Explaining Unemployment Duration in Australia

© 2006 The Economic Society of Australiadoi: 10.1111/j.1475-4932.2006.00341.x

THE ECONOMIC RECORD, VOL. 82, NO. 258, SEPTEMBER, 2006, 298–314

298

Blackwell Publishing, Ltd.Oxford, UKECORThe Economic Record0013-0249&ãgr; 2006. The Economic Society of Australia.March 2006823Original Article

UNEMPLOYMENT DURATIONECONOMIC RECORD

Explaining Unemployment Duration in Australia*

NICK CARROLL

Economics Program, Research School of Social Sciences, Australian National University, Canberra, ACT, Australia

What influences the probability that someone will leave unemploy-ment? Informed by a search-theoretic framework and allowing for exitsto not in the labour force and employment, in this paper I examine whatinfluences the probability that somebody will leave unemployment. Theunemployment data used are derived from the retrospective work historyinformation from the Household, Income and Labour Dynamics inAustralia Survey. The results show that variables that increase wageoffers and lower reservation wages are associated with shorter unemploy-ment durations, and that exit rates from unemployment appear to remainsteady initially with duration before declining relatively sharply.

I Introduction

What influences the probability that someonewill leave unemployment? Does the probability ofexit from unemployment fall the longer somebodyhas been in unemployment? Is the negativerelationship commonly observed between lengthof spell and exit probability related to ‘lowexiters’ remaining in unemployment, or becauselong unemployment duration leads to morelimited job opportunities?

A number of authors have found that unemploy-ment affects people’s happiness and life satisfaction(Winkleman & Winkleman, 1998; Clark & Oswald,1994), wages (Ruhm, 1991; Arulampalam, 2001),and probability of unemployment in future periods(Arulampalam

et al

., 2000).

1

In addition, if un-employment is scarring, exit rates may fall as thecurrent spell of unemployment lengthens (Machin& Manning, 1999). While in this literature thereare issues about whether unemployment causesworse outcomes (state dependence), or whetherpeople with certain characteristics become un-employed (unobserved heterogeneity), there appearsto be a general finding that unemployment hasnegative impacts on the people who experience it.

While unemployment experiences, compared toemployment experiences, appear to be associatedwith worse outcomes, some unemployment expe-riences are short and singular, while others arelonger and repeated. This paper investigates thevariables associated with exits from unemploy-ment using a duration modelling framework. Theanalysis enables us to see what characteristics areassociated with exit from unemployment and to

* The author is grateful for assistance from GaryBarrett, Alison Booth, Tue Gorgens, Paul Miller and twoanonymous referees; and for comments received fromparticipants at a workshop at Research School of SocialSciences (RSSS) in July 2004 and at the AustralianNational University/University of Western Australia(ANU/WA) PhD conference in November 2004. Thispaper uses the Household, Income and Labour Dynamicsin Australia (HILDA) database, but the views expressedshould not be attributed to The Department of Familyand Community Services (FACS) or the MelbourneInstitute and neither organisation accepts responsibilityfor the accuracy of the research findings.

JEL classification: J64

Correspondence

: Nick Carroll, Labour Market StrategiesGroup, Department of Employment and WorkplaceRelations, GPO Box 9879, Canberra, ACT 2601, Australia.Email: [email protected]

1

Ruhm (1991) specifically examined the impact oflay-offs on future wages.

2006 UNEMPLOYMENT DURATION

299

© 2006 The Economic Society of Australia

investigate, holding certain characteristics con-stant, how exit rates vary with duration.

Owing to data limitations, research in Australiainto unemployment spells using duration model-ling has tended to be quite descriptive in nature orto focus on demographic characteristics and limitedlabour market and personal characteristics (seeBrooks & Volker, 1986; Chapman & Smith, 1992;Borland, 2000). Outside the duration modellingframework, Knights

et al

. (2002) used a randomeffects incidence methodology to examine statedependence and found evidence of negative dura-tion dependence for youth in Australia. Hui (1991)estimated a structural model of the reservationwage for youths and found evidence to supportpositive duration dependence.

2

This paper makes three contributions to theliterature. First, it uses an important new longitu-dinal data source (Household, Income and LabourDynamics in Australia (HILDA) Survey) from arecent period to investigate the factors associatedwith exit from unemployment. In so doing it usesmore flexible baseline hazards than have typicallybeen used in unemployment duration modelling inAustralia. Second, the paper examines severalfactors associated with unemployment exit thathave not previously been examined in Australianresearch. In particular, employment experience,unemployment experience and risk aversion areincluded in estimation. Risk aversion is potentiallyimportant, given the job-seekers’ choice aboutwhether to accept a certain wage offer in the currentperiod (hence exiting unemployment) or wait foran inherently uncertain wage in a future periodthat may be higher (hence lengthening duration).Third, the paper uses and analyses a flexibleunderlying baseline hazard. Australian research inthis area has typically assumed that changes in thehazard rate are monotonic and have not been flexibleenough to allow for non-monotonicity.

The first result from this paper is that factorsexpected to raise a person’s wage offers (such asemployment experience) and lower their reserva-tion wage (such as ineligibility for some benefits)are associated with higher exit rates from un-employment. The second result is that exit rates fromunemployment do not fall with duration initially,

but after a period of approximately 4 months theythen appear to decline by a large amount. Becauseof this, modelling that assumes a monotonic rela-tionship between spell length and exit rates willtend to underestimate the decline in the hazard atlonger durations (because it will be a weightedaverage of the declining and non-declining com-ponents).

3

The final key result is that, even withthe number of variables used in the estimationpresented here, they ‘explain’ less than one-fifth ofthe decline in the baseline hazard. This suggeststhat either unobserved characteristics are poten-tially important or that some scarring is occurring.

The remainder of this paper is structured as fol-lows. Section II briefly describes the econometrictools, Section III describes the data used, SectionIV gives the results from the full model, SectionV presents two extended models focusing on therole of risk and previous wages, Section VI exam-ines the baseline hazard, and section VII providesconclusions. In Appendix I, descriptions of thevariables are provided.

II Estimation of Survival Time Models

Survival data are used for the analysis in thispaper. In the case of unemployment survival data,a person is observed to enter unemployment andthe spell is observed until exit. Two key conceptsare the hazard rate (the proportion of an ‘at-risk’group that leaves a particular state over areference period) and the survival rate (theproportion of the initial group that remains in thestate at the reference period).

I use a proportional hazards model, wherebyexplanatory variables move a baseline hazard upand down by a fixed proportion. The proportionalhazard model takes the following form:

h

i

(

t

) =

h

0

(

t

)exp(

z

i

(

t

)

′

β

) (1)

where

h

i

(

t

) is the hazard for person

i

,

h

0

(

t

) is thecommon baseline hazard,

z

i

(

t

) are the observablecharacteristics, and

β

and

h

0

are the parameters tobe estimated.

