RDP 2024-09: The ‘Clean Energy Transition’ and the Cost of Job Displacement in Energy-intensive Industries 3. Methodology

Determining the cost of job displacement for a given worker poses important methodological challenges. Ideally, we would like to compare outcomes for the same worker in two states of the world: first, a situation where the worker loses employment and second, a counter-factual situation where the same worker continues to be employed in that same job. As this is evidently infeasible, we follow the job displacement literature and match observationally identical displaced and non-displaced workers (Jacobson, Lalonde and Sullivan, 1993_[6]; Lachowska, Mas and Woodbury, 2020_[7]; Schmieder, von Wachter and Heining, 2023_[8]). Following this procedure, we estimate displacement effects in an event study framework using the matched sample of displaced and non-displaced workers. The precise implementation of the event study approach follows Barreto, Grundke and Krill (2023_[9]) and OECD (2024_[1]).

Defining the treatment

The analysis focuses on job displacement as a result of mass-layoff events. This approach ensures that firm separations are plausibly exogenous, involuntary and unrelated to the performance of workers or their individual career plans. As such, it prevents unobservable differences in performance between workers remaining employed and those laid off from affecting the estimates of job displacement costs.

To identify mass layoff events in our datasets, we follow the displacement literature by defining mass layoffs as events in which employment in a establishments with at least 30 employees declines by at least 30% between one year to the next (Jacobson, Lalonde and Sullivan, 1993_[6]).^[13] Our definition of mass-layoff events includes complete plant closures. To avoid contaminating our measures of mass layoffs with restructuring events (e.g. mergers and acquisitions), we impose the restriction that no more than 30% of displaced employees move to the same establishment following the event (Hethey-Maier and Schmieder, 2013_[10]). This tends to be particularly important in sectors that have experienced important pro-competitive reforms during the sample period, such as those in the transport sector of several included countries.

Having defined mass layoff events, we define treated workers as those who separate from their employer in the year a mass layoff takes place and are not recalled to the same employer over the six subsequent years. In practice, we focus on workers 18 to 50 years old at the time of displacement to limit the influence of early retirement programs.^[14] To identify separations among workers who had stable employment trajectories in their original employer, we also restrict the analysis to workers with at least 2 years of tenure in the year of mass layoff. By considering only the first observed mass layoff event for each worker, treatment can happen only once over the sample period, consistent with the idea that displacement represents a permanent shock to labour market trajectories.

Balancing treatment and controls

In our main analysis, we compare the outcomes of workers who are displaced between one year and the next as a result of mass layoff (treated) with those of non-displaced workers (control) who satisfy the same restrictions in the year immediately before displacement. We allow non-displaced control workers to be co-workers of displaced workers and to separate from their employer in subsequent years for any reason except a mass layoff event.^[15]

As displaced and non-displaced workers may differ in their observable characteristics, we match each displaced worker to an observationally identical non-displaced worker through a 2-step matching procedure (“statistical twinning”). First, we use exact matching by baseline year, industry (1-digit NACE) and energy-intensive sector and sex. This ensures that displaced workers are matched only with workers in the same (energy-intensive) sector, of the same sex, and from the same year. In the second step, we estimate a propensity score separately for each cell using a probit model of job displacement on observable characteristics, including log daily wages in the three years prior to displacement, log employer size, age, and tenure (all included contemporaneously). Based on the estimated propensity scores, we apply nearest neighbor one-to-one matching without replacement to assign each displaced worker to a unique control worker. After the implementing this matching procedure, matched treatment and control workers have similar observable characteristics as shown by standardized differences below a value of 0.1 (Austin, 2011_[11]) (see 3).

