Week 3: Measurement and the Nature of Innovation

[jamesallenevans]

Please post your memo that engages any of the week's readings about a theoretically interesting empirical case regarding artificial intelligence, innovation, and/or growth you anticipate will become the basis of your final project.

“Connecting to Power: Political Connections, Innovation, and Firm Dynamics” Ufuk Akcigit, Salome Baslandze, and Francesca Lotti. Econometrica, 2023, 91(2): 529-564.
“Barriers to Creative Destruction: Large Firms and Non-Productive Strategies” 2023. Salome Baslandze, in The Economics of Creative Destruction: New Research on Themes from Aghion and Howitt.
“Combination and Structure” and “Mechanisms of Evolution” Download “Combination and Structure” and “Mechanisms of Evolution”, Chapters 2 (27-43) and 9 (167-189) from The Nature of Technology: What it is and How it Evolves. 2009. Brian Arthur. Free Press: New York.
“Assembly theory explains and quantifies selection and evolution”. 2023. Abhishek Sharma, Dániel Czégel, Michael Lachmann, Christopher P. Kempes, Sara I. Walker & Leroy Cronin. Nature.
“Modularity, Architectural Innovation, and New Venture Success”. Likun Cao, Ziwen Chen, and James Evans. Preprint.

Post by Thursday @ midnight. By 12pm Friday, each student will up-vote (“thumbs up”) what they think are the five most interesting memos for that session. These memos should: 1) test out ideas and analyses you expect to become part of your final projects; and 2) involve a custom (non-hallucinated) theoretical and/or empirical demonstration that will result in the relevant analytical visualization. Some of the top-voted memos will form the backbone of discussion in Friday's discussion sessions.

[zararania]

Zara Rania Suran

Research Question:
Does the pharmaceutical industry exhibit the "leadership paradox" described by Akcigit, where dominant firms substitute genuine R&D innovation with political connections to protect their market position? Specifically, I investigate whether heavy lobbying expenditures allow incumbent drugmakers to reduce the regulatory wedge (τ) associated with FDA approval times (known to be notoriously slow for novel drugs).

I ask whether this dynamic reduces market contestability, allowing incumbents to survive via "red tape reduction" rather than through the creative destruction of inferior therapies, effectively slowing the aggregate growth of medical productivity.

Theoretical Framework:
I ground my analysis in the political economy model of Akcigit, Baslandze, and Lotti (2023), which posits a trade-off between investing in innovation (to raise productivity A) and political connections (to reduce the cost wedge τ). In pharma, FDA regulations act as this cost wedge; while vital for safety, they create a barrier to entry. I test the hypothesis that as drug firms grow, they switch strategies from innovation-led growth to connection-led growth. By paying the fixed cost of lobbying (w
p), incumbents lower their variable regulatory costs.

This creates an anti-competitive environment where connected firms persist not because their drugs are clinically superior, but because they can navigate bureaucracy faster than more innovative, yet unconnected, challengers (Sharma et al., 2023). For example, in the case of Eteplirsen (a drug for Duchenne muscular dystrophy, or DMD), aggressive lobbying and pressure from patient advocacy groups funded by the manufacturer, Sarepta Therapeutics, reportedly contributed to the FDA granting accelerated approval despite the drug's initial failure to demonstrate clinical efficacy in trials.

Research Design:
I will construct a panel dataset merging 25 years of FDA drug approval records with firm-level lobbying data from OpenSecrets. To measure the red tape wedge, I will regress "Time to Approval" (submission to market) on lobbying expenditures, controlling for drug complexity. To tests for the leadership paradox, I will analyze the relationship between market rank (revenue share) and "Innovation Intensity" (clinical value of drugs per R&D dollar). If market leaders show high lobbying but low innovation intensity, it confirms that political capital is acting as a non-productive strategy (Baslandze 2023) that blocks the diffusion of better technologies and dampens the industry's long-term growth trajectory.

Data Visualization:

Statically, the first graph demonstrates the red tape wedge, showing that incumbents use political connections to purchase immediate regulatory speed (τ), effectively bypassing the efficiency hurdles that should filter out lower-quality entrants. Dynamically, the second graph illustrates the leadership paradox, revealing that as firms grow into dominant market leaders, they substitute genuine innovation with lobbying, a shift that stifles the long-term creative destruction process necessary for aggregate productivity growth. Together, these findings would suggest that political influence distorts the market selection mechanism, favoring the fittest lobbyists over the fittest technologies and ultimately slowing the trajectory of medical advancement.

