The study uses a cross-sectional survey design, estimating a reflective, multi-path structural model with indirect pathways through BPR, OA, and DI and one interaction term, via standardized composite scores, path coefficients, and bootstrap inference. Associational language reflects a theorized, not demonstrated, direction of influence (Section 9).

The target population is managerial and operational employees with visibility into their firm’s digital initiatives across six Middle Eastern markets: Egypt, Saudi Arabia, the UAE, Jordan, Morocco, and Qatar. In total, 336 respondents were nested within 50 firms, not 336 independent firms. The unitofobservationistheindividualrespondent; sampling was organized around these 50 firms, with multiple eligible respondents surveyed within each where available, so respondents are nested within—not each representing—an independent organization. The constructs are of theoretical interest at the firm level, but every item asks a respondent to rate firm-level phenomena from their own individual vantage point; what is actually collected is therefore respondent-level perceptual data about firm-level practices, not audited firm-level records, and the 336 observations are not 336 independent firm-level data points. Firm-clustered standard errors (Section 6.6) address the resulting non-independence of observations within firms, but—as Section 6.6 makes explicit—that correction does not by itself license firm-level conclusions; the hypotheses and results are stated at the level they were actually measured, the individual respondent’s perception.

Analyzing at the respondent level, rather than aggregating each firm’s responses into a single score, was deliberate: aggregation would discard the within-firm variation in how digital initiatives are experienced across roles—as the near-zero within-firm agreement statistics in Section 6.6 show, that variation is substantive, not noise. A key-informant design sampling one respondent per firm would avoid the nesting problem but at the cost of that variation; retaining multiple respondents and modeling the dependence through firm-clustered standard errors and, throughout Section 7, firm-block percentile bootstraps (Section 6.6), was judged the better trade-off.

Purposive sampling was used at the respondent level: employment in one of the 50 firms; middle-management level or above, or an operational role directly involved in digital or process decisions; and sufficient tenure to observe process-change activity. Recruitment combined online panel distribution (68%) and direct company HR-channel distribution (32%) during January-March 2026, yielding 351 responses from 520 invitations. Purposive rather than probability sampling suits the need for informed respondents, at the cost of representativeness at both levels (Section 9).

SurveyInstrument,Language,andTranslation

The questionnaire was administered in both English and Arabic, since respondents across the six markets vary in language preference; respondents chose the version they were more comfortable with. The instrument was originally developed in English, forward-translated into Arabic, and then independently back-translated from Arabic into English; the back-translation was compared against the original English wording and discrepancies were reconciled before fielding, following standard cross-language survey practice. The specific translator qualifications and pilot-sample size were not centrally logged and are not reported here as a precise figure.

The raw file (n = 351) was screened for duplicates, out-of-range codes, careless responding, and excessive missingness (Table 1): six duplicate patterns, five straight-line responders, and four respondents with over 30% missing items were removed (351 − 6 − 5 − 4 = 336), four out-of-range Likert codes were recoded to missing, and remaining item-level missingness (304 cells, under 6% of the matrix) was mean-imputed; extreme firm-age and experience values were winsorized at the 99th percentile rather than deleted (Section 7.10). This same n = 336 analytic file was used consistently to produce every table and figure in this paper.

Table 1. Data screening summary.

A distinction runs through everything that follows: items ask each respondent to rate firm-level phenomena from their own individual vantage point, so what is collected is respondent-level perceptual data about firm-level constructs, not verified firm-level records (Section 6.2). Given the near-zero ICCs reported in Section 6.6, every path coefficient in this study should be read at that level.

All six constructs were operationalized as reflective latent variables on five-point Likert scales: each construct’s items are treated as manifestations, or reflections, of a single underlying concept rather than distinct sub-components that jointly define it—the appropriate specification when indicators are interchangeable expressions of the same latent tendency, so high inter-item correlation is expected and removing any one item should not change its meaning. DC and BPR (5 items each) received closer scrutiny under this logic: their indicators could plausibly be separable drivers that jointly compose the construct rather than correlated manifestations of it—a distinction not formally tested here, acknowledged as a limitation (Section 9).

