## wpiea2021031-print-pdf

## Source details

**Canonical URL:** [wpiea2021031-print-pdf](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021031-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021031-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021031-print-pdf.pdf.json)

---

### Legal framework
- Legal termination and statutory severance:
  - Open-ended contract terminations: employees with tenure longer than one year and no serious/gross misconduct are granted a minimum legal severance equal to one fifth of monthly salary per year of tenure, plus an additional two fifteenths after ten-year tenure.
  - Sector collective agreements can top up these amounts.
  - Terminations lawful only if justified by a “real and serious cause” (economic or personal).
    - Economic dismissals lawful only to “safeguard” firms, not to improve profitability.
    - Personal dismissals lawful only for misconduct or lack of adaptation.
    - Burden of proof on employers; employers must prove no other position available in the firm (worldwide in the period studied) when dismissal motivated by economic reasons or lack of adaptation.

- Prud’hommes councils and appeals:
  - Employees can file complaints before Prud’hommes councils; judges are equal numbers of employee and employer representatives.
  - Empirical recourse and outcomes:
    - For economic dismissals in 2006, recourse rate between 1% and 2%; for disciplinary dismissals between 17% and 25%.
    - Among claims reaching Prud’hommes (1998–2012), 62% resulted in acceptance of employee claims.
    - In 1996–2003, “60% of cases end up with a trial, among which 75% lead to a worker’s victory”.
  - Appeals:
    - Appeal rates between 60% and 67% in 2004–2013.
    - From 2006 to 2016, only 45% of Prud’hommes councils decisions about compensations for dismissal were confirmed by Appeal courts.
    - Compensation decided at the Appeal court (Appeal-level compensation) is used as preferred measure because appeals suspend Prud’hommes decisions and are frequently not fully confirmed.

- Appeal courts organization and judge mobility:
  - Institutional facts:
    - 36 Appeal courts and 210 Prud’hommes councils.
    - Each Appeal court has different chambers; at least one social chamber treats cases from Prud’hommes. Paris court has fourteen social chambers.
    - Each social chamber has one president assisted for each judgment by two councillor-judges.
  - Status and mobility:
    - Ordonnance Organique of 22 December 1958 determines placement: Appeal court judges are “placed judges” assigned to a given Court or Chamber.
    - Promotions decided yearly by a National Commission of Advancement; First President placed by decree following recommendation by the National Council of the Judiciary.
    - Mobility/turnover rules and facts:
      - Promotions awarded only to judges in a given position for less than 5 years in same jurisdiction (7 years from 2017).
      - Prohibition to stay in same specialized function in same jurisdiction more than ten years altogether.
      - Organic law 2001-539 of June 25th, 2001 includes geographical mobility requirements to achieve the first grade of the remuneration schedule.
      - Every year 20% of positions are re-assigned among judges (Conseil de la Magistrature, rapport d’activité 2016).
    - The First President sets objective criteria driving case distribution between chambers, independently of judges’ identity (articles R312-42 and R312-42-1).

### Identification, judge assignment, and empirical strategy
- Random allocation assumption:
  - Identification rests on random allocation of judges to cases due to: i) judges inherit a large backlog, ii) judges are mobile, iii) plaintiffs/defendants have limited information about judge identity.
  - Random component exploited is allocation across judges within court × social chamber × year.
  - Supporting facts:
    - Average waiting time before judgments is about two years (667 days); only 10% of cases judged in less than 300 days.
    - Judge mobility network is dense (documented in Figure 10).
    - Replacements for absent presidents can occur without notice; chamber that will judge a case often unknown before judgment.

- Construction of judge-specific bias measures (leave-in / leave-one-out):
  - Leave-in difference for judge j in chamber k, year t:
    - ̄ε_jkt = (1/n_jkt Σ_{i∈(j,k,t)} y_i) − (1/n_kt Σ_{i∈(k,t)} y_i)
  - Judge j overall bias:
    - ̄ε_j = Σ_{(k,t)∈(K,T)(j)} (n_jkt / n_j) ̄ε_jkt
  - Leave-one-out, case-specific judge bias for regression:
    - ̄ε_ij = Σ_{(k,t)∈(K,T)(j)} Σ_{i′, i′≠i} (n_jkt / (n_j − 1)) ̄ε_{i′jkt}
    - ̄ε_ijkt = (1/(n_jkt − 1) Σ_{i′∈(j,k,t), i′≠i} y_{i′}) − (1/(n_kt − 1) Σ_{i′∈(k,t), i′≠i} y_{i′})
    - By definition: Σ_{i∈j} ̄ε_ij = ̄ε_j.

- Randomization tests and balance checks:
  - Regressions of judge-specific differences on worker and firm characteristics show absence of correlation with observables, supporting lack of selection on observables.
  - Acknowledged limitation: cannot test correlation with unobservables, but randomization tests are reassuring.

### Data construction, extraction, and sample
- Primary extraction and linkage:
  - Gathered 145,638 Appeal court rulings published by the Ministry of Justice.
  - From rulings extracted case variables, firm name and address, decision narrative, and compensation amounts.
  - Retrieved SIREN identifier from text or automatic search using firm name and address; SIREN allows merging with DADS and FICUS-FARE.
- Automated extraction and validation:
  - Variables extracted using a Python program based on keyword extraction and natural language processing.
  - Validation: on a manually-filled subsample of about 2,500 observations selected at random, correlation between manually-filled and automatically-filled compensation amounts = 94%.
- Administrative datasets used:
  - Matched employer-employee data: DADS Postes Déclarations Administratives de Données Sociales from 2002 to 2015.
  - Tax data: FICUS-FARE company accounts files from 2002 to 2016.
- Sample construction and restrictions:
  - Exclusions applied: firms in liquidation at judgment date; missing presiding judge’s name/surname, total compensation amount, or monthly wage; public sector employers; cases judged by judges who judged less than 50 cases.
  - Final sample used to estimate judge fixed effects: 37,149 cases and 159 judges (the 159 presidents cover 93.3% of cases among the universe analyzed).
  - Average judgments per president: mean 450 cases; median 339.

### Descriptive statistics of judgments and judge-related measures
- Compensation statistics (case-level, count 37,149):
  - Average compensation for wrongful dismissal (months of salary): mean 4.3 months.
  - Total amount including other indemnities represents 10.5 months of salary.
  - Appeal courts’ average compensation for unfair dismissal: 12,288e.
  - Prud’hommes average compensation for unfair dismissal: 7,236e.
- Comparisons between Appeal courts and Prud’hommes:
  - Appeal court amount same as Prud’hommes in 45% of cases, higher in 38%, lower in 17%.
  - Appeal courts overall more favorable to workers than Prud’hommes.
- Case and appeal behavior:
  - Worker appeals in 58% of cases (case-level mean worker who appealed indicator reported as 0.6101 in some samples).
  - Histogram of compensation conditional on positive shows a mass around six months of salary (reflecting legal minimal threshold for certain workers).
- Variance and explanatory power:
  - Only 13.6% of variance explained by salary and seniority alone; adding many covariates raises explained variance to 32.9%.
  - Therefore, 67% of variance of dismissal compensation remains unexplained when controlling for wide covariates.
- Judge bias measures and correlations:
  - Two bias measures: frequency of granting positive compensation; amount of compensation granted.
  - Amounts granted for unfair dismissal positively correlated with amounts under other motives: on average one month of salary granted for unfair dismissal associated with one third of additional monthly wage granted for other motives.
  - Main variable of interest: total compensation for contract breach (for unfair or other motives).

### Case-level summary statistics (selected exact figures)
- Final sample attrition path (counts):
  - Initial severance pay data: 145,638 cases.
  - After restricting to firms not liquidated at judgment date: 123,304 cases.
  - Cases with non-missing president name and surname: 117,989 cases; 1,039 judges.
  - Cases with non-missing total amount of compensation: 84,151 cases; 878 judges.
  - Cases with non-missing monthly wage: 61,728 cases; 731 judges.
  - Elimination of public sector employer cases: 39,843 cases; 652 judges.
  - Cases restricted to judges with at least 50 cases: 37,149 cases; 159 judges.
- Table 2 highlights (count 37,149):
  - Total amount in euro: mean 29,794; min 0; median 15,724; max 963,154; sd 50,056.
  - Total amount in months of salary: mean 10.47; min 0; median 7.84; max 76.26; sd 11.12.
  - Positive total amount indicator: mean 0.8901.
  - Amount for unfair dismissal in euro: mean 12,288; min 0; median 3,000; max 530,000; sd 24,193.
  - Amount for unfair dismissal in months of salary: mean 4.32; min 0; median 1.55; max 73,176.10.
  - Positive amount for unfair dismissal indicator: mean 0.5801.
  - Prud’hommes amount: mean 7,326; min 0; median 0; max 277,200; sd 17,649 (count 27,725).
  - Amount demanded by worker: mean 44,458; min 1; median 25,000; max 985,536; sd 64,439 (count 19,371).
  - Worker’s seniority in months: mean 81.66; median 50.00.

### Judge pro-worker bias: core correlations (selected coefficients, significance preserved)
- Correlation with dismissal qualification (dependent = wrongful dismissal indicator; # obs 9,138):
  - Column (1): coefficient 0.508 *** (standard error 0.141).
  - Column (2) with case controls: 0.493 *** (standard error 0.133).
- Correlation with compensation (total compensation in months; # obs 9,138):
  - Column (1): coefficient 0.852 *** (standard error 0.241).
  - Column (2) with case controls: 0.838 *** (standard error 0.241).
- Share of variance explained (Table 4 excerpt; # obs 9,138):
  - Qualification regressions: Adj.R2 increases from 0.014 to 0.030 when adding controls and judge bias.
  - Compensation regressions: Adj.R2 increases from 0.013 to 0.111 when adding case characteristics and judge bias.
  - Dispersion of judge bias explains only a small share of variance conditional on observables.

