Data
Sample
| Item | Value |
|---|---|
| Raw responses | 282 |
| Dropped for missing data | 38 |
| Analytic sample | 244 |
| Distinct firms | 219 |
| Firms appearing in more than one wave | 48 (approximately 22%) |
| Observations from multi-wave firms | 111 |
| Firms appearing once | 171 |
| Waves | 2022–2025 |
| Firm size range | 1–300 employees |
Missingness was tested before deletion rather than after; the test was consistent with data missing completely at random, which is what licenses listwise deletion here.
Composition
| Wave | N |
|---|---|
| 2022 | 32 |
| 2023 | 71 |
| 2024 | 73 |
| 2025 | 68 |
Sector composition in the first wave: manufacturing 56%, IT 22%, services 22%.
Geographic distribution: Capital area 38%, Yeongnam 31%, Chungcheong 18%, Honam 13%.
Measures
| Construct | Operationalization |
|---|---|
| Technology (T) | Digital transformation awareness, self-reported |
| Environment (E) | Smart-system infrastructure score |
| Organization (O) | Firm size, log-transformed |
| Training exposure | Programme participation intensity |
| Outcome | Post-minus-pre DT competency, single five-point self-report item |
Test–retest reliability on the multi-wave subset: T r=.78, E r=.82.
The outcome is a single self-report item answered by the same respondent who supplies the T and E predictors. The study’s own argument is that single self-report items under-capture training effects — which means the instrument that motivates the argument is also the instrument the estimates depend on. Resolving this is the first item of the pending revision, and it changes the dependent variable rather than merely adding a robustness check.
Structural caveat
The distinction between 219 firms and 244 observations matters for method selection. Methods that require within-firm variation are identified only on the 48 firms observed more than once. Applying them to the pooled sample and relying on year fixed effects does not recover firm-level identification.