Applying Jarvis, MacKenzie, and Podsakoff’s [23] conceptual criteria offers further, if not definitive, support: the DC and BPR items read as manifestations of one underlying disposition, dropping any single item would not alter the construct’s meaning, and the items covary positively and substantially (Section 7.5)—consistent with a reflective structure, though only a partial substitute for confirmatory tetrad analysis [24], a priority for future replication.

OA (4 items), DI (5 items), and BP (4 items) more straightforwardly satisfy the reflective criterion. DL was reduced from 4 items to 2 (Section 7.4): the retained pair captures a specific facet of digital leadership, sponsorship and modelling, rather than the full construct, a caveat carried through Sections 7.9 and 8.3. Full item wording is in Appendix A (Table A1); no reverse-worded items were identified.

Because multiple respondents were drawn from the same firms, the 336 respondent-level observations are not statistically independent, as respondents within a firm likely share firm-level influences. Three corrections for this nesting appear in this paper. First, the structural model was re-estimated with standard errors clustered by firm ID (50 clusters; Section 7.10); this analytic correction addresses within-firm dependence for the eight base paths, and every base path significant under the unclustered estimate remained significant under firm-clustered inference. Second, and more directly responsive to the nesting itself, every percentile bootstrap reported in this paper—the base structural paths and the DL main effect and interaction (Sections 7.7 and 7.9), the item-level loadings and HTMT ratios (Section 7.5), the H8 and H9 indirect effects and their decomposition (Section 7.8), and the exploratory serial pathway (Section 7.12)—was re-run as a firm-block bootstrap: each resample draws 50 firms with replacement (not individual respondents) and retains every respondent belonging to a sampled firm intact, so the within-firm dependence structure is preserved inside each resample rather than only corrected for afterward. Results were materially unchanged from the earlier respondent-level bootstrap on every substantive conclusion—the same paths are significant, H9 remains restricted to a significant OA-mediated route, and H10’s interaction is, if anything, more clearly centered on zero ([−0.146, 0.152], firm-block, versus [−0.151, 0.164] previously)—which is itself evidence that firm-level non-independence was not driving any of the paper’s conclusions. Third, as a robustness cross-check, a two-level (respondents nested in firms) multilevel re-specification remains a natural next step (Section 10), though it was not run here given the firm-block bootstrap already directly addresses the dependence concern.

Beyond correcting standard errors, this nested design does not require respondents’ reports to converge into a shared firm-level perception, which would call for within-firm agreement statistics such as r_wg or ICC(1)/ICC(2) [25][26]. Computed on the five reflective composites, ICC(1) estimates ranged from −0.05 to −0.01 (ICC(2): −0.52 to −0.08)—near-zero, since variance components cannot be negative—indicating firm membership accounts for essentially none of the between-respondent variance. This is the empirical basis for the key-informant, individual-level framing adopted throughout this paper [27][28]; divergence within a firm is treated as substantive, not noise, and is the reason every construct, hypothesis, and result here is defined at the level of the individual respondent’s perception rather than the firm (Sections 1, 5, 6.2).

This study’s structural model was estimated using genuine PLS-SEM, implemented in the Python plspm package (v0.5.7), with Mode A outer estimation and the PATH weighting scheme. Analysis proceeded in five stages: data-quality checks; measurement-model estimation (outer loadings, alpha, composite reliability, AVE, a Dijkstra-Henseler rho_A via a corrected congeneric-reliability formula, plus HTMT and VIF diagnostics); structural-model estimation, with inference based primarily on a firm-block percentile bootstrap (2000 resamples for the base model, resampling 50 firms with replacement and keeping each sampled firm’s respondents together; not bias-corrected-and-accelerated), alongside R²/f² and a Q² approximated via 10-fold cross-validated redundancy; firm-cluster-robust standard errors plus outlier-trimmed and control-variable re-estimations; and mediation via bootstrapped direct/indirect/total effects [22], with moderation (H10) tested via the two-stage approach [29]—the moderation and mediation bootstraps, refitting the full model per resample, used 1000 resamples given compute limits. All statistics were computed from this single, version-controlled analysis pipeline, applied consistently across every table and figure. As Section 6.6 details, all percentile bootstraps reported in Section 7—base paths, measurement-model loadings and HTMT, indirect effects, the interaction term, and the exploratory pathways—use firm-block resampling, so the firm-clustered correction and the bootstrap inference are now consistently dependence-aware throughout.

