What Works in Work Inclusion: A Weighted Review of the Causal Evidence

20 November 202615 min read
Share this article

InkludX Evidence Review, Paper 2 of 3. Job-first support (IPS), employer responsibility, benefit design, quotas, subsidies, discrimination and job design — each rated by strength of evidence, with the null results included.

InkludX Evidence Review on Work Inclusion · Paper 2 of 3 · November 2026

In brief

  • The best-evidenced approach is to place people in ordinary jobs first, then support them in the job. Individual Placement and Support (IPS) raises the share of people with mental illness who get a job from about 28% to about 49%. The effect is smaller in European welfare states and for common mental disorders, and some trials in that group find no significant effect.
  • Making employers responsible for sick employees keeps people in work, but it can also deter hiring. Dutch reforms cut disability applications sharply. The same reforms protected people who already had a job and may have made employers more reluctant to hire disabled newcomers.
  • Discrimination at the hiring stage is real, large and stubborn. Across field experiments, applicants who disclose a disability get substantially fewer positive responses: about 44% fewer in a meta-analysis, about 28% once outliers are excluded. In a large UK experiment, adding evidence of productivity did not close the gap.
  • Quotas and wage subsidies work only when they are designed to bite. Their effects are real but small, and depend on the size of the levy, how the subsidy is targeted, and the support around it.
  • The most promising frontier is how work is designed. In the United States, the rise of working from home explains most of the post-pandemic increase in full-time employment among people with physical disabilities.

How to read this paper

Paper 1 set out our four-tier hierarchy of evidence. Here each finding carries a strength rating:

  • High: several RCTs or a meta-analysis that point the same way.
  • Moderate: one strong trial or credible quasi-experiments.
  • Low: surveys, single small studies or modelling.

Studies older than three years are included only where they remain the defining evidence on a question, and they are dated. We report results that found no effect alongside positive ones.


1. Place, then train: Individual Placement and Support

IPS puts people straight into a search for an ordinary paid job, matched to their preferences. An employment specialist works inside the mental health team and keeps supporting the person once they are hired. It is the opposite of the traditional "train, then place" model of vocational rehabilitation.

What the strongest evidence says.

  • Meta-analysis. A 2022 meta-analysis of 32 randomised trials found that 48.8% of IPS participants were employed during follow-up, against 28.3% of controls (de Winter et al., 2022). IPS was effective in every subgroup. Strength: high.
  • Cochrane review. A Cochrane network meta-analysis of 48 trials (2017) found supported employment, especially when augmented with other interventions such as skills training, the most effective approach for getting and keeping competitive work for people with severe mental illness. It found no difference between interventions in hospital admissions (Cochrane).

The European reality check. The same meta-analysis found IPS relatively less effective in European studies and for depression and other common mental disorders. The authors suggest a "benefits trap" in high-welfare countries may be part of the reason. Trials in the Nordic countries and the UK confirm the pattern:

TrialFindingStrength
Norway, multicentre RCT (2019)37.4% vs 27.1% in competitive work at 18 months, against high-quality usual care: about +10 points (Reme et al.)Moderate–high
Norway, young adults (18–29) not in work or education (2020)About 48% vs 8% employed within 12 months, against traditional vocational rehabilitation (Sveinsdottir et al.). Small trialModerate
Denmark, three-arm RCT, 720 participants (2019)Over 18 months, 411 hours in work or education with IPS and 488 with enhanced IPS, against 341 with usual services. Adding cognitive and social-skills training gave nothing extra (Christensen et al.)High
Denmark, IPS adapted for mood and anxiety disorders (2017)44.4% vs 37.8%, not significant (Hellström et al.)Moderate
Norway, Supported Employment, 3,737 participantsNo more people in work than NAV's existing follow-up. Early gains faded (OsloMet / Frisch Centre)High
UK Health-led Employment Trials, 9,785 participants, mild to moderate conditions (2023)3–4 points more in work at best, null in one group, no effect on earnings (DWP)High

Our reading.

