When R&D Taxonomies Travel: Target-Population Validity in the Selection of Young Technology-Oriented SMEs

Technical Summary: When R&D Taxonomies Travel: Target-Population Validity in the Selection of Young Technology-Oriented SMEs

Problem Statement
Technology-transfer and innovation-support systems rely on sector taxonomies (e.g., OECD R&D intensity classifications) to filter large pools of firms for costly, detailed due diligence. The paper identifies a critical validity gap: a taxonomy may be descriptively valid (accurately measuring R&D intensity in its original reference population) but selection-invalid when applied to a different target population. Specifically, whole-industry R&D aggregates are often dominated by large incumbents, whereas support programs frequently target young Small and Medium-sized Enterprises (SMEs). Applying a taxonomy calibrated on a "whole-industry" population to a "young SME" population constitutes an untested transport that may distort the sector ordering and exclude viable candidates (false negatives) or include non-viable ones.

Methodology
The study utilizes a proprietary panel from the Industrial Bank of Korea (IBK) covering 116,193 young operating SMEs (aged 0–6 years) across 54 two-digit industries during the 2012–2015 window. The analysis evaluates downstream outcomes from 2016–2019.

Three sector screening rules are constructed and compared at nearly identical candidate capacities (approx. 29,000 firms):

1. OECD Whole-Industry R&D: The standard benchmark selecting industries based on international whole-industry R&D intensity.
2. SME-Population R&D: The same R&D construct recalibrated using R&D intensity data specifically from the observed SME population in the early window.
3. SME Multi-Signal Screen: An expanded measure combining SME R&D intensity with registered patent stocks and recognized intangible assets.

The primary downstream outcome is a "joint-success" criterion: firms must achieve at least 10% annualized sales growth (2016–2019) and maintain a credit rating of BBB or better. Secondary outcomes include high growth alone, rating thresholds, and later entry into formal R&D activity.

The empirical strategy employs linear probability models and logit specifications with extensive controls (firm age, size, leverage, prior growth, profitability, legal form, and prior technological activity). Crucially, standard errors are clustered by the 54 industries, and Webb wild-cluster bootstrap inference is used to account for the fact that screen assignment occurs at the industry level, not the firm level.

Key Results

1. Superiority of Population Alignment: At equal candidate capacity, both SME-calibrated screens outperform the OECD whole-industry screen. The SME R&D screen identifies 274 more joint-success firms than the OECD screen, and the multi-signal screen identifies 196 more. In terms of precision, the SME screens achieve ~26.4% compared to the OECD's ~25.0%.
2. Statistical Significance: After controlling for firm-level characteristics and prior technological activity, the SME R&D and multi-signal screens are associated with 2.23 and 2.39 percentage-point higher probabilities of joint success, respectively. The OECD screen's adjusted coefficient (0.85 points) is statistically imprecise.
3. Mechanism Decomposition: The performance gain is driven almost entirely by target-population alignment (recalibrating R&D on the SME population) rather than construct expansion (adding patent/intangible signals). Moving from OECD to SME R&D yields significant gains; moving from SME R&D to the multi-signal screen yields negligible marginal improvement in precision or recall.
4. Downstream Implications: The SME screens are significantly associated with a higher probability of firms entering formal R&D activity later, even among those with no early R&D. However, the screens do not predict high sales growth in isolation; their value lies in identifying firms that combine growth with financial viability.
5. Architecture Trade-offs: A dual-route architecture (admitting firms selected by either OECD or SME rules) increases recall but expands the candidate pool. The study calculates that this dual route is only cost-effective if the cost of evaluating an additional candidate is less than ~11% of the value of avoiding a false negative.

Significance and Claims
The paper introduces the concept of target-population validity as a distinct requirement in technology-transfer screening, situated between descriptive validity (does the measure work in its reference population?) and selection validity (does the measure select candidates who succeed downstream?).

The authors claim three primary contributions:

1. Theoretical: It distinguishes between descriptive, target-population, and selection validity, arguing that a taxonomy can be descriptively accurate yet selection-invalid due to population mismatch.
2. Empirical: It provides a decomposition showing that population alignment is a more potent remedy for selection failure than construct expansion. The study suggests that the perceived weakness of R&D screening often stems from measuring the construct in the wrong population, not from the construct itself being too narrow.
3. Practical: It connects measurement validity to screening architecture, demonstrating that while dual-route systems (polyarchies) increase recall, they are only efficient under specific cost-benefit conditions.

The paper maintains a modest stance, clarifying that it evaluates simulated screening rules rather than the causal effect of an actual policy change. It acknowledges limitations regarding the specific IBK bank population (excluding pre-revenue ventures) and the use of bank ratings as a proxy for viability rather than direct investor returns. The core argument is that before adopting a taxonomy, organizations must verify whether the taxonomy's reference population aligns with their specific target population.