Digital spending drags returns at first but pays off: Chinese listed tourism firms see an initial fall in ROA from digital transformation that turns positive around two years later as improved cost governance and higher asset turnover recover performance, with cost-stickiness relief driving gains in hotels and scenic-area operators and asset-turnover gains more important for travel agencies.

Main Finding

Digital transformation (DT) in Chinese listed tourism firms produces a J‑curve: a significant short‑run drag on profitability (ROA) that reverses to positive returns after about two years. The net effect is decomposed into two competitive mediating channels: reductions in cost stickiness (CS) and increases in total asset turnover (TAT). Both mediators produce positive indirect effects that partially offset a larger short‑term direct implementation burden, with the cost‑stickiness channel dominating in heavy‑asset sub‑sectors (hotels, scenic areas) and the asset‑turnover channel stronger in asset‑light sub‑sectors (travel agencies, retail). Robust inference relies on one‑year‑lagged mediators to avoid mechanical accounting overlap; results hold when excluding COVID crisis years.

Reference: Longlong Liu & Enyi Zhou (2026). How digital transformation shapes tourism firm profitability: cost stickiness and asset productivity channels in Chinese listed tourism enterprises. DOI: https://doi.org/10.21203/rs.3.rs-10401922/v1

Key Points

• Dynamic/J‑curve effect (H1): contemporaneous DT intensity is associated with a large negative ROA effect (example estimate: DIGt → ROAt β ≈ −7.24, p < .001). Effect attenuates at 1 year (β ≈ −2.37), reverses at 2 years (β ≈ +3.87, p < .001), and remains weakly positive at 3 years.
• Dual mediation (H2, H3): DT operates through two separable channels:
  • Cost stickiness (CS): DT significantly reduces CS (DIG → CS β ≈ −0.182, p < .001). Lower CS is strongly associated with higher ROA (CS → ROA β ≈ −116.24, p < .001), producing a positive indirect effect.
  • Total asset turnover (TAT): DT raises TAT (DIG → TAT positive), and higher TAT increases ROA, producing an additional positive indirect effect.
• Competitive/offsetting mediation: Positive indirect effects via CS and TAT partially offset a larger negative direct implementation burden in the short run (masking‑type mediation).
• Sub‑sector heterogeneity (H4): the CS pathway is dominant in capital‑intensive sub‑sectors (hotels, scenic areas), while the TAT pathway dominates in asset‑light sub‑sectors (travel agencies; retail shows mixed patterns).
• Identification strategy addresses an important inference problem: because CS, TAT and ROA are derived from the same financial statements, contemporaneous mediation risks mechanical overlap; the authors therefore present primary inference using one‑year‑lagged mediators (mediator at t−1 predicting ROA at t).
• Sustainability link: TAT improvements imply greater revenue per unit of installed capacity (resource efficiency), relevant to SDG 12 (responsible consumption/production), and SDGs 8 and 9 (economic growth and industrial innovation).

Data & Methods

• Sample: Balanced panel of 28 Chinese A‑share listed tourism firms (hotels, scenic areas, travel agencies, duty‑free/retail) over 2015–2024 (N = 280), representing ≈80% of qualifying listed tourism firms in China.
• Key variables:
  • Outcome: ROA = net income / average total assets (%).
  • DT intensity (DIG): ln(1 + frequency of DT‑related terms) from annual report text (five dimensions: AI, big data, cloud, blockchain, digital application).
  • Cost stickiness (CS): firm‑year Weiss (2010) measure (difference in log cost‑response ratios for revenue declines vs increases); higher = more stickiness.
  • Total asset turnover (TAT): operating revenue / average total assets.
  • Controls: firm size (ln assets), leverage, firm age, ownership concentration, state ownership dummy, revenue growth; continuous variables winsorized at 1st/99th percentiles.
• Estimation:
  • Two‑way fixed effects (firm and year), firm‑clustered standard errors. Diagnostics (F test, Hausman) support fixed effects over pooled OLS or random effects.
  • Dynamic total‑effect tests for DIG at lags 0–3 years.
  • Parallel mediation framework (CS and TAT as simultaneous mediators). Primary mediation inference uses one‑year‑lagged mediators (CSt−1, TATt−1) to sever contemporaneous accounting overlap.
  • Indirect effects and confidence intervals obtained by bootstrap with 5,000 firm‑clustered resamples (bias‑corrected CIs).
  • Robustness checks include excluding 2020–2022 (COVID crisis years); results remain consistent.

Limitations noted by authors: - Modest absolute sample size (28 firms), though it covers most listed tourism firms in China. - DT measure is disclosure‑based (text frequency), which reflects strategic salience and may imperfectly capture actual expenditures or technology intensity. - Observational panel; while lagged mediator specification mitigates mechanical overlap, residual endogeneity remains a possibility (authors use robustness checks but do not rely on IV/GMM identification).

Implications for AI Economics

• Timing matters for returns to AI/digital investments: economists and practitioners should model multi‑year payoff profiles (J‑curve) when valuing AI projects or constructing firm investment policies. Short‑horizon performance measures will understate DT/AI value.
• Mechanism decomposition is essential: aggregate performance estimates mask opposing internal channels. Empirical studies on AI should explicitly model mediators (cost governance, asset utilization/productivity) and use lag structures to avoid mechanical accounting artifacts.
• Sector heterogeneity influences which AI/digital interventions will be most effective:
  • Heavy‑asset sectors (hotels, scenic areas): prioritize AI tools that improve cost governance — demand forecasting, workforce optimization, dynamic staffing — because lowering cost stickiness yields larger profitability gains.
  • Asset‑light sectors (travel agencies, online retail): prioritize AI that increases throughput and conversion (distribution algorithms, recommendation engines, dynamic pricing) to raise asset turnover.
• Sustainable productivity link: AI that increases utilization (TAT) can be framed as resource‑efficiency policy (lower environmental burden per service unit), enabling joint economic and sustainability objectives (SDG alignment). AI economics research should quantify environmental co‑benefits of productivity improvements where possible.
• Empirical practice recommendations:
  • Use disclosure‑based DT measures as feasible proxies but supplement with direct expenditure or implementation indicators where available.
  • Employ lagged mediators and bootstrap clustered inference when mediators and outcomes derive from the same accounting statements to avoid mechanical mediation bias.
  • Consider heterogeneity by industry/sub‑sector when estimating benefits of AI/digital adoption and when advising policy (e.g., subsidies, tax incentives) or corporate strategy.
• Directions for further research: cross‑country replication, larger firm samples, richer measures of DT (capex/software spend, firm‑level AI adoption indicators), causal identification (IVs, natural experiments), firm‑level environmental impact measurement to quantify SDG tradeoffs and co‑benefits.

If you’d like, I can: - Extract the main regression and mediation coefficients into a one‑page table; or - Draft suggested empirical specifications (including possible instruments or GMM approaches) for causal identification in follow‑up work.