Unemployment Duration in the United Kingdom An Incomplete Data Analysis Ralf A. Wilke University of Nottingham
1. Motivation The determinants for the length of unemployment and out of the labour market duration has high political relevance: Public spending for unemployment benefits/jsa, publicly sponsored training and regional labour market policies Spending for public administration (Jobcentres) Education decisions A lot of theoretical and empirical work has been done to analyse determinants of unemployment duration: In particular influence of unemployment insurance/benefits
While from a theoretical point of view the results are often clear, empirical studies regularly contradict. This can be for several reasons: Choice of the data set (type of data, period etc.) Different sets of covariates/regressors Specification or choice of the econometric model Random sampling errors Measurement error in the data Definition of Unemployment Incomplete Information in the data
Several UK data sets and methods have been used to analyse unemployment duration: Family Expenditure Survey (survey, stock, yearly) British Labour Force Survey (survey, stock, quarterly, small) British Household Panel Survey (survey, panel, monthly, large) JUVOS administrative unemployment count data (administrative, spell, daily, huge)
Empirical Literature: Duration Models: Boeheim/Taylor (2000): BHPS, 1991-1997, discrete time hazard rate model, competing risks Kalwij (2004): JUVOS, 1982-1998, multiple spell mixed proportional hazard model, single risks McVicar/Podivinsky (2002): JUVOS, 1995-2001, aged 16-30, discrete time hazard rate model, competing risks, evaluation of the new deal programme Analysis of the stock: Blackaby/Bladen-Hovell/Symons (1991): FES, 1980-1986 Manning (2005): LFS, 1985-1998, evaluation of JSA It is difficult to compare results because of different sets of covariates/regressors or dependent variable.
Further references There are several variants of the JUVOS which have been used for applied work on unemployment: Merged with employment/tax records (Gregory/Jukes, 2001) Merged with other admin data about training (Dolton/ONeill, 2002) Merged with survey data (Bryson/Kasparova, 2003) This data is not freely accessible and access criteria are unclear. No public documentation of the data available.
Contribution of this paper: Defines concepts of unemployment duration in the JUVOS taking into account data limitations. Previous work considered one claim period as one unemployment spell. Creates a competing risks data structure although the information in the data comes from one administrative source only. Performs an econometric analysis of the determinants of unemployment duration. Important result patterns are sensitive with respect to the definition of unemployment.
2. Data 5% cohort of unemployment compensation claimants in the UK. (Scientific Use File) Daily information about claim periods for unemployment benefits/jsa from the 1980s until summer 2007. Data is used for operational processes in the DWP and Jobcentre Plus (main purpose is not academic work). Contains basic individual characteristics: age, sex, marital status, occupation, region. Variable indicating the end reason of a claim spell.
Example of the data structure: one individual with two claim spells Only claim periods in the data: Unemployment without receipt of UB/JSA is not recorded. This implies that the true unemployment duration is not observable. Long term unemployed with employed spouse underreported as UB is means tested. (Machin, 2004, Manning, 2005)
The reason for leaving variable provides important information about the post claim labour market state of the unemployed Use reasons for leaving variable to determine post unemployment labour market state. Often not unique; Missing values Construct a competing risks data structure Create a set of employment history variables Use data after 1996 (JSA)
3. Definition of Unemployment 5 different concepts to define unemployment duration in the JUVOS. merge subsequent claim periods if conditions are met. Concept 1: lower bound The resulting unemployment duration should not include periods other than unemployment. Concept 5: upper bound Claim periods and gaps are not combined if information in between is for sure employment or ALMP. Resulting spells possibly include nonemployment and employment periods.
4. Empirical Analysis Number of unemployment spells and distribution of destination states.
Median unemployment duration Analyse the distribution of unemployment duration: Unconditional (Kaplan-Meier, 1958) Conditional (Cox, 1972)
Result patterns for duration until job finding: Uncertainty due to random sampling is much less important than uncertainty due to missing information in the data. Older unemployed require more time to find a job. Mainly driven by longer long-term unemployment periods of females. Singles are slower. Lower skilled occupations are related with longer unemployment. Previous unemployment is not clearly associated with longer unemployment. Past participation in ALMP is associated with long unemployment (weak labour market type). Previously successful transition to employment is a strong positive sign about the type of the individual. Worsening situation in 2006 and 2007.
Results patterns for duration until ALMP assignment: younger unemployed more quickly assigned Older unemployed and married females less quickly Unemployed with low skilled occupations are more quickly assigned Previous unemployment and past ALMP participation lead to much quicker assignment (bad labour market types) Successful previous job finding leads to slower assignment Increase in ALMP assignment after 1998 (New Deal)
5. Conclusion The empirical results suggest that data preparation steps have a strong influence on several resulting statistics and results. Missing information and partial identification of the unemployment duration cause more uncertainty for the results than random sampling errors. However, several interesting and robust result patterns help in better understanding determinants of unemployment duration. Merged administrative data (employment, other benefits) would contribute to the precision of the results. Should be made accessible to independent research
Kaplan Meier Survival Function Estimates
Cox regression results from six estimations: failure employment
Cox regression results from six estimations: failure employment
Cox regression results from six estimations: failure employment
Cox regression results from four estimations: failure ALMP
Cox regression results from four estimations: failure ALMP