Event study design

We rely on an event study design to compare the outcomes of displaced and non-displaced workers before and after displacement separately for each energy-intensive sector and the rest of the economy, using the equation below:

where y_itc is the outcome of worker i belonging to cohort c of displaced workers and matched controls at time t. The coefficients of interest ${\theta }_k$ capture the change in outcome of displaced workers relative to the evolution of the respective outcomes for non-displaced workers in the same sector, where k indexes event time such that k=1 in the first post-displacement year and k=0 in the last year before displacement. The coefficients are normalized to k=-2, such that the effects are measured relative to that time. The worker fixed effect ${\alpha }_i$ controls for time-invariant unobserved worker heterogeneity (see below), ${\lambda }_t$ is a calendar year fixed effect, ${\gamma }_k$ a time since event fixed effects and ${X^{\prime }}_{it}$ contains a cubic in age. Finally, r_itc is the idiosyncratic error term. Standard errors are clustered at the worker level.

The outcomes considered are annual earnings relative to the pre-displacement average, a dummy for being employed, the number of days worked, the log daily wage, the firm wage premium, and various mobility outcomes, such as the likelihood of changing the pre-displacement sector, occupation and region. Annual earnings are defined as the sum of labour payments (potentially, from different employers) in a given year divided by average pre-displacement annual earnings. The employment dummy is equal to one if a worker has at least one day of dependent employment in a given year and zero otherwise. Recall that in the event a worker is not observed in dependent employment in a given year, zero earnings are imputed. This may overstate the actual costs of job loss to the extent that some displaced workers move to the public sector or become self-employed. Days worked are defined as the total number of days in dependent employment in a given year conditional being employed at least one day irrespective of hours worked. Log daily wages are constructed as the natural logarithm of annual earnings divided by days worked at the main employer. Firm wage premia measure the average wage premia paid to all employees in a firm net of worker characteristics and is estimated using an AKM two-way fixed effects model (Abowd, Kramarz and Margolis, 1999_[12]).^[16] Finally, the probability of changing sector, occupation or region is measured using a dummy which is equal to one if the observed value after displacement differs from its pre-displacement value and zero otherwise.

Decomposing earnings losses

To provide an indication of the different components behind displacement costs, expressed as differences in annual earnings between displaced and non-displaced workers, we also include a decomposition exercise into several components that make up annual earnings. These are i) being out of work for an entire year, (ii) fewer days worked conditional on being employed at some point during the year and (iii) lower daily wages upon re-employment. Fewer days worked conditional on being employed at least one day during the year may reflect a combination of the return to work during the year (as the first job after displacement starts after 1 January), the instability of employment following displacement (for example if workers are more likely to be re-employed on temporary contracts) or permanent exits during the year from employment (for example workers moving to inactivity or self-employment after 1 January). Lower wages upon re-employment, measured as daily wages, reflect a combination of hourly wages and working time. Daily instead of hourly wages are used to enhance the cross-country comparability of the results.

In practice, we decompose differences in annual earnings y for a given worker into the components that can be attributed to the probability of being employed in the year p, the number of days worked N_D, and the daily wage w – mostly following Schmieder, von Wachter and Heining (2023_[8]). Taking expectations over the samples of displaced and non-displaced workers, we can express the earnings losses of a displaced worker (D) related to the control group (S) in each time period after the event as:

Rearranging terms gives:

where the first term gives the contribution of changes in daily wages to changes in annual earnings relative to the control group, while the second and third term capture the contribution of days worked and employment probability, respectively. The contribution of the probability of being employed captures periods out of dependent employment lasting a full calendar year, in which annual earnings are imputed to be zero. The contribution of days worked captures periods of non-employment shorter than a full calendar year, which reflects a combination of non-employment and job instability. The contribution of daily wages captures changes in daily wages following displacement. Finally, the term $\mu$ is a residual which captures the change in the covariances between days worked, employment probability and daily wages and can be broadly interpreted as the selection into employment.

To shed light on the underlying drivers of wage losses, the role of daily wages for earnings losses can be further decomposed into a worker- and a firm-related component (Lachowska, Mas and Woodbury, 2020_[7]). We do this by decomposing the treatment effect on wages $\text{Δ}E\left[w\right]$ into the sum of changes in firm wage-premia $\text{Δ}E\left[\text{ψ}\right]$ plus changes in worker-related components $\text{Δ}E\left[\rho \right]$. While the former captures changes in the generosity of a firms' wage policies, the latter captures differences in human capital and match quality. The term $\mu$ is a residual which captures the change in the covariances between employment, days worked and daily wages which arise due to selection into employment. In practice, this component is very small and omitted for presentational purposes.