Because all constructs were collected from a single respondent at one point in time, common method variance and single-informant bias are relevant threats. Procedural remedies were applied—separated construct blocks, no grouping by hypothesized relationship, respondent anonymity [30]—reducing but not eliminating these risks. Full-collinearity VIF diagnostics did not indicate problematic collinearity, and a Harman single-factor test [30] provided no indication of a dominant common factor: six factors emerged with eigenvalues above 1.0, the largest at 23.5% of variance, well below the 50% threshold. Neither diagnostic is definitive; a marker-variable technique [31] was not implemented, so common method bias cannot be ruled out given the single-source design.

This study did not undergo formal institutional ethics board (IRB) review; no such review was mandated by the authors’ affiliated institution for this anonymous, non-clinical organizational survey. Participation was voluntary and anonymous: respondents were informed of the study’s academic purpose, no identifying information was collected, no incentives were offered, and consent was implied by voluntary completion; respondents could withdraw at any point by not submitting the questionnaire. Data were stored securely and analyzed only in de-identified, aggregated form. Authors should confirm with their institution whether retroactive ethical clearance is available, since indexed journals increasingly require a formal ethics statement.

DataandCodeAvailability

The anonymized, respondent-level analytic dataset (n = 336) underlying Tables 2-16 and Figures 1-10 is held by the authors and available upon reasonable request; the analysis code is not currently available in a public repository. The full survey instrument (English and Arabic versions) and the data-cleaning and imputation steps are described in Sections 6.3 and 6.4 (Table 1). The PLS-SEM model specification and bootstrap settings used are described in Section 6.7.

See Table 1 (Section 6.4). Across the 27 items, means ranged from 3.39 to 3.48, SDs from 0.76 to 0.87, skewness from −0.32 to 0.20, and kurtosis from −0.47 to 0.30—within conventional guidelines for approximate univariate normality.

Table 2 and Figure 2Figure 2 summarize the respondent-level and firm-level composition of the sample.

Table 2. Respondent and firm profile.

Note. Panel A percentages are respondent-level and rounded; they do not equal Panel B’s firm-level percentages, since firms vary in respondents contributed. Firm size and industry reflect each respondent’s own report, not a verified firm-level record, and can vary among respondents from the same firm; country was verified consistent within each firm. The small per-country firm counts in Panel B (Morocco: 4; Qatar: 2) preclude formal multi-group or country-level analysis (Section 9).

Figure 2. Respondent-level distribution across country, firm size, managerial level, and digital maturity (n = 336); Table 2, Panel B gives the firm-level country distribution (n = 50).

Items were adapted from established organizational-capability and process-management item pools and reworded for the enterprise context; construct blocks were separated and item order was not grouped by hypothesized relationship [30]. Table A1 (Appendix A) reports complete wording.

In the genuine PLS-SEM estimation, two DL indicators loaded well below the preferred 0.70 threshold [32] in the original four-item block: |loading| = 0.510 (DL1), 0.465 (DL2), 0.518 (DL3), 0.217 (DL4)—all other constructs’ indicators loaded above 0.70 (Table 3). DL1 and DL2 were removed, reducing DL to a two-item construct capturing sponsorship and modelling; DL3 and DL4 loadings then rise sharply (to the algorithm’s ±1.0 boundary and 0.553), a feature of Mode A estimation in small blocks, so alpha and AVE (Table 3) remain interpretable. Because DL is the central moderator, this is a genuine measurement limitation (Section 9); H10’s non-support (Section 7.9) should be read with this limitation in mind, since it may partly reflect measurement constraints rather than a true absence of the effect.

Retaining a two-indicator solution, rather than removing the construct, follows established guidance that at least two reflective indicators are needed to separate true-score variance from measurement error [33]. The retained items (DL3, DL4) capture active sponsorship and modelling; removed items tapped vision articulation and resource allocation.

Per-item outer loadings and their bootstrap intervals for every retained indicator—not only DL’s—are reported in Table 3 below, so the measurement model can be assessed construct-by-construct rather than relying on the summary reliability statistics alone.

Table 3. Measurement model: item-level outer loadings and construct reliability.

