In the public aggregates, no: when a Danish municipality's long term unemployment rose between 2009 and 2021, its rate of new firms showed no detectable change, yet person by person the unemployed became self employed at 5.6 times the rate of employees, mostly because they change status far more often. No public table can join those two facts, because none says who the founders were, which job they left, how long they had been out of work, or how exposed that job was to generative AI. That join is what the project's registry data makes possible, and this page shows where each public table stops.
The same 1,274 municipality years give a negative link or none, depending on what is compared with what
What you see: five estimates of one number from the same data: how many more new firms per 1,000 people in the labour force go with one percentage point more long term unemployment, 98 municipalities, 2009 to 2021. Each bar is a 95 percent confidence interval.
How to read it: a dot left of the dashed zero line means more long term unemployment goes with fewer new firms; a bar that crosses zero means the data cannot tell the sign.
What to take from it: the negative link comes from comparing different places, and municipality effects alone remove it (0.07). Adding year effects absorbs the national rise in long term unemployment from 0.5 to 1.5 percent between 2009 and 2011, which is why the main interval is wider. A pattern across places says little about what unemployed people do.
The page uses a few terms from Danish labour statistics and from panel econometrics. These five statements and the glossary are enough to read every chart.
The question on this page is the first step of the project's own question, and the 2019 paper is the reason duration and home industry appear in every act below.
“We find that unemployment duration affects whether ventures are launched in ‘home’ or in external industries, and moderates the extent to which founders’ industry experience and the attractiveness of external opportunities relative to those in the ‘home’ industry shape industry choice.” Nikiforou, Dencker and Gruber (2019), Necessity entrepreneurship and industry choice in new firm creation, Strategic Management Journal 40(13), 2165 to 2190, doi 10.1002/smj.3075. Abstract, published under CC BY.
The paper analysed primary data on 576 necessity entrepreneurs in Greece, combined with three secondary data sets. The long term unemployed were less likely than the short term unemployed to stay in their home industry, and more susceptible to the sunk cost fallacy (sticking with a course because of what has already been invested in it). When the long term unemployed had high industry specific experience, they were more likely to stay in their home industry, so unemployment duration changes how experience and the attractiveness of other industries shape the choice (DTU Entrepreneurship, July 2020).
The PhD project, The New Landscape of Necessity Entrepreneurship, funded by the Swiss National Science Foundation, asks how AI and automation shape the move from unemployment into entrepreneurship, with surveys, interviews and Danish administrative registers (DTU job advertisement 8416). Before using registers, I wanted to know how far the free public tables go on the same question.
If unemployment pushed people into founding, a municipality whose long term unemployment rose should see more new firms. This is the cheapest test of the project's question with public data, and the first one a reader of StatBank would run.
I joined three tables for Denmark's 98 municipalities from 2009 to 2021: new enterprises (DEMO9), long term unemployment as a percentage of the labour force (AULK04, the mean of the twelve monthly values) and the labour force at the end of November (RAS400). The outcome is new firms per 1,000 people in the labour force, which averages 11.5 across the 1,274 municipality years.
DEMO9 also has a column of employees, but it counts the staff of the new firms themselves: 8,786 full time persons for 27,024 new firms in 2009. I therefore divide by the labour force instead.
Switch the view to remove each municipality's own average and each year's national swing; the dots move to where only the changes within a municipality remain. Pick a municipality to draw its 2009 to 2021 path.
What you see: each dot is one municipality in one year: long term unemployment across, new firms per 1,000 people in the labour force up. The line is the fitted slope for the view you chose.
How to read it: in the first view, the dots of one municipality sit together, so the slope mostly compares places. In the second, each dot shows how far that municipality was from its own average and from that year's national level, with the overall averages added back so the axes keep their units.
What to take from it: the pooled slope is −0.84; following each municipality over time, it is 0.04 (95 percent interval −0.57 to 0.66).
