This report was published by the former Department of Families, Community Services (FaCS).
- Executive summary
- 1 Introduction
- 2 The duration of unemployment benefit spells in theory
- 3 Labour market experience of Aborigines and Torres Strait Islanders
- 4 The model
- 5 The data
- 6 Patterns of utilisation of unemployment benefits
- 7 Hazard function estimation
- 8 Aborigines and Torres Strait Islanders
- 9 Conclusions
- Appendix: Variable means for spells data used in hazard estimation
This paper reports the results from an analysis of the duration of spells of unemployment- related income support for all persons and for persons of Aboriginal and Torres Strait Islander origin. The theoretical framework envisages individuals choosing whether to remain on income support depending on the value of alternative opportunities. In this framework, the duration of an unemployment spell depends mainly on how job opportunities evolve over time. The empirical model based on this framework is a regression model in which the exit rate from unemployment is a function of individual characteristics. This model is estimated using data from the Department of Community and Family Services Longitudinal Data Set.
The research has found that the exit rate declines over the duration of an unemployment spell but that most individual characteristics have a relatively small effect. As there are few indicators of long expected duration apart from having already been unemployed for a long time, policy intervention is best directed to the long-term unemployed. Using the data on all Aborigines and Torres Strait Islanders, the research has found that their exit rate is lower than for other persons. However, in contrast to most other indicators of labour market performance, the difference between Aborigines and Torres Strait Islanders and other persons in terms of exit rate must be regarded as quite small. Thus, the much higher unemployment rate of Aborigines and Torres Strait Islanders can be attributed to the much larger proportion of their number that becomes unemployed.
This paper reports the results from an analysis of the duration of spells of receipt of unemployment-related income support in Australia. This analysis has been made possible by the development of a longitudinal data set of fortnightly records for income support clients by the Department of Family and Community Services (FaCS). The data set is unique in that it is one of the few in which there are adequate numbers of Aboriginal and Torres Strait Islander persons to analyse whether their pattern of spells on income support differs from the patterns of the rest of the population. A central theme of the paper is to investigate whether differences in the experiences of Aboriginal and Torres Strait Islander peoples can help us understand their relative exclusion from mainstream employment.
Although the effect of the amount of payment unemployed persons receive has been the focus of most economic analysis, the data used here do not permit this effect to be identified. Rather, the focus is on the frequency and duration of spells and the impact of a limited range of personal characteristics such as age, sex, marital status, number of children and country of birth on the duration of spells experienced.
The following two sections contain an overview of the literature relating to the theory and evidence of the duration of unemployment spells and a discussion of the labour market situation facing Indigenous Australians. The model and data used in the analysis are then described. Descriptive analyses and results from the estimation of hazard or exit rate models for spells of unemployment receipt are presented in turn for the representative sample and for the full file of Aboriginal and Torres Strait Islander client records. The major findings and avenues for future work are summarised in the concluding section.
Whether or not a person claims unemployment benefits depends first on whether they are eligible for benefits and, if they are, whether claiming benefits is the best alternative for them. One of the principal eligibility criteria is, of course, being unemployed. There is a large body of literature examining unemployment experience and other welfare payment experience independently of each other, as well as the common issue of the utilisation of unemployment benefits. These are largely based on static models of participation, but attention on entry and exit rates is increasing with the availability of longitudinal data.
The incidence of unemployment among different subsections of the labour force has been well documented. Miller & Le (1999) provide the most comprehensive recent analysis of the characteristics of the unemployed in Australia. Those with low levels of educational attainment, English language difficulties, disabilities, a previous history of unemployment and those from the younger and older age groups are all disproportionately represented in unemployment. Structural changes, such as in employment by skill category, occupation, industry and spatial region can also contribute substantially to unemployment for those from the ‘declining’ sectors. Unemployment is extremely high among Aboriginal and Torres Strait Islander peoples, as discussed further below. The dynamics of unemployment have also been extensively studied, although not as thoroughly in Australia as in other developed countries. Factors that increase the likelihood of being in unemployment typically have the effect of increasing spell duration. There is general consensus that unemployment displays ‘negative duration dependence’—a person’s chance of leaving unemployment declines the longer she has been unemployed(see, for example, Stromback & Dockery 2000).
Studies of welfare utilisation have been motivated by the so-called ‘welfare explosion’ witnessed in many countries in recent decades and by the associated issues of welfare dependency and disincentive effects. When a person becomes unemployed, their further time in unemployment is often modelled through ‘search’ models. Critical parameters in these models are the wages they might earn, the rate at which they receive offers, the cost to them of continuing to search for employment, and their reservation wage (the lowest acceptable wage). The specification is usually such that higher unemployment benefits reduce the cost of continued search and/or increase the reservation wage, leading to longer unemployment spells.
Other models follow a standard discrete choice model where the individual maximises their benefit in choosing between welfare payments and paid work. The disincentive effect of welfare payments has been borne out empirically in many studies, most convincingly in countries in which benefits (or insurance) are not open-ended and exit rates from unemployment increase sharply as the expiration of benefit eligibility approaches (see Meyer 1990, Katz & Meyer 1990, Moffit 1992, Fortin & Lacroix 1997, Holmlund 1997).
The international findings are not expected to translate directly to the Australian situation due to differences in unemployment income support. Most developed countries have an unemployment insurance system in addition to welfare benefits. Unemployment insurance usually provides earnings-related income support for a limited period, while government-provided welfare supplements the unemployment insurance and provides for those who are not entitled to insurance benefits. With a finite-duration unemployment insurance system, the exit rate from unemployment is strongly influenced by the time limit. Thus, the exit rate tends to rise sharply as this limit is approached.
Unemployment insurance can also affect the duration of welfare spells by strengthening the incentive to return to work to re-establish eligibility for unemployment insurance. Among the Australian studies, Webster (1999) finds the replacement ratio to be one of the important ‘shift variables’ for the Australian Beveridge curve, with a 10 per cent increase in ratio of unemployment benefits to minimum award wages increasing the level of unemployment for a given level of vacancies by 5 per cent. 1 Stromback & Dockery (2000) find that the receipt of income support markedly reduces the exit rate from unemployment (increases the expected duration of spells).
The degree of exclusion of Australia’s Indigenous population from mainstream economic and labour market activity is striking. This exclusion had force in law until the 1967 national referendum granted full citizenship rights for Indigenous Australians. Only from the following 1971 Census of Population and Housing were Indigenous people included in the data collection (Smith 1995, p. 29). Subsequently, a more thorough statistical picture of the contrast between the labour market fortunes of Indigenous and non-Indigenous Australians has emerged. This evidence indicates that gains in equality in legal and administrative status for Aborigines have not meant equality in labour market outcomes.
Aboriginal and Torres Strait Islander peoples comprise just 2 per cent of the Australian population. Hence, none of the mainstream labour force or population surveys based on representative sampling offer adequate sample sizes to analyse the labour market experience of this group. Thus, most of the analysis of Indigenous population has relied on Census data. Table 1 compares the labour force status of Indigenous and non-Indigenous Australians in Census data. It can be seen that, according to the standard labour market constructs, Indigenous Australians have faced unemployment rates varying from two and a half to four times greater than for other Australians over the past 30 years, despite a far lower participation rate.
Note: All calculations exclude those who did not indicate their labour force state. Sources: Committee of Review of Aboriginal Employment and Training Programs 1985, Table 3.5 p. 46; Taylor & Hunter 1998, Table 2, p. 13.
Even Census data contain serious technical and conceptual deficiencies for the purposes of describing Indigenous labour market experience. The true unemployment rate for Indigenous persons is likely to be higher in absolute terms and relative to the wider population due to non-response bias. The classification of persons as Aboriginal or Torres Strait Islander relies on self-identification. This has always presented limitations to the use of administrative data from government departments for wider statistical purposes, particularly as the rate of self- identification has been sensitive to administrative rules. The Census counts show the Indigenous population growing 33 per cent from 265 371 in 1991 to 352 970 in 1996 (Taylor & Hunter 1998, p. 10), clearly a result of data collection or identification effects. The Indigenous population is actually estimated to be expanding at around 2 per cent a year, twice the rate of the wider population.
