conditional frequencies are based on unadjusted effects, meaning that they do not control for the effects of other events, we calculated them only for events that the regression analysis identified as the strongest predictors of transitions from each initial coverage status. As a result, we can be certain that we are restricting our attention to those events with the strongest net effects.

Children with ESI. On average, 41.8 million children were covered by ESI at any one time from June 1993 through May 1994. ⁽³²⁾ For this group of children, Table 18 reports the average number who experienced individual events in the next month and, for each event, the percentage who retained their ESI for at least the next four months or lost their coverage and became uninsured, enrolled in Medicaid, or obtained other insurance. In the final three columns Table 18 gives the actual number of transitions implied by the average monthly number of events listed in the first column and the four-month transition probabilities reported in the previous three columns. For the group of children as a whole--most of whom experienced no events--94 percent remained covered by ESI, 3 percent became uninsured, and between 1 and 2 percent enrolled in Medicaid or obtained other insurance. ⁽³³⁾ For children who experienced an event, however, we find as many as 31 percent losing their ESI, with most of these becoming uninsured. For example, among the nearly 200,000 children covered by ESI whose fathers lost employment in an average month, 23 percent became uninsured in the next four months, 4 percent enrolled in Medicaid, and an additional 4 percent obtained other insurance. In combination with the average monthly frequency of children's fathers losing employment, these transition probabilities imply that about 46,000 children moved from ESI to uninsured, 8,000 moved from ESI to Medicaid, and 7,000 moved from ESI to other insurance.

The greatest number of transitions from ESI to uninsured is associated with children whose family income fell markedly in the past month. While declines in family income were associated with a very modest transition probability--only 7 percent--declines in income were the most common event, occurring nine times as often as fathers losing employment, for example. Children whose fathers lost employment were the most likely to lose ESI in the next four months--23 percent did so--but they accounted for only one-third as many transitions as children who experienced a marked reduction in family income.

Children with marked reductions in family income also accounted for the most transitions from ESI to Medicaid and ESI to other insurance. In the regression analysis, however, declines in family income had significant coefficients for transitions from ESI to either uninsured or other insurance but not to Medicaid, suggesting that we should discount the 47,000 transitions to Medicaid. Similarly, marked increases in family income were nearly as common as marked decreases and were also associated with large numbers of transitions of all three types, but the regression analysis indicated that a decline in income had a significant effect only on transitions to other income.

Further underscoring the importance of looking at both the effect of a given event on the probability of a transition and the frequency of that event, we see that the father's and mother's loss of employment had roughly similar gross effects on the probability of a child moving from ESI to Medicaid, but the mother's loss of employment was associated with three times as many transitions as the father's, owing primarily to the greater frequency of employment loss among the mothers than among the fathers of ESI children. Likewise, a parent's leaving the family was associated with the largest conditional probability of a child leaving ESI for Medicaid (and also the largest net effect in the regression analysis), yet because of the relatively low frequency of ESI children losing parents, the associated transitions are less than 7,000 or the second lowest among all of the events reported in Table 18.

Children Without Insurance. Transitions out of uninsurance occur at a much higher rate than transitions out of ESI. In the first row of Table 19 we see that 29 percent (100 minus 71) of the 9.0 million children who were without insurance in an average month from June 1993 through May 1994 became insured over the next four months. Roughly equal fractions moved into ESI and Medicaid while a much smaller fraction obtained other insurance. Because transitions out of uninsurance were so common, we see rather substantial transition rates associated with individual trigger events. More

than two-thirds of the children who experienced a parent joining the family became insured in the next four months, with 31 percent obtaining ESI and 37 percent enrolling in Medicaid. Likewise, nearly half of the children whose parents increased their hours of work to 30 or more became insured, with 30 to 33 percent obtaining ESI, 12 to 16 percent obtaining Medicaid, and 2 to 3 percent obtaining other insurance.

