Slipping into and out of Poverty: the Dynamics of Spells

This paper examines the dynamics of poverty. Previous analyses have examined either fluctuations in the male heads' eamings or the frequency of poverty periods over a fixed time frame. Our approach depends on a definition of spells of poverty. Using this methodology we find that the majority of poor persons at any time are in the midst of a rather long spell of poverty. The methodology also allows us to estimate that less than 40 percent of poverty spells begin because of a drop in the heads' eamings, while 60 percent of the spells end when the heads' eamings increase. Thus, researchers must focus on household formation decisions and on the behavior of secondary family members. There has beeti a dramatic resurgence recently of discussion of the "underclass." The discussion is retniniscent of debates about poverty during the 1960s, when the notions of a "culture of poverty," particularly as popularized by Michael Harrington's Other America (1962), dominated both intellectual and policy thinking. The idea of an underclass seems inconsistent, however, with much of the research on the dynamics of poverty during the 1970s. That research, using new longitudinal data, seemed to show that the bulk of the poor were poor for only a few years. The research also showed that the poor were a very heterogeneous group, including a small minority of persistently poor. The persistence of poverty is of interest both for understanding the phenomenon and for developing policy. Claims about dependency and separate life styles among the poor rest on assumptions about the long-term nature of poverty. Questions The authors are both associate professors at the John F. Kennedy School of Government, Harvard University. [Manuscript received August 1985; accepted Septemher 1985.] 2 I THE JOURNAL OF HUMAN RESOURCES about the allocation of resources can better be answered when the characteristics of the poor are understood. To answer all these questions it is important to be able to describe the experience both of people who ever slip into poverty and of people who currently are poor. We shall show that the distinction between the ever-poor (or the newly poor) and the poor at a particular time is crucial in understanding poverty and in shedding light on the question of culture, dependency, and allocation of resources. The availability of longitudinal income and poverty data now spanning more than a decade makes possible comprehensive analyses of the durations, beginnings, and endings of spells of poverty. The preliminary analyses we report in this article lead us to conclude that the seemingly inconsistent findings on permanent and transitory poverty from the sixties and seventies can indeed be reconciled. Our primary finding is that although many people have very short spells of poverty, the few with very long spells account for the bulk of all poverty and represent the majority of the poor at any given time. We also report some analyses of events leading to the beginnings and endings of spells of poverty which help to explain the ways in which the poor slip into poverty and escape it. /. DESCRIBING DYNAMICS—PREVIOUS RESEARCH There are three primary approaches which have been followed in recent years to describe the dynamics of various types of behavior. These include statistical methods which model the level of some variable such as income, allowing for a complex lag or error structure to capture dynamics; methods using spell durations and exit probabilities; and finally tabulations of the frequency of the event over some fixed time frame. The bulk of the literature on the dynamics of poverty uses the first and third methods. We employ the second. We begin by briefly considering the advantages and disadvantages of each approach. Any model of income estimated with longitudinal data implicitly or explicitly provides a model of intertemporal dynamics. Typically the dynamics are subsumed in the error structure. In their classic paper Lillard and Willis (1978) model the level of eamings of a group of prime-age men and pay close attention to the error structure, allowing for both permanent and transitory components in the error. After estimating such a model it is possible to examine the frequency and duration of periods of poverty by asking what fraction of the population is likely to be below the poverty line and for how long, based on the estimated structure of eamings. This approach has appeal. First it largely mirrors the famous Friedman theoretical decomposition of permanent and transitory income. Moreover it deals explicitly with the problem that the poverty line is an arbitrarily defined standard, around which income can fluctuate randomly. Permanent income can be Bane and Ellwood I 3 estimated and the poor can be decomposed into the groups which are permanently and transitorily poor. And the expected durations in poverty can be inferred for any arbitrary poverty line. Both Levy (1977) and Gottschalk (1982) also emphasize the importance of the permanent/transitory decomposition, although they use different methodologies from those of Lillard and Willis. Although the Lillard and Willis approach has great appeal for ascertaining the income dynamics of prime-age males, it has shortcomings as a method for understanding the nature and dynamics of poverty for the entire population. It is exceptionally difficult to cope with the fact that "poverty" is a concept that applies to families—and that family membership changes. One can certainly speak of permanent and transitory components in eamings for prime age males, but how should one treat the income pattern that results when a family splits up and the former wife who was previously out of the labor force goes to work to help support her children? How should one characterize changes in income and economic status caused by the household formation choices of young people who leave well-to-do homes and are poor for a period while they make the complete transition to the labor market? In principle the income of each family member could be modeled individually, allowing for simultaneous influences from and to family structure, and allowing for life cycle changes. In fact such models are difficult to develop. Altematively, one could model not just personal income hut family income relative to the poverty line for each individual. When the membership of the individual's family changed, both family income and family needs would be adjusted to reflect a new situation. This income-to-needs ratio might be hypothesized to have a permanent and a transitory component just as in the case of individual eamings. Yet this approach also has weaknesses. Changes in the income-needs ratio caused by the departure or entrance of another family member, those caused by variations in earnings of the individual, and those caused by variation in other sources of income are all treated equivalently. While it may make sense to talk of a permanent component in earnings, it is far less clear that there is a permanent component of income-to-needs where family situations arc changing rapidly. The notion of a permanent income is easier to implement empirically for ablebodied prime-age males than it is for women, children, the elderly, or the disabled. Both Lillard and Willis and Gottschalk limit their analyses to prime-age males, and Levy did not actually implement a permanent income