The conceptual problems relate to the appropriateness of the standard labour force status constructs for Aboriginal and Torres Strait Islander persons. To the non-Indigenous population, favourable labour market status is largely synonymous with paid employment in the market sector. Many Aboriginal and Torres Strait Islanders live in remote communities with very limited integration with the mainstream economy, and thus opportunities for what would normally be considered market work are largely non-existent. Further, traditional Indigenous roles relating to culture or a hunting and gathering existence, which are clearly legitimate economic roles in the context of Indigenous societies, do not fit well into our perceptions of a labour market:
…non-Aboriginal ideas of what constitutes ‘real’ work, ‘real employment’ and the ‘real’ labour force have informed official assessments of Aboriginal labour market activities. The validity of associated labour force indicators needs to be critically evaluated then, especially in respect to their conceptual limitations (Smith 1995, p. 35).
This has created a difficult challenge for policy makers in addressing the labour market outcomes of Indigenous Australians. Improving outcomes involves an inevitable tension between the need for integration into mainstream economic activities to achieve the goal of equality in economic opportunity, and the goals of self-determination and preservation of traditional society and culture. This tension is particularly relevant to the formulation of labour market programs designed to improve the labour market outcomes of Aborigines and Torres Strait Islanders. In 1985, the Committee of Review of Aboriginal Employment and Training Programs questioned whether ‘…the assumption implicit in the programs that work for wages and salaries is the most appropriate and practicable basis for earning a livelihood’ was valid of all Aboriginal people. The Committee recommended that assistance measures respect the desire of Indigenous Australians for varying mixes between their own traditional lifestyle and appropriate components of the wider market-based economy (pp. 5–6).
The Community Development Employment Project (CDEP) is a scheme that attempts to chart such a delicate balance. It arose originally in response to requests from communities for bulk payment of their unemployment benefit entitlements. The CDEP allows remote communities to administer the unemployment benefits, which would otherwise have been payable to persons within the community, as wages for designated activities. The community determines the work to be undertaken, though guidelines provide that the work should be meaningful to the community and contribute toward self-sufficiency. The program thus promotes self-determination as well as partly ameliorating the potential for the negative effects of welfare dependency.
The CDEP has grown to be the largest employment program for Aboriginal and Torres Strait Islander peoples, with an estimated 33 600 individuals ‘employed’ through the 280 community organisations in 1999 (DEWRSB 1999).
Returning to the problem of applying standard constructs to measuring Indigenous labour market status, whether CDEP participants are treated as employed, on labour market programs or on welfare thus has a significant impact on any statistics used to describe the aggregate picture. According to the 1994 National Aboriginal and Torres Strait Islander Survey (NATSIS), the CDEP accounted for over one-quarter of Indigenous employment (Taylor & Hunter 1997, p. 296).
Data from recent censuses indicate that Indigenous employment has been growing at a far higher rate than overall employment. Much of this growth is clearly due to expanded CDEP and other program employment, and to positive employment policies within the public sector. According to Altman & Taylor (1995), Indigenous employment in the private sector increased between 13 to 22 per cent between 1986 and 1991, depending on the measure used. This compares to a 37 per cent increase in private and public sector employment as measured from Census data, indicating a growing reliance on public sector employment among Indigenous workers. However, the growth in private Indigenous employment has been much higher than for the rest of the population (8.4 per cent), suggesting some real gains in relative employment status (Altman & Taylor 1995, p. 78).
Notwithstanding the limited statistics and the problems of its interpretation, it is abundantly clear that Indigenous Australians who wish to participate in the mainstream labour market face severe disadvantages. Beyond the static aggregate measures that estimate the extent of this disadvantage, the data problems outlined have precluded further econometric analysis of the nature of this disadvantage. The availability of the FaCS Longitudinal Data Set (LDS) offers a much-needed opportunity to undertake such an analysis.
The capacity to distinguish between Aborigines and Torres Strait Islanders living in metropolitan and rural areas also permits us to overcome two major hurdles in applying standard labour market constructs analysis to data on Indigenous peoples. First, the statistical impact of the CDEP is removed. According to the 1994 NATSIS, 80 per cent of all Indigenous employment in rural areas was in the public sector, particularly community organisations (Taylor & Hunter 1997, p. 299). Second, those individuals living in metropolitan areas can be assumed to ‘have chosen or have felt compelled to accept’ mainstream economic activity as the basis for their livelihood to a greater extent than those who remain in isolated communities (Committee of Review of Aboriginal Employment and Training Programs 1985, p. 6). Thus, mainstream indicators of labour market status can be accepted as being more relevant.
In common with most of the literature, it is assumed that the choice to claim or not to claim benefits can be modelled as a discrete choice problem in continuous time. Several approaches to the derivation of empirical models can be found in the literature. What follows should be seen as example of how an empirical model could be rationalised using a very simple framework.
Consider an individual who is entitled to a base level support of b per unit of time if she has no other income. For the sake of notational simplicity, the time subscript is suppressed. If she has an unearned income, a, the benefit is reduced at a rate of q for every dollar. Thus, working zero hours she receives a benefit g = b-θa <b. If she works positive hours, h>0, at an hourly rate of w, the benefit is subject to the same claw-back rate. Thus, the benefit received is g = b-θ(a+wh). Ignoring taxation, her net income, y, is given by y = b+(1-θ)(a+wh)if b>θwh and y = (a+wh) if θ(a+wh)³b, that is, if her income exceeds the level at which the benefit cuts out.
The individual’s kinked budget constraint is illustrated in Figure 1. The slope of the segment AB reflects the higher effective marginal rate of tax associated with the benefit reducing with earned income. The point where the benefit cuts out is labelled B. This is at best a highly stylised description of the much more complex underlying reality. The budget constraint along AB may contain several kinks and discontinuities that arise from the precise rules governing the determination of benefits.
In the case of married persons, the situation is more complex. In general, each person’s decision depends on the decision taken by the spouse. The simplest approach to this problem takes the labour supply decision of a spouse as given and assumes that the income of the spouse is equivalent to own unearned income. This approach is also broadly consistent with the Australian income-testing regime. Although this regime makes a distinction between own and spouse income, the two types of income have a similar effect on reducing benefit entitlements.
Figure 1: The choice between claiming or not claiming unemployment benefit
Conditional on new information received the individual’s decision between claiming g >0 and not claiming depends on a comparison of the utilities associated with the two alternatives.
The indirect utility from working can be represented by the function:
(1) v = v(θ, a, h, x) + ε v
This function has two parts. In the deterministic part, the unobserved wage rate is replaced by its determinants x, a vector of individual characteristics. The random term reflects the arrival of new information and it is this stochastic component of the utility function that drives the model.
Conditional on remaining on income support, the individual’s direct utility function can be represented as:
(2) u = u(b, θ, r, a, h, x) + εu
Let μd = v(.)-u(.) denote the difference between the deterministic components of the utility function and ms = ε u - εvthe difference between the stochastic components. The arrival of new information is modelled as a Poisson process with a constant arrival rate n. On the assumption that the error terms ε vand ε u are temporally independent, the instantaneous rate of exit from being a benefit recipient is given by:
(3) λ(z) = nProb[μs≤ μd(z)]= nF[μd(z)]
where z = (θ,b,a,x).
It is straightforward to show that the level of unearned income (a) has a negative effect on the exit rate from income support, providing that leisure is a normal good. The exit rate is also declining in the level of benefit (b) and in the claw-back rate. The predicted effect of personal characteristics (x) depends on their effect on the wage rate. Since the wage rate is increasing in age, at least up to a point, the expectation is that the exit rate would increase with increasing age. Similarly, other characteristics that are associated with a higher wage rate, for example, being married, having children and being native rather than foreign born, would also be expected to increase the exit rate. However, these predictions depend critically on the assumed constancy of the Poisson parameter n. In general, n would depend on the same personal characteristics as the wage rate. Considering this largely precludes any predictions about the effect of personal characteristics. For example, while the wage rate is increasing in age, the rate at which new information arrives is most likely declining in age, making the net effect ambiguous.