### Empirical specifications for firm performance
- Benchmark OLS:
  - Y_{ij(i)t} = α_0 + α_1 bias_{ij(i)} + α_2 X_{it} + η_{ij(i)t}
  - bias_{ij(i)} = (̄ε_{ij} − ̄ε)/σ_ε is judge j’s normalized bias.
  - X_{it} includes Appeal court fixed effects, year fixed effects, leave-one-out average industry annual growth rate of sales, and indicator for economic dismissals.
  - Standard errors clustered at judge level.
- Heterogeneous effects by firm financial health:
  - Y_{ij(i)t} = β_0 + β_1 bias_{ij(i)} × low_i + β_2 bias_{ij(i)} × high_i + β_3 X_{it} + ν_{ij(i)t}
  - low_i = 1 if return on assets or leverage below median; high_i = 1 if above median.
- Outcomes:
  - Survival indicators for t = 1,2,3 years after judgment.
  - Symmetric growth rates computed as ΔY_{ij(i)t} = 2 (Y_{ij(i)t+1} − Y_{ij(i)t−1}) / (Y_{ij(i)t+1} + Y_{ij(i)t−1}).
- Instrumental-variable strategy:
  - Regress performance on share of compensation for wrongful dismissal in firm payroll, instrumenting payroll share with judge bias to evaluate impact of unexpected compensation shocks.

### Firm-level sample and descriptive statistics (selected exact figures)
- Sample used to estimate effect of judge bias on firm performance:
  - 82,320 observations; 13,995 firms; 159 judges (sample with judge fixed effects).
  - Firms with less than 100 employees: 35,888 observations; 4,486 firms; 129 judges.
- All firms (< 100 employees) summary (count 4,486):
  - Nb of workers: mean 20.08; sd 20.55; median 13.
  - Sales (K euros): mean 4,788.61; sd 7,429.92; median 1,929.91.
  - Share of firms < 10 years: mean 0.24.
  - Survival at t+30: mean 0.92.
  - Wrongful dismissal indicator: mean 0.52.
  - Amount in annual payroll (%) (when >0): mean 10.75; median 4.06.
  - Judge pro-worker bias: mean -0.03; sd 0.76; min -2.76; max 2.22.
- Small firms (< 10 employees) summary (count 1,902):
  - Nb of workers: mean 4.98; sd 2.38; median 5.
  - Sales (K euros): mean 1,231.21; sd 2,136.92; median 684.44.
  - Share of firms < 10 years: mean 0.35.
  - Survival at t+30: mean 0.89.
  - Wrongful dismissal indicator: mean 0.52.
  - Amount in annual payroll (%) (when >0): mean 19.28; median 10.56; sd 24.56.
  - Judge pro-worker bias: mean 0.00; sd 0.77.

### Main results: judge bias effects on firm performance
- Reduced-form (pro-worker judge bias normalized):
  - First-year effects:
    - One standard deviation increase in judge pro-worker bias reduces employment growth by 1.8 percentage points (first year) for low-performing firms (Low Roa).
    - Sales growth for low-performing firms falls by similar order in first year.
    - High-performing firms not significantly impacted at one-year horizon.
  - Two- and three-year effects (full sample driven by low-performing firms):
    - Impact on low-performing firms’ employment approximately doubled in the third year compared with first year.
    - Employment effects induced by drop in permanent jobs only; temporary jobs not affected.
    - One standard deviation increase in judge pro-worker bias reduces sales growth by 1.4 percentage points one year after judgment and by 4.7 percentage points three years after judgment.
    - Survival for low-performing firms drops by 0.7 percentage points two years after judgment and by 1 percentage point three years after judgment for one standard deviation increase in judge pro-worker bias.
    - Employment effects within 3-year horizon are not solely driven by firm death: surviving low-performing firms also face significant negative impacts on employment and sales growth.
  - Entries and exits:
    - Effects on number of entries and exits non-significant at any horizon; interpretation: two counteracting forces yield no net significant change.
- Size heterogeneity:
  - Small low-performing firms (< 10 employees, Low Roa):
    - One standard deviation increase in judge pro-worker bias reduces employment growth and sales by 6 percentage points at 3-year horizon.
    - Survival reduced by 3 percentage points at 3-year horizon.
    - All employment effects driven by drops in permanent jobs.
  - Firms with 10 employees and more:
    - Employment not significantly impacted even if low-performing.
    - Sales impact for low-performing firms ≥ 10 employees about half that for small low-performing firms.
  - High-performing firms, even small, not significantly impacted.
- IV evidence (instrumenting compensation share by judge bias):
  - First stage: judge bias strongly correlated with share of compensation for wrongful dismissal in payroll (Table 19).
  - Second stage (all firms < 100 employees):
    - An increase in amount of compensation of one percent of the payroll reduces employment by 3 percentage points at 3-year horizon for low-performing firms.
    - Sales growth: increase in compensation of one percent of payroll reduces sales growth by 4 percentage points at 3-year horizon for low-performing firms.
    - High-performing firms not impacted by this revenue shock.
  - Table 21: point estimates for firms below 100 employees same as for firms below 10 employees; larger reduced-form impact on small firms arises because compensations represent a higher payroll share for small firms.

### Dispersion of judge bias, counterfactuals, and welfare implications
- Dispersion and upper-bound cost:
  - Dispersion of judge bias explains less than 0.3% of variance of compensations conditional on observables (Table 4).
  - Approximation yields upper bound of cost of risk associated with dispersion of judge bias at most equal to 1.5% of average compensation depending on risk aversion (Appendix D numeric exercise).
  - Conclusion: actual dispersion of judge bias likely has negligible effects on selection of cases going to Appeal courts.
- Non-linearity and counterfactual exercises:
  - For small firms < 10 employees with Low Roa, no evidence of non-linearity in effects of judge bias (quadratic terms not significant).
  - Counterfactuals (1,000 bootstrap replications) capping judge bias at percentiles:
    - Reducing dispersion of judge bias has very small and non-significant effects on firm survival and employment growth for small low-performing firms.
    - Setting all biases to the mean yields point estimates close to zero and not significantly different from zero.
    - Capping bias of pro-worker judges to mean has larger but still small and not statistically significant impact.
  - Caveat: cannot exclude possibility that all judges are biased; setting all biases to the mean does not ensure absence of bias in interpretation of labor laws.

### Robustness and placebo checks
- Placebo tests:
  - Placebo regressions show absence of significant correlation between judge bias and growth rates of employment and sales between two years and one year before judgment for all firms and small firms.
  - Interpretation: post-judgment effects not driven by pre-judgment selection or anticipation.
- Alternative financial split:
  - Replacing Roa with Roe yields same pattern: negative impacts concentrated among low-performing firms, larger for small low-performing firms.
- Subsample of large Appeal courts:
  - Restricting to large Appeal courts (Aix-en-Provence, Paris, Versailles) yields similar results, supporting random allocation assumption.

### Appendices, extraction details, and compensation types
- Appendix A: caps on dismissal compensation in selected European countries (exact rules listed for Italy, Germany, Austria, Belgium, Denmark, Spain, Finland, Netherlands, Portugal, United Kingdom, Sweden, France).
- Appendix B: computation of judge bias; estimator expressions showing ˆε_j equals ̄ε_j.
- Appendix C: judge mobility improves ranking precision.
- Appendix D: derivation and numeric application of risk premium associated with judge-bias dispersion; inputs:
  - Standard deviation of compensations for wrongful dismissal = 56,385 euros.
  - Mean compensation = 31,461 euros.
  - Dispersion explains 0.3% of variance; implied risk premium ≈ 0.0048 times coefficient of relative risk aversion.
- Appendix E: extraction procedure using NLP on ~145,000 text documents; structure split into five blocks; manual checks on 2,560 observations; manual/automatic correlation = 94%.
- Compensation types tracked (more than twenty categories aggregated):
  - indemnité pour licenciement sans cause réelle et sérieuse; non-respect of dismissal procedure; rappel de salaire; préjudice moral et financier; indemnité compensatrice de préavis; article 700 costs; indemnité compensatrice de congés payés; heures supplémentaires, etc.
- Prevalence (original ~145,000 cases):
  - A positive amount awarded in 60% of cases.
  - Among positive-amount cases: dismissal deemed unfair in 61%; paid leave compensation 47%; advance notice 40%; salaries 13%; overtime 7%; moral damage 2%; harassment 2%; discrimination 0.3%.
  - One or several other types awarded in 93% of positive-amount cases.

*Source: wpiea2021031-print-pdf*

### 2.1   Legal framework

### 2.1   Legal framework

### Legal termination and statutory severance
- Following the termination of an open-ended contract, employees with a tenure longer than one year and who did not commit any serious or gross misconduct (faute grave or faute lourde) are granted a minimum legal severance payment calculated as one fifth of monthly salary per year of tenure, plus an additional two fifteenths after ten-year tenure.
- These amounts can be topped up if the professional branch to which the firm belongs has signed a collective agreement setting higher payouts.
- Under French law, terminations of open-ended employment contracts are lawful if they are justified by a “real and serious cause”, either economic or personal.
  - Dismissals for economic reasons are lawful only to “safeguard” firms, but not to improve their profitability.
  - Dismissals for personal reasons are lawful only in case of misconduct or lack of adaptation to the job.
  - For both types of dismissal, the burden of proof is on the side of employers.
  - Employers must prove that there is no other position available in the firm (worldwide in the period studied) for dismissed employees when the dismissal is motivated by economic reasons or by lack of adaptation to the job.

### Prud’hommes councils and appeals
- When the employee deems her dismissal wrongful, she can file a complaint before the Prud’hommes councils (courts of first instance).
- Prud’hommes judges are employee and employer representatives, with an exact equality between the numbers of councilors representing employers and those representing employees.
- Empirical evidence on recourse and outcomes:
  - For economic dismissals in 2006, the rate of employee recourse to Prud’hommes in case of dismissal is between 1% and 2%, while for disciplinary dismissals it is between 17% and 25% (Serverin and Valentin, 2009).
  - Among claims that reached the judicial stage at Prud’hommes council from 1998 to 2012, 62% resulted in the acceptance of the employee’s claims (Desrieux and Espinosa, 2019).
  - In the 1996-2003 period, “60% of cases end up with a trial, among which 75% lead to a worker’s victory” (Fraisse et al. (2015)).
- Appeal dynamics:
  - Appeal rates are between 60% and 67% in the 2004-2013 period (Guillonneau and Serverin, 2015).
  - From 2006 to 2016, only 45% of Prud’hommes councils decisions about compensations for dismissal were confirmed by Appeal courts.
  - Because appeals suspend the application of Prud’hommes decisions and are frequently not fully confirmed, compensation for wrongful dismissals decided at the Appeal court level is used as the preferred measure of compensation.

