Technology Adoption and Knowledge Diffusion

Core Questions

The literature organized around this project asks two linked questions. First, why do apparently profitable technologies diffuse slowly, and how much of the delay is true adoption cost rather than unobserved heterogeneity in returns? The cleanest contribution in the context is the structural logic in Suri (2011), Selection and Comparative Advantage in Technology Adoption, which explicitly separates comparative advantage in adoption from differences in net returns, and the broader adoption-friction literature around Jaffe and Stavins (1994), The energy paradox and the diffusion of conservation technology, Atkin, Chaudhry, Chaudry, Khandelwal and Verhoogen (2017), Organizational Barriers to Technology Adoption: Evidence from Soccer-Ball Producers in Pakistan*, and Oliva, Jack, Bell, Mettetal and Severen (2019), Technology Adoption under Uncertainty: Take-Up and Subsequent Investment in Zambia. Together these papers suggest that slow diffusion is rarely a simple puzzle of irrationality: it is usually a composite of fixed implementation costs, complementary investment requirements, learning failures, and uncertainty about realized returns. The harder empirical task is to distinguish a large fixed cost from a technology that is simply better suited to some agents than others; the project’s core question is therefore not only whether adoption is slow, but what margin of the slow response is structural cost, what margin is selection, and what margin is delayed learning.

Second, the literature asks how knowledge moves across firms, workers, regions, and time, and how fast it depreciates once created. The most informative papers in the database show that knowledge flows are localized, networked, and often fragile: citations cluster geographically, mobility matters, and spillovers decay with distance and time Thompson and Fox-Kean (2005), Patent Citations and the Geography of Knowledge Spillovers: A Reassessment Almeida and Kogut (1999), Localization of Knowledge and the Mobility of Engineers in Regional Networks Azoulay, Zivin and Wang (2010), Superstar Extinction^*. On the depreciation side, the key result is that learning is not simply cumulative forever; it can fade, reverse, or become obsolete, as in Benkard (2000), Learning and Forgetting: The Dynamics of Aircraft Production and Caballero and Jaffe (1993), How High Are the Giants' Shoulders: An Empirical Assessment of Knowledge Spillovers and Creative Destruction in a Model of Economic Growth. This creates a direct bridge to adoption: if knowledge diffuses slowly or decays rapidly, the effective private return to adopting a technology depends on both current information and the survival of tacit know-how. That logic is especially visible in GPT settings, where the long lags in adoption coexist with the need for continuous complementary investment Brynjolfsson, Rock and Syverson (2020), The Productivity J-Curve: How Intangibles Complement General Purpose Technologies Brynjolfsson, Rock and Syverson (2019), The Productivity J-Curve: How Intangibles Complement General Purpose Technologies Helpman and Trajtenberg (1994), A Time to Sow and a Time to Reap: Growth Based on General Purpose Technologies.

Key Findings

A striking pattern across the adoption literature is that many apparently high-return technologies are not underused because agents dislike productivity; they are underused because adoption is embedded in a broader production system. The clearest evidence comes from field and firm studies that uncover organizational, financial, and behavioral barriers. Atkin, Chaudhry, Chaudry, Khandelwal and Verhoogen (2017), Organizational Barriers to Technology Adoption: Evidence from Soccer-Ball Producers in Pakistan* shows that even when a cutting technology is close to universally beneficial, take-up can remain low for a long time because implementation disrupts internal organization and incentive alignment. Bresnahan, Brynjolfsson and Hitt (2002), Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence and Bloom, Sadun and Reenen (2012), Americans Do IT Better: US Multinationals and the Productivity Miracle reinforce that the payoff to IT depends on workplace redesign, management practices, and skill mix, while Buera and Shin (2013), Financial Frictions and the Persistence of History: A Quantitative Exploration and Mukherjee (2020), Access to Formal Banks and New Technology Adoption: Evidence from India identify financial constraints as a key bottleneck to adoption. Behavioral frictions also matter: Duflo, Kremer and Robinson (2009), Nudging Farmers to Use Fertilizer: Theory and Experimental Evidence from Kenya and Hanna, Duflo and Greenstone (2016), Up in Smoke: The Influence of Household Behavior on the Long-Run Impact of Improved Cooking Stoves show that small purchase or usage costs can be large once present bias, procrastination, or imperfect follow-through are admitted into the decision problem. The important implication is that measured low uptake does not imply low ex ante returns; it often reflects a wedge between the technology and the organizational or behavioral capacity needed to use it well.