For the purpose of empirical estimation, the hazard function is assumed to be of the proportional form:
(4) λ(z) = λ(t) exp(z’ β)
where λ0(t) is the baseline hazard (that is, the hazard when z = 0) and β is a vector of parameters to be estimated. In general, the elements of z might vary over the duration of a spell. However, in the empirical work reported here the initial values of the z variables are used unless otherwise noted.
The empirical work uses a flexible (non-parametric) hazard function; the Cox proportional hazard model as implemented in the software used (SAS). Specific parametric forms are subject to a number of problems. First, the theoretical basis for a particular form is rarely clear. Second, the empirical hazard is often irregular and may not be well approximated by a simple parametric form. Thirdly, misspecification of the baseline hazard can lead to inconsistent estimates of β. Fourthly, the estimates are sensitive to unobserved heterogeneity if this is ignored or incorrectly specified. While the Cox model overcomes these problems, the proportionally assumption may be an invalid restriction. Our use of SAS for estimation entails further limitations. First, it was not possible to allow for unobserved heterogeneity, that persons with the same values of z may have a different hazard. Second, because SAS does not recover the underlying base hazard from the estimated parameters and the data, a useful piece of information is missing from the results.
The LDS is a 1 per cent representative sample of the fortnightly records of income support clients of FaCS. The sample was selected from those in receipt of income support at a point in time, and then expanded by moving both forwards and backwards in time through the fortnightly records of FaCS clients. For each new fortnight added to the panel (moving both forwards and backwards in time), all newly appearing income support clients had an equal (1 per cent) chance of being selected. Length bias (oversampling those with long spells) is thus present for those selected at the initial time point, but this effect is diminished as the length of the panel increases. Once selected into the sample, clients’ records over the full panel are known, including ‘null’ records for fortnights when they were not recipients. Persons’ activity and circumstances while not in receipt of income security are, however, not known. This means, for example, that the destinations of persons who exit spells of income support receipt are not known.
At the time this analysis was undertaken, the panel spanned the 105 fortnights from 23 June 1995 to 18 June 1999, almost four years. The data cover persons in receipt of most types of income support payments, but the focus of this paper is on spells of unemployment- related benefits. The two major types of unemployment benefit are known as Newstart Allowance and its counterpart for 18–21 year olds, Job Seeker Allowance. Also included are several variants or predecessors of these payment types: Youth Training Allowance, Newstart Mature Age Allowance and Mature Age Allowance. In what follows, all the above types of benefits are simply referred to as unemployment benefits (UB).2
Spell data are derived from the series of cross-section records. There are two potential means of defining continuous spells of benefit receipt for individuals. One is simply as a series of consecutive fortnightly payment dates for the one individual. The other is through the ‘duration on current payment’ variable in the data set. In either case, a spell is deemed to end in the fortnight after which there is no record for an individual for the following two fortnights.
Relying on the duration of current payment variable, the start of the spell would be taken to be the date corresponding to the end date minus the duration of current payment. This method has the advantage that the start date of all spells observed to take place during the observation period, and hence the duration, would be measured. The previous discussions between FaCS and other researchers using this data set concluded that the duration of current payment should not be used in defining and measuring spells for two main reasons. First, the duration of current payment is not always reset to zero when persons re-apply for UB after a short absence (for example, a few weeks of work). In such cases, their case might be treated as a re-application rather than as a new application, for administrative convenience. Second, in many instances the duration on current benefit variable indicated a spell of shorter length than the number of continuous payments. This should not occur and FaCS has confirmed problems with the integrity of the duration on current payment variable.
The alternative method is to take as the start date the first fortnight there is a record for the individual in question. By using this method, the start date of spells that began before the observation period cannot be determined. The break of up to two fortnights treats short interruptions as part of a current spell in line with the administrative practice. This practice also suggests that a variation of this method should be used to define the end of the spell. This variation would take a spell to end if either there are no records for the individual for at least two fortnights or if there are two consecutive fortnights for which the relevant benefit variables are zero. However, this variation could not be implemented, because correctly measured benefit variables were not available at the time this part of the analysis was undertaken.
The data set derived by the above method resulted in 45 147 spells of unemployment benefit receipt by 26 747 individuals. Table 2 shows that around 57 per cent of these individuals experienced one spell only in the period, but that these accounted for 34 per cent of all spells recorded. Thus, about two-thirds of spells in the sample are experienced by repeat recipients, suggesting a significant degree of churning into and out of UB receipt.3 The maximum number of spells experienced by any one individual was nine (six people). In the analysis that follows, no distinction is made between first and repeat spells.
|No. of spells||No. of individuals experiencing that number of spells||Total no. of spells|
|1||15 385||57.5||15 385||34.1|
|3||2 880||10.8||8 640||19.1|
|4||1 128||4.2||4 512||10.0|
|Total||26 747||100.0||45 147||100.0|
Around 19 per cent of the spells (8 378) were in existence at the beginning of the observation period. Of all spells that began before the beginning of the observation period, only those that are sufficiently long will be observed. These ‘old’ spells were omitted from the analysis. Effectively, this means that the sample comprises all spells that began during the observation period June 1995 to July 1999. In another paper using this data set, Barret (2000) also excludes ‘old’ spells.4A further 16 per cent of all spells (7 364) were right-censored—that is, ongoing at the end of the observation period. For spells that began and ended within the observation period, the mean duration was 14.2 fortnights, just over half a year. Clearly, this represents an underestimate of the actual mean duration because the excluded spells (the continuing spells) are more likely to be relatively long ones. For continuing spells, the mean elapsed duration was 41.6 fortnights (just over one and a half years). Including only newly commencing spells, the mean and quartiles from non-parametric estimation of the survival function are shown in Table 3. This takes into account the right-censoring of the continuing spells. The estimated mean duration is around 23 fortnights. The estimated survival functions are similar for UB spells for males and females, with males on average having slightly longer average spells.Likewise, those with a partner have longer spells than singles, as have older persons and those born in a non-English-speaking country. These estimates were obtained using the lifetable or actuarial method for estimating the hazard and survival function.
|Mean||25% quantile||50% quantile||75% quantile|
|Country of birth: Australia||22.2||5||11||27|
As discussed earlier, a deficiency with the data is the lack of information on the destination of exits. It is possible, however, to identify those who moved onto other payment types. To allow some time for transition, we investigated whether individuals were receiving another payment type two fortnights after the end of a UB spell. Spells with end dates in the last or second last fortnight of the observation period are censored in the sense that we cannot determine whether or not the individual moved on to a different income support payment. For the other completed spells, 12.4 per cent terminated because the person moved onto another form of income support payment. The remainder, presumably, gained work, left the labour market or ceased to claim unemployment benefit for other reasons. The most common categories of income support to move into were Disability Support Pension (3.2 per cent), Youth Allowance (2.4 per cent), Age Pension (1.4 per cent), Sole Parent Pension (1.1 per cent) and various categories of Parenting Allowance (1.1 per cent). The duration of spells that did end in a transition to another form of income support are considerably longer, with mean duration of 23.6 fortnights, compared to 14.8 for other completed spells.5
Figure 2 shows the non-parametric estimates of the hazard function up to 52 fortnights (two years). Comparing males and females reveals that the shape of the hazard function is the same but that females have a higher hazard than males at most duration times. In general, however, the differences are quite small. The hazard initially rises from just under 5 per cent to a peak in fortnight 4 or 5. Around 7.5 per cent of those still receiving unemployment benefits at this point are no longer on UB in the following fortnight. From then on, the hazard declines monotonically with duration. The graphs of the hazard function by age, country and birth, and marital status all display the same unimodal hazard declining after a peak at four or five fortnights.
