Estimation of apparent and inactive unemployment by structural time series models

Transcription

1 Faculty of Informatics and Electronic Economy Poznań University of Economics and Business Summary of dissertation Estimation of apparent and inactive unemployment by structural time series models Author: mgr Kamil Wilak Supervisor: dr hab. Elżbieta Go lata, prof. nadzw. UEP Poznań 2018

2 Summary The present thesis is an attempt to fill a gap concerning research on the scale of the phenomenon of apparent unemployment, which involves people registering as unemployed with district labour offices (DLO) in spite of not meeting the criteria of the official definition of an unemployed person. Under current regulations, a person can register as unemployed if he or she is not engaged in paid employment, is looking for a job, and is ready to start work. However, not all persons registered as unemployed meet the above criteria. Some of them are not interested in finding a job and only register to take advantage of the unemployment benefit or health insurance. In the thesis, the author distinguishes two groups of persons that do not meet the definition of an unemployed person: apparently unemployed persons and inactively unemployed persons. The former category includes people registered as unemployed but engaged in paid employment (Kwiatkowski, 2013). People classified as inactively unemployed are registered as unemployed and are economically inactive, i.e. neither willing nor ready to start work, nor looking for a job. The main goal of the thesis is to estimate the level of apparent and inactive unemployment in Poland. The quantities to be estimated are the number of apparently and inactively unemployed and the corresponding percentages of the total number of persons registered in DLOs. The estimation is to be performed for domains defined by cross-classifying sex, age group (18-24, 25-34, 35-44, 45-59/64), and province. The reference period includes all quarters between 2001 and The main goal of the thesis was broken down into specific objectives: (1) to identify apparent and inactive unemployment, (2) to ensure consistency between different sources of information about registered unemployment, (3) to estimate bias and autocorrelation due to the rotation panel design in the Labour Force Survey, (4) to construct structural time series models and (5) to assess the quality of obtained estimates. The study presented in the thesis was aimed at verifying four research hypotheses: (a) apparently and inactively unemployed persons account for a large percentage of registered unemployment, (b) apparent and inactive unemployment varies considerably in terms of demographic variables, (c) apparent and inactive unemployment is characterised by considerable spatial variation, (d) estimates of apparent and in- 1

3 active unemployment based on structural time series models are of better quality than direct estimates. So far, studies of apparent and inactive unemployment have only provided annual estimates of the phenomenon at the national level (Czapiński i Panek, 2014). This thesis is an attempt to provide more detailed estimates of apparent and inactive unemployment not only by selecting finer domains of interest but also by increasing the frequency of observation. The obtained results illustrate the development of the phenomenon in time and help to identify the demographic structure of people representing each of the two categories of unemployment. The level of apparent and inactive unemployment was estimated using unit-level data from the LFS. In order to distinguish between apparently and inactively unemployed in the LFS dataset, the author proposed an identification procedure based on definitions recommended by the International Labour Organisation (GUS, 2012). Direct estimation is the starting point for estimating the level of apparent and inactive unemployment. To overcome the inconsistency between LFS-based estimates and unemployment registration data from DLOs, the calibration approach was adopted (Deville i Särndal, 1992; Lundström i Särndal, 2001). Weights obtained after applying the calibration procedure were used to construct the estimators. There is a considerable variation in the percentage of apparently unemployed in terms of sex. Men are more likely to be categories as apparently unemployed than women. The average percentage of apparently unemployed for men in was 13.1%, compared to 7.7% for women. Even more pronounced differences exist when the place of residence is taken into account. The average percentage of apparently unemployed in rural areas is equal to 16%, compared to 5.8% in urban areas. There are also marked differences between men and women in terms of inactive unemployment. The average percentage of inactively unemployed for women in the reference period is 35.2%, while for men it is equal to 19.1%. When applied to the economic-demographic domains defined in the study, calibration estimation turned out to yield estimates of acceptable precision. The lowest variation in estimates of apparent unemployment was observed in rural domains, where CV for the reference period was on average equal to 6.4%. The 2