### Overview of Appeal court’s organization
- Institutional facts:
  - There are 36 Appeal courts and 210 Prud’hommes councils.
  - Each Appeal court has different chambers; at least one social chamber treats cases coming from the Prud’hommes council. The Paris court has fourteen social chambers.
  - Each social chamber has one president; the chamber president presides over the chamber’s trials and is assisted, for each judgment, by two councillor-judges.
- Status, placement, and mobility of judges:
  - Status and mobility are determined by the Ordonnance Organique of 22 December 1958: Appeal court judges are “placed judges” assigned to a given Court or Chamber in a specific position according to annual decisions by higher judicial authorities.
  - Promotions are based on merit and decided every year by a National Commission of Advancement.
  - The First President of the Appeal court is placed by a decree following recommendation by the National Council of the Judiciary.
  - Mobility and turnover rules and facts:
    - Promotions awarded only to judges in a given position for less than 5 years in a same jurisdiction (7 years from 2017).
    - Prohibition to stay in the same specialized function in the same jurisdiction more than ten years altogether.
    - Organic law 2001-539 of June 25th, 2001 includes geographical mobility requirements to achieve the first grade of the remuneration schedule.
    - Every year 20% of positions are re-assigned among judges (Conseil de la Magistrature, rapport d’activité 2016).
  - The First President of the Appeal court sets objective criteria driving case distribution between chambers, independently of judges’ identity (articles R312-42 and R312-42-1 of the Judiciary Organisation Code).

### Assignment of judges to cases and implications for identification
- Two institutional features imply important random components in case allocation to judges:
  - Backlog and timing: it takes a judge on average two years from appointment to rule on all cases assigned to the social chamber prior to her arrival. The average spell of a judge in a social chamber is equal to about 2.5 years, so the identity of the president who will judge a case assigned to a social chamber is generally unknown when the case is allocated.
  - Replacement and multi-chamber complexity: when a president is absent (vacation, sickness, vocational training, etc.) she is replaced by another chamber president; in larger Appeal courts the social chamber that will judge the case is not known before the judgment.
- Consequences for empirical strategy:
  - Plaintiffs and judges cannot reliably manipulate assignment because of the backlog, mobility, and uncertainty about the presiding judge until the day of judgment.
  - These quasi-random components are leveraged to identify judge bias in the analysis.

### Data construction and linkage
- Primary extraction:
  - Gathered 145,638 Appeal court rulings published by the Ministry of Justice.
  - From rulings, extracted case variables, firm name and address, decision narrative, and compensation amounts.
- Additional compensation types:
  - When dismissal is ruled wrongful, employees may receive additional compensations (moral and financial damages, compensation for unpaid wages, unpaid notice, unpaid overtime, compensation for harassment or discrimination, etc.). Judges’ assessment can create correlations across these amounts.
  - Appendix E lists dozens of possible additional compensations.
- Automated extraction and validation:
  - Variables extracted using a Python program based on keyword extraction and natural language processing.
  - Validation: on a manually-filled subsample of about 2,500 observations selected at random, the correlation between manually-filled and automatically-filled compensation amounts is equal to 94%.
- Firm identifier and administrative linkage:
  - Retrieved the SIREN identifier either directly from text or via automatic search using firm name and address on online registries.
  - SIREN allows merging the rulings compensation dataset with administrative social security and tax data (matched employer-employee data and FICUS-FARE).

### Administrative datasets used
- Matched employer-employee data:
  - DADS Postes Déclarations Administratives de Données Sociales from 2002 to 2015 reporting detailed payroll information for private firms.
  - Enables tracking firm wage bill and number of employees over time.
- Tax data:
  - FICUS-FARE company accounts files from 2002 to 2016, containing sales, net income, EBITDA and enabling construction of financial indicators (leverage ratio, return on assets, etc.).

### Sample construction and restrictions
- From initial 145,638 rulings, applied exclusions:
  - Exclude cases where firm was in liquidation at judgment date (compensations paid by public agency Agence de Garantie des Salaires).
  - Exclude cases missing presiding judge’s name/surname, total compensation amount, or monthly wage.
  - Exclude public sector employers.
  - Exclude cases judged by judges who judged less than 50 cases.
- Final sample:
  - 37,149 cases and 159 presidents.
  - The 159 presidents who judged more than 50 cases cover 93.3% of cases among the universe analyzed.
  - Each of these presidents judged 450 cases on average with a median equal to 339.

### Descriptive statistics of judgments and judge-related measures
- Compensation statistics:
  - Average amount of compensation for wrongful dismissal granted by Appeal courts is equivalent of 4.3 months of salary.
  - Total amount, including other possible indemnities, represents 10.5 months of salary.
  - Appeal courts’ average compensation for unfair dismissal: 12.288e.
  - Prud’hommes average compensation for unfair dismissal: 7.236e.
- Comparisons between Appeal courts and Prud’hommes:
  - Amount given at Appeal court is the same as the amount decided at Prud’hommes in 45% of cases, higher in 38% of cases, and lower in 17% of cases.
  - Appeal courts are overall more favorable to workers than Prud’hommes.
- Case and appeal behavior:
  - The worker appeals in 58% of cases.
  - Histogram of compensation (conditional on positive) shows a mass around six months of salary, reflecting legislation instituting a minimal threshold of six months of salary for workers with more than 24 months of seniority employed in firms with more than 11 workers when dismissal is wrongful.
- Variance and explanatory power:
  - There is important dispersion of compensation conditional on seniority in both tribunals.
  - Only 13.6% of the variance is explained by salary and seniority alone.
  - Adding many other covariates raises explained variance to 32.9%.
  - Therefore, 67% of the variance of dismissal compensation remains unexplained when controlling for a wide range of covariates.
- Judge bias measures and correlations:
  - Two types of variable used to evaluate judge bias:
    - Frequency at which the judge grants a positive compensation to the worker (for unfair dismissal or any other motive).
    - Amount of compensation granted.
  - Amounts granted for unfair dismissal are positively correlated with amounts granted under other motives: on average one month of salary granted for unfair dismissal is associated with one third of additional monthly wage granted for other motives.
  - Main variable of interest used throughout the analysis is the total compensation for contract breach (for unfair or any other motive).

*Source: wpiea2021031-print-pdf - 2.1   Legal framework*

### 4.2  Empirical strategy

### 4.2 Empirical strategy

### Identification and random allocation assumption
- Identification rests on the assumption that the allocation of judges to cases is random. This relies on three institutional features: i) judges inherit a large backlog, ii) judges are mobile and iii) defendants and plaintiffs have limited information about the identity of the judge.
- The random component exploited is the allocation of cases across different judges within court, social chamber and year, relying on differences between decisions of presidents belonging to the same social chamber within the same year.
- Example chronology: in year 2014 and social chamber 1 of the Paris Appeal court, cases may be allocated to president A in the first part of the year or to president B in the second part of the year.
- Features supporting randomness:
  - Large backlog: the average waiting time before judgments is about two years (667 days), and only 10% of cases are judged in less than 300 days.
  - Judge mobility across chambers (documented in Figure 10): the network of judges is dense, indicating high mobility and enabling comparisons across chambers.
  - If a judge is absent the day of judgment, replacement can occur without notice to parties; plaintiffs and defendants do not know which social chamber will judge their case before judgment.

### Construction of judge-specific measures of bias
- For each social chamber × year pair (k,t) and judge j, the leave-in difference (equation (1)) is:
  - ̄ε_jkt = (1/n_jkt Σ_{i∈(j,k,t)} y_i) − (1/n_kt Σ_{i∈(k,t)} y_i)
  - where i∈(j,k,t) means case i is judged by judge j in chamber k and year t; i∈(k,t) means case i is judged in chamber k and year t; y_i is the outcome of case i; n_jkt is the number of judgments of judge j in chamber k during year t; n_kt is the number of judgments in chamber k during year t.
- Judge j’s overall bias is the weighted average across observed (k,t) pairs (equation (2)):
  - ̄ε_j = Σ_{(k,t)∈(K,T)(j)} (n_jkt / n_j) ̄ε_jkt
  - where (K,T)(j) is the set of all chamber×year pairs observed for judge j; n_j is total judgments of judge j.
- Leave-one-out, case-specific judge bias used in outcome regressions (equations (3) and (4)):
  - ̄ε_ij = Σ_{(k,t)∈(K,T)(j)} Σ_{i′, i′≠i} (n_jkt / (n_j − 1)) ̄ε_{i′jkt}
  - ̄ε_ijkt = (1/(n_jkt − 1) Σ_{i′∈(j,k,t), i′≠i} y_{i′}) − (1/(n_kt − 1) Σ_{i′∈(k,t), i′≠i} y_{i′})
  - By definition: Σ_{i∈j} ̄ε_ij = ̄ε_j.

### Randomization tests and balance checks
- Randomization tests regress judge-specific differences on worker and firm characteristics of corresponding cases.
- Main finding: absence of correlation between observable characteristics of cases and judge-specific differences, supporting lack of selection on observables.
- Acknowledgement: cannot test correlation with unobservables, but randomization tests are reassuring for identification.

---

### 4.3 Results

### Overview
- Judges’ subjectivity affects both (a) the qualification of dismissal (wrongful or lawful) and (b) the compensation amount granted to the worker.
- Two pro-worker bias indices are constructed: one from dismissal qualification (indicator) and one from amount of compensation (as proportion/months of monthly wage).

### Qualification of dismissals
- Distribution: Figure 11 presents the histogram of judges’ pro-worker bias across cases (per equation (3)), showing variability of biases.
- Relation between judge bias and probability dismissal deemed wrongful:
  - Local polynomial fit: judge pro-worker bias is positively related to the probability that dismissals are deemed wrongful.
  - Being assigned to one of the 10% most pro-worker judges vs one of the 10% least pro-worker judges increases the probability that the dismissal is deemed wrongful by about 4 percentage points, corresponding to an increase of 7% in the probability that the dismissal is deemed wrongful.
  - Table 3 OLS: coefficients significant at 1% level. Computation note: multiplying the point estimate in column (3) of Table 3 by the difference of pro-worker bias from the 1st to the 9th decile (respectively equal to -0.46 and 0.36) yields an increase of 4.1 percentage points.
- Contribution to variance:
  - Table 4: adjusted R^2 increases from 2.7% to 3.0% when controlling for judge bias once case controls, court fixed effects and year fixed effects are accounted for.
  - Interpretation: dispersion of judge fixed effects explains only a small share of variance in dismissal qualification; a large share of variation remains unexplained.
- Allocation and balance checks:
  - Adding controls does not significantly change judge-bias coefficient estimates (p-value = 0.25 across Columns (1) and (2) of Table 3).
  - Regressions of judge fixed effects on case characteristics (Table 5) and on firm characteristics (Table 6) show no significant relationships; F-test rejects joint significance of explanatory variables in the judge-bias regressions.