Figure 2: Non-parametric estimates of the hazard function for UB spells
(a) by gender
(b) by age group
(c) by marital status
(d) by country of birth
The effects of covariates on the hazard function are estimated using the Cox proportional hazards model. This method allows for an arbitrary baseline hazard, but the disadvantage is that the baseline hazard is not estimated. The data permit only a relatively small number of covariates to be included in the estimation. Hence, there are omitted variables that we know from previous research to be important in determining the rate of exit from spells of unemployment. Of particular importance is educational attainment. Though this is available as a variable in the data set, it is not known or not specified in over 90 per cent of cases.
The personal characteristics that determine the wage rate include age, sex, country of birth, race, marital status, children, housing circumstances and unearned income. Also included are additional variables that may proxy the effect of unobservable characteristics on the wage rate. This set of variables includes whether and from whom a person rents the place where they live, partner’s activity (receiving unemployment benefit or not) and previous labour market history (previous receipt of income support and repeat spell). Some of these variables could also be expected to affect the arrival of new information and thus the chance of finding a job.
The identification of the benefit effect depends on exogenous variations in the benefit level over time or between individuals. The benefit level did not change over the period in question apart from cost of living adjustments. The variation between individuals is entirely related to differences in individual circumstances, of which marital status, number and age of children and housing arrangements exert the largest influence. For that reason it is uncertain whether the effect of the benefit level can be identified—separated from the influence of these other characteristics. It should be noted that most studies of duration of welfare spells have nevertheless included a benefit level variable. We follow this approach by including the potential benefit rate measured by the sum of maximum basic rate, maximum rate assistance rate, maximum pharmaceutical allowance rate, remote area allowance and the add-on-amount. However, this variable is closely related to the determinants of the benefit level—age, marital status, number and age of children and the amount of rent paid.
Unearned income is directly measured in the data set. In addition to the direct measurement, the amount of unearned income is also represented by a set of home ownership variables. Home ownership constitutes ownership of an asset, which in most case yields an imputed income that is not included in the measured unearned income.
As discussed, partner’s income could affect the choice between claiming and not claiming, in the same way as unearned income could affect it. For each individual with a partner, there is a matching record of that person’s partner. That record includes the partners’ earned and unearned income. If the partner is also in receipt of income support, the amount of that support is also recorded.
The variables included in the data set but not used in the analysis include activity (job search or other), type of breach of the conditions for income support and the penalty imposed, if any. The principal reasons for omitting these are that it would be unreasonable to assume that these variables are constant during a spell and the proportional hazard framework is not well suited to analysis of their effect.
Several of the covariates used in the analysis vary over the duration of the spell. In this paper, the estimation does not take this into account. Instead, the initial value of a variable is taken to represent the values of the variable for the duration of a spell. This limitation of the analysis is a consequence of the software used (SAS). Consequently, several activity status variables available in the data set, such as participation in training or employment assistance programs, were not included. These variables are highly likely to be related to the duration of the spell, and hence any identified effect would be spurious unless time-varying variables are constructed.
The proportional hazard model was estimated from the 36 769 spells that began during the observation period. The results are given in Table 4 separately for persons with and without partners. Positive coefficients indicate that the hazard function or exit rate is increasing with that variable, and the expected duration of spells is decreasing. A negative coefficient implies a falling hazard and longer expected duration of the UB spells. The magnitude of the coefficients gives the approximate per cent effect of a unit change in the variable. Taking age as an example, if the hazard for a person in the omitted age group (25–44) is 5 per cent per fortnight, the hazard for a person aged 15–19 with a coefficient of 0.2463 is exp(ln 0.05+0.2463)*100=6.4 per cent – that is, 28 per cent higher.
|(1) No benefit variables||(2) With benefit variables|
|15 to 19||0.2463 ***||0.2134 ***|
|20 to 24||0.1325 ***||0.1297 ***|
|25 to 44||—||—|
|45 to 54||-0.3187 ***||-0.3191 ***|
|55 to 59||-0.4813 ***||-0.4804 ***|
|60 to 64||-0.6132 ***||-0.6388 ***|
|Main Eng-spkg country||0.1154 ***||0.1184 ***|
|Female||0.0423 ***||0.0418 ***|
|Indigenous||-0.1824 ***||-0.1750 ***|
|Home Owner||-0.1223 ***||-0.0835 **|
|Other owner||-0.4659 **||-0.4143 **|
|Government||-0.4567 ***||-0.4030 ***|
|Site/other fees||-0.1856 **||-0.1284|
|No rent paid||-0.0952 ***||-0.0503 **|
|Sharer||-0.4399 ***||-0.4399 ***|
|Has child aged:|
|0 to 6||0.1254||0.1040|
|7 to 12||0.1331||0.1032|
|13 to 15||-0.2923||-0.2794|
|over 15||-0.4769 ***||-0.5101 ***|
|no depdt. children||—||—|
|Partner receives UB||—||—|
|Benefit ($ eligible)||—||-0.0006 ***|
|Benefit squared||—||2.9x10-7 **|
|Unearned income||0.0013 ***||0.0012 ***|
|Rent paid||-0.0012 ***||-0.0012 ***|
|(Rent paid)2||2.0x10-6 ***||1.8x10-6 ***|
|Previous IS receipt||-0.1084 ***||-0.1388 ***|
|Repeat spell||-0.1034 ***||-0.1340 ***|
|N (spells)||27 027||27 027|
|Censored spells||4 525||4 525|
|Model χ2 (-2 Log L)||1 650 ***||1 650 ***|
|Degrees of freedom||30||32|
***, ** and * denote significance at the 1, 5 and 10 per cent levels, respectively.
—: denotes ‘omitted’
|(1) No benefit or partner income variables||2) With benefit variables||(3) With benefit and partner income variables|
|15 to 19||0.2823 ***||0.2980 ***||0.3018 ***|
|20 to 24||0.1346 ***||0.1360 ***||0.1306 ***|
|25 to 44||—||—||—|
|45 ot 54||-0.1961 ***||-0.1992 ***||-0.2002 ***|
|55 to 59||-0.6044 ***||-0.6076 ***||-0.6022 ***|
|60 to 64||-0.6693 ***||-0.6827 ***||-0.6739 ***|
|Main Eng-spkg country||0.1105 ***||0.1096 ***||0.1157 ***|
|Non-Eng-spkg country||-0.1785 ***||-0.1773 ***||-0.1667 ***|
|Indigenous||-0.2692 ***||-0.2705 ***||-0.2578 ***|
|Purchaser||0.2209 ***||0.2386 ***||0.2316 ***|
|Lives elsewhere||0.3972 ***||0.4047 ***||0.4101 ***|
|Government||-0.2555 ***||-0.2283 ***||-0.2210 ***|
|No rent paid||-0.0816||-0.0460||-0.0435|
|Sharer||-0.9168 ***||-0.9184 ***||-0.9184 ***|
|Has child aged:|
|0 to 6||-0.0657 **||-0.0673 **||-0.0539 *|
|7 to 12||-0.0354||-0.0374||-0.0310|
|13 to 15||-0.0638 *||-0.0664 *||-0.0681 *|
|no depdt. children||—||—||—|
|Partner receives UB||-0.0661 **||-0.0634 *||-0.0473|
|Benefit ($ eligible)||—||-0.0005 ***||-0.0004 **|
|Partner income||—||—||0.0003 ***|
|(Partner inc.)2||—||—||-1.2x10-7 ***|
|Rent paid||-0.0004 ***||-0.0004 ***||-0.0004 ***|
|(Rent paid)2||9.3x10-7 ***||9.1x10-7 ***||9.4x10-7 ***|
|Previous IS receipt||-0.0058||-0.0173||-0.0196|
|Repeat spell||-0.1099 ***||-0.1247 ***||-0.1185 ***|
|N (spells)||9 742||9 742||9 741|
|Censored spells||2 071||2 071||2 071|
|Model χ2 (-2 Log L)||481 ***||496 ***||539 ***|
|Degrees of freedom||31||33||35|
***, ** and * denote significance at the 1, 5 and 10 per cent levels, respectively.
—: denotes ‘omitted’
Considering first the results for single persons in Table 4a, the hazard for young people is higher than the omitted prime age group (25–44). Persons in the 45–54 and 55–59 and 60–64 groups are found to have a lower hazard. In fact, the exit rate decreases linearly with age across these ranges, rather than the ‘U’ shaped relationship often observed. Persons born overseas in a main English-speaking country have a higher exit rate than their Australian-born counterparts. However, there is no difference between Australian born and those born in a non-English speaking country. The male-female difference in the hazard is quite small, with females having a slightly higher exit rate than males. A larger difference is found between the Indigenous and non-Indigenous persons. This is taken up in detail in a following section. Several of the home ownership and rental status indicators have a large and significant effect on the hazard. Sharer status in particular is associated with a much lower hazard than persons in other living arrangements.