Note. (Panel A) All items above 0.70 except the DL block discussed in Section 7.4. Loadings and 95% CIs above are from a firm-block percentile bootstrap (2000 resamples, resampling 50 firms with replacement and keeping each sampled firm’s respondents together) on the retained PLS-SEM model, so these intervals are dependence-aware (Section 6.6). ¹DL3/DL4’s bootstrap draws include sign flips across resamples—a known artifact of Mode A estimation for a two-indicator block, not a data error—so each resample was sign-corrected to the full-sample orientation before computing these two CIs; DL4’s corrected interval still crosses zero, evidence that this indicator’s loading is genuinely unstable in a two-item scale (Section 9). (Panel B) The retained constructs met the reported reliability and convergent-validity criteria (α, CR > 0.70; AVE > 0.50; Figure 3Figure 3) with one exception: DL’s Dijkstra-Henseler rho_A (0.551) falls below the 0.70 threshold even though its alpha and CR do not—a genuine reliability concern specific to the two-item scale (Section 9). rho_A elsewhere is computed via a corrected congeneric-reliability formula referencing principal-axis-factoring loadings (Section 6.7), consistent with Dijkstra and Henseler’s [34] logic.

Discriminant validity was assessed using HTMT ratios of correlations (Table 4; Figure 4Figure 4).

Figure 3. Internal consistency reliability and convergent validity by construct (Table 3, Panel B).

Table 4. Discriminant validity (HTMT).

Note. HTMT ratios are reported here as absolute values throughout, per conventional practice, since a heterotrait/monotrait ratio is a discriminant-validity diagnostic and is not meaningfully signed; the underlying DL row/column correlations were negative in the original computation (−0.028 to −0.158), which follows mechanically from DL’s small, negative average correlations with every other construct (Section 7.5) rather than indicating a computation error, but every HTMT ratio and every bootstrap interval reported for it—point estimates and CIs alike—is computed and presented on the unsigned (absolute-value) scale, so no HTMT figure in this paper is negative. A single threshold is used consistently throughout this paper: HTMT < 0.90 [35] (Figure 4Figure 4’s caption, which previously cited a 0.85 cutoff, has been corrected to match). All ratios fall below 0.90 (0 of 15 pairs exceed it); the highest (0.571, DC-BPR) confirms these theoretically adjacent constructs are empirically distinguishable. Bootstrap 95% CIs for these ratios (2000 resamples, firm-block, dependence-aware, unsigned scale) are: DC-BPR [0.494, 0.645]; DC-OA [0.107, 0.331]; DC-DI [0.283, 0.511]; DC-BP [0.012, 0.286]; BPR-OA [0.275, 0.528]; BPR-DI [0.308, 0.511]; BPR-BP [0.005, 0.213]; OA-DI [0.310, 0.498]; OA-BP [0.197, 0.397]; DI-BP [0.241, 0.466]; and, for the five DL pairs (whose underlying correlations are small and negative, Section 7.5), the unsigned bootstrap CIs are DC-DL [0.002, 0.184], BPR-DL [0.002, 0.154], OA-DL [0.004, 0.189], DI-DL [0.020, 0.294], and BP-DL [0.017, 0.288]—each interval’s lower bound sits near zero, consistent with DL’s near-null average item-level correlation with the other five constructs, so these five ratios should be read as statistically indistinguishable from zero rather than as a meaningful discriminant-validity signal either way. All 15 upper bounds remain well below the 0.90 threshold, so the discriminant-validity conclusion above now rests on both point estimates and dependence-aware bootstrap intervals, none of which is negative.

Figure 4. HTMT ratios of correlations (Table 4); all values fall below the 0.90 threshold used consistently throughout this paper (Section 7.5).

All full-collinearity VIF values (Table 5; highest = 1.49, for BPR) are well below Kock’s [36] 3.3 threshold, indicating no multicollinearity concern; as a secondary diagnostic, they cannot alone rule out common method bias.

Table 5. Full-collinearity VIF.

Seven of the eight hypothesized base paths (H1 - H7; Table 6) were significant under both naive and percentile-bootstrap inference (Figure 5Figure 5); the DL main effect was not. H10, the DC × DL interaction, reached naive significance (p = 0.041) but its bootstrap 95% CI crossed zero with a negligible effect size (f² = 0.012)—not robustly supported (Section 7.9). DC → BPR was the strongest association, approaching a large effect size (f² = 0.308), just below Cohen’s [37] 0.35 cutoff.