### Compensation for wrongful dismissal
- Distribution: pro-worker bias based on amount granted (as proportion of monthly wage); Figure 12 histogram shows significant heterogeneity.
- Relation between judge bias and compensation:
  - Polynomial fit: being assigned to one of the 10% most pro-worker judges rather than one of the 10% least pro-worker judges increases the amount by about 2 months of salary.
  - Table 7 OLS: with controls, an increase in judge pro-worker bias by one point increases the amount of compensation in months of salary by 0.8 points.
  - Implication: being assigned to one of the 10% most pro-worker biased judges versus one of the 10% least 10% increases compensation by 2.1 months of salary.
  - Deciles for this bias: the judge bias of the 1st decile is equal to -1.28 and that of the 9th decile to 1.25.
- Contribution to variance:
  - Table 4: adjusted R^2 increases from 10.8% to 11.1% when controlling for judge bias once case controls, court fixed effects and year fixed effects are accounted for.
  - Interpretation: judge bias explains only a limited share of the large dispersion of compensation conditional on observables.
- Allocation and balance checks:
  - Addition of case controls does not significantly change coefficients (p-value = 0.71 across Columns (1) and (2) of Table 7).
  - Tables 8 and 9: pro-worker biases are not correlated with case or firm observables; the amount received at Prud’hommes, the seniority, and the worker’s salary are positively correlated with compensation at Appeal court, but regressing judge bias on the same characteristics yields no significant relationship.
- Correlation across indices:
  - Judges who often qualify dismissals as wrongful are also those who, conditional on granting a positive compensation, grant the highest compensations. The two pro-worker bias indices are highly and positively correlated (Figure 13).

### Summary conclusion of section 4
- Significant judge biases exist that influence both the probability that dismissals are deemed wrongful and the compensation amounts.
- However, judge-bias dispersion explains a very limited share of the conditional dispersion of dismissal qualification and compensation, suggesting many case-specific features not observable in the data are important.

---

### 5 Effects on firm performance — Empirical strategy

### Sample and descriptive statistics
- Focus: firms with fewer than 100 employees the year before the Appeal judgment; for-profit firms in private sector excluding agriculture.
- Sample selection: appeal court rulings 2006–2016; firms going to court no later than 2012 to analyze outcomes up to three years after the judgment.
- Matched employer-employee data availability: from 2002 to 2015.
- Table 11 firm-level descriptors (sample restricted to firms under 100 employees):
  - Average number of workers: about 20 employees.
  - 24% of firms are less than 10 years old.
  - 52% of firms end up paying a positive compensation for wrongful dismissal.
  - For firms paying a positive compensation amount:
    - Average share of firms’ annual payroll: 10.7%
    - Median: 4.1%
  - Survival probability: 99% one year after the judgment and 92% three years after.
- Very small firms (below 10 employees), Table 12:
  - Probability of wrongful dismissal: identical to larger firms.
  - Share of compensation in annual payroll (conditional on being positive): about 20.9% for very small firms versus 10.7% for others.
  - Firm age: 35% of very small firms have less than 10 years vs 25% for larger firms.
  - Survival probability three years after judgment: 89% for very small firms versus 92% for larger firms.

### Empirical specifications
- Benchmark OLS specification (equation (5)):
  - Y_{ij(i)t} = α_0 + α_1 bias_{ij(i)} + α_2 X_{it} + η_{ij(i)t}
  - Y_{ij(i)t} is the outcome of interest for firm i assigned to judge j, t ≥ 0 years after the judgment.
  - bias_{ij(i)} = (̄ε_{ij} − ̄ε)/σ_ε is judge j’s normalized bias (difference between judge’s bias and average judge bias scaled by standard deviation).
  - ̄ε_{ij} is the leave-one-out mean of residuals for all other cases judged by judge j (defined in Section 4.2).
  - X_{it} includes Appeal court fixed effects, year fixed effects, the leave-one-out average industry annual growth rate of sales and an indicator variable for economic dismissals.
  - Standard errors clustered at the judge level.
  - Identification requires mean independence of η_{ij(i)t} and bias_{ij(i)} (random assignment of judges to cases).
- Heterogeneous effects by firm financial health (equation (6)):
  - Y_{ij(i)t} = β_0 + β_1 bias_{ij(i)} × low_i + β_2 bias_{ij(i)} × high_i + β_3 X_{it} + ν_{ij(i)t}
  - low_i = 1 if the financial variable (return on assets or leverage) of firm i the year before judgment is below the median; high_i = 1 if above the median.
  - X_{it} includes same variables as before plus low_i and high_i indicators.
- Outcomes:
  - Indicator variables for firm survival within t = 1,2,3 years after judgment.
  - Symmetric growth rates for total, temporary and permanent employment and sales, computed as (equation (7)):
    - ΔY_{ij(i)t} = 2 (Y_{ij(i)t+1} − Y_{ij(i)t−1}) / (Y_{ij(i)t+1} + Y_{ij(i)t−1})
- Standard errors:
  - Clustered at the judge level following Abadie et al. (2017), since randomization occurs primarily at the judge level.

### Instrumental-variable strategy for compensation shocks
- To quantify the impact of the shock on compensation induced by judge bias on firm performance:
  - Regress performance indicators on the share of compensation for wrongful dismissal in firm payroll, instrumenting this payroll share with the judge’s bias.
  - This IV approach evaluates the impact of unexpected shocks to compensation (in payroll share) on firms.

*Source: wpiea2021031-print-pdf — section 4.2–5.2*

### 5.3  Results

### 5.3 Results

### 5.3.1 Reduced form estimates (All firms below 100 employees; small firms < 10 employees)
- First-year effects (Tables 13–15):
  - Pro-worker judge bias has a significant negative impact on employment growth the first year after the judgment only for firms with low return on assets.
  - A one standard deviation increase in judge pro-worker bias reduces employment growth by 1.8 percentage points (first year) for low-performing firms.
  - Sales growth for low-performing firms falls by a similar order of magnitude in the first year.
  - High-performing firms (returns on assets above the median) are not significantly impacted at the one-year horizon.
- Two- and three-year effects:
  - Effects strengthen over time and become statistically significant for the full sample, driven entirely by low-performing firms.
  - The impact on low-performing firms’ employment is approximately doubled in the third year compared with the first year.
  - Employment effects are induced by a drop in permanent jobs only; temporary jobs are not affected.
  - A one standard deviation increase in judge pro-worker bias reduces sales growth by 1.4 percentage points one year after the judgment and by 4.7 percentage points three years after the judgment.
  - Survival for low-performing firms drops by 0.7 percentage points two years after the judgment and by 1 percentage point three years after the judgment for a one standard deviation increase in judge pro-worker bias.
  - Employment effects within a 3-year horizon are not solely driven by firm death: surviving low-performing firms also face significant negative impacts on employment and sales growth (Table 16).
- Entries and exits:
  - Effects of judge bias on the number of entries and exits are non-significant for either type of firm at any time horizon (Tables 13–15).
  - Interpretation: two counteracting effects—(i) pro-worker bias reduces employment (reducing entries/exits) and (ii) pro-worker bias decreases the share of permanent jobs (increasing job turnover)—yield no significant net change in entries and exits.
- Small firms (< 10 employees) versus medium-sized (≥ 10 employees):
  - Small low-performing firms (return on assets below the median) are strongly impacted:
    - A one standard deviation increase in judge pro-worker bias reduces employment growth and sales by 6 percentage points at the 3-year horizon for firms with less than 10 employees (Table 17).
    - This 6 percentage points effect is about twice as high as for all low-performing firms below 100 employees.
    - All employment effects are driven by drops in permanent jobs; temporary jobs do not change significantly.
    - Survival: a one standard deviation increase in judge pro-worker bias reduces the survival rate by 3 percentage points for low-performing small firms at the 3-year horizon.
  - Firms with 10 employees and more:
    - Employment of firms with 10 employees and more is not significantly impacted by judge bias even if they are low-performing (Table 18).
    - Sales impact for low-performing firms ≥ 10 employees is about half of that for small low-performing firms.
  - High-performing firms, even if small, are not significantly impacted by judge bias.
- Summary:
  - Judge pro-worker bias reduces employment growth, raises employment instability, lowers sales growth, and reduces survival predominantly for small, low-performing firms. Larger firms and high-performing firms are not significantly impacted.

### 5.3.2 IV estimates (instrumenting compensation share by judge bias)
- First stage (Table 19):
  - Judge bias is strongly correlated with the share of compensation for wrongful dismissal in the firm payroll.
- Second stage (Table 20; all firms < 100 employees):
  - An increase in the amount of compensation of one percent of the payroll reduces employment by 3 percentage points at the 3-year horizon for low-performing firms.
  - The employment effect arises from the growth of permanent employment; temporary employment is not significantly impacted.
  - Sales growth: an increase in compensation of one percent of the payroll reduces sales growth by 4 percentage points at the 3-year horizon for low-performing firms.
  - High-performing firms are not impacted by this revenue shock induced by judge bias.
- Table 21 finding:
  - Point estimates for all firms below 100 employees are the same as for firms below 10 employees: a transitory shock on revenue equal to one percent of payroll has a similar impact on small and medium-sized firms.
  - The larger reduced-form employment impact of pro-worker judges on small low-performing firms arises because dismissal compensations represent a higher share of the payroll for small firms (Tables 11 and 12).
- Interpretation:
  - The same amount of compensation has very different effects depending on firms’ financial capacity (size and return on assets); smaller low-performing firms are more affected due to weaker financial capacity.
  - Possible “cleansing” effect: pro-worker bias may destroy structurally weakest firms, potentially improving overall efficiency if destroyed jobs are reallocated at low cost to high-performing firms (left for future research).