4 best estimates of inactive unemployment were obtained for women, where the average CV value was 3.3%. In comparison, direct estimates at the province level are characterised by a much higher variance. The biggest variation is observed in Zachodniopomorskie province, where CV amounted to as much as 32.2%. In order to improve the precision of estimates of apparent and inactive unemployment across provinces, special attention was paid to the impact of the rotation panel design in the LFS on direct estimates. In the literature, two problems are typically recognised in this context: rotation group bias (RGB) and sampling error autocorrelation (SEA) (Bailar, 1975). Estimates obtained on the basis of subsequent subsamples of respondents participating in the survey for the first, second, third and fourth time (in consecutive waves) differ considerably. These differences are particularly evident in the case of apparent unemployment, where estimates based on the first wave are systematically higher than those obtained in the following waves. However, based on the analysis conducted by the author, it was not possible to determine which of the survey waves contained data of the best quality. This question would require additional investigation. The second problem associated with the rotation panel design was examined using the method of estimating sampling error autocorrelation coefficients proposed by Pfeffermana Pfeffermann i in. (1998), which was adjusted by Wilak (2015) to fit the sampling design used in the LFS. The obtained results indicate that there is significant first-order sampling error autocorrelation in direct estimates, while for most domains there is also fourth-order autocorrelation. Given the high variance of direct estimates across provinces, apparent and inactive unemployment was estimated using methods of small area estimation (Rao i Molina, 2015). Based on the review of literature, the actual method used was an indirect estimation approach known as borrowing strength over time (Pfeffermann, 2002; Schaible, 2013; Van den Brakel i Krieg, 2015). The approach involves using data from previous surveys to construct time series models, which exploit the fact that both phenomena of interest exhibit trend and seasonality, and LFS data can be used to generate such models. The development of the two types of unemployment over time was described by means of structural time series (STS) models (Harvey, 1989; Durbin i Koopman, 2012). The structure of these models accounts for trend and 3

5 seasonality, as well as variance and autocorrelation of direct estimates, which are modelled. The constructed STS models were expressed as state space models and could therefore be estimated by applying the Kalman filter and smoothing (Koopman i Durbin, 2000). Consistency between indirect estimates and direct estimates at a higher level of aggregation was achieved by applying the statistical benchmarking procedure (Denton, 1971). The quality of indirect estimation was evaluated in terms of precision and bias. For one thing, the indirect estimates are less variable than direct estimates. The largest gains in precision expressed by CV values were obtained for apparent unemployment in Opolskie province (on average by 38.2%), while in the case of inactive unemployment, in Kujawsko-pomorskie province (on average by 31.2%). The use of indirect estimation made it possible to eliminate outliers which occurred in series of direct estimates for some provinces. The least variable estimates of apparent unemployment were obtained for Podlaskie province, where the average value of CV was 9.6%. In the case of inactive unemployment, estimators with the lowest level of variation were obtained for Warmińsko-mazurskie province, with mean CV equal to 5.9%. The level of estimation bias was assessed using three diagnostic procedures: the bias diagnostic, the coverage diagnostic and the goodness-of-fit diagnostic (Brown i in., 2001). Unfortunately, results of these procedures were not conclusive. Results of the bias diagnostic indicate that indirect estimates are biased at the province level. However, based on coverage diagnostic results, the indirect estimates can be regarded as unbiased. Finally, when the goodness-of-fit diagnostic is taken into account, it can be concluded that estimates of apparent unemployment are unbiased, except for one province, while estimates of inactive unemployment are biased for 9 out of 16 provinces. By achieving each of the five specific objectives, the author managed to accomplish the main goal of the thesis, i.e. to estimate apparent and inactive unemployment in Poland. All four research hypotheses were confirmed. In particular, it was demonstrated that apparent and inactive unemployment account for a large percentage of registered unemployment (first hypothesis): in some domains of interest the two kinds of unemployment together made up half of registered unemployment. The second hypothesis, which states that apparent and inactive unemployment varies 4