Table 6. Structural path estimates (naive SE; 95% CI from a firm-block percentile bootstrap—resampling 50 firms with replacement and retaining each sampled firm’s respondents intact, so within-firm dependence is preserved in every resample; 2000 resamples for H1 - H7 and the DL main effect, 1000 for the H10 interaction, per Section 6.6).

Figure 5. Standardized path coefficients with 95% percentile-bootstrap CIs, ordered by magnitude (Table 6); both the DL main-effect and the H10 interaction intervals span zero.

Table 7. Explanatory power (R²).

R² values (0.116 - 0.237; Table 7, Figure 6Figure 6) indicate modest, not comprehensive, explanatory power; a higher R² does not by itself establish practical importance, assessed via f² in Table 6.

Figure 6. Explanatory power (R², left) and predictor-level effect sizes (f², right) against Cohen’s [37] benchmarks; the two answer distinct questions and are not interchangeable.

Table 8. Mediation results.

For H8 (Table 8), the direct (β = 0.210, p < 0.001) and indirect (β = 0.120, via BPR both directly to DI and through OA) effects are identically signed, and the direct path is significant under percentile bootstrap (Table 6, H2); following Zhao et al. [22], this is complementary (partial) mediation, with the indirect route carrying VAF ≈ 36% of the total effect (0.330; Figure 7Figure 7). Its constituent paths—DC → BPR (H1) and BPR → DI (H3)—are each independently bootstrap-significant, and a dedicated single-estimand, firm-block bootstrap of the indirect path itself (2000 resamples, dependence-aware) confirms this directly: SE = 0.025, 95% CI [0.073, 0.172], p < 0.001—the interval excludes zero, so H8’s indirect effect is significant as a single estimand under dependence-aware inference, not merely inferred from its jointly significant components.

For H9 specifically, the direct BPR → BP effect is small, negative, and non-significant under firm-block bootstrap (β = −0.097, SE = 0.070, 95% CI [−0.241, 0.033], p = 0.150). Following the reviewer suggestion, the indirect effect is decomposed into its three constituent routes, each bootstrapped as its own single estimand with its own confidence interval and p-value rather than inferred from component paths (Table 8): BPR → OA → BP (β = 0.070, SE = 0.021, 95% CI [0.032, 0.114], p < 0.001); BPR → DI → BP (β = 0.044, SE = 0.016, 95% CI [0.014, 0.079], p < 0.001); and the serial route BPR → OA → DI → BP (β = 0.023, SE = 0.008, 95% CI [0.011, 0.041], p < 0.001, under the same model2 specification; a base-model version, β ≈ 0.020, is reported separately in Section 7.12). All three routes are individually significant, but only the first—BPR → OA → BP—is the route H9 hypothesizes. Per Zhao et al.’s [22] classification, a small, non-significant direct effect alongside this significant, positive indirect route is consistent with indirect-only mediation through organizational agility specifically, and the H9 conclusion is restricted to this route alone, per the hypothesis’s original scope. The combined indirect effect (β = 0.134, SE = 0.028, 95% CI [0.085, 0.193], p < 0.001, “through OA and DI”) is a separate, directly bootstrapped single estimand describing the total BPR → BP indirect pathway via both mediators jointly; it is not presented as an arithmetic sum of the three routes above, each of which is its own independently bootstrapped estimand with its own resample distribution. The DI-only and serial routes are significant but exploratory, outside H9’s formal scope, and are discussed further in Section 7.13.

Figure 7. Decomposition of the total effect of DC on DI into direct (β = 0.210) and BPR-mediated (β = 0.120) components (H8; total = 0.330, VAF ≈ 36%). Depicts H8 only; H9’s OA-specific pattern is in Table 8.