  • IPS is the best tool we have for people with severe mental illness, and for young people with a weak attachment to work. That is where trials compare it with genuinely passive alternatives and find large effects.
  • It is not proven for the fastest-growing group: people with common mental disorders, many of whom are already employed or on sick leave. In the UK trials, which covered people with mild to moderate conditions, the effects were small.
  • The size of the effect depends on what IPS is compared with. When the alternative is already individual and work-focused, as in the large Norwegian Supported Employment trial, the gain disappears. That is not a failure of the idea. It shows the active ingredient: individual, job-first support. A well-run welfare system can deliver that ingredient under other names.

Does it pay?

  • Norway. For people aged 18–40 starting a time-limited health-related benefit, adding IPS is estimated to return 4.76 to society and 1.34 to the public purse for every unit spent (Park et al., 2026).
  • Netherlands. A nationwide study found IPS was, on average, cost-saving compared with traditional vocational rehabilitation, with wide uncertainty (Vukadin et al., BMJ Public Health, 2024).

Strength: moderate. These results come from modelling and non-randomised data.

2. Act early, at the workplace: employer responsibility and gatekeeping

The Netherlands made employers pay up to two years of sick pay and required a documented return-to-work plan, the "Gatekeeper" protocol, before anyone could apply for disability benefit. It is the most studied reform of its kind.

  • Applications fell sharply. Stricter gatekeeping cut disability benefit application rates by 40% within a year, most among people with hard-to-verify impairments and milder conditions. The change in who received benefits came almost entirely from people choosing not to apply and from people resuming work during the sick-pay period (Godard, Koning & Lindeboom, 2024). Those deterred from applying had worse health and worked less than people who never applied, so screening has a cost. Strength: moderate–high.
  • Taking away employers' financial stake raised inflow. When the Netherlands stopped charging small firms more for their former employees' disability benefits (experience rating), inflow rose by about 7% and outflow fell by 12% (de Groot & Koning, 2016). Strength: moderate.
  • The reforms protected people already in work, and may have deterred hiring. Making employers pay sickness benefits lowered disability benefit receipt and raised employment after a health shock (Hullegie & Koning, 2018). An earlier version of the same study concluded that the reforms mainly protected people who already had a job and, on descriptive evidence, may have reduced hiring opportunities for people with disabilities (Hullegie & Koning, 2015). A 2025 study found that putting more financial responsibility on employers had no significant effect on how much they did to accommodate long-term sick workers (Jansen et al., 2025). Strength: moderate.

Our reading. Keeping people in work and recruiting people from outside are two different problems. Measures that make employers responsible for their own staff keep people in work. The same measures can make an employer warier of taking on someone who looks like a risk. A system that relies only on employer responsibility will protect the people who already have jobs and leave outsiders behind.

Partial sick leave and workplace adjustments: the evidence is mixed for mental health.

  • Finland. Early part-time sick leave raised sustained return to work by 8.0 percentage points among people off sick with mental disorders or musculoskeletal diseases. The gain in sustained return to work came from the mental-disorder group; there was no difference for musculoskeletal disease (Viikari-Juntura et al., 2017).
  • Norway. Graded sick certificates are linked to shorter absences (Markussen, Mykletun & Røed, 2010).
  • Mental health. A Dutch study found no such benefit for people with mental health problems (Kools & Koning, 2019). A Cochrane review found workplace interventions helped workers with musculoskeletal disorders, but no clear benefit for mental health problems (van Vilsteren et al., 2015).

Strength: moderate. The results point in different directions: graded work helped people with mental disorders in Finland, but not in the Netherlands, and workplace interventions in the Cochrane review helped mainly musculoskeletal conditions. Our tools were largely designed with physical conditions in mind, and for mental health we do not yet know reliably which ones work.

3. Benefit design: incentives work, exits are not inclusion

  • Many recipients can work, and incentives help. Using a sharp change in Norwegian benefit rules, Kostøl and Mogstad (2014) show that many disability recipients have considerable capacity to work, and that financial work incentives can unlock it (NBER). Strength: moderate–high.
  • Reassessment can raise employment. Reassessing Dutch disability recipients aged 30–44 under stricter rules raised employment by 6.7 percentage points (Garcia-Mandicó et al., 2020). Strength: moderate.
  • But many who lose benefits do not find work. In Hungary, of people who lost disability benefit after reassessment, 58% worked in the ordinary labour market. Another 36% ended up with neither work nor benefits. Among those who had no job before the reform, only 33% found work (Bíró et al., IFS, 2023). Strength: moderate–high.