Two estimation methods with different assump-tions about the baseline hazard are used to pro-duce the results below. In the first estimation it isassumed that all people have the same shapedbaseline hazard, but the baseline hazard is notrestricted to be of a certain shape (semiparametric

2

This paper builds on both the Knights

et al

. (2002)and Hui (1991) papers by including all people aged15 years and older, including those people who flowfrom unemployment to not in the labour force in the ‘at-risk’ group, allowing for more flexibility in the baselinehazard and using survival analysis.

3

It should also be noted that the coefficients will alsobe affected by mis-specifying the baseline hazard and itmay be reflected in the violation of the proportionalhazards assumption.

300

ECONOMIC RECORD SEPTEMBER


estimation). In the second estimation the time axisis divided into a finite number of intervals and aseparate baseline hazard parameter is estimatedfor each interval. The approaches that are used arevery flexible methods for estimating the baselinehazard, which avoid the problems associated withimposing a parametric functional form.

4

In the piecewise constant model, more intervalsare preferred to less, all else being equal, becauseit allows for more flexibility in the shape of thebaseline hazard. However, to ensure meaningfulestimates of the changes in hazard rate by dura-tion interval, it is important to ensure that there issufficient sample size. Fourteen intervals are cho-sen to ensure that there are at least 40 exits perinterval. The following intervals (in thirds of amonth) employed are: 1, 2, 3, 4, 5, 6, 7–9, 10–12, 13–15, 16–18, 19–24, 25–30, 31–36, 37+.

5

The standard partial log likelihood function forthe semiparametric estimation (sometimes calledthe Cox regression) is used (see Lancaster (1990)for more details about the log likelihood functionsgiven in this section):

(2)

where

j

indexes the ordered failure times

t

j

,

D

j

is the set of

d

j

observations that fail at

t

j

,

d

j

is thenumber of failures at

t

j

and

R

j

is the set ofobservations

m

that are at risk at time

t

j

. Thestandard likelihood function for the piecewiseconstant estimation with

N

unemployment spellsis used:

(3)

where

k

i

is the observed length of the

i

th spell,

δ

i

= 1 if the spell is not right censored (0otherwise). In maximising the log likelihood the

γ

(

t

) = ln[

h

0

(

u

).

du

] are treated as parameters tobe estimated.

The independent competing risks framework isused to examine exits to employment separatelyfrom exits to not in the labour force. It is assumedthat there are two latent survival times, one foremployment and one for not in the labour force,and the destination observed is the minimum ofthe latent survival times. With this assumption,exits from unemployment to not in the labourforce are treated as censored spells for exits toemployment in estimation.

6

The likelihood func-tions in Equations 2 and 3 are used for estimation.

An assumption of the independent competingrisks framework is that the risk of exit to employ-ment and not in the labour force are independent,conditional on the covariates. Three points shouldbe noted regarding this assumption. First, becausethe result is conditional on the covariates, if themodel is well specified, then this assumption isnot particularly strong, because the included covari-ates are effectively controlled for. Second, becausemost exits are to employment, the size and sig-nificance of the coefficients are similar when allexits from unemployment are examined to whenonly exits from unemployment to employmentare examined. Third, those papers that have under-taken dependent competing risks have consistentlyfound no evidence to support the hypothesis ofcorrelation between the exit-specific heterogeneityterms (see Han & Hausman, 1990; van den Berg

et al

., 2003).In estimation below, unobserved individual level

heterogeneity is allowed for. The method sug-gested by Heckman and Singer (1984) is used,whereby the unobserved heterogeneity distributionis estimated non-parametrically by a discretemultinomial distribution.

III Household, Income and Labour Dynamics in Australia Survey

The data used for this study are from theHILDA panel database. The survey is primarilyadministered in the second half of each year, with

4

Barrett (2000) (drawing from the literature) notesthat mis-specification of the baseline hazard is a majorsource of error in drawing inferences concerning boththe presence of duration dependence and the impact ofcovariates.

5

The number of nodes is less than the 21 intervalsused by Barrett (2000), although Barrett (2000) had asample size approximately five times larger. In addition,the number of nodes is more disaggregated than the fivenodes used by Meyer (1990). The coefficients wereunaffected by use of more disaggregated intervals(results available from author). However, the estimatesof the baseline hazard were significantly more noisywith less aggregated data.

L z d zmm D

j ii Rj

D

j j

( ) log exp( )β β β= −

∈ ∈=

∑ ∑∑1

L k

z k t z t

ii

N

i

i it

k

i

i

( , ) log( exp( exp[ ( )

( ) ])) [ ( ) ( ) ]

β γ δ γ

β γ β

= − −

+ − +

∑

∑−

=1

= 0

exp

1

1

′ ′

6

In the discrete setting it is possible to undertakeestimation with a variety of assumptions about how riskof exit may vary within discrete periods (seeNarendranathan & Stewart, 1993). However, thisassumption is beyond the scope of this paper.

tt+∫ 1


301


the first wave being collected in the second half of2001. The first two waves of HILDA were usedfor this study (2001 and 2002). In wave 1, 7682households were sampled comprising 13 969members. The HILDA survey has a wide varietyof variables on health, family background, workhistory, demographics and educational andtraining history, which allows us to control forsome heterogeneity.

The data used for the analysis of spells in theHILDA database are drawn from a calendar. Inthis calendar the respondent is asked to providetheir employment status in thirds of a month overthe past 12–18 months (depending on when theinterview takes place). The unit of analysis forthe remainder of the paper is thirds of a month.From Table 1 the sample has 2402 spells of un-employment, with 1542 exits to employment or notin the labour force (referred to as fails) and 860spells where the end of the spell is not observedin the sample period (in other words the spell isright censored).

7

Out of the 2402 spells, 1757 ofthese spells begin during the sample period, 471spells are left censored (for a description seesubsection (ii) of Section III) and 174 spells areleft truncated (also see subsection (ii) of SectionIII).

Figure 1 presents the raw hazard rates. There isa general decline in the hazard rates over time,from 6 to 9 per cent at the start of the spell to 2to 5 per cent after 6 months. Note that the sample

used to estimate the hazard rate is small at longerdurations and therefore less reliable and thus anyconclusions about the peak after 75 periods(2.1 years) should be tentative. Another interestingpoint from Figure 1 is the saw shape in the lines.There are peaks in the hazard rate after threeperiods (1 month), six periods (2 months) and soon. This most likely reflects the nature of thequestions in HILDA and the fact that they werebased on recall.

(i) Can the Calendar Data from Different Interview Dates Be Joined Successfully?

Ideally calendars reported at different interviewdates (i.e. the wave 1 and the wave 2 interviewdates) would always be consistent and therewould be no large outflow at the time of the join(the point where the calendars in wave 1 andwave 2 intersect). However, from Figure 2 there isa considerable issue at the join, with a large fallin the number of people who remain unemployedand large increases in the inflow and outflow fromunemployment. The number remaining unemployedfalls from 500 per period to approximately 250at the join, while the inflow and outflow tounemployment increase from less than 50 tonearly 250 per period at this time.