### 5.4 Effects of the dispersion of judges bias
- Concept and identification:
  - Setting all bias estimates to the mean simulates eliminating judge-related dispersion in dismissal compensation.
  - The dispersion of judge bias explains less than 0.3% of the variance of compensations conditional on observable worker and firm characteristics (Table 4).
  - An approximation implies an upper bound of the cost of the risk associated with the dispersion of judge bias at most equal to 1.5% of the average compensation, depending on the degree of risk aversion (see Appendix D).
  - Conclusion: the actual dispersion of judge bias likely has negligible effects on the selection of cases going to Appeal courts.
- Non-linearity analysis:
  - For small firms below 10 employees with return on assets below the median (the group significantly impacted), no evidence of non-linearity in the effects of judge bias:
    - Visual inspection of augmented component-plus-residual plots (Figure 14) and quadratic terms in reduced form (Table 22) show no non-linear effects.
  - Implication: mean-preserving spread changes in judge bias should have no significant impact on average firm outcomes.
- Counterfactual exercises (bootstrap; Figure 15):
  - Method: cap judge bias at several percentiles; 1,000 bootstrap replications to predict outcomes three years after judgments for counterfactual distributions.
  - Results:
    - Reducing dispersion of judge bias has very small and non-significant effects on firm survival and employment growth for small low-performing firms.
    - Setting all biases to the mean yields point estimates for differences between actual and counterfactual outcomes very close to zero and not significantly different from zero at standard confidence levels.
    - Capping the bias of pro-worker judges to the mean has a larger impact than mean-capping all judges but remains small and far from statistically significant.
    - Similar null result when capping the bias of pro-employer judges to zero.
- Caveat:
  - Study cannot exclude the possibility that all judges are biased (i.e., setting all biases to the mean does not ensure absence of bias in interpretation of labor laws).

### 5.5 Robustness checks
- Placebo tests:
  - Placebo regressions (Table 23) document the absence of significant correlation between judge bias and growth rates of employment and sales between two years and one year before the judgment for all firms and for small firms (both high- and low-performing).
  - Interpretation: effects of judge bias on firm performance after the judgment are not driven by pre-judgment selection or anticipation of judge bias.
- Alternative financial measures:
  - Replacing return on assets with return on equity (Table 24) yields the same pattern: judge bias significantly impacts low-performing firms only, with larger effects for small low-performing firms.
- Large Appeal courts subsample:
  - Restricting the sample to large Appeal courts (Aix-en-Provence, Paris, Versailles) where identity of president is less predictable yields similar results (Table 25), supporting random allocation of cases to judges and robustness of findings.

*Source: 5.3 Results (sections 5.3.1–5.5) of wpiea2021031-print-pdf*

### 2017. doi: 10.1093/qje/qjw035. URLhttp://dx.doi.org/10.1093/qje/qjw035.

### wpiea2021031-print-pdf - 2017. doi: 10.1093/qje/qjw035. URLhttp://dx.doi.org/10.1093/qje/qjw035.

### Sample selection and data construction
- Initial severance pay data: 145,638 cases.
- After restricting to firms not liquidated at judgment date: 123,304 cases.
- Cases with non-missing president name and surname: 117,989 cases; 1,039 judges.
- Cases with non-missing total amount of compensation: 84,151 cases; 878 judges.
- Cases with non-missing monthly wage: 61,728 cases; 731 judges.
- Elimination of public sector employer cases: 39,843 cases; 652 judges.
- Cases restricted to judges with at least 50 cases (threshold for judge fixed-effects): 37,149 cases; 159 judges.
- Final sample used to estimate judge fixed effects: 37,149 cases and 159 judges.

### Case-level summary statistics (Table 2)
- Total amount in euro: mean 29,794; min 0; median 15,724; max 963,154; sd 50,056; count 37,149.
- Total amount in months of salary: mean 10.47; min 0; median 7.84; max 76.26; sd 11.12; count 37,149.
- Positive total amount indicator: mean 0.8901; min 0; max 1; sd 0.31; count 37,149.
- Amount for unfair dismissal in euro: mean 12,288; min 0; median 3,000; max 530,000; sd 24,193; count 37,149.
- Amount for unfair dismissal in months of salary: mean 4.32; min 0; median 1.55; max 73,176.10; count 37,149.
- Positive amount for unfair dismissal indicator: mean 0.5801; min 0; max 1; sd 0.49; count 37,149.
- Other amount in euro: mean 17,506; min 0; median 6,197; max 963,154; sd 38,024; count 37,149.
- Prud’hommes amount: mean 7,326; min 0; median 0; max 277,200; sd 17,649; count 27,725.
- Amount demanded by worker: mean 44,458; min 1; median 25,000; max 985,536; sd 64,439; count 19,371.
- Higher amount than prud’hommes indicator: mean 0.3800; min 0; max 1; sd 0.49; count 27,725.
- Lower amount than prud’hommes indicator: mean 0.1700; min 0; max 1; sd 0.37; count 27,725.
- Same amount as prud’hommes indicator: mean 0.4500; min 0; max 1; sd 0.50; count 27,725.
- Worker who appealed indicator: mean 0.6101; min 0; max 1; sd 0.49; count 33,767.
- Economic dismissal indicator: mean 0.1600; min 0; max 1; sd 0.36; count 37,149.
- Worker’s seniority in months: mean 81.66; min 0; median 50.00; max 5,3887.20; sd (displayed as 27,147) — note: count 27,147.

### Judge pro-worker bias: correlation with dismissal qualification and compensation
- Correlation between judge pro-worker bias and dismissal qualification (dependent = wrongful dismissal indicator):
  - Column (1) coefficient: 0.508 *** (standard error 0.141), Year FE Yes, Court FE Yes, # obs 9,138, F test 12.91.
  - Column (2) coefficient (with case controls): 0.493 *** (standard error 0.133), Year FE Yes, Court FE Yes, Case controls included, # obs 9,138, F test 13.82.
  - Controls in column (2): indicator for economic dismissal, worker’s wage, worker’s seniority. Top fifth percentiles of judge pro-worker bias trimmed. Standard errors clustered at judge level.
- Share of variance of compensations explained by judge bias (R-squared and Adj. R-squared; # obs 9,138):
  - Qualification of dismissal regressions:
    - (1) R2 0.020; Adj.R2 0.014 (fixed effects only).
    - (2) R2 0.024; Adj.R2 0.018 (add judge pro-worker bias).
    - (3) R2 0.033; Adj.R2 0.027 (add case characteristics).
    - (4) R2 0.037; Adj.R2 0.030 (case characteristics + judge bias).
  - Compensation in months of salary regressions:
    - (5) R2 0.019; Adj.R2 0.013.
    - (6) R2 0.022; Adj.R2 0.016.
    - (7) R2 0.114; Adj.R2 0.108.
    - (8) R2 0.117; Adj.R2 0.111.
  - All regressions include Court and Year fixed effects; case controls include firm size (>11 workers), Prud’hommes compensation, salary, seniority.

- Correlation between judge pro-worker bias and compensation for wrongful dismissal (dependent = total compensation):
  - Column (1) coefficient: 0.852 *** (standard error 0.241), Year FE Yes, Court FE Yes, # obs 9,138, F test 12.47.
  - Column (2) coefficient (with case controls): 0.838 *** (standard error 0.241), Year FE Yes, Court FE Yes, Case controls included, # obs 9,138, F test 12.10.
  - Controls in column (2): indicator for economic dismissal, wage, seniority. Bottom and top fifth percentiles of judge bias trimmed. Standard errors clustered at judge level.

### Randomization tests (balance checks)
- Case-level randomization test for judge bias with respect to dismissal qualification (Observations 9,128):
  - Amount at Prud’hommes (in months): coefficient for Dismissal deemed wrongful 4.594 *** (0.537); coefficient for Judge’s pro-worker bias 0.0983 (0.106).
  - Legislation threshold applied: Dismissal -0.022 *** (0.007); Judge bias 0.001 (0.001).
  - Seniority: Dismissal -0.0125 (0.039); Judge bias 0.0006 (0.005).
  - Number of employees: Dismissal -0.001 * (0.001); Judge bias -0.000 (0.000).
  - Worker’s salary: Dismissal 0.000 (0.000); Judge bias 0.000 * (0.000).
  - Economic dismissal: Dismissal 0.061 *** (0.008); Judge bias -0.001 (0.001).
  - Time between dismissal and Appeal Court: Dismissal -0.005 (0.007); Judge bias -0.002 (0.001).
  - Joint F-Test: Dismissal 0.0000; Judge bias 0.2291.
  - Note: Independent variables (except ’Legislation threshold applies’ and ’Economic dismissal’) divided by 1000 for display clarity. Court and year fixed effects included; standard errors clustered at judge level.

- Firm-level randomization test for judge bias with respect to dismissal qualification (Observations 4,847):
  - Number of workers in t-1: Dismissal -0.153 (0.131); Judge bias 0.032 (0.023).
  - Sales in t-1: Dismissal 0.000 (0.001); Judge bias -0.000 (0.000).
  - Total wages in t-1: Dismissal -0.009 (0.010); Judge bias -0.002 (0.002).
  - Value added in t-1: Dismissal 0.007 (0.007); Judge bias 0.001 (0.001).
  - Net income in t-1: Dismissal 0.011 (0.012); Judge bias 0.000 (0.002).
  - Debt in t-1: Dismissal 0.002 (0.003); Judge bias 0.000 (0.000).
  - Cash in t-1: Dismissal -0.014 ** (0.007); Judge bias -0.000 (0.001).
  - Joint F-Test: Dismissal 0.2313; Judge bias 0.8956.
  - Note: All independent variables divided by 1000 for clarity. Court and year fixed effects included; standard errors clustered at judge level.

- Case-level randomization test for judge bias on total compensation (Observations 4,948):
  - Dependent variable: Compensation in monthly wages (column 1) and Judge pro-worker bias (column 2).
  - Amount at Prud’hommes (in months): Compensation coef 0.536 *** (0.083); Judge bias coef -0.002 (0.002).
  - Legislation threshold applied: Compensation coef 0.116 (0.386); Judge bias coef 0.013 (0.027).
  - Seniority: Compensation coef 0.019 *** (0.004); Judge bias coef 0.000 (0.000).
  - Number of employees: Compensation coef -0.000 (0.000); Judge bias coef -0.000 (0.000).
  - Worker’s salary: Compensation coef -0.000 *** (0.000); Judge bias coef 0.000 (0.000).
  - Economic dismissal: Compensation coef 1.116 (0.683); Judge bias coef -0.025 (0.030).
  - Time between dismissal and Appeal Court: Compensation coef 0.001 (0.001); Judge bias coef -0.000 (0.000).
  - Joint F-Test: Compensation 0.0000; Judge bias 0.7458.
  - Note: Covariates include Appeal court fixed effects, year fixed effects, leave-one-out industry sales growth, and indicator for economic dismissals. Standard errors clustered at judge level.