[erindner13]

Erik Rindner

Research Question
Does AI-driven team construction optimize on-field performance at the expense of unmeasured qualities like team chemistry and managerial philosophy in professional baseball?
Recent advances in artificial intelligence have expanded its role beyond in-game decision-making and training and into organizational strategy and team construction. In professional baseball, AI-assisted scouting, player valuation models, and analytics-driven roster construction are now widely used to identify players who best fit a team’s needs based on measurable performance indicators. At first glance, this shift appears unambiguously beneficial: AI promises to reduce uncertainty, improve efficiency, and help teams assemble rosters that maximize on-field performance. However, building a successful baseball team has never been purely a numbers problem. Managers and executives have long emphasized qualities such as leadership, chemistry, and alignment with organizational philosophy—factors that are difficult to quantify and often evaluated through human judgment.
This raises a deeper question about what is gained and what may be lost when AI systems increasingly guide team-building decisions. AI-driven scouting tools are typically trained on historical performance data and designed to identify statistically optimal player profiles. When widely adopted, these systems may encourage teams to value similar traits and converge on a shared definition of what a “successful” roster looks like. While this convergence may improve average performance, it may also crowd out alternative philosophies of team construction and reduce the role of managerial intuition and experiential knowledge. In this sense, AI may not simply assist decision-makers but reshape organizational priorities. This concern aligns closely with Arthur’s argument that technologies evolve by recombining existing components and, in doing so, redefine the problems they are used to solve. Once AI-based scouting becomes embedded in baseball organizations, future decisions are made within the constraints it imposes, creating path dependence in how teams are built. Similarly, Simulating Subjects highlights how AI systems trained on past human judgments can standardize future behavior, producing feedback loops that narrow diversity in decision-making. Applied to baseball, AI-driven evaluation may replicate historically successful player profiles while undervaluing intangible contributions that fall outside the data.
Methods
I propose a comparative research design examining patterns of player selection and team performance before and after the widespread adoption of AI-driven scouting tools. Using draft data, roster compositions, and publicly available performance metrics, teams could be categorized based on their reliance on analytics and AI-assisted decision-making. Key outcomes would include variance in player profiles, stylistic diversity across teams, and performance volatility across seasons. These outcomes would then be compared to measures of team success in high-variance contexts where unmeasured qualities like leadership and cohesion may matter more. The key outcome of interest is not simply whether AI-driven teams perform well on average, but whether increasing optimization comes at the cost of diversity and human judgment in team construction. AI may be highly effective at identifying the statistically “best” roster, yet still incomplete in capturing what makes teams resilient and adaptive. In an ideal setting, AI would serve as a decision-support tool rather than a decision-maker, complementing managerial philosophy rather than replacing it entirely.

[Arthur-Langlois]

Internalizing Institutional Knowledge and Anti-Competitive Behavior

My research question can be simplified to the following: What are the channels through which firms engage in anticompetitive practices?

Background & Motivation
Building from the insights of Akcigit et al. (2020), establishing a causal link between politician hiring and anti-competitive behavior in Italy, we extend these findings to study the specific mechanisms at play in determining firm's anti-competitive behavior stemming from the employment of government officials. In other words, it points to more precise suggestions for regulation than the literature has focused on. In particular, this project aims to draw a clearer path towards 1) regulating anti-competitive behavior 2) to spur and incentivize innovation instead. For instance, previous work only suggests regulating the hiring of politicians, which may not be useful: firms could easily circumvent these rules by hiring ex politicians that have the same institutional knowledge. Instead, legislators could restrict the oversight of some politicians on certain sensible projects, such as ensuring a process is run by multiple people/or automated instead of direct oversight approval. We hypothesize that hired government officials have leverage over four particular mechanisms through direct oversight, political pressure, institutional knowledge, or interpersonal relationships more broadly.

Data

In particular, we study four potential mechanisms suggesting anti-competitive behavior (or unfair practices):

• Court judgments (% of favorable court judgments)
• Firm acquisitions (Acquisition per employee)
• Taxation (Total tax paid in year $t$)
• Relevant Permit (New Permit per Employee)

We combine replication material from Akcigit et al. (2020), and match them with publicly available data from ISTAT. Note that the permit per employee are specific to any individual industry (construction permits for real estate, liquor license for food & beverage groups...etc), that would be identified ex ante.