Each row below is a two way fixed effects regression with municipality and year effects and standard errors clustered by municipality, estimated with PanelOLS in the Python package linearmodels. The units are the same as in the chart.
| Specification | Estimate | 95 percent interval | Years | Observations |
|---|---|---|---|---|
| Main estimate: long term unemployment, same year | 0.04 | −0.57 to 0.66 | 2009 to 2021 | 1,274 |
| Long term unemployment of the year before | 0.42 | −0.12 to 0.97 | 2010 to 2021 | 1,176 |
| All gross unemployment in full time equivalents (AUS08) | 0.02 | −0.21 to 0.24 | 2009 to 2021 | 1,274 |
| All gross unemployment in full time equivalents, year before | 0.01 | −0.21 to 0.24 | 2009 to 2021 | 1,274 |
| New firms per 1,000 salaried employees working in the municipality (RAS305) | 0.00 | −0.58 to 0.57 | 2009 to 2021 | 1,274 |
| Municipalities weighted by their labour force | 0.04 | −0.40 to 0.48 | 2009 to 2021 | 1,274 |
| Leaving out 2020 and 2021 | −0.01 | −0.62 to 0.60 | 2009 to 2019 | 1,078 |
The shaded row is the main estimate. The lagged row loses 2009 because AULK04 starts in January 2009. The largest estimate, 0.42 for the year before, still has an interval from −0.12 to 0.97.
The two way fixed effects estimate is 0.04 new firms per 1,000 people in the labour force per percentage point of long term unemployment (95 percent interval −0.57 to 0.66, 1,274 municipality years), against an average of 11.5. The negative link in the pooled data comes from stable differences between municipalities, so reading it as a statement about unemployed people would be the ecological fallacy.
The estimate is aggregate and not causal. A municipality's new firms mix founders who were unemployed with founders who left jobs, and a downturn can push the first group in while it holds the second group back.
Places cannot answer the project's question, because the people who found are hidden inside the totals. The next act therefore looks at the closest public table to people: how many of the unemployed become self employed each quarter.
The project's question is about individuals, and FLOW1 is the public table closest to them: it counts people whose labour market status changed from one quarter to the next.
For each year I divided the number of people aged 16 to 69 who went from unemployed to self employed by everyone who was unemployed at the start of the quarter, and did the same for employees. Pooled over 2021 to 2024, the rate is 1.41 percent per quarter for the unemployed against 0.25 percent for employees.
Most of that gap comes from how often the unemployed change status at all. Among people who leave their status in a quarter, the unemployed choose self employment less often than employees who leave employee work.
| Per quarter, 2021 to 2024 | Unemployed | Employees |
|---|---|---|
| Leave their status (unemployment, or employee work) | 52% | 4.5% |
| Of those who leave, become self employed | 2.7% | 5.7% |
| All, become self employed | 1.41% | 0.25% |
Click a year to redraw the rates for that year; the first chip pools 2021 to 2024.
What you see: for each age group, the share of people who were self employed one quarter later, among those who started the quarter unemployed (dark dot) or as employees (open dot).
How to read it: the longer the line between the two dots, the larger the gap between the two groups at that age.
What to take from it: at every working age the unemployed move into self employment far more often than employees, with the peak at ages 45 to 49 (1.91 percent per quarter, pooled).
The rate for 2020 is lower, 0.91 percent, because the pandemic swelled the number of unemployed, not because fewer became self employed: 3,519 did, more than in any year from 2022 to 2024. The unemployed base was 387,172 person quarters in 2020, against 251,791 at most in 2022 to 2024.
Most new self employed still come from jobs, because there are far more employees than unemployed people: in the three quarter changes FLOW1 reports for 2024, 3,164 moves into self employment came from unemployment, against 19,776 from employee jobs.
The 2019 paper's key variable is time spent unemployed. StatBank's AULK09 counts, each month, the people whose current gross unemployment spell has lasted at least 26, 39, 52, 78, 104, 130 or 156 weeks, which shows how large the long term group is at any time.
Drag the slider under the chart to move the dashed line to any month; the table below it gives every duration for that month.
What you see: the number of people in a spell of at least 26, 52 or 104 weeks, each month from July 2008 to May 2026. Each line starts when the register has enough weeks of history to count it.
How to read it: the gap between two lines is the number of people whose spell is between those two durations.
What to take from it: in May 2026, 11,418 people had been unemployed for at least a year, against 41,995 at the peak in February 2011.
No public table follows these people into their next status, so the duration result of the 2019 paper cannot be checked here. AULK09 counts spells and FLOW1 counts moves, and nothing links the two.
Each quarter in 2021 to 2024, 1.41 percent of the unemployed became self employed, against 0.25 percent of employees. FLOW1 does not say which industry they entered, what job they had before, or how long they had been unemployed.