The continuous variables are included as second order polynomials to allow for any non- linearities. The higher the amount of rent paid rent the lower is the hazard, but with a positive quadratic term this effect becomes smaller at higher levels of rent. In this and most other cases, however, the coefficient on the quadratic term is small so the effect is approximately linear in the relevant range.
In theory, the exit rate should be increasing in unearned income on the grounds that leisure is a normal good. The higher the unearned income the more leisure a person can afford and the more likely it is that they will choose to work few hours and claim UB. The estimated coefficient on unearned income, however, is positive, implying that the exit rate is increasing in unearned income. To reconcile the theoretical predictions with the estimated effect, a possible explanation is that the unearned income variable also captures the effect of other omitted variables that have a positive effect on the exit rate and, when combined, they dominate the predicted negative effect. Including a measure of the potential level of benefits as in column 2 of the table reveals that the benefit level reduces the hazard but leaves the effect of other individual characteristics unchanged. The point estimates of the effect of the benefit amount imply that every $100 reduces the hazard by about 6 per cent.
Comparison of the estimates pertaining to persons with partners to single persons reveals both similarities and differences in the impact of individual characteristics. The age effect is similar for both groups but the effect of birthplace and gender differ. There are more marked differences in the estimated coefficients on the variables representing type of home ownership and rental status. For both groups, however, living in government rented accommodation is associated with a much lower hazard. The effect of children differs even more between the two groups. For single persons, having a child over 15 is associated with a significantly lower hazard. In the case of persons with a partner, having a child under the age of six is associated with a significantly lower hazard.
As previously stated, home ownership represents the ownership of an asset that in most cases yields only imputed rather than actual income. Thus, home ownership would be predicted to have the same effect as unearned income – those who own their home can afford more leisure and are therefore more likely to choose few hours and claim UB. In the case of single persons, some of the estimated coefficients on the home ownership variables correspond to this prediction. It is difficult to explain, however, why the ‘Other owner’ category should have the largest negative effect on the exit rate. This category includes several types on joint ownership that one would expect to be associated with lower asset values than the basic ‘Home owner’ category. In the case of persons with a partner, home ownership is associated with higher exit rate. The same result was obtained for Indigenous persons with a partner, as will be shown in the next section. As before, the most likely explanation is omitted variables. Home ownership is associated with omitted and possibly unobservable variables that tend to increase the exit rate. Evidently, these effects dominate and result in a positive coefficient on the home ownership variables. That single persons are different would then have to be attributed to home ownership being much less common among single persons. Single persons are much younger and many of those who do not yet own their own home will do so later in life. Thus, many of those not owning a home at an early age do not possess the unfavorable characteristics that older non-home owners do.
Living in government rental accommodation is strongly associated with lower individual and neighbourhood socioeconomic status. The effect of this is highly significant and is to prolong the duration of spell on benefit. Living in shared accommodation has an even stronger effect of the same sign, and is particularly strong for persons with partners. Across all renters, the duration of spells increases with the amount of rent paid.
If a person’s partner is also in receipt of unemployment benefits in their own right, their rate of exit from unemployment is lower, suggesting a concentration of unemployment within family units. For single persons only, the duration of spells falls with the average fortnightly amount of unearned income received and increases if they were in receipt of another form of income support prior to moving onto unemployment benefits.
Repeat spells are identified by the individual having experienced a previous spell during the four-year observation period. This implies that many spells that began early during the period are not identified as repeat spells even though this should really have been the case. This censoring of repeat spells is not taken into account in the estimation. The coefficient implies a reduction of the hazard but, by definition, the average duration of spells must be relatively short for a person who experiences multiple spells within the four-year observation period. Therefore, the coefficient is likely to understate the effect that would be observed given a longer panel. Further, repeat spells may occur for contrasting reasons. Some persons may experience repeat spells because of a tendency to move in and out of employment rapidly. We would expect spells for these persons to be short. Repeat spells may also be expected for the ‘hardcore unemployed’, but in this case we would expect the spells to be of long relative duration.
Column 2 of Table 4b adds the amount of benefit to the model. As was the case for single persons, the amount of benefit is estimated to have a small but significant negative effect on the hazard. As was also the case for singles, the inclusion of the benefit variables does not have much of an effect on the estimates of the other variables.
The third column of Table 4b reports the estimates obtained when partner’s income was included as well as the amount of benefit. Partner income is defined as the sum of partner’s earned and unearned income. The estimated coefficient is highly significant and the positive coefficient implies that the larger is partner’s income the higher is the hazard. In theory, partner’s income should be equivalent to unearned income. Both are expected to increase the demand for leisure and thus to result in a lower hazard. However, the estimates contradict this prediction.
In terms of the magnitudes, the most important determinants of the duration of UB spells appear to be living in shared or governmental rental accommodation and age. As is evident from the previous discussion, however, the interpretation of some of the results is highly speculative and relies on bold and untested assumptions. In the main, one can only note the effect of a particular variable and speculate about the likely reason for the direction and magnitude of the effect.
These limitations are mainly a consequence of three related factors. The first is the omission of many variables found to be important determinants in previous research. Secondly, instead of these variables, many idiosyncratic variables that are specific to this data set are included to account for the variation between individuals. As stated, this makes the interpretation of the findings somewhat speculative. Furthermore, and this is the third factor, it raises doubts that the included variables account for all the variation between individuals. Regrettably, it was not possible to allow for any remaining and unobservable heterogeneity in the estimation.
The results can be compared to a previous analysis of the hazard rate for spells of unemployment using the Australian Bureaus of Statistics’ 1994–97 Survey of Employment and Unemployment Patterns (SEUP) (Stromback and Dockery 2000). The SEUP data measure the length of job search spells but not all job search spells result in a spell of unemployment benefit.6 Some persons are not eligible for unemployment benefit and, of those who are, not all apply. Thus, spells of UB can be regarded as a subset of spells of job search. Another difference between the data sets is that the SEUP data are a stock sample dominated by persons classified as jobseekers when the sample was drawn. Since most jobseekers were unemployed when the sample was drawn, the SEUP sample is subject to length bias—those with long duration of unemployment are over-sampled.
The magnitude of the hazard rate estimated from SEUP data is slightly lower than in the LDS data, with persons leaving spells of job search at a rate of around 6 per cent per fortnight in the first month, falling to below 3 per cent per fortnight after a year. However, even a small difference in the hazard rate translates into quite a large difference in the survival function. Thus, job search spells are very much longer than spells of unemployment benefit. The SEUP median spell is estimated to be 17 fortnights compared to 12 fortnights in the LDS data (see Table 3). The different sampling schemes provide one reason for this difference. The SEUP job search spells are length-biased. They reflect the length of spells of currently unemployed persons while the LDS data refer to persons who begin a spell of unemployment. However, apart from this reason, non-UB job search spells may well be longer than UB job search spells. As previously discussed, many UB spells end by transfer to another type of income support rather than by finding a job. In addition, UB spells can end by the loss of eligibility due to an increase in partner’s income. Job search spells, however, end in either of two ways: by finding a job or giving up the search.