Our reading. Tightening eligibility moves people off benefits. It moves people into work mainly when they already had a foothold in the labour market. For people without one, it mostly shifts hardship elsewhere. Incentives to work, and the chance to combine partial benefits with partial work, are the better-evidenced levers.

4. The demand side: discrimination is the barrier we talk about least

Discrimination is large and consistent across countries. In correspondence experiments, researchers send matched job applications that differ only in whether the applicant discloses a disability.

StudyFindingStrength
Meta-analysis, 2005–2020 (Lippens, Vermeiren & Baert)Applicants with a disability about 44% less likely to get a positive response. Based on only 13 estimates, with a wide uncertainty range; the category includes obesity and HIV, and excluding outliers gives about 28%Moderate
Sweden, about 2,000 applications (Bjørnshagen, Rooth & Ugreninov, 2023)Non-disabled applicants got 33% more callbacks than equally qualified wheelchair usersModerate–high
UK, about 4,000 vacancies (Antinyan, Burn & Jones, 2024)A 15% callback gap for wheelchair users. Adding references and extra skills widened the gap rather than closing itModerate
Norway and Denmark (Ugreninov, Krogh & Bjørnshagen, 2024)Wheelchair users needed about 1.7 (Norway) and 2.4 (Denmark) times as many applications per callback. Conference abstractLow–moderate
Norway, young applicants (Bjørnshagen, 2021)Disclosing past mental health problems cut the chance of an interview invitation by about 27%Moderate

Employers say yes in principle and hesitate in practice:

  • Norway. 94% of employers in NAV's 2025 survey agree that inclusion is a social responsibility (NAV).
  • Great Britain. Only 35% of employers are confident about recruiting disabled people or people with long-term health conditions (DWP employer survey 2024).
  • Germany. Asked why few people with severe disabilities are employed, more than half of establishments named a "lack of suitable tasks", by far the most common answer. Only 8% rated a higher compensatory levy as a suitable measure (IAB, 2025).
  • Systematic review. A review of 47 studies found the barriers employers cite most are an expectation of low productivity, an expectation of high cost, and lack of knowledge (Nagtegaal et al., 2023).

Strength: low–moderate. These are attitudes, not behaviour.

Accommodation is cheaper than employers expect. US data from the Job Accommodation Network show that 61% of accommodations cost nothing, and the median one-off cost is $300 (JAN). Strength: low. The sample is self-selected, but the gap between expected and actual cost is consistent.

Quotas: real effects, but small, and they depend on enforcement.

  • Austria. Firms at the quota threshold employ about 0.05 more disabled workers (about 20% more) (Lalive, Wuellrich & Zweimüller, 2009).
  • Hungary. When the non-compliance levy was raised steeply, firms above the threshold hired 0.24–0.29 more disabled workers (Krekó & Telegdy, 2022).
  • Germany. About a quarter of employers obliged to meet the quota fill none of their places (Paper 1).

Strength: moderate. A quota changes behaviour only when not complying costs more than complying.

Wage subsidies work, but watch who gets them.

  • Denmark. The introduction of the Flexjob scheme raised the target group's employment probability by 33 percentage points (Datta Gupta & Larsen, 2010). When the subsidy to public employers was cut, hiring shifted towards people who already had a job (Datta Gupta, Larsen & Thomsen, 2015).
  • Norway. An evaluation for the government judged that the temporary wage subsidy brings the highest social return of NAV's labour-market measures (Oslo Economics / Frisch Centre, 2024).
  • Belgium. Telling employers that an applicant came with a wage subsidy did not raise callbacks (Baert, 2016).

Strength: moderate. Money helps once an employer is already engaged. It does not overcome prejudice at the application stage.