If the unemployed found at this rate, the project's next question is which jobs they come from, and whether generative AI is changing those jobs. Act three asks what the public record shows about that.
The project asks how AI shapes the move into entrepreneurship. Before any link to people, the first public fact is where generative AI exposure sits in the Danish job stock, and whether exposed jobs are already shrinking.
I joined the ILO's 2025 generative AI exposure index (Gmyrek and others 2025, ILO Working Paper 140), which scores each of 427 ISCO-08 occupations by how much of its tasks generative AI could perform, to Statistics Denmark's count of full time employees in each DISCO-08 occupation (LONS20), 2013 to 2025. Of 423 Danish four digit codes, 412 match, covering 93 percent of employees in 2025.
Full time employees by ILO exposure group, as a share of matched employees
What you see: one bar per year, split into the ILO's six exposure groups, from occupations not exposed on the left to gradient 4, the most exposed, on the right.
How to read it: gradients 3 and 4 are occupations whose average task score is at least 0.5; the number at the right end of each bar is their combined share.
What to take from it: the two most exposed groups held 18.6 percent of full time employees in 2013, 18.4 percent in 2022 and 18.6 percent in 2025.
ChatGPT was released at the end of November 2022, so 2022 is the last full year before generative AI tools were widely used. For each occupation with at least 500 full time employees at the start of a period, I compared its share of all matched employees at the end of the period with its share at the start.
Switch the period to compare the years after ChatGPT with the three years before it. Type or tap an occupation to read its numbers.
What you see: each circle is one occupation, sized by its employees at the start of the period: ILO exposure score across, change in its share of employees up. Dark circles are gradients 3 and 4.
How to read it: if exposed occupations were shrinking, the dark circles on the right would sit below the dashed line.
What to take from it: across 263 occupations the correlation between exposure and the 2022 to 2025 change in share is 0.02, the same as for 2019 to 2022 (0.02).
Weighted by employees, the correlation is 0.21 after ChatGPT and 0.09 before, pulled up by large exposed occupations that grew, such as software developers; the rank correlation after ChatGPT is 0.07.
Across 263 occupations, exposure and the 2022 to 2025 change in employment share correlate at 0.02, as they did before ChatGPT (0.02). Exposure is a technical potential, not job loss, and LONS20 leaves out private firms with fewer than ten full time employees and the self employed, the group the project studies.
The job stock shows no AI shift yet, and the stock is the wrong unit anyway: the project needs the job each founder left. That is the last panel, and it is empty in the public record.
Each act above stopped at the same place: the public tables count groups, and the project's question is about the path of one person from a job, through unemployment, into a firm.
| What the project needs | Closest public table | What it lacks | What registry data adds |
|---|---|---|---|
| Previous occupation and industry of each founder | FLOW1 status moves; LONS20 employees by occupation | No link from a founder to the job they left | Employer, industry and DISCO code of the last job, linked by person (registers such as IDAN and AKM) |
| Unemployment spell length per person | AULK09 counts of spells | Spells are counted, never followed into self employment | Weekly benefit records (DREAM), so the exact spell before founding |
| Founding in the home industry or a new one | DEMO11 new enterprises by industry | Firms by industry, with no founder attached | The industry of the new firm next to the industry of the founder's last job |
| Education and age of each founder | FLOW1 by age band | Age bands only, no education for founders | Highest education (UDDA) and exact age |
| AI exposure of the job each founder left | The ILO index joined to LONS20, act 3 | Exposure of the job stock, not of the people who leave it | The ILO score of each founder's last occupation |
These are the places where a reader could be misled by the numbers above, stated before they find them.
Three Python scripts in the repository produce everything on this page. fetch.py downloads the eleven StatBank tables through the StatBank API and the ILO scores from the authors' repository (commit ce5fc25), and records the date. analyse.py computes every estimate and writes the CSV files in data/. render.py writes the numbers into this page and stops if any number in the text is not computed.
git clone https://github.com/bolgacg/founding-after-unemployment cd founding-after-unemployment python3 -m venv .venv && .venv/bin/pip install -r analysis/requirements.txt python3 -I analysis/fetch.py raw .venv/bin/python -I analysis/analyse.py raw data python3 -I analysis/render.py analysis/page.html data/results.json index.html
A rerun on a later day fetches newer data, so some numbers may change as Statistics Denmark revises its tables.