The shape of the hazard is also different in the two data sets. The SEUP data showed that the hazard is monotonically declining from the first month. In contrast, the LDS data indicate an increasing hazard up to 4–6 fortnights and declining monotonically thereafter. The difference could arise because the short spells in LDS are atypical of all job search spells. Persons only claim unemployment benefit if they expect to be out of work for some time, meaning that there will be a considerable number of very short spells of job that do not appear as spells of unemployment benefit. Thus, in the LDS data few spells are observed to end after a short period.
The effects of personal characteristics are estimated to be very similar in both data sets. The exit rate is decreasing in age, the difference between males and females is small, the effect of country of birth is the same, the labour market status of a spouse has a large effect (in SEUP a working spouse increases the exit rate—in LDS a spouse on unemployment benefit decreases the exit rate) and previous unemployment/income support is associated with a lower exit rate. In many cases, the magnitude of the estimated effect of these variables is also similar. The SEUP data also show that the exit rate is lower for those receiving income support and unemployment-related income support in particular.
Another pertinent comparison is with the analysis of the duration on single parent pension (SPP) undertaken by Barrett (2000) using the LDS. In this case, the comparison of greatest interest is between single persons on unemployment benefit and single parent pension beneficiaries. This comparison reveals that SPP recipients have (i) a lower hazard, which is (ii) monotonically declining at a slow rate. That the hazard is lower, and the duration longer, is not surprising, since these persons are under no obligation to engage in job search. Likewise, that the hazard does not decline much over the duration of a spell could be attributed to differences in exit destinations of UB and SPP recipients. If it were the chance of finding a job that decreases over the duration of a spell (due to a scarring effect), this effect would be less pronounced for SPP recipients since their spells are more likely to end for other reasons (such as re-partnering).
More difficult to rationalise are the significant differences in the effects of individual characteristics on the exit rates of the two types of beneficiaries. Although some of the estimated effects are of the same sign and magnitude (for example, age, unearned income, home ownership variables) in most cases the estimated signs differ. Thus, sex, age of youngest child, Aboriginal status, repeat spell and rental status have a different impact. In the case of Aborigines and Torres Strait Islanders, those on UB have a 16 per cent lower exit rate and those on SPP an 18 per cent higher exit rate than other persons. As regards the effect of the benefit level, Barrett’s estimates range from a 3 to 13 per cent lower hazard for each $100 of benefit. Our estimates all fall at the lower end of this range and to that extent the findings are similar. However, in the case of SPP, the estimates of the benefit effect are at best weakly significant.
Hitherto, persons of Aboriginal and Torres Strait Islander origin have been represented by a dummy variable implying an exit rate proportional to that of non-Aborigines. Based on this representation it was found, as expected, that the exit rate for Aborigines and Torres Strait Islanders was significantly lower. To investigate this result in more detail, this section reports results of a separate analysis of Aborigines and Torres Strait Islanders. This is based on a census file of all fortnightly records for Indigenous clients receiving unemployment-related benefits. The file was generated by the FaCS for the same observation period as in the preceding analysis. Although the range of explanatory variables in the data is very restrictive, the number of observations for Aboriginal and Torres Strait Islanders provides a unique opportunity for the analysis of labour market dynamics for this group. The Indigenous census file generated a sample of 116 147 spells of UB receipt experienced by nearly 65 000 individuals.7
|No. of spells||
No. of individuals experiencing Total no. of spells that number of spells
|Total no. of spells|
|1||32 124||50.1||32 124||27.7|
|2||18 725||29.2||37 450||32.2|
|3||8 492||13.2||25 476||21.9|
|4||3 416||5.3||13 664||11.8|
|5||1 030||1.6||5 150||4.4|
|Total||64 148||100.0||116 147||100.0|
Comparison of Tables 5 and 2 shows that the usage of the UB system is remarkably similar for Indigenous and non-Indigenous persons, although just over 7 per cent more Aboriginal and Torres Strait Islander people experienced repeat spells. Eleven per cent of the Indigenous spells were left-censored and 22 per cent right-censored.
As shown in Table 6, however, Indigenous persons are much more likely to experience UB spells than non-Indigenous persons are. Over the four-year period covered by the data, 28 per cent of Indigenous persons of working age experienced an UB spell compared to 18 per cent for all persons.
|Indigenous persons||All Persons|
|Working-age population (15 and over)||231 660a||14 600 000 b|
|Number of persons with a UB spell||64 148||2 674 700|
|Proportion of persons with a UB spell||0.28||0.18|
a: figure for 1996, taken from Taylor & Hunter 1998, Table 1
b: average figure for 1995 to 1998 ABS time series data, 3201.0 Table 9.
Due to the problems with the application of standard labour force status concepts to Aborigines and Torres Strait Islanders living in remote areas and of the impact of CDEP participation on the statistics, results are reported separately for Indigenous persons living in the major metropolitan areas. This is determined via the postcode variable at the commencement of the spell. This approach is by no means a definitive solution. However, it does allow us to single out a group for which the data are free of the impact of the CDEP and who face the same external labour market conditions as most other Australians. Thus, a more legitimate comparison can be made between the results for this group and those for the 1 per cent sample of the general population. For just under one-quarter of the Indigenous sample (27 779 spells), the individual’s initial address was identified as being in a metropolitan area. Figure 3 compares the non-parametrically estimated hazards for Indigenous persons living in metropolitan and non-metropolitan regions with those for non-Indigenous UB recipients.
Figure 3: Non-parametric estimation of the hazard function for UB spells; Indigenous persons in metropolitan and non-metropolitan area (FaCS one per cent sample and Indigenous census file)
For all Aboriginal and Torres Strait Islanders, the mean duration of spells is 26.2 fortnights, 3.4 fortnights longer than for the one per cent sample. It is lower for Aboriginal and Torres Strait Islanders in metropolitan areas, but still higher than for UB recipients as a whole. The Indigenous file included only records for unemployment-related benefits, so it is not possible to investigate the frequency or patterns of transition from UB spells into other payment types. However, the means for the variable ‘Previous IS receipt’ indicate that Aborigines and Torres Strait Islanders are twice as likely to have entered a spell of UB receipt straight from an episode of income support receipt. It would seem likely that they are also more likely to move onto other income support categories from spells of UB receipt.
|All Indigenous||Mean||25% quantile||50% quantile||75% quantile|
Table 8 reports the results for estimation of the proportional hazard model for UB spells by Aboriginal and Torres Strait Islander clients. Separate models were estimated for those living in metropolitan and non-metropolitan areas and for single persons and persons with partners. Table 8a shows that the vast majority of Indigenous persons living in metropolitan areas are single persons (87 per cent). Thus, it seems appropriate to focus mainly on the single group. The signs and order of magnitude of most of the coefficients are similar to those reported for all persons. The main differences are that among Indigenous single persons, older age, sex, type of home ownership and rental status have no significant effect on the hazard, with the exception of sharer status. For those with partners, however, type of home ownership matters, but the presence of children has no effect.
Although non-metropolitan Indigenous persons face a very different labour market, the estimates of the effect of individual characteristics are very similar to the estimates for those in metropolitan areas. One of the few differences that may be noted is the significant effect of home ownership variables for single persons. Among all Indigenous persons, owning or purchasing a home is associated with a larger hazard. This may reflect that home ownership among Indigenous persons is limited to a far smaller proportion of the population and hence is more strongly correlated with other favourable labour market characteristics. The effect of remote communities may have some influence here, but the result is particularly strong for married Indigenous persons within the major metropolitan regions. In addition, in contrast to the results for the wider population, we find no negative effect of living in government rental accommodation for Aborigines and Torres Strait Islanders, but living in shared accommodation is associated with a longer duration of spells.