model. In 1981, however, males aged 22-64 made up only 15.5 percent of the officially defined poor (U.S. Bureau of the Census 1982). The experience of the rest of the poor may not be well described by the permanent income notion. Perhaps even more fundamentally, all deviations from permanent income tend to be treated as random and behaviorally equivalent in those models. Typically all "disturbances" in income lead to the same temporal path of income in the future. But all changes in family income are not likely to lead to the same sort of long-run dynamics. The worker who is poor because he was temporarily laid 4 I THE JOURNAL OF HUMAN RESOURCES off from his job is unlikely to have the same prospect of long-term poverty as one who lost his job when he became disabled. Quite often these disturbances are of great interest in their own right. Indeed if dynamics are being considered, the changes themselves really are the driving force. Presumably those interested in understanding poverty are interested in knowing what sorts of adverse events lead people into poverty, whether the duration of a poverty stay varies depending on how it began, and how (if ever) families escape poverty. When the events leading into and out of poverty are a source of considerable interest it seems strange and unfortunate to treat these as homogeneous and largely unobservable disturbance terms. Of course these could be modeled explicitly, but at great cost in complexity. Another approach has been adopted by Duncan (1984), Coe, Duncan, and Hill (1982), Coe (1978), Rainwater (1982), and others. They look at the proportion of the number of persons who are poor by some definition over a fixed time frame—typically eight or ten years. One can tabulate how many people were poor for, say, ten out often years, or five out often, or one out often. The approach is very simple to use—it need involve no more than simple tabulation. Changing family structures cause no problems. The unit of analysis is typically the individual and his or her poverty status at any time and the poverty status of his or her family at that time. For purposes of tabulation it does not matter if poverty status changes because family structure changes or because income of a family changes. The approach also has some of the appeal of methods based more explicitly on permanent income notions, because those who were transitorily poor will seemingly show up as people with very little poverty over the period, those who have very low permanent income will be poor for most of the period, and those whose incomes fluctuate back and forth across the poverty line will be the intermediate group. But in this case again, no attention is focused on the events which lead people into and out of poverty. It is very difficult to trace processes whereby persons may gradually or suddenly escape from poverty. But more importantly, this method can be misleading. Consider an extreme example. Suppose all poverty occurs in spells lasting exactly ten years. If we were to ask how many persons who were poor over a ten-year survey period remained poor the entire time, only those people who happened to begin their ten-year spell in the first year of the survey would be counted. Those who began spells in the year prior to or the year after the survey began would have nineyear episodes in the survey. Those who began nine years before or after the survey started would have one-year episodes. Thus even though all spells lasted exactly ten years, because completed spells cannot be observed in the survey, one will find that roughly equal numbers of people were poor for one survey year, two survey years, and so forth. Obviously any conclusion that only a small number of poor persons remained in poverty for a long time would be quite misleading because of the censored spells. Bane and Ellwood I 5 We propose instead to model spells of poverty. While we acknowledge and address the problems caused by the crossing of an artificial threshold', we think a spell approach provides a simple and compact way to understand the dynamics of poverty. One advantage of using spells is that information can be summarized in a comprehensible manner. Indeed Lillard and Willis and many others impose the artificial poverty line on their structurally estimated income dynamics, and report durations of spells and the probability of moving from poverty. All we propose is to examine those issues directly. Other researchers have looked at movements into and out of poverty, both to estimate the amount of movement and to examine the characteristics of those who do and do not move. Hill (1981) and Levy (1977) have calculated exit probabilities for those who enter and exit from poverty. Boskin and Nold (1975), Hutchins (1981), Plotnick (1983), and Wiseman (1976) have explored movements on and off of welfare. We follow this line of research, and extend it in three ways: by looking at a variety of distributions, by allowing for duration-dependent exit probabilities, and by identifying beginning and ending events. None of the research has reported the full set of distributions that are important for understanding the dynamics of poverty: completed spell distributions for people beginning a spell of poverty and for those poor at a given time; and the uncompleted spell distribution for people poor at a given time. The importance of the distinctions between these distributions has been emphasized by several scholars. Kaitz (1970), Salant (1977), Clark and Summers (1979), and Akerlof and Main (1982), for example, have pointed out that while most people who become unemployed are in that state for only a short period of time, the bulk of unemployment is long term. As we explain in our next section, we believe these distinctions are important for understanding poverty as well. Moreover, little of the research on flows into and out of poverty has looked explicitly at differences in exit probabilities by time in poverty. Levy (1977) reported no differences in exit probabilities over time, but his analysis included all the people who were poor in his beginning year, regardless of how long they had been poor before being observed. Much of the research on movements on and off of welfare, for example Plotnick's (1983) event history analysis, as well as Hutchins's (1981) and Wiseman's (1976) assumes constant exit probabilities. It's worth noting that a permanent income model predicts that exit probabilities will decline with duration because those who are temporarily poor will leave early, leaving behind those who will never exit. Finally, there has been little research on the events associated with movements into and out of poverty. Levy (1977) and Gottschalk (1982), for example, dismiss the importance of family composition changes, noting that most people do not 1 It is worth noting that nearly all logit or probit models use an artificial threshold to model behavior. 6 I THE JOURNAL OF HUMAN RESOURCES change their family composition very often and that most income changes are not associated with family compositions changes. While this is true generally, we do not believe it is so true for the poverty population, especially looking over rather long periods of time, and thus we deal with it here as an important topic for empirical investigation. A key innovation of this study is our characterization of the events which lead to the beginnings and endings of spells of poverty.