New estimation methodology for the Norwegian Labour Force Survey
Labour Force Survey (LFS) is an important source of the labour market statistics that provides information about the participation of people aged 15 and over in to the labour market and people outside of the labour market. It is a rotating panel sample survey that is carried out in accordance with the European Union (EU) Council Regulation. Statistics produced are subject to both sampling and non–response errors. Sampling errors are monitored through standard errors, which are provided alongside with the point estimates for the key variables. In that respect, finding an efficient estimator is one of the main goals for the LFS. This requires data sources that includes good auxiliary variables. Thus we aim to find an estimation methodology which better utilises the auxiliary information in the light of a new available data source, namely A–ordningen. In this regard, we compare the regular generalized regression estimator (GREG) and the (multiple) model–calibration estimator, which has been shown to be optimal among a class of calibration estimators, in terms of efficiency by using the Norwegian LFS data. Standard errors are estimated by using the Jackknife linearisation (JL) variance estimator. Overall, for the data used, the (multiple) model–calibration estimators have been more efficient than than the GREG estimators. Thus the former has been chosen to be used in the production of the Norwegian labour force statistics.
Non-response may lead to biased estimates if it is not properly handled in the estimation under a non–uniform response mechanism (i.e. not missing completely at random (MCAR)). We discuss two types of weighting procedures. One of them involves a separate step for non–response adjustment, and the other one handles with non–response as a part of calibration. We have observed, for the data used, that the two–step estimators have provided higher standard errors without reducing non– response bias more. Thus it has been decided to use a one–step (multiple) model– calibration estimator in the production of the Norwegian labour force statistics.
Equal– and unequal–weighted averages of monthly estimates have also been compared in order to investigate the effects of each on quarterly estimates. The former was used by the previous estimation methodology (see Section 4). The latter is proposed being used in the new estimation methodology (see Section 12.4).
The new estimation methodology has been examined with regards to whether or not it satisfies the EU precision requirements. The requirements are met for the data used.
A stratified one–stage cluster sampling is used to select sample units for the Norwegian LFS. We observe that the cluster effect may be ignored in the variance estimation if good auxiliary variables are used in the estimation. This facilitates the computation of variance estimates, especially for changes in statistics over time, for which the variance estimation may be more tedious in rotating panel surveys. The cluster effect is also ignored in the variance estimation procedure previously used.
Author: Melike Oguz-Alper