Most of the other parameters have the same sign and are of the same order of magnitude as for the population as a whole. This suggests that the lower hazard for Aborigines and Torres Strait Islanders is due to their having characteristics which are associated with a longer benefit duration. However, inspection of the average values of the covariates reveals that the differences between the two groups are usually quite small. Thus, the observed difference in the hazard must be due to a number of small differences in average values of the covariates and their estimated effect.
|(1) No benefit variables||(2) With benefit variables||(3) No benefit variables||(4) With benefit variables|
|15 to 19||0.4336 ***||0.4600 ***||0.4025 ***||0.3851 ***|
|20 to 24||0.1488 ***||0.1647 ***||0.2011 ***||0.1977 ***|
|25 to 44||—||—||—||—|
|45 to 54||-0.1830 **||-0.1749 **||-0.1881 ***||-0.1864 ***|
|55 to 59||-0.5264 ***||-0.5216 ***||-0.0541||-0.0392|
|60 to 64||0.1890||0.0840||0.4414 **||0.4119 *|
|Main Eng-spkg country||-0.5430||-0.5941||0.1816||0.1818|
|Home owner||0.3234 ***||0.3149 ***||0.0965||0.1078|
|Purchaser||0.5503 ***||0.5457 ***||0.2313||0.2602|
|Lives elsewhere||1.5219 ***||1.5434 ***||0.4286||0.4552|
|No rent paid||0.1526 *||0.1598 *||-0.0497 *||-0.0060|
|Sharer||-0.6744 ***||-0.6627 ***||-0.3387 ***||-0.3597 ***|
|Has child aged:|
|0 to 6||-0.0117||-0.0126||0.3428 ***||0.3292 ***|
|7 to 12||-0.1049||-0.0990||0.1760||0.1568|
|13 to 15||0.0042||0.0057||-0.1389||-0.1224|
|no depdt. children||—||—||—||—|
|Partner receives UB||-0.0814||-0.0749||—||—|
|Benefit ($ eligible)||—||0.0001||—||-0.0009 ***|
|(Benefit)2||—||-1.4x10-6 **||—||6.2x10-7 ***|
|Unearned income||0.0012 ***||0.0012 ***||0.0006 *||0.0006 *|
|Rent paid||0.0006 ***||0.0005 **||-0.0007 ***||-0.0007 ***|
|Previous IS receipt||-0.1721 ***||-0.2187 ***||-0.1897 ***||-0.1883 ***|
|Repeat spell||-0.1894 ***||-0.2270 ***||-0.0931 ***||-0.1320 ***|
|N (spells)||3 261||3 261||21 791||21 791|
|Censored spells||772||772||4 443||4 443|
|Model χ2 (-2 Log L)||155 ***||175 ***||1 074 ***||1 219 ***|
|Degrees of freedom||29||31||29||31|
***, ** and * denote significance at the 1, 5 and 10 per cent levels, respectively.
—: denotes ‘omitted’
|(1) No benefit variables||(2) With benefit variables||(3) No benefit variables||(4) With benefit variables|
|15 to 19||0.3108 ***||0.3268 ***||0.4268 ***||0.3869 ***|
|20 to 24||0.0928 ***||0.0990 ***||0.1387 ***||0.1360 ***|
|25 to 44||—||—||—||—|
|45 to 54||-0.1453 ***||-0.1435 ***||-0.2288 ***||-0.2308 ***|
|55 to 59||-0.2119 ***||-0.2115 ***||-0.0970 **||-0.1032 **|
|60 to 64||0.0815||0.0812||0.3127 ***||0.2775 ***|
|Main Eng-spkg country||0.4666||0.4410||-0.0439||-0.0599|
|Non-Eng-spkg country||-0.2245 **||-0.1902 *||-0.0620||-0.0380|
|Female||0.0301||0.0307||-0.0429 ***||-0.0416 ***|
|Home owner||0.2371 ***||0.2446 ***||0.1232 **||0.1549 ***|
|Purchaser||0.3977 ***||0.3839 ***||0.4407 ***||0.4385 ***|
|Lives elsewhere||0.0321||0.0577||0.4140 ***||0.4381 ***|
|Government||0.0546 *||0.0781 **||-0.0183||0.0030|
|Board/lodging||0.0404||0.0675 *||-0.0747 ***||-0.0263 *|
|No rent paid||-0.1101 ***||-0.0641 **||-0.0863 ***||-0.0267 *|
|Sharer||-0.4869 ***||-0.4758 ***||-0.3654 ***||-0.3837 ***|
|Has child aged:|
|0 to 6||-0.0644 ***||-0.0563 **||0.1938 ***||0.2205 ***|
|7 to 12||-0.1001 ***||-0.0955 ***||0.0752||0.0868|
|13 to 15||-0.0515 *||-0.0501 *||-0.1387 *||-0.1395 *|
|no depdt. children||—||—||—||—|
|Partner receives UB||-0.2408 ***||-0.2393 ***||—||—|
|Benefit ($ eligible)||—||-0.0002||—||-0.0008 ***|
|(Benefit)2||—||-5.4x10-7 **||—||3.1x10-7 **|
|Unearned income||0.0014 **||0.0013 **||0.0004||0.0003|
|(Unearn inc.)2||1.3x10-6 **||1.2x10-6 **||4.6x10-11||4.3x10-1|
|Rent paid||0.0000||-0.0001||-0.0011 ***||-0.0011 ***|
|Previous IS receipt||-0.2019 ***||-0.2305 ***||-0.1888 ***||-0.2010 ***|
|Repeat spell||-0.1497 ***||-0.1815 ***||-0.0976 ***||-0.1380 ***|
|N (spells)||18 754||18 754||59 959||59 959|
|Censored spells||5264||5264||13 491||13 491|
|Model χ2 (-2 Log L)||327 ***||691 ***||3 071 ***||3 491 ***|
|Degrees of freedom||30||32||29||31|
—: denotes ‘omitted’
The LDS has provided a unique opportunity for researchers to examine the temporal nature of the utilisation of unemployment benefits in Australia. The analysis here provides a summary of the structure of spells of unemployment benefit. In the four-year period for which the 1 per cent sample is available, we find that the average duration of a spell on unemployment benefits is almost 11 months. Two-thirds of the spells observed are accounted for by clients who experienced repeat spells. Thus, the average stay on unemployment benefits in Australia is long and there is a high degree of churning – exit and re-entry to unemployment benefits. Given that the spells of repeat clients are of longer average duration, it can be surmised that a large proportion of time on unemployment benefits is experienced by a small proportion of the population.
The hazard or exit rate from UB spells initially rises to peak in the fourth or fifth fortnight, and then falls monotonically with duration. This finding does not necessarily reflect duration dependence, but can arise because of heterogeneity among persons. That is, persons with a high but constant probability of exiting UB leave after a short time, while those with a low probability leave after a long time. In contrast, estimates of the hazard from spells of job search over a similar period show a uniformly declining hazard. We believe this is likely to reflect that many persons do not claim benefit if they expect to be unemployed for a short period. This may arise because of administrative rules, the cost of applying for benefits or the desire to avoid any possible stigma attached with receiving welfare payments, factors that may outweigh the gains from benefit receipt for short spells.
The factors estimated to have the greatest impact on the duration of UB spells are accommodation status and age. However, the results for individual covariates are difficult to interpret since there are a number of important omitted variables that cannot be observed. Hence, many of the estimated effects may not reflect a direct effect, but a proxy for the effect of unobservable characteristics on the wage rate or arrival rate of new information. Unearned income and partner’s income are two obvious examples. The expected direct effect would be to increase the demand for leisure and reduce the exit rate. Alternatively, and this effect seems to dominate, unearned (partner’s) income may be associated with positive human capital characteristics that increase the exit rate.
The main policy implication of this paper follows from the result that the expected duration of a benefit spell is not strongly related to observable individual characteristic. This implies that there is limited scope for targeted policies that seek to reduce the duration of benefit spells for those most likely to experience long spells. Instead, in view of the declining exit rate, those most likely to experience long spells are those whose spell has already lasted a long time. In other words, disadvantage cannot be identified in advance of unemployment. Instead, long elapsed duration is the principal predictor of remaining long duration. Thus, policies that seek to reduce the duration of unemployment should in the first instance be directed towards the long-term unemployed. Note that this implication is not dependent on whether the declining exit rate is a result of true duration dependence or heterogeneity.