Empirical Strategy
The first stage of my analysis provides suggestive evidence that our potential mechanisms of interest are correlated with anticompetitive behavior (as proxied by Total Factor Productivity), justifying the use of such dependent variables (Figure 1).

The second stage –analogous to Akcigit et al. (2020)–, I use a regression discontinuity design to assess the causal impact of the election of a supported candidate in local elections:

$y_{i,T(m)} = \alpha + \beta Win_{i,T(m)-1} + f(margin_m) + (\sigma_1 X_{i,T(m)} + \sigma_2 X_{i,m} + \sigma_3 X_{i,T}) + v_{i,T(m)}$

where $T(m)$ is the time of marginal election m, $y_{i,T(m)}$ it the outcome for firm $i$ at $T(m)$, $Win_{i,T(m)-1}$ indicates a firm connected to a marginally wining party in election $m$, $margin_{m}$ is the victory margin, and $f(.)$ a smooth polynomial. I provide an estimation using simulated data in Figure 2.

Figures

[nbcodes57]

Nick Baumgartner

Research Question:
Do larger firms systematically allocate human capital away from productive innovation roles toward protective legal and regulatory positions?

Background and Motivation:
Baslandze (2021) synthesizes emerging evidence that large incumbent firms increasingly rely on non-productive strategies such as political connections, strategic patenting, and anti-competitive acquisitions to preserve market dominance rather than competing through innovation. The review highlights a fundamental tension in firm strategy: as market share grows, Arrow's replacement effect reduces innovation incentives (incumbents cannibalize their own rents), while the Gilbert-Newbery "preserving-the-market" effect increases incentives to defend existing positions through barriers to creative destruction.
This study focuses on the outcomes—political connections formed, patents filed, acquisitions completed—rather than the internal decisions that enable these strategies. How do firms build capacity for non-productive strategies? Before hiring external politicians or filing defensive patents, firms must first develop internal expertise in legal strategy, regulatory navigation, and policy influence. Hiring decisions therefore reveal firms' strategic priorities and may serve as a leading indicator of the shift from innovation-based to barrier-based competition. This project aims to examine the first stage of human capital allocation by looking at the composition of job postings on LinkedIn between ‘productive’ and ‘unproductive’ roles.

Research Design:
Data is sourced from a dataset that scraped 123,000 job postings on LinkedIn from 2023-2024 found on Kafggle containing job titles, job descriptions, companies, and firm sizes. I have tagged the ‘productive’ and ‘unproductive’ roles using keywords such as "research," "scientist," "engineer," "developer," "R&D", for the former and "legal," "attorney," "regulatory," "compliance," "policy," "government relations," "IP counsel", for the later. I will then calculate protective jobs / productive + protective jobs to find the share of jobs fitting into non-productive categories and explore this data across firm sizes.

Once my initial analysis is complete, I aim to conduct robustness checks such as alternating keyword specifications, controlling for industry, manually validating random samples, as well as sourcing new job data to verify there aren't systematic biases on LinkedIn.

Below is an initial visualization based on the methodologies above. It shows that the unproductive job posting share actually decreases with company size, going against my initial hypothesis. That being said, there is a lot of room to build more robust analysis and pull in new data sources.

[genevese]

Geneva Kirk Drayson

Research question: To what extent do AI agents relax the traditional disadvantages faced by entrants pursuing modular innovation in established industries, and which industries are therefore most versus least vulnerable to incumbent displacement?

Research design.
I propose a comparative, industry-level research design that combines a structured cross-industry analysis with simulated entry scenarios to test whether AI agents substitute for the knowledge bases and organizational capabilities that Henderson and Clark (1990) and Christensen (1997) argue entrants typically lack. Building directly on Modularity, Architectural Innovation, and New Venture Success, the design focuses on modular innovation—defined as recombining existing modules without changing system architecture—and asks whether AI agents alter the entrant–incumbent asymmetry in this specific mode of innovation.

The core of the design is a two-stage framework. First, I construct an industry typology based on three dimensions drawn from the modular innovation literature: (1) codifiability of knowledge (tacit vs. explicit), (2) modularity of production or service delivery, and (3) dependence on complementary assets (regulatory approval, distribution networks, trust). Second, for each industry category, I simulate an entrant equipped with AI agents capable of tasks such as rapid learning from documentation, process optimization, code generation, and decision support. These simulations do not assume architectural innovation or market creation; they strictly model incremental, modular improvements consistent with the Henderson–Clark definition.

