Tasks, activities and skills: compare representations
Tasks and work activities
Grouping posting-derived task incidences into O*NET detailed work activities holds the activity vocabulary fixed while using postings to locate activities in occupations. These are the mean recovery results behind Figure 1G.
| Representation | SST (N = 604) | EMSI (N = 429) |
|---|---|---|
| Published O*NET lists | 11.22% | 17.75% |
| Posting tasks | 15.68% | 27.41% |
| Posting tasks grouped into DWAs | 14.78% | 25.72% |
Figure results (CSV) · Native versus posting-derived DWA profiles, common samples and paired intervals (CSV; SI Tables S10–S11)
The native-DWA comparison uses 590 SST and 422 EMSI origins with all representations available, rather than the larger samples above. Compare representations within the same sample.
Download the comparison results
Results, data definitions and code
Task-linked skills and abilities · CSV · Current figure work-activity comparison · CSV · Direct ESCO SkillMatch comparison · CSV
Capture and share columns are proportions (multiply by 100 for percentages). SkillMatch uses its paired extraction sample.
What people do at work provides information about the skills their jobs require. In these comparisons, profiles built from posting tasks recover more observed occupational transitions than profiles built from broader O*NET-linked skills and abilities or ESCO skills extracted with SkillMatch.
Detailed tasks support a measure of potential skills-based mobility.
Intermediate and generalized work activities
| Model | SST | EMSI | CPS + SIPP | NLSY97 + PSID · 3 years | NLSY97 + PSID · 5 years | NLSY97 + PSID · 10 years |
|---|---|---|---|---|---|---|
| Posting tasks | 15.68% | 27.41% | 21.15% | 14.79% | 13.10% | 10.47% |
| IWA | 12.99% | 21.45% | 19.06% | 11.75% | 10.76% | 8.95% |
| GWA | 9.80% | 13.45% | 16.40% | 9.91% | 8.73% | 6.60% |
Intermediate (IWA) and generalized (GWA) work activities are broader groupings of tasks. This table retains the standalone comparisons listed in SI Tables S17 and S19. The current DWA comparison is above.
Tasks or directly extracted ESCO skills?
On postings with both extractions, the posting-task model also captures more transitions than the 97-feature ESCO representation extracted by SkillMatch.
| Benchmark | Posting tasks | ESCO skills · SkillMatch |
|---|---|---|
| SST (N = 604) | 15.7% | 9.0% |
| EMSI (N = 429) | 27.5% | 13.0% |
| CPS + SIPP (N = 109) | 21.4% | 13.4% |
| NLSY97 + PSID · 3 years (N = 77) | 14.8% | 8.9% |
| NLSY97 + PSID · 5 years (N = 76) | 13.2% | 7.5% |
| NLSY97 + PSID · 10 years (N = 65) | 10.5% | 6.6% |
Code and data
Download SkillMatch validation code and aggregate inputs. Run python3 replicate.py to reconstruct all 24 model–benchmark results and origin-level comparisons. The package also includes commented profile-construction functions. Validation runs without database access; reconstructing the profiles requires restricted posting and extraction inputs. No posting records or unreleased task profiles are distributed.
For the main model: Python and Stata validation · Worked notebooks.
Methods: the paired SkillMatch experiment
Sampling, weights, and review
The frame contains non-null posting IDs in each monthly task-extraction table. Skill-only postings absent from that frame cannot enter. The annual occupation target is the smaller of the available count and the larger of 10% (rounded up) or 200. Allocate one observation to each nonempty month, then distribute remaining places proportionally to residual monthly capacity using largest remainders. Deterministic hash draws select postings. A repeated posting in another month is a separate observation. Each observation receives its occupation-month population count divided by its sample count as weight.
The paired sample requires nonempty task and skill IDs before support and review filters. Weights correct sampling rates, not extraction noncoverage. Pre-existing task-review exclusions are applied; unreviewed tasks remain. SkillMatch features do not receive that task-specific LLM review. These are comparisons of full pipelines, not extraction quality alone.
Profiles, occupation coverage, and validation
RCA divides each feature’s weighted incidence share within an occupation by its pooled weighted share across supported construction occupations. Task construction supports 1,012 occupations; SkillMatch supports 1,013. Their intersection with O*NET supports 922 specialties. The comparison pools these into 797 six-digit groups. A benchmark group enters only if all its reference specialties are supported. Benchmark origins/candidate codes are SST 604/671, EMSI 429/697, CPS + SIPP 109/401, and NLSY97 + PSID 77/325, 76/325, and 65/325 at three, five, and ten years. Candidate counts include the origin, which is excluded from its own list.
Group scores average specialty-pair similarities without renormalizing the pooled vectors. Positive scores are ranked after rounding to 12 decimals, breaking ties by destination code. All original outcome mass remains in the denominator, including destinations outside the candidate pool. Survey origins retain the frozen minimum of 40 transitions. Comparisons with O*NET use a tolerance of 10⁻¹². No uncertainty intervals are estimated here.
Advertised requirements are not observed worker competence. Historical transition capture is a descriptive validation of a measure of potential skills-based mobility; it does not estimate hiring probabilities or separate licensing, credentials, geography, wages, and other constraints.
Comparison rule: This representation experiment now uses the same per-origin cutoff as the paper’s comparisons using the same number of recommendations, applied to every model. Download the runnable calculation and aggregate inputs. These linked features are projections of tasks through O*NET, not direct SkillMatch extraction.
Another representation: task-linked skills and abilities
All models use the same corrected posting-task data and matched benchmark populations. SST measures the share of recorded transitions captured by the recommended destinations. EMSI measures the fraction of its original top-20 destinations recovered. The columns have different meanings and should be compared within each benchmark.
| Model | SST share captured · equal length | EMSI destinations recovered · equal length |
|---|---|---|
| Posting tasks | 15.7% | 27.4% |
| Task-linked skills | 7.0% | 8.9% |
| Task-linked abilities | 6.1% | 8.0% |
| O*NET published lists | 11.2% | 17.8% |
Reading the evidence: SST and survey percentages are the share of recorded transitions captured; EMSI is the fraction of its original top-20 destinations recovered. Each origin receives equal weight. These are historical benchmark comparisons, not estimates of hiring probabilities.
Explore the posting-task model → · Compare with published O*NET related occupations
Model construction: code, data and review decisions
Worked notebooks · Paper code and released scores · Excluded task–occupation links (CSV) · Methods and audit