The DC × DL interaction was tested with the two-stage approach [29]: standardized DC and DL scores were multiplied to form the interaction term, and the model refit with DC × DL → BPR added. Under naive standard errors, the interaction is marginally significant (β = 0.099, SE = 0.048, t = 2.049, p = 0.041). Under the more defensible, firm-block percentile bootstrap (1000 resamples, dependence-aware), the SE nearly doubles (0.095 vs. 0.048) and the 95% CI is [−0.146, 0.152] (Figure 8Figure 8)—it includes zero, and the interval is if anything more tightly centered on zero (point estimate = −0.003) than under the earlier respondent-level bootstrap. The interaction added essentially nothing to fit (ΔR²(BPR) = 0.009) and its effect size is negligible (f² = 0.012); H10 is not robustly supported, read alongside DL’s two-item limitation (Section 7.4).

Figure 8. Percentile-bootstrap distribution (n = 1000 resamples) of the DC × DL → BPR interaction; the naive point estimate (β = 0.099, p = 0.041) sits inside a distribution whose 95% CI, [−0.146, 0.152], straddles zero.

Table 9. DL measurement robustness: four-item vs. retained two-item scale.

DL1 and DL2 loaded at 0.510 and 0.465 in the four-item block (Table 9), both below the 0.70 threshold; including them collapses AVE to 0.198 and alpha to 0.540, though CR stays just above 0.70. The four-item DL → BPR effect remains non-significant (β = 0.083, p = 0.082). Model B (the retained two-item scale) is the quantitatively justified specification: its reliability and AVE are stronger, and H10’s non-support (Section 7.9) does not depend on which DL specification is used.

Three robustness checks were conducted against the base model: firm-level clustering, outlier sensitivity, and control-variable specification (the four- versus two-item DL comparison in Section 7.9 is a measurement sensitivity analysis, not a fourth check). As Section 6.6 details, all three checks below apply only to the eight base structural paths, not to the indirect (H8, H9) or interaction (H10) estimates.

Firm-levelclustering. Standard errors were re-estimated clustered by firm ID (50 clusters) to address the nested structure (Section 6.6).

Table 10. Cluster-robust standard errors.

Cluster-robust SEs were not uniformly larger than naive SEs (Table 10), and every base path significant under naive inference remained significant under firm-clustered inference—firm-level non-independence corrects standard errors but does not overturn the base-model conclusions. SEs were computed for the base eight-path model; the H10 interaction’s naive-versus-bootstrap discrepancy (Section 7.9) is a separate concern.

Outliersensitivity. The model was re-estimated after excluding the most extreme respondents by Mahalanobis distance on the six LV composite scores (top 2.5%: 8 of 336 excluded, n = 328 retained).

Table 11. Outlier sensitivity.

Eight of 336 respondents (the top 2.5% by Mahalanobis distance on the six LV composite scores) were excluded, leaving n = 328 (Table 11). All eight base paths retained the same sign and stayed within 0.02 of the base estimates, with no change in significance.

Control-variablerobustness. Firm size, firm age, industry, and digital maturity were entered as controls on the endogenous constructs they theoretically influence.

Table 12. Control-variable robustness.

Every substantive structural path significant without controls remains significant (p < 0.05, same sign; Table 12) after adding Firm_Size, Firm_Age_Years, Industry dummies, and Digital_Maturity; none of the controls reach significance, and all coefficients remain within 0.02 of the base estimates. Country was not entered as a control given small per-country cells (Section 7.2), addressed only through firm-level clustering. Together, these checks demonstrate the stability of the base-path association estimates across specifications, not proof of causality; the design remains cross-sectional and correlational.

All four Q² values (Table 13) exceed zero, indicating predictive relevance for every endogenous construct (BPR strongest at Q² ≈ 0.23, BP weakest at Q² ≈ 0.10, consistent with BP’s low R² in the base model). Because plspm has no native blindfolding routine, Q² was approximated via 10-fold cross-validation: full-sample outer weights were fixed and each construct’s inner-model regression was cross-validated, training on 9 folds and predicting the held-out fold’s LV score against the training-set mean as benchmark—a construct-level approximation, not a full per-fold re-estimation.

Table 13. Predictive relevance (Q²).

The model estimates BPR → OA (β = 0.343), OA → DI (β = 0.250), and DI → BP (β = 0.237), implying a three-step serial association not formally hypothesized but following arithmetically from H4 - H6; each component path is individually bootstrap-significant (Table 6), and their product is reported descriptively (Table 14). This pathway runs through digital innovation as well as agility, and is accordingly outside the OA-specific scope of H9 (Section 5); it is reported here as a supplementary, exploratory finding rather than evidence for or against H9.