6 considerably in terms of demographic variables, was also corroborated: the most pronounced differences are associated with sex. Women are more likely than men to be inactively unemployed, while a reverse pattern can be observed in the case of apparent unemployment. The results also confirm the third hypothesis about the spatial differentiation of the two types of unemployment. Spatial correlations are particularly evident in the case of apparent unemployment, which is considerably higher in the eastern part of Poland than in the western part. The fourth hypothesis, which states that estimates of apparent and inactive unemployment based on structural time series models are of better quality than direct estimates, was only partially confirmed. Taking into consideration the measures of precision expressed in terms of variance, the indirect estimates clearly outperform direct estimates; however, results of bias diagnostics are inconclusive. In conclusion, it must be admitted that the thesis does not exhaust the subject. Future research will focus on the following problems that need to be addressed: (i) detailed analysis of bias due to the rotation panel design of the LFS, (ii) construction of multivariate STS models to estimate apparent and inactive unemployment, (iii) implementation and application of the benchmarking procedure proposed by Pfeffermann i Tiller (2006), (iv) estimation of apparent and inactive unemployment for the following years. 5

7 Literatura Bailar, B. A. (1975). The effects of rotation group bias on estimates from panel surveys. Journal of the American Statistical Association, 70(349): Brown, G., Chambers, R., Heady, P. (2001). Evaluation of small area estimation methods an application to unemployment estimates from the UK LFS. In Statistics Canada Symposium, Achieving Data Quality in a Statistical Agency: A Methodological Perspective. Czapiński, J. Panek, T. (2014). Wykluczenie spo leczne. In Diagnoza Spo leczna Warunki i jakość życia Polaków, pages Warszawa. Denton, F. T. (1971). Adjustment of monthly or quarterly series to annual totals: An approach based on quadratic minimization. Journal of the American Statistical Association, 66(333): Deville, J.-C. Särndal, C.-E. (1992). Calibration estimators in survey sampling. Journal of the American statistical Association, 87(418): Durbin, J. Koopman, S. J. (2012). Time series analysis by state space methods. Oxford University Press,. GUS (2012). Aktywność Ekonomiczna Ludności Polski. IV kwarta l G lówny Urz ad Statystyczny, Warszawa. Harvey, A. C. (1989). Forecasting, structural time series models and the Kalman filter. Cambridge University Press. Koopman, S. J. Durbin, J. (2000). Fast filtering and smoothing for multivariate state space models. Journal of Time Series Analysis, 21(3): Kwiatkowski, E. (2013). Ekonomiczne aspekty zasi lków dla bezrobotnych. In Góral, Z., editor, Bezrobocie i polityka zatrudnienia, pages Wolters Kluwer Polska. Lundström, S. Särndal, C. E. (2001). Estimation in the Presence of Nonresponse and Frame Imperfection. Pfeffermann, D. (2002). Small Area Estimation-New Developments and Directions. International Statistical Review, 70(1):

8 Pfeffermann, D., Feder, M., Signorelli, D. (1998). Estimation of Autocorrelations of Survey Errors With Application to Trend Estimation in Small Areas. Journal of Business & Economic Statistics, 16(3): Pfeffermann, D. Tiller, R. (2006). Small Area Estimation with State Space Models Subject to Benchmark Constraints. Journal of the American Statistical Association, 101(476): Rao, J. Molina, I. (2015). Small Area Estimation. John Wiley & Sons, Inc, Hoboken, NJ, USA. Schaible, W. L. (2013). Indirect estimators in US federal programs, volume 108. Springer Science & Business Media. Van den Brakel, J. A. Krieg, S. (2015). Dealing with small sample sizes, rotation group bias and discontinuities in a rotating panel design. Survey Methodology, 41(2): Wilak, K. (2015). Autokorelacja b ledów oszacowań w Badaniu Aktywności Ekonomicznej Ludności. Wiadomości Statystyczne, (6):