Approximately half of the unemployment periodsrecorded between July and November 2001were reported inconsistently between waves 1 and2 (there is an overlap between the calendars ofwave 1 and wave 2 of up to 5 months – ‘theseam’ – where the consistency of answersbetween interview dates can be checked). Twenty-seven per cent of people gave answers betweenwaves that were never consistent (where the cate-gories are unemployment and not unemployment),40 per cent of people were consistent the entiretime and the remaining one-third were consistentfor between one and eight of the nine periodsexamined.

Fortunately the seam and join issues areresolved in a relatively straightforward manner.The basic rule applied to the seam is that the datareported closest to the interview data are usedwhen there is some inconsistency (assuming thatrecall is better over shorter periods than longerperiods). More importantly, where greater than 20per cent of observations are inconsistent betweeninterview dates, the spells reported in wave 1 andwave 2 are treated separately (because the qualityof the join is considered too low). When this ruleis applied the problem of the large inflow and out-flow at the join is resolved.

7

While left censoring is a key issue in durationanalysis, right censoring is less of an issue. This isbecause as long as right censoring is random (i.e. theprobability that a spell of length

t

’s exit will not beobserved is random) then the spell enters into thedenominator of time at risk at each period prior tocensoring. However, because the spell is not observed toend, the spell does not contribute to the numeratorthrough an exit (see Lancaster, 1990).

Table 1Spells Affected by Left Censoring/Left Truncation

(Unit: Spells)

No fail Fail Total

Left truncated 122 52 174Left censored 92 379 471Begin in scope 646 1111 1757Total 860 1542 2402

302



(ii) Are There Issues with Left Censoring and Left Truncation?

Left censoring occurs when the first observationof the spell is sometime after it has begun, but it

is not known how long the spell has been goingwhen it is observed. The problem with left-censored unemployment spells is that theircharacteristics may be quite different to other

Figure 1Raw Hazard Rate Over Time

Figure 2Numbers Entering, Staying and Exiting from Unemployment


303


unemployment spells. One possibility is toexclude the left-censored data from analysis andundertake analysis with the inflows tounemployment. Another possibility is to use stocksample techniques to include the data (seeLancaster, 1990). However, stock sampling is notexamined in most unemployment duration studiesand is beyond the scope of this research.

Left-truncated data are where the spell does notbegin in the scope of the calendar, but the date itbegan is known. These data are included in thelikelihood function here as the conditional proba-bility of exit given the spell has lasted to length

x

(see Lancaster, 1990). In HILDA, the start date isavailable for spells that are in progress at thewave 1 interview date (15–18 months after thebeginning of the reference period). If the spellbegan prior to the reference period and is continu-ous until the wave 1 interview date then the spellstart date is available (left-truncated data). Other-wise, if the spell began prior to the referenceperiod and finished before the interview date thenthe start date is not available (left-censored data).Because left-truncated spells have been continuousbetween the start of the reference period and the surveydate, the longest spells in HILDA are likely to beprimarily left truncated rather than left censored.

(iii) What Variables Are Included in the Analysis?

Explanatory variables available in HILDArelated to wages include educational background,employment experience, location, country of birth,long-term disability, parental employment status(which may affect opportunities and preferencesfor human capital) and previous unemploymentexperience. Explanatory variables available relatedto non-wage income (through intrahouseholdincome sharing and government transfers) andtherefore reservation wages include marital status,children, disability and location.

8

A number of the explanatory variables may beendogenous. Therefore it is important to use otherinformation from the survey to back-cast the vari-able to prior to the beginning of the spell.

Back-casting is only done for spells that beginprior to the wave 1 interview date, otherwise thevariable is as at the wave 1 interview date. Loca-tion is one example of a time-varying endogenousvariable. People living in high unemploymentareas may be less likely to find work and theunemployed, who may have less income, maymove to areas with cheaper housing but wherethere are fewer job prospects. HILDA has a ques-tion that asks how long has the respondent livedat their current address and using this informationit is possible to confirm their location prior to theunemployment spell and partially overcome thisendogeneity issue. Other variables treated in thisfashion are number of children, marital status, andemployment and unemployment experience.

9

It is not possible to back-cast the risk-aversionand lagged wage variables, which are potentiallyendogenous. In specifications that include thesevariables, the sample is restricted to include spellsbeginning after the wave 1 interview date (whenthe data are collected). A number of variables areprimarily time-invariant over the short periodstudied or not endogenous to unemployment andtherefore they are recorded as at the wave 1 inter-view date. These variables are educational qualifi-cations, parental employment status, sex, countryof birth and disability.

Dummies to control for the time of intervieware also included. A dummy variable indicates ifthe spell began in wave 1, and two dummy vari-ables indicate what part of the duration month thespell ended (monthly reporting dummies), to control forthe saw shape in the raw hazard (see above).

IV Estimation of Full Model

The primary focus of this paper is what affectsthe probability that someone will move fromunemployment to employment. A commontheoretical framework to investigate this issue isthe search framework. In this framework, anindividual sets a reservation wage (based on costsand benefits of search and unemployment), drawsoffers from a wage distribution and accepts offersabove the reservation wage and declines offersbelow the reservation wage. Thus, the higher thewage offers relative to the reservation wage thehigher the exit rate, and alternatively, the lower

8

A reservation wage variable for the stock ofunemployed is collected in HILDA at the interview date.However, these data are not used because the sample ofreservation wages for the inflow to unemployment is toosmall, and because the reservation wage variable isendogenous to the unemployment experience and thus itis not appropriate to use this reservation wage variablefor estimation with the stock of unemployed (given theirvarying unemployment durations).

9

Where a person moves location it is not possible togive this person’s previous location, for this group amissing location is given. However, it is possible toback-cast children, marital status, and employment andunemployment experience.

304



the reservation wage relative to the wage offersthe higher the exit rate.

The proportional hazards competing risksframework is used in estimation (see Section II).The results from the piecewise baseline hazard

(with and without unobserved heterogeneity) andthe Cox estimations (see Table 2) are presentedhere. These three sets of results are presented toshow the robustness of the findings to alternativeassumptions about the baseline hazard and

Table 2Affect of Covariates on Exit Rates – Exits to Employment – Hazard Ratio

(1)Cox partial likelihood

(2)Piecewise

without heterogeneity

(3)Piecewise with heterogeneity

Means forspecification

(1)

Male 0.825* 0.774* 0.659** 0.54[2.08] [−2.53] [−2.85]

Non-English-speaking COB 0.678* 0.547** 0.427** 0.14[2.02] [−2.91] [−2.59]

English-speaking COB 1.318 1.304 1.132 0.11[1.56] [1.34] [0.43]

Length resident (English) 0.992 0.994 0.992 3.69[1.08] [−0.74] [−0.64]

Length resident (non-English) 1.006 1.015 1.022 2.27[0.78] [1.53] [1.40]

Non-university qualification 1.007 0.971 0.912 0.44[0.11] [−0.39] [−0.85]

University qualification 1.398** 1.396** 1.842** 0.16[3.69] [3.24] [4.24]