The outcome variables are predicted entrant performance relative to incumbents, measured as (a) time-to-parity with incumbent offerings, (b) cost reduction achievable through modular substitution, and (c) scope of functionality replicable without incumbent-specific organizational routines. This design is informed by the reading’s claim that modular innovation fails for entrants because of missing organizational capabilities; AI agents are explicitly modeled as a partial replacement for those capabilities, allowing a direct test of the assumption rather than an abstract critique.

[verachaudhry]

Impact of Revolving Door Policies on Firm Innovation and Productivity
Vera Chaudhry

Research Question:
How do revolving door policies impact firm production and innovation?

The Akcigit et. al. paper, “Connecting to Power” discusses interesting empirical work about the influence of local Italian politicians being employed by Italian firms, and studies the effects of these political connections on innovation. In the study, the paper only looks at firms that employ politicians that are actively in office. However, there are many examples where politicians are employed by firms when they are not in office. In some cases, firms employ politicians that are out of office. This is usually what happens in the United States (vs. Italy where local politicians can be dually employed). These are sometimes called “revolving door policies,” meaning the ‘door’ between politics and private firms is open and there is a flow of workers between the two. How does firm production and innovation change as the number of industry revolvers (ie politicians that work for a firm) change over time?

Research Design:

This research project looks at how ‘revolving door’ numbers have changed over time for each industry, and how this has impacted production and innovation for firms over time. Specifically, my data analysis reveals that different firms have been growing the % of industry revolvers over time, while others have been decreasing, and others have remained more constant. Using this fact, this paper will then assess whether innovation and production for an industry has changed over time in a parallel trend to the changes in the number of industry revolvers.

As a further investigation, this project can then look at firm specific innovation (proxied by patents filed) for firms that hired former politicians for specific roles (like government affairs roles) vs. firms that did not hire politicians for specific roles. This will aim to capture the value of an ex-politicians political connections and impact in the industry for a firm. The research project will also look at the difference in firm output for parallel firms that did and did not hire former politicians.

The theoretical model will follow the same structure as the Akcigit et. al. model (2020), except “politicallly connected firms” will represent U.S. firms that hire ex-politicians.

Data visualization:

This data from OpenSecrets(represented in the chart and the graph) shows the varying trends in % of industry revolvers per industry. Note that some industries, like Health Services and real Estate have increased their numbers of industry revolvers in the last few years, while others, like Air Transport and Oil & Gas, have declined. This data visualization represents the first step of a multi-step project (Notably, this data only shows the changes in number of industry revolvers. Based on this data, we see that within industries, the % of industry revolvers changes over time. The second step of this project is to use measures for innovation (patents) and production (firm output) to see if this changes in line with the changes in % of industry revolvers. This is not yet causal, but is the first step a potential project design. The patent data will have to be regrouped to do this analysis, but this can be my next step of the project. First, I think its important to see how "revolving door" policies are dynamic over time).

Example in the media:

[eknez]

Name: Ethan Knezevich

Paper Addressed: Modularity, Architectural Innovation, and New Venture Success

Central Question(s): As Emerson Pugh said, “If the human brain were so simple that we could understand it, we would be so simple that we couldn’t,” so how can the important topic of modularity be confidently approached when the human capacity to understand such minutia is inherently - and forever - limited? Moreover, to what extent can complex modular units be considered units of decomposable objects when their use and inner-workings begin to surpass the human capacity to understand the degree to which each component contributes to the final product, especially when human understanding is constantly evolving and/or incomplete?

Emerson Pugh’s observation—“If the human brain were so simple that we could understand it, we would be so simple that we couldn’t”—captures a fundamental tension that sits at the heart of modularity. Modularity presupposes that complex systems can be meaningfully decomposed into parts whose internal coherence and external interfaces are sufficiently stable to permit analysis, recombination, and innovation. Yet the very systems for which modularity is most attractive—brains, economies, technological ecosystems, and increasingly machine intelligence—are those whose internal complexity exceeds the limits of human comprehension. This raises a paradox: how can modularity be approached with confidence when the act of understanding is itself irreducibly constrained?