The knowledge-diffusion branch gives a similarly nuanced picture. Keller (2004), International Technology Diffusion and Buera and Oberfield (2020), The Global Diffusion of Ideas place international diffusion in a quantitative framework where trade, cross-border ideas, and domestic absorption jointly determine productivity paths. At the micro level, Thompson and Fox-Kean (2005), Patent Citations and the Geography of Knowledge Spillovers: A Reassessment and Jaffe, Trajtenberg and Henderson (1993), Geographic Localization of Knowledge Spillovers as Evidenced by Patent Citations show that patent citations are not randomly spread in space; geographic localization is strong enough to be an empirical object in its own right. Mobility and networks matter as well: Almeida and Kogut (1999), Localization of Knowledge and the Mobility of Engineers in Regional Networks finds that engineer movement is tightly linked to spillovers, and Azoulay, Zivin and Wang (2010), Superstar Extinction^* uses premature deaths to isolate the loss of coauthor spillovers. Recent frontier evidence in the database pushes the same message into new settings. Dahlke, Beck, Kinne, Lenz, Dehghan, Wörter and Ebersberger (2023), Epidemic effects in the diffusion of emerging digital technologies: evidence from artificial intelligence adoption studies diffusion of AI and finds that early adoption exhibits epidemic-like patterns, suggesting that proximity and network structure are not just background conditions but part of the diffusion technology itself. Nagler, Schnitzer and Watzinger (2022), Fostering the Diffusion of General Purpose Technologies: Evidence from the Licensing of the Transistor Patents* is especially important because it shows that licensing can accelerate diffusion of a GPT and that cross-technology spillovers intensify once barriers to transfer fall. That is a particularly useful bridge between the two axes: diffusion accelerates not merely because the invention exists, but because channels for transferring knowledge become cheaper and more legible.

Identification & Credibility

The strongest adoption papers identify frictions by changing the price, timing, or feasibility of adoption while holding technology quality fixed, so the researcher sees how much behavior moves on the extensive margin. Jaffe and Stavins (1994), The energy paradox and the diffusion of conservation technology is a canonical example of this logic in energy efficiency: the puzzle is gradual diffusion of a technology with clear private savings, and the identification challenge is to separate hidden costs and information problems from genuine heterogeneity in returns. Celhay, Gertler, Giovagnoli and Vermeersch (2019), Long-Run Effects of Temporary Incentives on Medical Care Productivity tackles the same issue with a randomized temporary incentive, which shifts the cost of early initiation rather than the underlying quality of care. Its estimand is about whether short-run adjustment costs suppress take-up, and the chief identifying assumption is that the temporary transfer affects behavior mainly through the cost channel, not through permanent changes in technology value. The paper is valuable precisely because it moves beyond reduced-form associations and asks whether a small, transitory wedge can release delayed adoption. Duflo, Kremer and Robinson (2009), Nudging Farmers to Use Fertilizer: Theory and Experimental Evidence from Kenya similarly identifies present bias by comparing observed behavior to a model in which procrastination, not low returns, explains delay; the cost is not just monetary but intertemporal. Oliva, Jack, Bell, Mettetal and Severen (2019), Technology Adoption under Uncertainty: Take-Up and Subsequent Investment in Zambia adds an uncertainty dimension by experimentally varying payoffs across stages of investment, which identifies real-option behavior and the option value of waiting. The threat in this kind of design is always that the shock changes beliefs about the technology rather than only the adoption cost; the best papers therefore carefully tie behavior to stage-specific responses, not just average uptake.

Several papers in the database tackle the hardest identification problem in this literature: distinguishing adoption costs from heterogeneity in returns. Suri (2011), Selection and Comparative Advantage in Technology Adoption is central because it makes that decomposition explicit in a structural framework. The maintained assumption is that observed adoption choices reflect a joint distribution of costs and benefits, and the empirical object is not simply average treatment effects but the shape of selection into adoption. Suri (2009), Selection and Comparative Advantage in Technology Adoption and Zeitlin, Caria, Dzene, Jánský, Opoku and Teal (2010), Heterogeneous returns and the persistence of agricultural technology adoption push on the same margin in agricultural settings, where adoption persistence can look like a demand problem even when treated units show high average returns. These papers are important because they remind us that high average returns do not imply low adoption costs for any given agent; what matters is whether the marginal adopter has a cost wedge large enough to offset those returns. Dhyne, Konings, Bosch and Vanormelingen (2020), The Return on Information Technology: Who Benefits Most? and Cirera, Comín, Cruz and Lee (2020), Anatomy of Technology in the Firm extend this logic to firms, showing that heterogeneous implementation capacity can sort adopters from non-adopters. The estimand is therefore a joint one: the extensive margin of adoption and the structural mapping from adoption to profitability. The main limitation of this branch is that return heterogeneity and cost heterogeneity are often observationally similar unless the design offers exogenous variation in costs, repeated choices, or high-dimensional information about user types.

