Break estimation in the Norwegian LFS due to the 2021 redesign. Documentation of the methods and some results

Abstract

In 2021 The Norwegian Labour force survey (LFS) went through a substantial redesign in accordance with the new regulation for integrated European social statistics (IESS). To ensure coherent labour market time series for the main indicators, the redesign's impact is modelled to make back calculated estimates adjusted for possible breaks due to the 2021 LFS-redesign. We pursue a structural time series approach in the tradition of Pfeffermann (1991), van den Brakel et al. (2009, 2015) and Elliott and Zong (2019). Breaks are estimated for the number of employed and unemployed persons. In addition to the 8 waves with monthly LFS data for employed and unemployed persons, we also include auxiliary time series for registered number of employees and unemployed, respectively, in the preferred models. The structural time series model contains unobserved components for trend, seasonality and irregularity, all of which are assumed to be the same for all waves. A smooth trend model is used. In addition, we account for rotation group bias and the autocorrelation structure brought about by the rotating panel design, as well as sampling error heterogeneity caused by changes in the (net) sample sizes over time. The auxiliary time series are decomposed into components for trend, seasonality and irregularity. Information from the auxiliary variables is used to obtain more precise break estimates by allowing the two trend components' error terms to be correlated. To correct for the effect of the COVID-19 pandemic, we allow the hyperparameters for the trend to be higher during the pandemic. We do this to counteract the contaminating effects the pandemic has on the estimate of the structural break following the redesign of the LFS. The effect of the redesign is modelled as separate level shifts for each wave. The final break estimates are based on modelling time series from 2006M1-2021M10. Information from a parallel survey with the new questionnaire carried out in the last quarter of 2020 for a small sample is also utilized in the time series model. The time series are modelled for four main domains: gender cross-classified by age 24 and below / 25 and above. The domain-specific break estimates are given as the average of the estimates of the break parameters for the 8 waves. These break estimates are divided into sub-groups using monthly time-varying sub-group splitting factors assuming a proportional distribution of the breaks. We find a positive break estimate of about 22,000 employed and 5,000 unemployed persons aged 15-74, but only the break estimate for employed persons is significant. The break estimates relative to the population are used to produce back-calculated monthly and quarterly time series for main indicators for the years 2006-2020.