Father’s unemployment status 1.278 1.163 1.112 0.13[1.76] [0.92] [0.47]

Mother’s employment status 0.941 0.926 0.971 0.44[0.86] [−0.99] [−0.26]

Children 0.812** 0.753** 0.739** 1.19[4.17] [−5.31] [−4.30]

Male*children 1.233** 1.326** 1.292** 0.60[3.55] [4.54] [2.98]

Male*father 0.717 0.818 0.821 0.07[1.80] [−0.93] [−0.61]

Marry 1.082 1.227* 1.365* 0.49[0.91] [2.29] [2.44]

Disability 0.722** 0.718** 0.768 0.17[3.07] [−3.09] [−1.79]

Employment experience 1.058** 1.059** 1.080* 12.97[3.20] [2.89] [2.47]

Employment experience2 0.999* 0.999 0.999 299.51[2.17] [−1.93] [−1.21]

Lagged unemployment 0.731* 0.648** 0.436** 0.11[2.23] [−3.15] [−3.85]

Observations 1400 1396 1396 1400Failures (exits) 769 767 767 769Log likelihood −3384.0 −2867.8 −2329.1 NAWald test 94.66 126.87 108.26 NA(Probability > χ2) (0.00) (0.00) (0.00)

The base is women, without qualifications, Australian born, unmarried, with no disability and with less than 25 per cent of labour marketexperience in unemployment. In raw data there is a saw shape with systematically different exit rates for different parts of the month.Additional dummy variables and stratification in Cox partial likelihood has been undertaken to control for this. Robust z-statistics arereported in brackets. Issues remain with the non-proportionality of these variables in the parametric regressions (see discussion in text).Standard errors adjusted for clustering on xwaveid. Cox partial likelihood has been stratified to adjust for monthly reporting in calendarand for wave 1 dummy. Age, Age2, regional unemployment rate and regional socio economic status are held constant. COB, country ofbirth; NA, not applicable.

2006 UNEMPLOYMENT DURATION 305


because these estimations allow for more flexi-bility in the baseline hazard. The results fromall estimations presented are consistent and in mostcases the signs and levels of significance agree.

The coefficients are presented as hazard ratios.So a coefficient of, for example, 0.5 for a dummyvariable is interpreted as lowering the exit ratefrom unemployment to employment by a half. Fora continuous variable, a coefficient of 0.5 impliesa unit change in the variable is associated with ahazard rate 1/2 as large and an n unit change inthe variable is associated with a hazard rate (1/2)n

as large.The preferred specification is the Cox partial

likelihood estimator because, first, it does not relyon the restrictions of the baseline hazard. Second,the tests for violation of the proportional hazardsassumption (see below) indicate that the wave 1dummy and the monthly reporting dummies vio-late the proportional hazards assumption. Themost convenient way to deal with this violation iswith the Cox partial likelihood estimator stratifiedby wave 1 and the monthly reporting dummies.10

The results described in this section are pre-sented in Table 2. There are 769 exits to employ-ment and a total of 1400 spells overall for thecontinuous time models and 1396 spells for thediscrete estimation (because the left-truncateddata are excluded). Looking at the means of thevariables in the estimation, 54 per cent of thespells are for males, 75 per cent of the spells arefor Australian-born people, 16 per cent are forpeople with university qualifications, and 11 percent are experienced by people who have spent25 per cent of their work experience in unemploy-ment (see column 4 of Table 2).11

(i) What Impact Do ‘Wage’ Variables Have on Exit Rates to Employment?

The first main result is that variables expectedto increase wages are associated with higher exit

rates from unemployment to employment (resultsfrom the preferred specification are presented inspecification 1 in Table 2). In particular, variablesexpected to increase wages through human capitaland productivity enhancement or screening(employment experience and university qualifications)are associated with higher exit rates toemployment.

The productivity variables would be expected toincrease the exit rates to employment becausethey increase the expected returns from workingrelative to not working. While they also increasethe reservation wage (because they are associatedwith higher future earnings), Mortensen (1986)shows that variables expected to increase thewage are associated with shorter unemploymentspells. Employment experience and universityqualifications significantly increase the exit rate toemployment at the 1 per cent level of significance.

Having university qualifications increases theexit rate by 40 per cent and shortens the medianduration by approximately 40 per cent.12 Oneextra year of work experience increases the exitrate by 6 per cent, while 10 extra years of workexperience increases the exit rate by 72 per cent.It should be noted that the addition of work expe-rience increases the exit rate at a decreasing rate(because the hazard rate of the square of employ-ment experience is less than 1).

Being born in a non-English-speaking country(‘non-English speaking COB’) is associated withlower exit rates from unemployment to employ-ment. People born in non-English-speaking coun-tries may have problems with communication andculture (that may affect productivity or discrimi-nation). Either way, this variable would beexpected to lower wages and thereby lower exitrates from unemployment to employment. Thisvariable significantly reduces the exit rate at the5 per cent level. The exit rates from unemploymentto employment increase as people born in non-English-speaking countries spend more time inAustralia (although not significantly at the 5 per10 The saw shape presented in the earlier hazard curve

in Figure 1 was held constant by dummy variables.It is not possible to stratify the parametric estimatorsby the monthly reporting dummies because this wouldresult in discontinuities in the baseline hazard. It ispossible to interact survival time with the monthly report-ing dummies to overcome the violation, but initial inter-actions with the parametric estimator did not resolve theproportional hazards violation.

11 These means were calculated for the continuoustime sample, but there is little difference in the meansfor the discrete time sample (because of the largeoverlap between samples).

12 The change in the median duration was calculatedfrom the survivor function for the Cox estimation for achange from school qualifications to post-schoolqualifications. The effect of the variable on medianduration may be sensitive to the value of the covariates.The covariate values are as follows: men, aged 40 years,with 5 years’ work experience, Australian born, with nodisability, unmarried, in wave 1, with 25 per centunemployment experience (38.2 per cent drop in medianduration) and without 25 per cent unemploymentexperience (39.9 per cent drop in median duration).

306 ECONOMIC RECORD SEPTEMBER


cent level). Overall, being born in a non-English-speaking country is associated with lower exitrates from unemployment initially, but as theimmigrant spends more time in Australia the sizeof this effect diminishes.13

The impact of disability on wages is likely tobe both through productivity (if disability affectsperformance) and discrimination. As with theother wage variables, disability takes the expectedsign (having a disability lowers the exit rate). Thecoefficient on disability is significant at the 1 percent level of significance. Disability lowers theexit rate from unemployment to employment by28 per cent per period and increases the medianduration by approximately 70 per cent.14

(ii) What Impact Do Non-wage Income Variables Have on Exit Rates to Employment?

The second main result is that variables asso-ciated with higher non-wage income (and hencehigher reservation wages) are associated with lowerexit rates from unemployment to employment.Two key sources of non-wage income aregovernment transfers and intrahousehold incomesharing. Two groups that in general receive highergovernment transfers are people with childrenand people with disabilities.15 Thus, these

variables would be expected to be associated withlower exit rates (although these variables may alsoaffect exit rates through productivity as well).

