Promoting the labour force participation and employment of older people in the EUs
Annex 3.2: multistage retirement projection model
This annex describes the model used to estimate the employment potential of older people and expected structural changes in the composition of the population of inactive older people. These estimates include dynamic country-specific projections of labour market activity and inactivity that consider current trends and country-specific socioeconomic determinants. The model considers four main factors:
- demographic change,
- the current pension reforms with forward-looking effects,
- the changing educational, sectoral and socioeconomic demographic structure,
- inactive subgroup trends.
In contrast to the Member-State-specific models of retirement used in the 2024 Ageing Report (208), the projections presented hereafter are based on a single-model specification calibrated for each Member State separately to replicate the retired and non-retired distributions from the Labour Force Survey (LFS) microdata and further extrapolate the subgroup inactivity trends to get projections of both the size and the composition of the inactive older population. The model used to project the demographic development of retired and non-retired people uses 2022 as the benchmark year and produces projections for 2030 and 2040. It is a multistage model for projecting future retirement patterns that also enables a demographic analysis of the non-retired population.
In the first stage, the future retirement age is determined from publicly available data and OECD pension reports for the 2022 baseline and projections for 2030 and 2040 based on the already adopted legal changes (209). Graph 3.21 shows the current and future retirement ages for the Member States.
In the second stage, the Eurostat baseline demographic projection is calibrated for each country, age and gender for 2022, 2030 and 2040, merged with the 2022 LFS microdata and integrated with the compiled data on future retirement ages from the first stage.
Graph 3.21: Future retirement age changes in Member States
Sources:
OECD and publicly available data.
In the third stage, the LFS sample from 2022 is projected to 2030 and 2040 by assuming the age development of the current educational and occupational structure, taking into account the structural change associated with ageing. In this phase, the changed educational and occupational structure of the older individuals in 2030 and 2040 is recorded in comparison with 2022. The LFS weights are decomposed and reweighted based on the projected demographic shares for each country, gender and age in order to achieve full consistency of the overall weighted results with the Eurostat baseline projections for 2030 and 2040 for each country, gender and age.
In the fourth phase, the logistic model with weights from the LFS is used to assess retirement behaviour separately for each country based on socioeconomic characteristics and the distance to the statutory retirement age. The following model is used for each country to assess the probability of being retired or not:
logit(retired) = β0 + β1*MAIN + β2*MAIN_1 + β3*MAIN_2 + β4*MAIN_3 + β5*MAIN_4 + β6*MAIN_5 + β7*MAIN1 + β8*MAIN2 + β9*MAIN3 + β10*MAIN4 + β11*MAIN5 + β12*Female + β13*i.isco+ β14*i.NACE + β15*yoe
where ‘MAIN’ indicates the difference between individual age and retirement age; ‘MAIN_X’ are dummies for ‘MAIN’ between 5 and – 5 to better capture transition and various early retirement patterns for each country; ‘Female’, ‘isco’ and ‘NACE’ are dummies for women, occupations and sectors; and ‘yoe’ represents years of education. The cut-off rate that determines retirement and non-retirement based on this model is dynamically optimised by an algorithm that guarantees the best fit of the model prediction to retirement behaviour in 2022 by each age and gender with additional robustness tests. This result is then projected onto the demographic projection data and the retirement age data for 2030 and 2040. The caveat of this type of specification is that it does not consider all the country-specific pension reform effects that go beyond the gender-specific changes in the pension age or required years of contribution, which contribute to the expected change in effective retirement age. Therefore, country-specific pension reform effects that go beyond these main major adjustments, such as changes to early retirement schemes or occupation-specific changes to retirement rules, remain unaccounted for.
The model of retirement dynamics is complemented by the projection of inactive subgroup trend analysis. The activity trends within highly heterogeneous groups are significantly influenced by a variety of factors, including targeted activation policies, economic incentives and changes in labour demand. However, unlike the modelling of retirement decisions – which can reasonably be captured by institutional, demographic and individual characteristics – the prediction of inactive subgroups based solely on these characteristics is implausible. This is because the individual idiosyncrasies leading to inactivity are far more pronounced, rendering models based directly on microdata and individual characteristics unreliable. Therefore, the model chosen for the projection of future inactivity and its composition extends the trends observed over the past 10 years in the inactive subgroups into the future. Simple linear ordinary least squares regression is used to estimate the trends for each subgroup. LFS microdata are used to decompose and weigh the subgroups of inactive older people by the following variables: country, age (in groups: 55 to 59, 60 to 64, 65 to 68) and gender. Trends for both the retired and non-retired subgroups are assessed. For each cohort in the non-retired segment, five subgroups are analysed that, collectively, encompass all inactive individuals within that cohort:
- inactive individuals with disabilities;
- individuals without disabilities who have never entered the labour market;
- individuals without disabilities who have been inactive for over 10 years;
- individuals without disabilities who have been inactive for 1 to 10 years;
- individuals without disabilities who have been inactive for less than 1 year.
Testing more detailed decompositions – by adding additional variables such as education or occupational characteristics, or by selecting more granular demographic subgroups – leads to issues with sample size and reduces the stability of the time trend, particularly in smaller Member States.