The results for Aborigines and Torres Strait Islanders reinforce this conclusion. Given what is known of the extent of the labour market disadvantage facing Aborigines and Torres Strait Islanders, the surprising finding from this paper is the similarity in the pattern of receipt of unemployment benefits between Indigenous and other Australians. By looking separately at Aborigines and Torres Strait Islanders living in the major metropolitan capitals, we know that this finding has not arisen because of the impact of peoples living in regions away from ‘mainstream’ economic activity or of participation in the Community Development Employment Program. The important message for policy-makers is that the greater utilisation of unemployment benefits among Aborigines and Torres Strait Islander peoples is primarily due to the higher proportion that enters spells of unemployment benefits rather than differential behaviour or incentive effects once receiving benefits.
While this paper provides a valuable analysis of the pattern of the utilisation of unemployment benefits over time, some limitations with the quality of the data set remain to be resolved. There are also some methodological issues relating the construction of variables and identification of their effects, notably with respect to payment levels. We have drawn attention to these limitations throughout the paper. Continued improvement in the data quality, extension of the panel duration and further development in methodology offers considerable potential to enhance the current analysis, particularly with respect to changes in the dynamics of unemployment utilisation over the economic cycle, between regional labour markets and in response to changes in policy parameters.
|One Percent Sample||Indigenous – metropolitan||Indigenous – non-metro|
|15 to 19||0.2007||0.0224||0.2493||0.0791||0.2504||0.0708|
|20 to 24||0.2973||0.0898||0.2636||0.2082||0.2434||0.1979|
|45 to 54||0.0830||0.1932||0.0406||0.0678||0.0521||0.0865|
|55 to 59||0.0185||0.0748||0.0067||0.0156||0.0105||0.0202|
|60 to 64||0.0058||0.0491||0.0011||0.0012||0.0021||0.0060|
|Main Eng-spkg country||0.0757||0.0953||0.0018||0.0009||0.0007||0.0005|
|No rent paid||0.2020||0.0527||0.1464||0.0754||0.2358||0.2199|
|Has child aged:|
|0 to 6||0.0020||0.3263||0.0072||0.4818||0.0091||0.4602|
|7 to 12||0.0013||0.1153||0.0035||0.0938||0.0052||0.1032|
|13 to 15||0.0017||0.1204||0.0033||0.1153||0.0045||0.1295|
|Partner receives UB||—||0.1966||—||0.2186||—||0.2621|
|Benefit ($ eligible)||$284.76||$265.51||$236.29||$218.89||$234.68||$227.49|
|(Benefit)2||$127 006||$108 654||$87 900||$75 406||$86 156||$79 647|
|- if have unearn inc.||$34.76||$47.89||$45.52||$307.85||$191.39||$62.74|
|(Unearn inc.)2||$879||$7 328||$191||$277 924||$136 509||$546|
|- if have unearn inc.||$7 039||$26 247||$12 417||$6.97m||$11.06m||$21 026|
|- if have ptnr income||$347.97||—||—||—||—|
|(Partner inc.)2||—||$936 228||—||—||—||—|
|- if have ptnr income||$2548854||—||—||—||—|
|- if renting||$155.92||$251.70||$148.07||$210.26||$132.13||$165.46|
|(Rent paid) 2||$17 610||$31 214||$17 336||$34 719||$11 633||$18 206|
|- if renting||$29 439||$73 186||$25 937||$53 254||$20 772||$35 174|
|Previous IS receipt||0.2400||0.2338||0.4349||0.4195||0.4346||0.4660|
|N (spells)||27 027||9 742||21 791||3 261||59 959||18 754|
—: denotes ‘omitted’
- That is, by 0.4 percentage points given an 8 per cent unemployment rate.
- The payment of unemployment benefits is normally conditional on meeting an activity test. However, as some categories of recipients are excluded from the activity test there is a case for excluding such persons from the analysis as well on the ground that their behaviour is different. Since the status of the relevant indicator variable was uncertain in the preliminary versions of the data set, this issue had to be left for future research.
- Dawkins, Harris & Loundes (2000) provide an extensive analysis of repeat spells using the same data set.
- In the statistical literature, the presence of ‘old’ spells is referred to as left-truncation. Rather than deleting these spells, unbiased estimates of the distribution of completed spells can be obtained if the contribution of these spells to the likelihood function is properly accounted for.
- ‘Old’ spells are included in this calculation.
- Spells of unemployment defined by the standard criteria cannot be identified in SEUP. The spells used in the analysis comprised periods for which a person was reported to be looking for work and was not working. This is a close approximation, but does not meet all the technical criteria of the standard definition of unemployment.
- In this data set, a person is an Aborigine or Torres Strait Islander if they identify themselves as such.
Altman, J. & Taylor, J. 1995, ‘Calculating Indigenous Australian employment in the private sector’, Labour Economics and Productivity, vol. 7.
Barrett, G. 2000, The dynamics of participation in the Sole Parent Pension, paper presented to Department of Family and Community Service Workshop, 21st March, Canberra. Committee of Review of Aboriginal Employment and Training Programs 1985, Report, AGPS, Canberra. Commonwealth of Australia 1994, Review of Aboriginal Employment Development Policy, Aboriginal and Torres Strait Islander Commission, Canberra.
Dawkins, P., Harris, M. & Loundes, J 2000, Repeated spells on benefits: an analysis of ‘Churning’ using the FaCS longitudinal administrative data set, paper presented to Department of Family and Community Service Workshop, 21st March, Canberra.
Department of Employment, Workplace Relations and Small Business (DEWRSB) 1999, Regional Labour Market Information for the Employment Services Request for Tender 1999, DEWRSB, Canberra.
Fortin, B. & Lacroix, G. 1997, ‘Welfare benefits, minimum wage rate and the duration of welfare spells: evidence from a natural experiment in Canada’, CIRANO Scientific Series 97s–25, Centre Interuniversitaire de Recherche en Analyse des Organisations, Montreal.
Holmlund, B. 1997, ‘Unemployment insurance in theory and practice’, Discussion Paper No. 380, Centre for Economic Policy Research, Australian National University.
Katz, L. F. & Meyer, B. D. 1990, ‘The impact of the potential duration of unemployment benefits on the duration of unemployment’, Journal of Public Economics, 41, pp. 45–72.
Meyer, B. D. 1990, ‘On unemployment insurance and unemployment spells’, Econometrica, 58, pp. 757–782.
Miller, P. W. & Le, A. T. 1999, A risk index approach to unemployment: an application using the Survey of Employment and Unemployment Patterns, Australian Bureau of Statistics Occasional Paper, ABS Catalogue No. 6293.0.00.001, February.
Moffit, R. 1992, ‘Incentive effects of the US welfare system: a review’, Journal of Economic Literature, 30, pp. 1–61.
Smith, D. 1995, ‘Measuring the contemporary labour force status of Indigenous Australians: issues, implications and options’, Labour Economics and Productivity, vol. 7, no. 1, pp. 28–48.
Stromback, T. & Dockery, A. M. 2000, Labour market programs, unemployment and employment hazards: an application using the 1994–1997 Survey of Employment and Unemployment Patterns, Australian Bureau of Statistics Occasional Paper, ABS Catalogue No. 6293.0.00.002, February.
Taylor, J. & Hunter, B. 1998, The job still ahead: economic costs of continuing Indigenous employment disparity, Aboriginal and Torres Strait Islander Commission, Australia.
—— 1997, ‘Promoting growth in Indigenous employment: the role of the private sector’, Australian Bulletin of Labour, 23, 4, pp. 295–313.
Webster, E. 1999, ‘Labour market programs and the Australian Beveridge Curve’, The EconomicRecord, 75, 231, pp. 405–416.