On the diffusion side, identification is typically built from quasi-exogenous shocks to geography, mobility, or network structure. Bai, Wang and Zhou (2021), Proximity and Knowledge Spillovers: Evidence from the Introduction of New Airline Routes leverages new airline routes to shift travel time and infer the causal effect of proximity on knowledge flows. The key assumption is that route changes affect citation patterns primarily through reduced travel time, not through unrelated changes in local economic activity. Azoulay, Zivin and Wang (2010), Superstar Extinction^* uses unexpected deaths of superstars to break coauthor networks, a powerful design because it preserves preexisting collaboration structure while perturbing its information content. Here the identifying variation is plausibly exogenous to collaborators' latent productivity trajectories, though one must still worry about grief, labor reallocation, and project substitution. Almeida and Kogut (1999), Localization of Knowledge and the Mobility of Engineers in Regional Networks relies on engineer mobility to trace spillovers, but that design is cleanest when mobility is not itself a response to unobserved shocks to both source and destination firms. Thompson and Fox-Kean (2005), Patent Citations and the Geography of Knowledge Spillovers: A Reassessment and Jaffe, Trajtenberg and Henderson (1993), Geographic Localization of Knowledge Spillovers as Evidenced by Patent Citations are measurement-oriented identification papers in the sense that they question whether citation distance truly proxies spillover distance; their value is in showing how sensitive localization estimates are to citation windows, benchmark sets, and counterfactual citation probabilities. In short, the diffusion literature is strongest when the variation comes from a plausibly exogenous break in the network or the transport cost of information, and weakest when mobility or proximity is chosen endogenously.

Methods

Methodologically, the project’s literature spans revealed preference, dynamic choice, structural estimation, and quantitative general equilibrium. The dynamic discrete-choice tradition is represented by Rust (1987), Optimal Replacement of GMC Bus Engines: An Empirical Model of Harold Zurcher, Gordon (2009), A Dynamic Model of Consumer Replacement Cycles in the PC Processor Industry, and the adoption-side structural papers in Suri (2011), Selection and Comparative Advantage in Technology Adoption and Suri (2009), Selection and Comparative Advantage in Technology Adoption. These papers are valuable because they recover latent costs from choice sequences rather than from cross-sectional correlations. The price of that discipline is strong functional-form and state-space assumptions: the researcher must specify what agents know, when they learn it, and how future payoffs enter current decisions. Holmes (2011), The Diffusion of Wal-Mart and Economies of Density offers a related revealed-preference approach in a spatial expansion setting, where store network design reveals the value of density and distribution cost savings. That type of method is attractive for adoption questions because it can infer implementation costs from observed placement patterns, but it typically identifies cost under a maintained model of profit maximization and a limited menu of feasible strategies.

The quantitative and model-based literature is strongest when it combines micro estimates with aggregate implications. Brynjolfsson, Rock and Syverson (2020), The Productivity J-Curve: How Intangibles Complement General Purpose Technologies and Brynjolfsson, Rock and Syverson (2019), The Productivity J-Curve: How Intangibles Complement General Purpose Technologies use a GPT framework to map intangible complementary investments into productivity dynamics and measurement. Benhabib, Perla and Tonetti (2021), Reconciling Models of Diffusion and Innovation: A Theory of the Productivity Distribution and Technology Frontier, Comín and Ferrer (2013), If Technology Has Arrived Everywhere, Why has Income Diverged?, Santacreu (2015), Innovation, diffusion, and trade: Theory and measurement, and Souza, Gaetani and Mestieri (2024), More Trade, Less Diffusion: Technology Transfers and the Dynamic Effects of Import Liberalization similarly build multicountry or industry models where adoption and diffusion jointly determine growth and convergence. The appeal of these papers is that they connect firm-level frictions to macro outcomes, but that connection relies on calibrated or estimated transition parameters that are often only partially identified by the data. The result is less a pure causal claim than a disciplined accounting of what adoption frictions would have to look like for the observed diffusion path to make sense. On the knowledge side, citation-based work remains foundational. Caballero and Jaffe (1993), How High Are the Giants' Shoulders: An Empirical Assessment of Knowledge Spillovers and Creative Destruction in a Model of Economic Growth and Benkard (2000), Learning and Forgetting: The Dynamics of Aircraft Production estimate decay and learning from patent citations and production histories, while Kim and Seo (2008), Depreciation and transfer of knowledge: an empirical exploration of a shipbuilding process and Sfekas (2008), Learning, forgetting, and hospital quality: an empirical analysis of cardiac procedures in Maryland and Arizona use learning-with-forgetting frameworks to recover depreciation in organizational settings. These methods are informative because they convert noisy performance trajectories into estimates of experience accumulation and decay, but they require careful assumptions about how output or citations map into latent knowledge stocks.