Having children for women and having a dis-ability are both associated with lower exit ratesfrom unemployment to employment.16 This resultis consistent with higher reservation wages beingassociated with lower exit rates from unemploy-ment to employment. There is very little impactfor men having children, perhaps suggesting thatmen have less access to the non-wage incomeassociated with children.

People who are married would be expected tohave greater access to non-wage income. There isno impact of marital status on the exit rate,although this may be related to non-observablecharacteristics of married people increasing theexit rate and swamping the reservation wageeffect.

(iii) Does Past Unemployment Experience Affect Exit Rates to Employment?

Lagged duration dependence is now examined;that is, whether previous unemployment affectsthe exit rate in the current spell. Lagged durationdependence may occur because of discriminationagainst people with unemployment histories, andbecause of erosion of human capital and workhabits, which all in turn lower wage offers andhence result in lower exit rates fromunemployment.

The results from this estimation are consistentwith unemployment experience before an indi-viduals’ current spell lowering the exit rates fromunemployment to employment. This result is sig-nificant at the 5 per cent level. The coefficientindicates that past unemployment experiencelowers the hazard rate by 27 per cent.17 This resultholds employment experience constant. The com-bined effect of being out of work and unemployedis the combined effect of the ‘lagged unemployment’and the ‘employment experience’ variables. Thus,there is an effect from not accumulating human

13 It may be expected that foreign-born people wouldhave a lower reservation wage and that this may swampthe effect of lower wage offers on the exit rate.However, the odds ratio is less than 1 for being born ina non-English-speaking country (indicating empiricallythat being in a non-English-speaking country is asso-ciated with a lower exit rate, at least initially). In addition,when we examine the effect of number of years in Aus-tralia for those people born in a non-English-speakingcountry (this variable takes a value of 0 for all peoplenot born in a non-English-speaking country) on the exitrate, we see that this variable is greater than 1, indicatingthat for each year spent in Australia that there is a higherassociated exit rate from unemployment.

14 A description of the covariates used for thiscalculation is given in footnote 12.

15 To give an historical example of relative benefitlevels, according to the Department of Family andCommunity Services, in the year to June 2001, the‘Disability Support Pension’ and the ‘Parenting PaymentSingle’ benefit were $A402 per fortnight, while the‘NewStart Allowance’ (unemployment benefit) was$A322 per fortnight. Figures are for people aged 21–59 years living away from home. In addition, parentswith children may be eligible for supplementary familypayments and all benefit recipients may be eligible forrent assistance and other miscellaneous supplementarypayments.

16 The lower exit rates from unemployment toemployment for women with children and people withdisabilities may also be related to lower wages offeredto these groups (potentially because of discrimination).

17 The lagged unemployment experience variable is anindicator variable. If a person spent more than 25 percent of their time since formal schooling (prior to theircurrent spell of unemployment) in unemployment thenthe explanatory variable takes a value of 1, otherwisethe variable takes a value of 0.



capital on the job, as well as a separate negativeimpact from being unemployed and so the totalimpact would be larger than 27 per cent (becauseit would need to take account of the loss inemployment experience as well).

While there is evidence of lagged durationdependence, unobserved characteristics (such asunobserved motivation) may be correlated withboth past unemployment experience and exitsfrom unemployment in the current spell. There-fore, in this case it may be that unobserved char-acteristics, rather then previous spells ofunemployment, are associated with lower exitrates from unemployment. The key question forresearch then becomes did the characteristicsdevelop over the previous unemployment spell, ordid they exist prior to the beginning of that previ-ous spell?

(iv) How Robust Are the Results?First, when evaluating the robustness of the

models in Section IV, the joint Wald test of allcoefficients equal to 0 is rejected at the 0.1 percent level of significance in all estimations inTable 2. Now turning to the results from someresidual analysis, the Cox model fits the data wellat short durations (plot available from author onrequest).

The test of the proportional hazards assumption(using the Schoenfeld residuals) indicates theassumption is violated in the non-stratified estima-tion (also available from the author upon request).In particular, the reporting (to deal with the sys-tematically different exit rates for different partsof the month) and wave 1 dummies (takes a valueof 1 if the spell begins in wave 1) appear toexhibit significant non-proportionality. Stratifica-tion by wave 1 and reporting dummies in Coxpartial likelihood estimation has been undertakento control for this.18 When estimation is stratified bythe reporting and wave 1 dummies, the proportionalhazards assumption is no longer violated. Hencethe Cox estimation stratified by wave 1 andreporting dummies is the preferred estimation.

Martingale residual analysis highlighted that themodels fitted the data well and that the resultswere not being driven by outliers. A test forwhether the male and female data should bepooled was undertaken and it was found that the

coefficient on the interaction between the ‘male’and ‘children’ variables was statistically signifi-cant at the 1 per cent level. However, it was alsofound that the joint test of significance of theother interactions between explanatory variablesand ‘male’ was not accepted at the 20 per centlevel of significance. Hence, in estimation the dataare pooled but an interaction term between ‘male’and ‘children’ is included.

(v) How Do These Results Compare to Earlier Results from the Literature?

The results presented in Section IV areconsistent with the findings presented in Borland’s(2000) review of the Australian literature. Ingeneral, exit rates to employment are positivelyrelated to educational attainment and jobexperience and negatively related to non-wageincome and reservation wages.19

One interesting and different result from theearlier literature is that when employment experi-ence is added to the estimation, age becomesinsignificant. This suggests that the impact of thisdemographic characteristic on exit rates may bethrough the impact on employment experience,rather than through another transmission mecha-nism. Another interesting result from this paper isthat past unemployment is associated with a lowerhazard rate.

V Inclusion of Lagged Wage and Risk Variables in Analysis

The results presented so far have providedconfirmation that variables associated with higherwages and lower reservation wages are associatedwith increased exits from unemployment to employ-ment. The results of two extended models that useadditional variables from the wave 1 interview are nowpresented (see Table 3). Because these variablesare potentially endogenous to unemployment, thesemodels are estimated on a subsample includingonly spells starting after the wave 1 interview.

18 With stratified estimation the baseline hazard isallowed to vary for different values of the variable, butthe coefficients on the other variables are the sameacross different values of the stratification variable.

19 Using a similar base category to Chapman andSmith (1992), a coefficient of 0.356 on educationalqualifications and 0.470 on being Australian born (asopposed to being born in a non-English-speakingcountry) was found in this paper. Chapman and Smith(1992) find a coefficient of 0.272 on educationalqualifications and 0.298 on being Australian born. Thedifferences in size can be explained by the slightlydifferent definitions used in the two papers and thepositive coefficient on time spent in Australia for peopleborn in non-English-speaking countries in the currentpaper.



The first additional variable included is a meas-ure of financial risk. The second additional variableincluded is wages earned in the year prior to thewave one interview date. The mean lagged annualwage is $A14 060 (this includes people who hadno earned income over the previous year). Thepercentage of the year worked is also included inthe lagged wage estimation to examine the wageeffect separately from the percentage of yearworked effect.