In a second step, wage losses upon re-employment are decomposed into a worker and a firm-related component. This is done by estimating the change in the firm-related component due to job displacement as measured by the firm fixed effect from an AKM two-way fixed effects model. The part of wage losses that is not firm-related may include worker-related wage losses due to the loss of human capital or wage losses due to a reduction in match quality.

Oaxaca-Blinder decomposition

The analysis of job displacement examines the outcomes of displaced workers compared to those who are not displaced but have similar characteristics within the same energy-intensive sector. Nevertheless, the characteristics of displaced workers in energy-intensive industries may still differ from those in other parts of the economy and from displaced workers in similar sectors in other analysed countries. Consequently, differences in the cost of job displacement across industries and countries might reflect variation in the composition of displaced workers or differences in the cost of job displacement for similar workers. Using Oaxaca-Blinder decompositions, we aim to measure the impact of worker composition on the differences in job displacement costs between industries and countries.

For this, we start by denoting the individual-level difference-in-differences estimate $\left(\Delta y_i\right)$ as:

where ${\overline{y}}_{i,after}^h$ indicates the average outcome for $h\in \left\{Displaced,Non-displaced\right\}$ after job displacement (1 to 4 years) and ${\overline{y}}_{i,before}^h$ the corresponding average outcome before job displacement (-3 to -1 years). The estimate individual-level difference-indifferences $\Delta y_i$ can in turn be characterised as a linear model of the observable characteristics of displaced and non-displaced workers in each sector:

where $X_j^s,S\in \left\{ERI,ROE\right\}$ is a vector of worker and firm characteristics measured before displacement (i.e. in the baseline period at k=0) for energy-intensive industries (ERI) and the rest of the economy (ROE) and ${\vartheta }_i^s$ is an error term. Naturally, this analysis extends to specific energy-intensive industries, i.e. energy supply, heavy manufacturing and transport, in place of ERI.

Using equations (1) and (2), the Oaxaca-Blinder decomposition of the difference in the difference-indifferences estimate between energy-intensive industries and the rest of the economy can be written as:

where the first component on the right-hand side captures the role of composition effect, i.e. the part that is explained by differences in observable characteristics between displaced workers in energy-intensive industries and the rest of the economy, and the second component captures the structural effect, i.e. unexplained differences in the cost of job displacement between energy-intensive industries and the rest of the economy holding composition constant. A similar decomposition can be used to shed light on the drivers of differences in the cost of job displacement in energy-intensive industries between countries.

Footnotes

We use an employment threshold of 30 instead of 50 as done by Jacobson et al. (1993) to allow for a sufficiently large sample in each of the sub-sectors. [13]

OECD (2024_[1]) presents a robustness check including workers aged 50-60. The results are qualitatively similar, with slightly larger earnings losses in both energy-intesive industries and the rest of the economy, while the diffience between the sectorsremain the same. For this reason, and in order to minimize the impact of differences in early retirement schemes across countries, our baseline sample includes workers up to 50 years of age at separation. [14]

The latter prevents “forbidden comparisons” of treated units with units that were treated in earlier periods (de Chaisemartin and D'Haultfœuille, 2020_[]; Callaway and Sant'Anna, 2021_[]) while the former avoids overestimating displacement effects when restricting the control group to workers who remain continuously employed with the same employer (Krolikowski, 2017_[]). [15]

In practice, we estimate the following AKM model: $w_{it}={\alpha }_i+{\psi }_{J\left(i,t\right)}+Y_t+\theta X_{it}+{\in }_{it}$, where w_it is the log wage of worker i in year t. Worker fixed effects for each worker i are captured by ${\alpha }_i$ while ${\psi }_{J\left(i,t\right)}$ captures the firm fixed effects which reflect employer-specific wage premia in each establishment (firm) J of worker i in year t. Year fixed effects are captured through Y_t, while X_it includes a cubic in age interacted with gender dummies. In the estimation of this model, we exclude post-displacement observations of treated and matched control units to avoid these transitions from impacting the estimation of firm effects. [16]