- Firm-level randomization test for judge bias on compensation (Observations 4,847):
  - Number of workers in t-1: Compensation coef 11.120 * (5.763); Judge bias coef 0.778 (0.537).
  - Sales in t-1: Compensation coef 0.042 (0.033); Judge bias coef -0.003 * (0.002).
  - Total wages in t-1: Compensation coef -0.586 (0.497); Judge bias coef -0.066 (0.043).
  - Value added in t-1: Compensation coef 0.315 (0.277); Judge bias coef 0.024 (0.022).
  - Net income in t-1: Compensation coef -0.713 (0.676); Judge bias coef 0.009 (0.043).
  - Debt in t-1: Compensation coef 0.055 (0.140); Judge bias coef 0.011 (0.010).
  - Cash in t-1: Compensation coef -0.161 (0.201); Judge bias coef -0.017 (0.013).
  - Joint F-Test: Compensation 0.1312; Judge bias 0.2241.
  - Note: Dependent variables and covariates include Appeal court fixed effects, year fixed effects, leave-one-out industry sales growth, and indicator for economic dismissals. Independent variables divided by 1000 for clarity. Standard errors clustered at judge level.

### Firm-level samples and descriptive statistics
- Sample used to estimate effect of judge bias on firm performance:
  - Sample with judge fixed effects: 82,320 observations; 13,995 firms; 159 judges.
  - Non-missing employment, wages, Roa: 40,280 observations; 5,035 firms; 129 judges.
  - Firms with less than 100 employees: 35,888 observations; 4,486 firms; 129 judges.
  - Note: Employment = headcounts on 31 December before the judgment year; Wages = gross monthly wage; Roa = Return on assets. Number of observations corresponds to number of cases × number of years in sample.

- All firms (< 100 employees) summary (counts and statistics):
  - Nb of workers: mean 20.08; min 1.00; median 13; max 299.00; sd 20.55; count 4,486.
  - Sales (in K euros): mean 4,788.61; min 0.00; median 1,929.91; max 64,175.00; sd 7,429.92; count 4,428.
  - Share of firms in manufacturing: mean 0.17; min 0.00; median 0.00; max 1.00; sd 0.384; count 4,486.
  - Share of firms in construction: mean 0.11; min 0.00; median 0.00; max 1.00; sd 0.314; count 4,486.
  - Share of firms in services: mean 0.34; min 0.00; median 0.00; max 1.00; sd 0.484; count 4,486.
  - Share of firms < 10 years: mean 0.24; min 0.00; median 0.00; max 1.00; sd 0.434; count 4,486.
  - Survival at t+10: mean 0.99; min 0.00; median 1.00; max 1.00; sd 0.11; count 4,486.
  - Survival at t+20: mean 0.96; min 0.00; median 1.00; max 1.00; sd 0.18; count 4,486.
  - Survival at t+30: mean 0.92; min 0.00; median 1.00; max 1.00; sd 0.26; count 4,486.
  - Wrongful dismissal indicator: mean 0.52; min 0.00; median 0.00; max 1.00; sd 0.50; count 4,486.
  - Amount in months of salary (when >0): mean 11.07; min 0.01; median 8.08; max 197.47; sd 12.25; count 3,009.
  - Amount in payroll (%) (when >0): mean 10.75; min 0.00; median 4.06; max 149.75; sd 18.56; count 3,805.
  - Judge pro-worker bias: mean -0.03; min -2.76; median -0.01; max 2.22; sd 0.76; count 4,486.

- Small firms (< 10 employees) summary:
  - Nb of workers: mean 4.98; min 1.00; median 5.00; max 9.00; sd 2.38; count 1,902.
  - Sales (in K euros): mean 1,231.21; min 0.00; median 684.44; max 24,990.51; sd 2,136.92; count 1,902.
  - Share of firms in manufacturing: mean 0.11; sd 0.32; count 1,902.
  - Share of firms in construction: mean 0.10; sd 0.30; count 1,902.
  - Share of firms in services: mean 0.35; sd 0.48; count 1,902.
  - Share of firms < 10 years: mean 0.35; sd 0.48; count 1,902.
  - Survival at t+10: mean 0.98; sd 0.15; count 1,902.
  - Survival at t+20: mean 0.95; sd 0.22; count 1,902.
  - Survival at t+30: mean 0.89; sd 0.31; count 1,902.
  - Wrongful dismissal indicator: mean 0.52; sd 0.50; count 1,902.
  - Amount in months of salary (when >0): mean 9.75; min 0.01; median 7.23; max 78.15; sd 9.78; count 1,299.
  - Amount in annual payroll (%) (when >0): mean 19.28; min 0.00; median 10.56; max 381.36; sd 24.56; count 1,620.
  - Judge pro-worker bias: mean 0.00; min -2.76; median -0.01; max 2.22; sd 0.77; count 1,902.

### Judge pro-worker bias and firm performance (short-run outcomes)
- 1 year after judgment (sample sizes reported):
  - Dependent variables and coefficients (Pro-worker bias main effect, standard errors in parentheses):
    - Survival within [t,t+1]: coef -0.001 (0.001); R2 0.025; # obs 4,486.
    - Growth rate between t−1 and t+1 of Employment: coef -0.009 (0.006); R2 0.037; # obs 4,486.
    - Growth rate between t−1 and t+1 of Employment cdi: coef -0.003 (0.007); R2 0.037; # obs 4,112.
    - Growth rate between t−1 and t+1 of Employment cdd: coef 0.001 (0.017); R2 0.030; # obs 4,112.
    - Change between t−1 and t+1 in the share of permanent jobs (Share cdi): coef -0.000 (0.002); R2 0.032; # obs 4,118.
    - Growth rate between t−1 and t+1 of Sales: coef -0.007 (0.005); R2 0.036; # obs 4,418.
  - Heterogeneity by Roa (interaction terms):
    - Pro-worker bias × Low Roa: Employment coef -0.018 ** (0.008).
    - Pro-worker bias × High Roa: coefficients reported; none significant at 5% except interactions noted.
  - Covariates include Appeal court fixed effects, year fixed effects, leave-one-out industry sales growth, and indicator for economic dismissals. Standard errors clustered at judge level. *, **, *** denote significance at 10, 5 and 1%.

- 2 years after judgment:
  - Dependent variables and coefficients (Pro-worker bias main effect, standard errors in parentheses):
    - Survival within [t,t+2]: coef -0.003 (0.003); R2 0.035; # obs 4,486.
    - Growth rate between t−1 and t+2 of Employment: coef -0.015 ** (0.008); R2 0.043; # obs 4,486.
    - Growth rate between t−1 and t+2 of Employment cdi: coef -0.014 (0.009); R2 0.037; # obs 4,112.
    - Growth rate between t−1 and t+2 of Employment cdd: coef 0.008 (0.017); R2 0.026; # obs 4,112.
    - Change between t−1 and t+2 in share of permanent jobs: coef -0.004 (0.004); R2 0.033; # obs 4,112.
    - Growth rate between t−1 and t+2 of Sales: coef -0.017 ** (0.007); R2 0.030; # obs 4,395.
  - Heterogeneity by Roa (interaction terms):
    - Pro-worker bias × Low Roa: Survival coef -0.007 * (0.004); Employment coef -0.033 ** (0.011); Employment cdi coef -0.024 ** (0.012); Sales coef -0.030 *** (0.009).
    - Pro-worker bias × High Roa: coefficients reported with no significant adverse effects noted at conventional levels.
  - Covariates and estimation details as in 1-year results. Standard errors clustered at judge level. *, **, *** denote significance at 10, 5 and 1%.

*Source: Appeal court rulings database; DADS; FICUS-FARE; SIREN.*

### section 4.2. Low roa firms denote firms with a return on assets below the median the year before the judgment. Covariate

### Judge pro-worker bias and firm performance (section 4.2)

### Main estimates: effects 3 years after the judgment
- Table 15 — Correlation between judge pro-worker bias and firm outcomes (all firms)
  - Pro-worker bias (α1 estimates):
    - Survival within [t,t+3]: -0.007** (0.003)
    - Employment growth rate between t−1 and t+3: -0.015* (0.009)
    - Employment cdi growth rate between t−1 and t+3: -0.018* (0.010)
    - Share cdi change between t−1 and t+3: -0.008* (0.005)
    - Sales growth rate between t−1 and t+3: -0.023** (0.008)
    - R2 reported across columns: 0.044, 0.046, 0.046, 0.031, 0.039, 0.028
    - # obs: 4486.000 (columns 1–2), 4112.000 (columns 3–5), 4398.000 (column 6)
  - Pro-worker bias × Low Roa (β1):
    - Survival: -0.010** (0.005)
    - Employment growth: -0.035*** (0.010)
    - Employment cdi growth: -0.031** (0.012)
    - Share cdi change: -0.005 (0.023)
    - Sales growth: -0.047*** (0.011)
  - Pro-worker bias × High Roa (β2):
    - Survival: -0.004 (0.004)
    - Employment growth: 0.003 (0.012)
    - Employment cdi growth: -0.006 (0.014)
    - Share cdi change: 0.006 (0.033)
    - Sales growth: -0.001 (0.012)

- Interpretation: Judge pro-worker bias is associated with lower survival probability, lower overall employment growth, declines in permanent employment (cdi), and lower sales growth at the 3-year horizon, with stronger negative correlations for firms with Low Roa.

### Conditional on surviving firms
- Table 16 — Estimates conditional on surviving to t+3
  - Pro-worker bias (α1):
    - Employment growth: -0.002 (0.006)
    - Employment cdi growth: -0.003 (0.007)
    - Employment cdd growth: 0.008 (0.022)
    - Share cdi change: -0.000 (0.004)
    - Sales growth: -0.009* (0.006)
    - R2: 0.040, 0.038, 0.033, 0.027, 0.029
    - # obs: 4149.000, 3797.000 (cols 2–4), 4062.000 (col 5)
  - Pro-worker bias × Low Roa:
    - Employment growth: -0.016** (0.008)
    - Employment cdi growth: -0.013 (0.008)
    - Employment cdd growth: 0.006 (0.025)
    - Share cdi change: 0.000 (0.004)
    - Sales growth: -0.027** (0.008)
  - Pro-worker bias × High Roa:
    - Employment growth: 0.011 (0.009)
    - Employment cdi growth: 0.006 (0.012)
    - Employment cdd growth: 0.010 (0.033)
    - Share cdi change: -0.001 (0.005)
    - Sales growth: 0.006 (0.009)

- Interpretation: Conditional on survival, negative employment effects are concentrated among Low Roa firms; overall effects for surviving firms are smaller and less uniformly significant.