Because risk and lagged wages are only availa-ble for analysis after the wave 1 interview, andbecause over half of the unemployment spells inthe calendar begin prior to the wave 1 interviewdate, the sample size is reduced in these estima-tions. For this reason, more caution should beused in interpreting the results. The results pre-sented in this section are again presented inhazard ratio format. The number of cases is givenin Table 3.

Table 3Comparison of Three Different Specifications (Cox) – Exits to Employment – Hazard Ratios

(1)StandardModel

(2)Risk

Model

(3)Lag Wage

Model

Means forspecification

(2)

Male 0.835 0.884 0.794 0.52[1.36] [0.89] [1.54]

Non-university qualification 0.989 0.989 1.022 0.44[0.11] [0.11] [0.19]

University qualification 1.408* 1.409* 1.102 0.15[2.43] [2.38] [0.61]

Children 0.749** 0.755** 0.748** 1.27[3.96] [3.70] [3.43]

Male*children 1.282** 1.261** 1.291** 0.61[2.89] [2.61] [2.64]

Marry 1.471** 1.459** 1.387* 0.43[3.28] [3.13] [2.49]

Disability 0.636** 0.655** 0.797 0.21[2.98] [2.73] [1.33]

Employment experience 1.040 1.030 0.992 12.84[1.56] [1.19] [0.29]

Employment experience2 0.999 0.999 1.000 295.16[1.29] [0.95] [0.05]

Lagged unemployment 0.637** 0.651* 0.651* 0.15[2.77] [2.57] [2.26]

Risk-seeking 0.685 0.576* 0.06[1.80] [2.33]

Lagged wage 1.011**[3.38]

14.06

Per cent of year worked 1.003 49.79[1.35]

Controls for country of birth yes yes yes NAand parental employment status

Observations 959 894 793 894Failures (exits) 453 423 367 423Log likelihood −2003.6 −1828.5 −1503.9 NAWald test 73.23 67.33 78.05 NA(Probability > χ2) (0.00) (0.00) (0.00)

The base is women, without qualifications, Australian born, unmarried, with no disability and with less than 25 per cent of labourmarket experience in unemployment. Standard errors adjusted for clustering on xwaveid. Cox partial likelihood has been stratified toadjust for monthly reporting in calendar and wave 1 effects. Estimation only undertaken on wave 2 data, because of endogeneity withexplanatory variables. Age and Age2 held constant. The means for male, non-university qualification and country of birth not includedhere are similar to the estimates presented in Table 4. Note the means for the wage variable are over a sample size of 791 rather than894, because of the number of missing values on this covariate. NA, not applicable.



The means of the variables for the extendedmodels with the smaller sample are presented incolumn 5 of Table 3. The means of all the vari-ables (university qualifications, marital status,disability, employment experience and lagged un-employment) are similar between Tables 2 and 3.

With the restricted sample, a key question iswhether the alternative sample leads to differentresults for the standard model. The major differ-ence between the standard models in Tables 2 andspecification 1 of Table 3 is that the coefficientsare less precisely estimated as the standard errorsare larger in Table 3 (because the estimations inTable 3 have a smaller sample size). In particular,the employment experience variable and the malevariable become insignificant. However, overallthe results from earlier estimations stand.

(i) Does a Person’s View of Risk Affect Their Exit Rates to Employment?

The paper now turns to investigate whether riskplays a role in the rate of exit from unemploymentto employment. Risk-seekers may be more likelyto decline a certain wage offer in the currentperiod, for the possibility of a higher uncertainwage offer in the future. Risk seekers would beexpected to have lower exit rates fromunemployment to employment (holding othercharacteristics constant) because they may bemore likely to turn down wage offers. On theother hand, risk-seekers may have higher post-unemployment wages.

The variable used to investigate risk-seeking inthis paper takes a value of 1 if an intervieweeindicates that they take substantial or above aver-age financial risks, otherwise the variable takes avalue of 0. Two key caveats should be noted wheninterpreting this risk-seeking variable. First, peoplewho are financial risk-seekers may not neces-sarily be risk-seekers when it comes to jobacceptance (risk-seeking in one domain of some-one’s life may not translate into risk-seeking in alldomains). Second, the risk-seeking variable iscorrelated with wealth and therefore the relation-ship between risk-seeking and unemployment maybe observed because wealth is related to both financialrisk-seeking and unemployment spell length.

Risk-seekers are more likely to be male (68 percent of males compared to 52 per cent offemales), with greater income (37 per cent earnedmore than $A30 000 in the previous year com-pared to 25 per cent of non-risk-seekers), univer-sity educated (24 per cent compared to 13 percent) and with more employment experience (41

per cent with more than 20 years’ employmentexperience compared to 33 per cent of non-risk-seekers).

Specification 2 from Table 3 presents the resultsfrom the estimation when the risk-seeker variableis included in estimation. The coefficient on therisk-seeking variable is less than 1 and thereforeindicates, as expected, that people who are preparedto take financial risks are more likely to delaytheir exit from unemployment to employment towait for a higher wage offer (hence they have alower exit rate).20 This result is not significant atthe 5 per cent level of significance, although thez-statistic is 1.80 (however, when the lagged wagevariable is added in the next section, the risk variablebecomes significant at the 5 per cent level). Thisresult suggests that on the supply side, a key issueis how the unemployed view future wage offers(and in turn how they view risk and uncertainty).

To the author’s knowledge there appears tohave been very little, if any, consideration of riskaversion in the unemployment duration literature.However, given the possibility that perceptions ofrisk may influence the choice about whether toaccept a certain wage offer in the current periodor to decline and wait for an uncertain, but poten-tially higher, wage in a future period, this may bean interesting area of future research.

(ii) Does Last Year’s Wage Explain the Variation in Exit Rates?

The role that previous wages may have on exitrates from unemployment is now investigated (seespecification 3 from Table 3). This variable maybe correlated with the wage distribution thatpeople face in the current period. Both period’swage distributions are likely to be affected by thesame unchanging unobservable characteristics,although some other characteristics may changesomewhat between years. This gives a furtherindication of the role that wage offers play inunemployment duration, beyond the indicatorspresented in Section IV. As noted above, thepercentage of the year that individuals work isincluded, to control for the wage effect separatelyfrom the percentage of the year worked effect.

20 An alternative measure of risk-seeking behaviour iswhether a person has ever smoked. This was included inestimation and while it was associated with a lowerhazard rate, it was not found to be significant. Anotherpoint to note is that the risk-seeking variable ismeasured at the start of (or prior to) the spell and thusthe length of spell will not affect it.



The coefficient on wages earned in the previousyear is significant at the 1 per cent level and indi-cates that the higher the wage earned in the previ-ous year (holding the percentage of the yearworked constant), the higher the exit rate fromunemployment. If there is a high correlationbetween last year’s wage offers and this year’swage offers, then this coefficient is consistentwith higher wage offers being associated withhigher exits from unemployment.