Table 14. Exploratory serial mediation through organizational agility and digital innovation.

The effect is small, as expected for a product of three coefficients each below 1.0 (≈0.020), and a dedicated single-estimand, firm-block bootstrap (2000 resamples, base model, dependence-aware) confirms it is statistically significant and excludes zero (SE = 0.007, 95% CI [0.009, 0.037], p < 0.001). It illustrates that OA is associated with BP both directly and indirectly through DI, so the “agility” and “innovation” routes discussed next are not a fully independent partition.

These four pathways (Table 15) are not an exhaustive decomposition of every indirect route (the exploratory serial pathway in Table 14 is a further one). The estimates are products of already-bootstrap-validated component paths (Table 6), but the products themselves—and the differences between the BPR-rooted and DC-rooted pairs—were not separately bootstrapped; Table 16 reports them descriptively rather than as a formally tested contrast:

Table 15. Selected indirect pathway comparison (descriptive point estimates).

Table 16. Descriptive comparison of selected pathways.

The descriptive gap between the agility- and innovation-oriented routes (Table 15; Figure 9Figure 9) is small at both levels and was not formally tested as a statistical contrast; it should not be read as evidence either route is reliably stronger, and—because only the BPR → OA → BP row is H9’s hypothesized route—it should not be read as evidence for or against H9 either, which rests on the OA route in isolation (Table 8).

H9’s conclusion in this revision—indirect-only mediation via the organizational-agility route specifically, now confirmed by its own dedicated firm-block bootstrap in Table 8 (β = 0.070, 95% CI [0.032, 0.114], p < 0.001)—is drawn from the model2 specification with the direct BPR → BP path estimated; the mediation specification (with the direct BPR → BP path) was not itself re-estimated across the controls and outlier-trimming checks (Tables 10-12), which were run against the eight-path base model. This is a scope limitation of the robustness battery, not evidence the base H9 result is fragile.

Figure 9. Point-estimate comparison of the two DC-rooted exploratory pathways (Table 15); the gap is small and descriptive only, not a formally tested contrast (Table 16), and neither row is part of the formal H9 test.

DC showed the strongest association with BPR (β = 0.485, f² = 0.308), consistent with DC providing resources relevant to process redesign [1][2]. The smaller direct DC-DI association (β = 0.210) suggests some digitally enabled innovation is assembled directly from digital components without internal process change [11], so DC’s association with DI is not exhausted by the BPR route (H8, VAF ≈ 36%). BPR is itself associated with both DI and, more strongly, OA, consistent with leaner, outcome-organized processes shortening sensing and acting loops [12][13].

Both OA and DI are significantly associated with BP. H9, read strictly, concerns only the OA route from BPR to BP; the descriptively similar-sized DI-involving route (BPR → DI → BP) is a supplementary, non-hypothesized finding, not part of the H9 test (Section 7.13), and the agility- versus innovation-oriented comparison was in any case not a formally tested statistical contrast. The exploratory serial pathway (Section 7.12) further indicates OA and DI are not fully independent, since OA also predicts BP through DI—a further reason to keep the H9-specific (OA) and supplementary (DI) findings analytically separate rather than pooling them into one mediation claim.

H10 hypothesized that digital leadership would strengthen the DC-BPR association; as detailed in Section 7.9, this reached naive significance (p = 0.041) but did not survive percentile-bootstrap validation (95% CI [−0.146, 0.152]; f² = 0.012), and the DL main effect was non-significant throughout every specification tested (Sections 7.9 - 7.10). This null result is reported as a genuine, defensible finding: the honest reading is that H10 is not supported.

Why might a theoretically plausible moderator fail to emerge? Measurement is one candidate: DL is estimated from only two items after DL1 and DL2 were dropped for weak loadings (Section 7.4), and its rho_A (0.551) falls below 0.70, limiting power to detect moderation; the four-item version does not rescue the interaction either (AVE collapses to 0.198; Table 9). Power is another: with per-country cells as small as 2 - 4 firms (Table 2) and a negligible effect size (f² = 0.012), the design may be underpowered. Finally, DL’s HTMT with every other construct is small (Table 4), and DL → BPR was never significant—consistent with DL being neither a reliable moderator nor a direct antecedent of redesign—a boundary condition for future research, not a failure of this study.