Because transitions out of uninsurance occurred at such a high rate in general, we must compare the transition rates associated with individual triggers to the rates for all uninsured children to properly interpret the numbers of transitions reported in the final three columns. For example, there were 75,000 transitions from uninsured to ESI among children whose family incomes fell markedly, but the 13.0 percent rate that these transitions represent is lower than the unconditional 13.6 percent rate among all uninsured children, and the regression coefficient in Table 15 indicated no relationship between such changes in income and transitions from uninsured to ESI. The implication is that none of these transitions should be attributed to the family income change. At the same time, however, the transition rate from uninsured to other insurance among children with a sharp drop in family income is nearly twice the average rate, and the regression coefficient in Table 15 showed a doubling of the odds of this particular transition when family income dropped markedly. Thus the 28,000 transitions from uninsured to other insurance reflect both a strong effect of a family's loss of income and the high frequency with which uninsured children experienced a sharp drop in family income. The mother's changing jobs is associated with a comparable transition rate from uninsured to other insurance, yet the lower frequency of such job changes translates into only 9,000 transitions.

The high rate of transitions from uninsured to ESI among children who gained a parent yields a relatively small number of transitions--10,000--compared to parents increasing their hours of work, which is associated with a comparable transition rate but six times as many transitions.

Children With Medicaid. Transitions out of Medicaid also occurred at a much higher rate than transitions out of ESI but not as high as transitions out of uninsurance. Table 20 reports the distribution of outcomes among all children with Medicaid and the subsets who experienced individual trigger events. About 15 percent of the 13 million children who were reported to be enrolled in Medicaid in an average month between June 1993 and May 1994 left Medicaid within four months. Of these, nearly 10 percent became uninsured while 5 percent obtained ESI and only one-tenth that number (.5 percent) acquired other insurance.

Four events were associated with particularly large movements from Medicaid to uninsured: the family's loss of AFDC benefits, the father's gaining employment, the father's increasing his hours to 30 or more, and a parent joining the family. In each case between 28 and 31 percent of the children recorded such a transition in the next four months. The numbers of transitions associated with these events ranged from about 8,000 for the family demographic change to 75,000 for the loss of AFDC. Transitions associated with income changes once again dominated the movements, and while the corresponding transition rates were not nearly as high as those associated with the other four variables, they were sufficiently above the average rate that the residual transitions are still high.

For transitions to ESI there were no events that clearly dominated the others with respect to rates of change. These rates ranged from 4.7 percent for all children with Medicaid to nearly 11 percent for those whose fathers gained employment. The actual numbers of transitions varied over a much broader range, of course. For the strongest net predictors from the regression analysis the numbers of transitions ranged from 11,000 to nearly 24,000. Children with family income changes had more numerous transitions, but these events had no net effects in the regression analysis, so we must discount their importance here.

Transitions from Medicaid to other insurance were infrequent generally. Consistent with the findings from the regression analysis the father's loss of employment was associated with the highest transition rate--at 2 percent--but children with large changes in family income appear to account for more transitions, even after we allow for the weaker association of income changes with this particular type of transition.

Children with Other Insurance. The overall rate at which children left other insurance was even higher than the exit rate from uninsurance. In Table 21 we see that an average of 30 percent of the children who were covered by other insurance at any one time between June 1993 and May 1994 left that coverage over the next four months. Most of the movement--22 percent--was to ESI, with 6 percent becoming uninsured and little more than 1 percent enrolling in Medicaid. Consistent with the interpretation of other insurance as generally privately purchased coverage, the high rate of movement from other insurance to ESI is consistent with the view that other insurance serves as an imperfect substitute for ESI.