Interestingly, when this wage history variable isincluded, the coefficients on most of the wage-related variables (university qualifications, disabil-ity, employment experience) become insignificantat the 5 per cent level and become smaller in sign(or switch sign). These correlations betweenwages in the previous year, exit rates to employ-ment and the wage-related variables provide fur-ther evidence that these variables affect exit ratesthrough the wage offer distribution.

To check if the drop in significance is causedby near multicolinearity between the lagged wagevariable and the wage related variables, the corre-lation coefficients were examined. There are highcorrelations between age and experience, and rela-tively high correlations between ‘lagged wage’and ‘employment experience’ (but still a pairwisecorrelation coefficient of less than 0.5). Overall,the pairwise correlation coefficients do not indi-cate a problem with near multicolinearity.

VI The Underlying Baseline HazardA usual finding when negative duration dependence

is observed is that it is not possible to distinguishwhether longer spells result in lower exit rates, orwhether there is unobserved heterogeneity leading to‘low exiters’ remaining in unemployment for longer.However, the shape of the baseline hazard is impor-tant, because falling exit rates by duration (holdingcharacteristics constant) indicate that unemploymenthas a scarring effect, which suggests adjustmentmay come with medium-term costs. Oneadditional point to note is that steady exit rateswith duration may occur because of factorsoffsetting unemployment scarring effects.21

(i) Do Hazard Rates from Unemployment to Employment Decline with Duration?

The results from this paper show that the exitrates from unemployment to employment do notinitially appear to decline with duration (for thefirst 4 months), but then hazard rates appear todecline relatively sharply.

Figure 3 shows the baseline hazard curves forthe raw data, the Weibull, the piecewise constant(and the confidence bands for the piecewise con-stant hazard). The Weibull has been the more tra-ditional of the measures to examine the baselinehazard in Australia; however, because it summa-rises duration dependence into one parameter,interest in this measure has waned. The preferredmeasure of the time pattern of hazard rates is thepiecewise constant hazard, which shows an ini-tially flat hazard curve before declining relativelysharply after 4 months. However, it should benoted that the only statistically significant timeparameter is on the 1 year+ duration effect.

The sharp decline in the hazard rate after4 months may reflect unobserved heterogeneity orstate dependence effects.22 Thus, as stated above,there is interest in understanding whether theheterogeneity developed prior to the start of the spell(unobserved heterogeneity), or during the spellitself (state dependence). If there is state depend-ence, the policy response (if any) will be deter-mined by the drivers of the state dependence(human capital depreciation, discrimination byemployers, or discouragement).

The non-monotonic hazard may occur becausepeople are more likely to forget short spells ofunemployment (<1 month). If the issue is areporting one then the ‘true’ distribution may bemonotonic (and a Weibull baseline hazard poten-tially could be used). However, estimation shouldthen take this into account and investigation hasshown that the estimates of the Weibull shapeparameter are sensitive to the inclusion of the firsttwo periods data.23 There may also be genuine

21 Machin and Manning (1999) highlight that thepresence of unemployment benefits that decline withduration and active labour market policies targeted atlong-term unemployed (amongst other factors) may leadto rising exit rates with duration. Thus, even minimalobserved negative duration dependence may be asso-ciated with scarring effects of unemployment (becausescarring is offset by institutional factors).

22 One method to control for the unobserved hetero-geneity is to integrate it out by restricting the distributionof the unobserved heterogeneity to be of a certainshape. Some studies have concluded that restrictingthe unobserved heterogeneity distribution is very sensitiveto the restrictions used and is not necessarily better thanexcluding the term where a flexible baseline hazard isused (see Narendranathan & Stewart, 1993).

23 Where the first two periods of data are excluded,the implied fall in the exit rate between 1 month and12 months is 48 per cent compared to 19 per cent whenthe first two periods of data are included.



reasons to believe that there are flat hazard ratesat the beginning of a spell. One example is thatjob search may require some learning (for examplecurriculum vitae (CV) creation, identification ofemployers to approach) and thus people’s job searchmay initially improve (leading to an initially ris-ing hazard rate), before any negative effects oflong-term unemployment occur.

(ii) How Much of the Decline in the Hazard is Explained by the Covariates?

The previous subsection showed that in generalthere appeared to be a downward slope in thebaseline hazard after 4 months of elapsedduration, but it is not possible to know whetherthis shape is because of state dependence orunobserved heterogeneity. A natural next questionis how much of the falling hazard rates areexplained with the variables included in theestimation (with a piecewise constant baselinehazard).

Given that demographic information, unemploy-ment history, educational qualifications, whetherAustralian born and a variety of family variableswere included, it may be expected that a signifi-cant part of the variation would be explained. Onthe other hand, it may be that the unobserved

characteristics (such as motivation, natural ability)are important drivers of the wage distribution andof exit probabilities.

Observed characteristics explain less than 15per cent of the decline in the hazard rate observedbetween 1 month and 12 months in the piecewiseconstant model (see Table 4). Table 4 shows howthe baseline hazard varies with duration. Usingthe duration parameters from the piecewise base-line hazard without covariates, the hazard declinesby 46 per cent over the first 6 months and a fur-ther 27 per cent over the second 6 months, whilewhen covariates are added, the hazard declines by42 per cent over the first 6 months and a further23 per cent the second 6 months.24 As observablecharacteristics only ‘explained’ a small proportionof the decline in the hazard rate, it is likely thatunobservable characteristics are particularlyimportant and that some scarring is occurring aswell.

This result highlights that, even holding con-stant employment experience, qualifications,

Figure 3Comparisons of Hazard Curves – Raw, Piecewise and Weibull Baseline Hazard

24 Note that the value of the hazard will vary with thevalues of the covariates, but the relative decline in thehazard will be constant across estimates.



previous unemployment experience and otherfactors, there is a large difference in exit ratesbetween people at longer durations than those atshorter durations. This also highlights how largethe unobserved component would have to be forthere to be no drop in the hazard rate with dura-tion for individuals.

VII ConclusionThis paper found that variables associated with

increased wages are associated with higher exitrates from unemployment to employment. Thesewage variables include employment experience,and educational qualifications. It was also foundthat variables expected to lower wages (throughimpacts on productivity or discrimination) areassociated with lower exit rates (non-English-speaking COB, disability, lagged unemployment).Variables that may be associated with increasednon-wage income (children, disability), and thushigher reservation wages, are associated withlower exit rates from unemployment.Reassuringly, results were robust to a variety ofdifferent estimation methods and specifications.

It was found that people who are risk-seekerswere more likely to experience longer spells ofunemployment. This finding is consistent withrisk-seekers being more likely to decline a certaincurrent wage for a potentially higher future uncertainwage. This is an interesting result that deservesfurther investigation, because the key decision ajob seeker may face is the choice about whetherto accept a certain wage offer in the currentperiod or to decline and wait for an inherentlyuncertain wage in a future period.