Open Questions & Gaps

The first major gap is the joint identification of adoption costs and return heterogeneity in a setting where both are endogenous. The literature has excellent cases that isolate one or the other, but far fewer designs that separately recover the distribution of costs, the distribution of returns, and the correlation between them. Suri (2011), Selection and Comparative Advantage in Technology Adoption, Suri (2009), Selection and Comparative Advantage in Technology Adoption, and Zeitlin, Caria, Dzene, Jánský, Opoku and Teal (2010), Heterogeneous returns and the persistence of agricultural technology adoption get closest, yet even these approaches tend to rely on strong structure to interpret selection into adoption. What is still unidentified is how much of non-adoption comes from fixed cost, how much from idiosyncratic comparative advantage, and how much from learning about returns after partial adoption. The best next step would be a panel design with repeated randomized cost shifters, detailed pre-adoption characteristics, and post-adoption outcome paths, so the same agent’s response can identify both the cost side and the return side of the decision problem.

A second gap concerns knowledge depreciation outside a few manufacturing or patent-intensive settings. Benkard (2000), Learning and Forgetting: The Dynamics of Aircraft Production, Caballero and Jaffe (1993), How High Are the Giants' Shoulders: An Empirical Assessment of Knowledge Spillovers and Creative Destruction in a Model of Economic Growth, Kim and Seo (2008), Depreciation and transfer of knowledge: an empirical exploration of a shipbuilding process, and Sfekas (2008), Learning, forgetting, and hospital quality: an empirical analysis of cardiac procedures in Maryland and Arizona show that forgetting exists, but the evidence base is still thin on how depreciation varies with task complexity, team turnover, codifiability, and digital tools. What remains unidentified is the elasticity of tacit knowledge loss with respect to organizational churn and technological change. This matters because depreciation is not just a side fact; it directly affects the private payoff to adoption and the urgency of re-adoption. A richer design would combine matched employer-employee data, task-level measures of codification, and quasi-exogenous turnover shocks to separate true obsolescence from mere staff replacement.

A third gap is the interaction between diffusion channels and adoption costs in GPT settings. Brynjolfsson, Rock and Syverson (2020), The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, Brynjolfsson, Rock and Syverson (2019), The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, Helpman and Trajtenberg (1994), A Time to Sow and a Time to Reap: Growth Based on General Purpose Technologies, and Nagler, Schnitzer and Watzinger (2022), Fostering the Diffusion of General Purpose Technologies: Evidence from the Licensing of the Transistor Patents* all suggest that complementary investments and transfer channels jointly determine adoption speed, but the literature rarely identifies their relative contributions cleanly. If diffusion is slow because information is scarce, the policy lever is information provision; if it is slow because complementary investments are expensive, the lever is financing or organization. The frontier evidence in Dahlke, Beck, Kinne, Lenz, Dehghan, Wörter and Ebersberger (2023), Epidemic effects in the diffusion of emerging digital technologies: evidence from artificial intelligence adoption and Souza, Gaetani and Mestieri (2024), More Trade, Less Diffusion: Technology Transfers and the Dynamic Effects of Import Liberalization is promising because it treats diffusion as networked and policy-sensitive, but the current database still leaves open which margin dominates in different sectors and countries. Designs that combine network shocks with direct measures of organizational readiness would help disentangle whether knowledge fails to arrive, arrives but is ignored, or arrives but cannot be implemented. That distinction is central to the project’s agenda and remains only partially answered by the current literature.