### Size heterogeneity: small vs medium firms
- Table 17 — Small firms (below 10 employees)
  - Pro-worker bias (α1):
    - Survival: -0.016** (0.007)
    - Employment growth: -0.016 (0.016)
    - Employment cdi growth: -0.027 (0.018)
    - Employment cdd growth: -0.012 (0.025)
    - Share cdi change: -0.017** (0.007)
    - Sales growth: -0.024 (0.017)
    - R2: 0.058, 0.059, 0.061, 0.069, 0.064, 0.053
    - # obs: 1902.000 (cols 1–2), 1750.000 (cols 3–5), 1893.000 (col 6)
  - Pro-worker bias × Low Roa:
    - Survival: -0.027** (0.011)
    - Employment growth: -0.064** (0.022)
    - Employment cdi growth: -0.058** (0.024)
    - Share cdi change: -0.023** (0.011)
    - Sales growth: -0.063** (0.021)
  - Pro-worker bias × High Roa: generally small and statistically insignificant

- Table 18 — Medium-sized firms (10 to 100 employees)
  - Pro-worker bias (α1):
    - Survival: 0.001 (0.003)
    - Employment growth: -0.014 (0.011)
    - Employment cdi growth: -0.011 (0.010)
    - Share cdi change: 0.000 (0.031)
    - Sales growth: -0.020** (0.009)
    - R2: 0.065, 0.082, 0.079, 0.051, 0.060, 0.054
    - # obs: 2581.000 (cols 1–2), 2359.000 (cols 3–5), 2502.000 (col 6)
  - Interactions with Roa show limited significant negative effects compared with small firms.

- Interpretation: Negative associations of judge pro-worker bias with firm performance are markedly stronger for small firms, especially those with Low Roa.

### Instrumental-variable evidence (judge bias as instrument)
- Table 19 — First-stage IV estimates (share of total compensation for wrongful dismissal instrumented by judge fixed-effect)
  - All size, All Roa: 1.910*** (0.262)
  - All size, Low Roa: 1.595*** (0.385)
  - All size, High Roa: 2.208*** (0.457)
  - All size, second specification: 2.267*** (0.500), 2.393*** (0.866), 2.924*** (0.928)
  - R2: 0.064, 0.064, 0.075, 0.075
  - F15.79, 13.17, 6.01, 5.05
  - # obs: 4486, 4486, 1902, 1902

- Table 20 — Second-stage IV estimates (Total amount = share of compensation in payroll, instrumented by judge bias)
  - Total amount (second stage):
    - Survival: -0.005** (0.002)
    - Employment growth: -0.011* (0.006)
    - Employment cdi growth: -0.013* (0.007)
    - Employment cdd growth: 0.000 (0.015)
    - Share cdi change: -0.006* (0.003)
    - Sales growth: -0.016** (0.005)
    - # obs: 4486.000, 4486.000, 4112.000, 4112.000, 4112.000, 4398.000
  - Total amount × Low Roa:
    - Survival: -0.008 (0.005)
    - Employment growth: -0.031** (0.013)
    - Employment cdi growth: -0.031** (0.015)
    - Employment cdd growth: 0.011 (0.023)
    - Share cdi change: -0.009 (0.007)
    - Sales growth: -0.041** (0.014)
  - Total amount × High Roa: coefficients small and not statistically significant

- Table 21 — Second-stage IV estimates for firms below 10 employees
  - Total amount:
    - Survival: -0.008** (0.004)
    - Employment growth: -0.008 (0.008)
    - Employment cdi growth: -0.013 (0.009)
    - Employment cdd growth: -0.006 (0.012)
    - Share cdi change: -0.008** (0.004)
    - Sales growth: -0.012 (0.009)
    - # obs: 1904.000, 1904.000, 1752.000, 1752.000, 1752.000, 1895.000
  - Total amount × Low Roa:
    - Survival: -0.014* (0.008)
    - Employment growth: -0.030* (0.016)
    - Employment cdi growth: -0.031 (0.019)
    - Employment cdd growth: 0.004 (0.022)
    - Share cdi change: -0.014 (0.009)
    - Sales growth: -0.036*** (0.018)
  - Total amount × High Roa: coefficients small and not statistically significant

- Interpretation: IV estimates corroborate that higher total compensations for wrongful dismissal (as driven by judge pro-worker bias) negatively affect firm survival, employment (especially permanent employment), and sales, with larger effects for Low Roa and for small firms.

### Robustness and nonlinearity checks
- Table 22 — Small low-performing firms (below 10 employees, Low Roa) with quadratic terms
  - Pro-worker bias:
    - Survival: -0.026** (0.012)
    - Employment growth: -0.058** (0.024)
    - Employment cdi growth: -0.055** (0.026)
    - Share cdi change: -0.025** (0.012)
    - Sales growth: -0.066** (0.023)
  - Pro-worker bias^2 (quadratic term): coefficients generally small and not statistically significant (e.g., survival -0.004 (0.006))
  - R2: 0.134, 0.111, 0.113, 0.109, 0.135, 0.104
  - # obs: 973.000 (cols 1–2), 911.000 (cols 3–5), 966.000 (col 6)

- Interpretation: Negative linear associations persist for small, low-performing firms; quadratic terms do not indicate strong nonlinear reversal.

### Placebo and pre-trend tests
- Table 23 — Placebo (growth between t−2 and t−1) — All firms and small firms
  - All firms:
    - Pro-worker bias × Low Roa: -0.005 (0.007) for employment growth between t−2 and t−1
    - Pro-worker bias × High Roa: -0.005 (0.005)
    - R2: 0.039, 0.047, 0.034, 0.042, 0.035
    - # obs: 4282.000, 3420.000, 3420.000, 3420.000, 4224.000 (panel structure as reported)
  - Small firms:
    - Pro-worker bias × Low Roa: -0.001 (0.011)
    - Pro-worker bias × High Roa: -0.005 (0.010)
    - R2: 0.083, 0.089, 0.067, 0.068, 0.075
    - # obs: 1843.000, 1477.000, 1477.000, 1477.000, 1829.000

- Interpretation: Placebo pre-judgment estimates do not show consistent significant pre-trends, supporting interpretation that effects appear after judgments.

### Alternative performance split: return on equity (Roe)
- Table 24 — Effects by Roe
  - All firms:
    - Pro-worker bias × Low Roe:
      - Survival: -0.011** (0.004)
      - Employment growth: -0.037** (0.011)
      - Employment cdi growth: -0.043*** (0.012)
      - Share cdi change: -0.013** (0.006)
      - Sales growth: -0.047*** (0.011)
    - Pro-worker bias × High Roe: coefficients small and generally not significant
    - R2: 0.044, 0.047, 0.046, 0.031, 0.039, 0.030
    - # obs: 4447.000, 4447.000, 4084.000, 4084.000, 4084.000, 4369.000
  - Small firms:
    - Pro-worker bias × Low Roe:
      - Survival: -0.025*** (0.007)
      - Employment growth: -0.062** (0.021)
      - Employment cdi growth: -0.087*** (0.020)
      - Share cdi change: -0.028** (0.010)
      - Sales growth: -0.063*** (0.017)
    - Pro-worker bias × High Roe: coefficients small and not significant
    - R2: 0.054, 0.057, 0.061, 0.069, 0.061, 0.058
    - # obs: 1887.000, 1887.000, 1735.000, 1735.000, 1735.000, 1878.000

- Interpretation: Splitting firms by Roe confirms pattern: negative post-judgment performance effects of pro-worker judge bias are concentrated among low-performing firms, especially among small firms.

*Source: DADS, FICUS-FARE, SIREN, Appeal court rulings database.*

### section 4.2. Low roe firms denote firms with a return on equity below the median the year before the judgment. Covariate

### wpiea2021031-print-pdf - section 4.2. Low roe firms denote firms with a return on equity below the median the year before the judgment. Covariate

### Judge pro-worker bias and firm performance — Large Appeal courts (Table 25)
- Dependent variables by column:
  - (1) Survival within [t,t+3] — indicator equal to one if the firm survives 3 years after the judgment.
  - (2) Growth rate between t−1 and t+3 of firm’s employment (symmetric growth rate).
  - (3) Growth rate between t−1 and t+3 of firm’s employment in permanent contract - cdi (symmetric growth rate).
  - (4) Growth rate between t−1 and t+3 of firm’s employment in temporary contract (symmetric growth rate).
  - (5) Change between t−1 and t+3 in the share of permanent jobs.
  - (6) Growth rate between t−1 and t+3 of sales (symmetric growth rate).

- Main estimates (upper part = coefficient α1 of equation (5)):
  - Pro-worker bias coefficients:
    - Column (1): -0.012**
    - Column (2): -0.035***
    - Column (3): -0.030**
    - Column (4): -0.039*
    - Column (5): -0.006
    - Column (6): -0.028**
  - R2 (first reported row): 0.0500.0580.0600.0500.0530.044

- Heterogeneity by firm profitability (bottom part = coefficients β1 and β2 of equation (6)):
  - Pro-worker bias × Low Roa. coefficients with standard errors:
    - Column (1): -0.010 (0.007)
    - Column (2): -0.038*** (0.010)
    - Column (3): -0.029** (0.013)
    - Column (4): 0.007 (0.026)
    - Column (5): -0.005 (0.009)
    - Column (6): -0.024* (0.012)
  - Pro-worker bias × High Roa coefficients with standard errors:
    - Column (1): -0.014*** (0.004)
    - Column (2): -0.032** (0.013)
    - Column (3): -0.030* (0.015)
    - Column (4): -0.084** (0.025)
    - Column (5): -0.006 (0.008)
    - Column (6): -0.032* (0.018)
  - R2 (bottom reported row): 0.0500.0580.0600.0520.0530.044

- Sample sizes:
  - # obs: 2074.000 2074.000 1907.000 1907.000 1907.000 2022.000

- Covariates and estimation details:
  - Covariates include Appeal court fixed effects, year fixed effects, the leave-one-out average industry annual growth rate of sales and an indicator variable for economic dismissals.
  - Standard errors clustered at the judge level.
  - *, **, and *** denote statistical significance at 10, 5 and 1%.
  - Low roe firms denote firms with a return on equity below the median the year before the judgment.
  - Source datasets: DADS, FICUS-FARE, SIREN, Appeal court rulings database.