It was also found that the shape of the hazardcurve may be flat at shorter durations, beforedeclining after 4 months (as illustrated by thepiecewise constant baseline hazard). Finally, whilethe duration dependence observed may be theresult of unobserved heterogeneity or statedependence, the paper showed that less than one-fifth of the decline in the hazard observed isexplained by the inclusion of explanatory vari-ables. This suggests that either unobserved charac-teristics are potentially important or that somescarring is occurring.

REFERENCES

Arulampalam, W. (2001), ‘Is Unemployment Really Scar-ring? Effects of Unemployment Experiences on Wages’,The Economic Journal, 111, F585–F606.

Arulampalam, W., Booth, A. and Taylor, M. (2000),‘Unemployment Persistence’, Oxford Economic Papers,52, 24–50.

Barrett, G. (2000), ‘The Effect of Educational Attainmenton Welfare Dependence: Evidence from Canada’, Jour-nal of Public Economics, 77, 209–32.

Borland, J. (2000), ‘Disaggregated Models of Unemploy-ment in Australia’, Melbourne Institute Working PaperNo. 16/00, Melbourne Institute of Applied Econmoicand Social Research, University of Melbourne, Vic.

Brooks, C. and Volker, P. (1986), ‘The Probability of Leav-ing Unemployment: The Influence of Duration, Destina-tion and Demographics’, The Economic Record, 62,296–309.

Chapman, B. and Smith, P. (1992), ‘Predicting the Long-Term Unemployed: A Primer for the CommonwealthEmployment Service’, in Gregory, R.G. and Karmel, T.(eds), Youth in the Eighties – Papers from the AustralianLongitudinal Survey Research Project, CEPR, Canberra,ACT; 263–82.

Clark, A. and Oswald, A. (1994), ‘Unhappiness andUnemployment’. The Economic Journal, 104, 648–59.

Han, A. and Hausman, J. (1990), ‘Flexible Parametric Esti-mation of Duration and Competing Risk Models’, Jour-nal of Applied Econometrics, 5, 1–28.

Heckman, J. and Singer, B. (1984), ‘A method for mini-mising the impact of distributional assumptions ineconometric models for duration data’, Econometrica,52, 271–320.

Hui, W. (1991), ‘Reservation Wage Analysis of Un-employed Youths in Australia’, Applied Economics, 23,1341–50.

Knights, S., Harris, M. and Loundes, J. (2002), ‘Dynamicrelationships in the Australian Labour Market: Hetero-geneity and State Dependence’, The Economic Record,78, 284–98.

Lancaster, T. (1990), The Econometric Analysis of Transi-tion Data, Cambridge University Press, Cambridge.

Machin, S. and Manning, A. (1999), ‘The Causes and Con-sequences of Long-term Unemployment in Europe’, in

Table 4How Do Hazard Rates Vary with Duration?

Simple piecewise (per cent)

Full piecewise (per cent)

1 month (3 periods) 100 1003 months (9 periods) 99 1066 months (18 periods) 54 5812 months (36 periods) 27 35Percentage of 12 months +

effect explained by explanatory variables

11.3

The simple model excludes all explanatory variables. The fullmodel includes all explanatory variables included inspecification 2 of Table 2.



Ashenfelter, O. and Card, D. (eds), Handbook of LaborEconomics, Vol. 3C, North Holland, Amsterdam; 3085–139.

Meyer, B. (1990), ‘Unemployment Insurance and Un-employment Spells’, Econometrica, 58, 757–82.

Mortensen, D. (1986), ‘Job Search and Labour MarketAnalysis’, in Handbook of Labor Economics, Vol. 2,North Holland, Amsterdam; 849–919.

Narendranathan, W. and Stewart, M. (1993), ‘Modellingthe Probability of Leaving Unemployment: CompetingRisks Model with Flexible Baseline Hazards’, AppliedStatistics, 42, 63–83.

Ruhm, C. (1991), ‘Are Workers Permanently Scarred byJob Displacements?’ American Economic Review, 81,319–24.

van den Berg, G., van Lomwel, A. and van Ours, J. (2003),‘Nonparametric Estimation of a Dependent CompetingRisks Model for Unemployment Durations’, IZA Dis-cussion Paper Number 898, Institute for the Study ofLabor, Bonn, Germany.

Winkleman, L. and Winkleman, R. (1998), ‘Why Are theUnemployed so Unhappy: Evidence from Panel Data’,Economica, 65, 1–15.



AppendixVariable Descriptions

Variable Description25

Age If spell begun prior to calendar then age at wave 1 interview date minus 2, if spell begun prior to wave 1 interview date then age at wave 1 interview date minus 1, if spell begun after wave 1 interview date then age at wave 1 interview date minus 1

Wave 1 = 1 if spell begun in wave 1 calendar, = 0 otherwiseMale = 0 for female, = 1 for maleRegional unemployment Regional unemployment rate (where region is capital city vs rest of state)SES decile Regional SEIFA 2001 Decile of Index of relative socioeconomic disadvantage

Refers to the region the person was immediately prior to beginning spell. Where the person moved between unemployment spell begin date and interview date the data will be missing. Regional disaggregation is finer than that used for regional unemployment.

Non-English speaking COB = 1 if born in a non-English-speaking country, = 0 otherwiseLength resident (non-English) Length of time spent in Australia for people born in non-English-speaking countriesEnglish-speaking COB = 1 if born in an English-speaking country (e.g. Australia), = 0 otherwiseLength resident (English) Length of time spent in Australia for people born in English speaking countries

(excluding Australia)Non-university qualification = 0 if no qualifications recorded, = 1 if non-university qualifications recordedUniversity qualification = 0 if no university qualifications, = 1 if university qualifications recordedFather’s unemployment status = 0 if father not unemployed for a total of 6 months growing up,

= 1 if father not unemployed for a total of 6 months growing upMother’s employment status = 0 if mother not in employment when respondent aged 14 years,

= 1 if mother in employment when respondent aged 14 yearsChildren Number of children as at the beginning of unemployment spell. Note: children’s

age at the interview date was used in calculationMarry Marital status as at the beginning of the unemployment spell (either married/de

facto married, or not married)Disability = 0 if no long-term disability recorded at wave 1 interview, = 1 if long-term

disability recorded at wave one interviewEmployment experience Number of years in employment prior to unemployment spellEmployment experience2 Number of years in employment squared prior to unemployment spellLagged unemployment = 1 if person spent more than 25 per cent of time since school in unemployment,

= 0 otherwiseAge2 Age squaredMale*children Interaction between male and number of childrenMale*father Interaction between male and father’s unemployment statusmark1 Observation recorded in first third of monthmark2 Observation recorded in second third of monthLagged wage Respondents wage (in thousands) in year prior to first interview datePer cent past year in job Percentage of the past year that the respondent has been in paid employment

(asked at wave 1 interview date)Risk-seeking = 1 if respondent takes substantial or above average financial risks, = 0 otherwise

(asked at wave 1 interview date)

COB, country of birth.

25 As at wave one interview date unless otherwise specified.

Documents

Explaining Unemployment Duration in Australia