The highest conditional exit rates from other insurance ranged from 63 to 74 percent (that is, the proportion retaining other insurance was between 37 percent and 26 percent). Interestingly, the father's gaining employment was associated with the highest exit rate from other insurance, but a substantial proportion of the children with this event became uninsured--26 percent compared to the 40 percent who obtained ESI. When fathers increased their weekly hours to 30 or more--which happened nearly three times as often as employment gains--54 percent of their children gained ESI, and less than 7 percent became uninsured. Children whose family income fell markedly had the most transitions from other insurance to ESI, but the transition rate of 21 percent was no higher than the unconditional average rate, implying that these transitions reflect no more than a small impact of the event per se. This contrasts with the smaller numbers of transitions associated with some of the employment changes, where many of the transitions truly reflect the impact of the events with which they correspond.

Both the father's gaining employment and losing employment are associated with high rates of children's movement from other insurance to uninsured, as is true as well of a parent leaving the family. All three of these events are relatively infrequent compared to the family's income falling markedly, which has a much weaker association with the transition but may still account for more transitions.

Transitions from other insurance to Medicaid were infrequent generally, and we found the regression results difficult to interpret. Compared to all children with other insurance there are very high transition rates associated with several trigger events, but we can infer from the magnitudes of the odds ratios in Table 17 that the standard errors were very large as well. As a result, we hesitate to attach much importance to the exit rates and transitions reported in Table 21.

Footnotes:

24. For example, Swartz et al. (1993) used data from the 1984 SIPP panel to estimate a multivariate hazard model showing the effects of personal characteristics on they rate at which adults exited spells of uninsurance. The only event included among the characteristics was the loss of private insurance. Short and Freedman (1998) used data from the 1990 SIPP panel to estimate discrete-time logit models of single women's exits from three types of coverage--Medicaid, private insurance, and uninsured--into each of the other two types. Predictors included both fixed and time-varying characteristics of the woman, her family, and the state labor market but no events.

25. The model assumes no ordering among the four categories, which is appropriate for our situation. That is, the results are indifferent to which outcome is assigned the value "1" on the dependent variable, which outcome is assigned the value "2," and so on.

26. We also included as a predictor a binary variable indicating whether or not the month in which the possible trigger events were observed was the first month of a four-month SIPP reference period, which could produce a spurious relationship between one or more trigger events and a transition in health insurance coverage. Generally, this variable was statistically significant but relatively weak in its net effect on the odds of observing a transition.

27. While the effects can be expressed as changes in probabilities, these effects vary across the range of the dependent variable. Researchers sometimes report the effects in logistic regression models as probabilities evaluated at the mean of the dependent variable, but this can easily lead to misinterpretation.

28. Letting Y=1 represent the occurrence of a transition and Y=0 the absence of a transition, and letting P represent the probability that Y=1, the odds of a transition occurring are defined as P/(1P). Thus if one in five or 20 percent of children experience a transition, the odds of the transition occurring are 1 to 4, or .25. If the two outcomes, transition versus no transition, are equally likely, the probability of either outcome is .5 while the odds favoring either outcome are 1.0.

29. This effect is constant across the range of the dependent variable.

30. The regression equations were estimated with the SUDAAN software, which can calculate standard errors that take account of the complex sample design of the SIPP. The standard error calculations used a pair of variables that the Census Bureau created and added to the SIPP file expressly to facilitate the estimation of standard errors. Because these variables are constant over time for individuals, they also help in correcting for the downward bias in the estimated standard errors that may arise from the use of multiple observations on the same individuals.

31. In Section E when we looked at the frequency of different types of events that preceded transitions we were seeing both their overall frequency of occurrence and, by comparison with children who reported no transitions, a measure of the strength of the relationship between each event and each type of transition.

32. The dates refer to the value of m-1 in the dataset we constructed. The universe of children in Table 18 is children who were covered by ESI in month m-1, that is, the month prior to month m.

33. It is possible for a child to have experienced more than one transition over the four-month period--for example, becoming uninsured and then enrolling in Medicaid. We count only the first transition in this table.

Where to?

Top of Page
Table of Contents

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 5, 2000