### Figures and descriptive evidence (selected highlights)
- Temporal case volumes:
  - Figure 1 panels show new Prud’hommes cases per year (2006–2014) with values spanning roughly 140000–180000 on the left axis and new Appeal Court cases coming from Prud’hommes per year spanning roughly 45000–60000 on the right axis.
- Compensation distributions and relationships:
  - Histogram of compensation amounts in monthly wages displays only amounts lower than 50 months of salary (conditional on amount being positive).
  - Scatter plots relate compensations (in monthly wage) to seniority for Prud’hommes and Appeal courts.
  - Relation between compensations set by Appeal courts and by Prud’hommes shown as a scatter plot (compensations in monthly wage).
- Judge-level distributions and networks:
  - Histogram of frequency of dismissals deemed unfair per judge (case-level observations equal number of different cases with computable pro-worker bias).
  - Histogram of mean compensation per judge (in months of salary).
  - Network of judges: nodes represent judges; an edge connects two judges if they shared the same social chamber at least once (network density measures judge mobility).
  - Allocation of cases exploited for identification illustrated (within Appeal courts, multiple social chambers and chamber presidents; judge assignment changes enable identification).
- Pro-worker bias measures:
  - Histograms and local polynomial fits for pro-worker bias with respect to dismissal qualification and compensation in months of salary.
  - Scatter plot of correlation between two indices of pro-worker biases (dismissal qualification vs compensation amount, conditional on being positive).
- Counterfactual exercise:
  - Counterfactuals cap judge bias at several percentiles; predicted outcomes estimated from equation (5) with 1,000 bootstrap replications; reported are mean and 95% CI of differences in predicted survival rate and employment growth rate at 3-year horizon for firms with fewer than 10 employees whose return on assets is below the median the year preceding the judgment.

### Appendix A — Caps on dismissal compensation in European countries (selected exact rules)
- Italy: fixed amount introduced in 2014 (Jobs Act) for new indefinite-duration contract with progressive employment protection, depends on seniority: from 4 months for less than 2 years of seniority to 24 months for 12 years of seniority. In 2018 the Italian Constitutional Court overruled this regulation.
- Germany: schedule depends on seniority and reaches 12 months of salary (and even 15 months if the worker is more than 50 years old with more than 15 years of seniority, and 18 months if more than 55 years olds with more than 20 years of service).
- Austria: schedule depends on seniority: less than 2 years = 6 weeks of salary; between 2 and 5 years = 2 months; between 5 and 15 years = 3 months; between 15 and 25 months = 4 months; beyond that = 5 months of salary.
- Belgium: minimum compensation is 3 weeks and the maximum 17 weeks of salary.
- Denmark: worker compensation capped at 1 year of salary for blue-collar; for white-collar compensation up to half of the wages received during the notice period, capped at 3 months for those under 30, at 4 months if more than 10 years of service and 6 months if they have more than 15 years of service.
- Spain: indemnity set at 33 days per year of seniority with a maximum of 24 months of salary, for contracts signed since the 2012 labor market reform.
- Finland: allowance between 3 (minimum) and 24 (maximum) months of salary, depending on several factors including seniority, age, length of unemployment period, or loss of income.
- Netherlands: schedule depends above all on age (1/2 month of salary per year of seniority up to 35 years old, 1 month per year of seniority between 35 and 45 years old, 1.5 month per year of seniority between 45 and 55 years old, 2 months per year of seniority beyond 55), with possible correction factor. Compensation received at time of dismissal must be deduced from these amounts.
- Portugal: court may grant between 15 (minimum) and 45 (maximum) days of salary per year of seniority with a minimum of 3 months.
- United Kingdom: for employees with more than two years of seniority the allowance consists of two components (i) a basic allowance which depends on seniority and capped at £ 14,250 and (ii) a compensatory allowance capped at one year of salary and limited to £ 78,335.
- Sweden: allowance is 16 months of salary for employees with less than 5 years of seniority, 24 months between 5 and 10 years, and 32 months for more than 10 years.
- France: since 2017 (Ordonnances), compensation for unfair dismissal is capped by an amount that depends on seniority varying from 1 month to 20 months for employees with 30 year or more of tenure, and cannot be less that 3 months of salary for employees with at least 2 years of seniority (at least 11 years for those working in firms with fewer than 11 employees).

### Appendix B — Computation of judge bias (key expressions)
- Model setup:
  - y_ijkt = η_kt + ν_ijkt (B1) assuming E(ν_ijkt | η_kt) = 0.
  - Chamber×year fixed effect defined by the expectation:
    - η_kt = E(y_ijkt | k,t) (B2)
  - Sample counterpart:
    - ˆη_kt = (1 / n_kt) ∑_{i∈(k,t)} y_i (B3)
  - Estimator of the judge fixed effect conditional on chamber×year fixed effect:
    - ˆε_j = (1 / n_j) ∑_{i∈j} ˆν_i (B4)
  - Equivalently:
    - ˆε_j = (1 / n_j) ∑_{i∈j} y_i − (1 / n_j) ∑_{(k,t)∈(K,T)(j)} (n_jkt / n_kt) ˆη_kt (B5)
  - Equation (B5) shows ˆε_j equals ̄ε_j defined in equation (2).

### Appendix C — Judge mobility and ranking intuition
- Mobility across social chambers improves the probability of correctly ranking judges by pro-worker bias.
- Example logic: with four judges A,B,C,D and limited chamber co-membership, lack of mobility can lead to ambiguous or incorrect rankings; increased judge mobility across chambers allows comparisons that mitigate erroneous rankings.

### Appendix D — Risk premium associated with the bias of judges (derivation and numeric application)
- Relative risk premium π associated with random e defined by:
  - E(u[w(1 + e)]) = u[w(1 − π)] (D6)
- First- and second-order approximations yield:
  - π' (1/2) E(e^2) ρ(w), where ρ(w) = − (w u''(w)) / u'(w)
- For two lotteries with e and e1:
  - π1 − π' (1/2)[E(e1^2) − E(e^2)] ρ(w) (D7)
- Variance of total compensation:
  - V = E[(w(1 + e1))^2] − [E(w(1 + e1))]^2 = w^2 E(e1^2) (D8)
- If judge biases explain share λ of the variance of total compensation:
  - π1 − π' λ (1/2) ρ(w) E(e1^2) or equivalently:
  - π1 − π' λ (1/2) ρ(w) V[w(1 + e1)] / w^2
- Numeric inputs and result reported:
  - Standard deviation of compensations for wrongful dismissal = 56,385 euros.
  - Mean compensation = 31,461 euros.
  - Dispersion of judges biases explains 0.3% of the variance of compensations for wrongful dismissals.
  - Implied risk premium ≈ (0.003)(1/2)(56385/31461)^2 ≈ 0.0048 times the coefficient of relative risk aversion.
  - Estimated coefficient of relative risk aversion for workers between 1 and 3 (references noted in source).

### Appendix E — Extraction of compensation amounts and variables from Appeal court rulings
- Universe: close to 145,000 text documents (Appeal court rulings over ten years), each a few pages long, some over a dozen pages.
- Approach: Natural Language Techniques (NLP) exploiting the well-established template structure of rulings.
- Rulings structure split into roughly five blocks:
  - i) brief header with case number, date of audience, identities of the parties, etc.;
  - ii) description of the history of the contractual relationship and parties’ claims;
  - iii) restatement of the decision appealed;
  - iv) main arguments and reassessment by the Appeal Court; and
  - v) conclusion ruling whether the dismissal is deemed wrongful, and assigning monetary awards, if any.
- Procedure: split blocks by tagging text with specific legal keywords used to mark boundaries of sections.

*Source: DADS, FICUS-FARE, SIREN, Appeal court rulings database (as reported in wpiea2021031-print-pdf section 4.2).*

### conclusion is generally introduced by the expression "Par ces motifs" (For these reasons) or

### wpiea2021031-print-pdf - conclusion is generally introduced by the expression "Par ces motifs" (For these reasons) or

### Compensation types awarded by Appeal court judges
- Compensations tracked include:
  - indemnité pour licenciement sans cause réelle et sérieuse (compensation for wrongful dismissal)
  - compensation for non-respect of the dismissal procedure
  - indemnité pour rappel de salaire (compensation for unpaid wages)
  - indemnité pour préjudice moral et financier (compensation for moral and financial damages)
  - indemnité compensatrice de préavis (compensation in lieu of notice period)
  - compensation under article 700 of the French Code of Civil Procedure (legal costs)
  - indemnité compensatrice de congés payés (compensation for unpaid annual leave)
  - allowance for heures supplémentaires (overtime hours)
- An employee may receive these different compensations concurrently.
- More than twenty categories were initially tracked before aggregating to detect substitution between different monetary awards.

### Prevalence and breakdown of awards in the sample
- Original sample of court decisions: 145,000 cases.
- A positive amount is awarded to workers in 60% of the cases, whatever the motive.
- Of the cases receiving a positive amount:
  - dismissal is deemed unfair in 61% of the cases.
  - compensation for paid leave: 47% of cases.
  - compensation for advance notice: 40% of cases.
  - compensation for salaries: 13% of cases.
  - compensation for overtime hours: 7% of cases.
  - compensation for moral damage: 2% of cases.
  - compensation for harassment: 2% of cases.
  - compensation for discrimination: 0.3% of cases.
- One or several of these other types of compensation are awarded in 93% of the cases with a positive amount paid to the worker at the end of the trial.

### Data construction and variable scope
- Extracted variables cover:
  - compensations for wrongful dismissals
  - worker seniority
  - wage
  - Appeal court
  - city of the Prud’hommes council
  - whether it was the worker who appealed
  - firm’s name and address
- Using firm’s name and address, the firm identifier (SIREN) is retrieved to link the compensation dataset to matched employer-employee data and financial variables.
- Extraction challenges noted:
  - some amounts are expressed in French francs (pre-Euro 2001) and require appropriate conversion.
  - tenure often not explicitly stated as a duration and can require recovery from mentions of hiring dates.
  - wages reported variably (annual, monthly, weekly, hourly); a large number of keywords are targeted to detect wage mentions.
  - some court rulings lacked extractable information, creating missing observations.

### Variable selection, validation, and sample attrition
- Manual checks conducted on a subsample of 2,560 observations, selected at random.
- Manual dataset creation undertaken as part of a project of Pierre Cahuc and Stéphane Carcillo, funded by the Chaire sécurisation des parcours professionnels.
- Ten research assistants each handled a given year and searched Appeal court rulings with the keywords ‘licenciement sans cause réelle et sérieuse’ and ‘indemnités’.
- Selection irregularities: assistants for years 2009, 2010 and 2012 mostly selected court rulings of September and October, and marginally November and December.
- The correlation between the compensation amount of the manually-filled and the automatically-filled datasets is equal to 94%.

*Source: wpiea2021031-print-pdf - conclusion is generally introduced by the expression "Par ces motifs" (For these reasons) or*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021031-print-pdf.pdf_