The "business case" is the weakest link. The widely quoted finding that leaders in disability inclusion earn 1.6 times the revenue of other firms comes from a self-selected benchmark sample and is purely correlational (Accenture, 2023). Strength: low. We recommend not relying on it. The stronger and more honest claim is narrower: in a French firm panel, greater effort to hire disabled workers did not hurt economic performance (France Stratégie, 2020). The fiscal and social returns in Section 1 and Paper 1 are better evidenced.

5. Design the job, not just the person

Remote work. Bloom, Dahl and Rooth estimate that the rise in working from home raised full-time employment of Americans with physical disabilities by 9–11%. That explains 68–85% of the post-pandemic increase. The gains came mainly from more disabled people entering or staying in the labour force, rather than from employers hiring more (American Economic Association summary, 2026). Strength: moderate. We read the summary, not the full paper, and the evidence is from the US only. In a UK survey, 79% of disabled respondents said they would not apply for a job without remote options (Work Foundation, 2026). Strength: low.

AI in recruitment. The EU AI Act classifies AI used for recruitment and selection as high-risk. It explicitly names the risk of perpetuating discrimination against persons with disabilities (AI Act, recital 57). An audit of GPT-4 ranking CVs found disability bias. A custom GPT instructed in disability-justice principles reduced it but did not remove it, and gave no improvement for depression (Glazko et al., 2024). Strength: low, but the stakes are rising quickly.

Our reading. When German employers say there are "no suitable tasks", that is a statement about how work is organised, not about the people applying. Flexible hours, remote options, task redesign and partial work arrangements change who can do a job. They are underused, under-researched and largely missing from policy.


Weighing it all: an evidence map

LeverEffect on employmentStrengthMain caveat
IPS for severe mental illnessLargeHighSmaller in European welfare states
IPS for common mental disordersSmall or noneHighThe growing group lacks a proven model
Employer responsibility and gatekeepingLarge drop in inflowModerate–highMay deter hiring of outsiders
Partial sick leavePositive in some studiesModerateResults differ by condition and country
Work incentives in benefitsPositiveModerate–highDepends on design
Stricter eligibilityExits, few jobsModerate–highIn Hungary, over a third ended up with neither work nor benefits
QuotasSmallModerateNeeds a levy that bites
Wage subsidiesPositiveModerateCan go to people who already have jobs
Anti-discriminationLarge problem, few tested remediesModerateProductivity signals did not help in the UK test
Job design and remote workPromisingLow–moderateEvidence mostly from the US

Conclusions

  1. Individual, job-first support is the active ingredient, whatever it is called. Scale it for severe mental illness and for young people. For common mental disorders, test before scaling.
  2. Treat keeping people in work and recruiting outsiders as separate problems. Employer responsibility solves the first and can worsen the second. Recruitment needs its own tools: subsidies, intermediaries, and enforcement against discrimination.
  3. Judge every reform by how many people end up in work, not by how many leave benefits.
  4. The demand side is under-treated. Discrimination is measured directly, in controlled experiments, and remedied rarely. Quotas without a levy that bites, and attitudes without employer capacity, change little.
  5. The biggest untested lever is how work is designed.

Paper 3 turns these findings into an agenda.


Sources

All sources are linked in the text. Studies older than three years are included as the defining evidence on their question: Cochrane reviews (2015, 2017); the Norwegian IPS trials (2019, 2020); the Danish trials (2017, 2019); Kostøl & Mogstad (2014); de Groot & Koning (2016); Hullegie & Koning (2018); Kools & Koning (2019); Viikari-Juntura et al. (2017); Markussen, Mykletun & Røed (2010); Garcia-Mandicó et al. (2020); Lalive, Wuellrich & Zweimüller (2009); Krekó & Telegdy (2022); Datta Gupta & Larsen (2010); Datta Gupta, Larsen & Thomsen (2015); Baert (2016); Bjørnshagen (2021); the Lippens, Vermeiren & Baert meta-analysis (2021, published 2023); France Stratégie (2020). The Supported Employment trial was run 2014–2017; the report summarised by OsloMet was published in 2021.

Tags

evidence-reviewipssupported-employmentemployersdiscriminationpolicy
Share this article