H8’s pattern—a significant direct DC-DI association alongside an indirect BPR-mediated one that is now confirmed by a dedicated single-estimand, firm-block bootstrap (β = 0.120, 95% CI [0.073, 0.172], p < 0.001; VAF ≈ 36%; Table 8)—points toward complementary (partial) mediation: some of DC’s association with DI does not require structural change, plausibly through digitally delivered offerings layered onto existing operations. H9’s pattern, restricted to its hypothesized organizational-agility route, is consistent with indirect-only mediation: the direct BPR → BP association is small, negative, and non-significant, while the OA-mediated route is now also confirmed by its own dedicated firm-block bootstrap (β = 0.070, 95% CI [0.032, 0.114], p < 0.001; Table 8), so BPR’s performance association—to the extent it runs through OA—appears to run through this downstream agility mechanism rather than directly. Read through the Dynamic Capabilities lens (Section 4), this is consistent with OA sitting at the “transforming/reconfiguring” stage the theory identifies as the point where a reconfigured resource base (BPR) becomes an operating capability (agility) that then shapes performance, rather than BPR translating into performance on its own; a supplementary, non-hypothesized route through DI shows a similar pattern (β = 0.044, 95% CI [0.014, 0.079], p < 0.001) and is reported separately (Section 7.13), not combined into the H9 conclusion. Neither establishes causal mediation given the cross-sectional design; the small but now-confirmed exploratory serial pathway (β = 0.021, 95% CI [0.009, 0.037], p < 0.001; Section 7.12) means the “agility” and “innovation” route language approximates a more interconnected structure even before that caveat. All of these bootstraps now use firm-block, dependence-aware resampling (Section 6.6), so the nesting caveat that motivated this as a priority follow-up in earlier drafts no longer applies.

This study does not propose a new theory of digital transformation; it offers two evidence-based contributions to existing perspectives.

The first is an integrative empirical examination of DC, BPR, OA, DI, and BP within a single structural model—testing whether the direct DC-DI association persists once BPR is included (H8), whether BPR’s association with BP through organizational agility specifically is direct or indirect-only (H9), and whether OA and DI are separable or interconnected, as the exploratory serial pathway (Table 14) and the supplementary DI-involving route (Table 15) suggest. This draws on Dynamic Capabilities [20] and Resource-Based View perspectives: DC → BPR fits the reconfiguration logic of Dynamic Capabilities without resolving its circularity problem [21], while the residual DC-DI association leaves room RBV alone cannot explain.

The second contribution is a boundary-condition finding rather than a confirmed effect: digital leadership, hypothesized as strengthening the DC-BPR association, did not hold up once naive inference was replaced with percentile-bootstrap validation (H10; Section 7.9). The main-effects structure (H1 - H7) is robust across the base-path robustness checks (Section 7.10), the mediation effects (H8, H9) are now each supported by their own dedicated single-estimand, firm-block bootstrap (Table 8); and the moderating role of digital leadership (H10)—measured with a two-item scale and per-country cells as small as 2 - 4 firms—remains a genuine scope boundary rather than a settled null. Both extend existing perspectives rather than propose a new theory; the cross-sectional, single-source design constrains the claims drawn from them (Section 9).

The findings suggest a chain of associations, not a causal sequence: stronger DC is associated with stronger BPR, so digital investment may be more consequential paired with attention to process redesign; BPR’s indirect-only route to BP through OA suggests redesign scoped only for efficiency may under-deliver on agility gains specifically, with a similar supplementary pattern observed for innovation gains. Given H10’s non-support, this study cannot responsibly recommend leadership-sponsorship investment as a lever, though the two-item measure’s limited power means the possibility cannot be ruled out. Figure 10Figure 10 summarizes the validated associations as an integrated framework.

Figure 10. Integrated framework summarizing the validated associations above, as a non-temporal map. “Digital readiness” is a conceptual precondition to DC, not a tested relationship; every other connection corresponds to a supported path in Table 6. DL is positioned at the DC-BPR link it was tested, but not found, to moderate.