Modular units need not be fully understood internally to be treated as modules; rather, they must exhibit reliable patterns of interaction at the level of inputs, outputs, and interfaces. In this sense, modularity operates through functional abstraction. We interact with components as if they were decomposable because doing so yields explanatory, predictive, or practical leverage, even when the true causal pathways remain opaque. Confidence in modularity, then, does not stem from completeness of understanding but from robustness across contexts: modules “work” often enough, and consistently enough, to justify their use. However, this constitutes a fundamental comprimise with the full-bodied concept.

Moreover, human understanding is not fixed. Indeed, while cognitive limits are permanent, the tools that extend cognition—formal models, computational methods, and collective knowledge systems—continually reshape what can be meaningfully abstracted. Modularity evolves alongside these tools, reflecting the current frontier of what can be coordinated, recombined, and reasoned about.

Modularity is neither a guarantee of full understanding nor a naive simplification, but a disciplined acknowledgment that progress often requires working productively with what we cannot fully grasp. In this way, modularity does not completely deny Pugh’s insight, but rather fuels a path toward deeper understanding of the structures that underlie what we consider technology.

As such, I intend to analyze - in both quantitative and qualitative terms - how human understanding directly influences modularity and the granularity of modular analysis on different technologies and decomposable objects, and how different paradigms of deconstructing objects may improve qualification.

Figure:

This figure uses an interesting paradigm that can potentially be applied to the decomposition of complex modular objects, whereby, instead of explicitly defining/finding modules, the probability that two or nodes are connected is calculated, thus allowing for the discovery of "hidden" modules/components in complex objects.

Source: https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/neuro.11.037.2009/full

[violetreed]

Violet Reed

Research Question:
Do the effects of pro-competition innovation policies on regional startup dynamism depend critically on the pre-existing level of political connections among incumbent firms?

Motivation:
This week's reading by Akcigit, Baslandze, and Lotti demonstrates that political connections entrench incumbents, reduce innovation, and worsen industry dynamics. Their finding of a "leadership paradox" - where connected market leaders innovate less - suggests that the local political economy isn't background noise but a primary boundary condition for policy effectiveness. A pro-entrant policy package introduced into a region with highly connected incumbents may be captured and rendered inert, while the same policy in a less politically captured region could successfully catalyze innovation. This refines my project's core question: Can policy counteract preservationist forces, or is its success predetermined by existing political ties?

Research Design:
I propose a triple-differences design that builds on the original DiD framework but incorporates political connections as a key moderating variable. Using methodology from Akcigit et al., I will construct a Regional Political Connection Index (RPCI) for each city-region from administrative data (e.g., local election records, corporate filings, and procurement databases from Brazil or India). The index will measure the pre-policy (2015-2018) density of ties between major firms and local politicians.

The model will estimate:
ΔDynamism = β₁(Treat × Post) + β₂(Treat × Post × High-RPCI) + controls

A negative and significant β₂ would indicate that pro-entrant policies are substantially less effective in regions where incumbents are politically entrenched. This directly tests whether political connections can nullify policy interventions aimed at boosting competition.

Data Visualization:

The simulated data illustrates the core hypothesis: pro-entrant policies drive significantly higher dynamism in regions where incumbents have low political ties, but have muted effects where political ties are high.

[lglickman-edu]

HireVue as an Architectural Technology that rewires hiring: Innovation Accelerator or Barrier to Creative Destruction?

Name: Lucas Glickman
Research Question: Do AI enabled automated hiring tools accelerate innovation by lowering talent matching frictions or do they slow creative destruction by strengthening incumbent firms’ ability to gatekeep and scale non productive selection strategies?

The hiring process is one of the most important allocation processes in the economy as it determines who gets access to opportunity, shapes which firms can scale, and governs how quickly labor reallocates from low-growth to high-growth. Therefore, I propose that hiring platforms should be treated as economic infrastructure that affects creative destruction, not as just a peripheral managerial tool. As such, online hiring has moved away from being a point solution, replacing an interview, to an architectural change recombining existing technology modules (video capture, standardized assessments and scoring) into a scalable hiring pipeline that yields new candidate data fitting both Arthur and Cao/Chen/Evans’ theories, where architectural recombination is often the safest innovation strategy.

In doing so, these selection systems risk becoming barriers to creative destruction as when firms become dominant, they shift from productive innovation to non-productive defenses that preserve existing rents. As discussed by Akcigit, incumbents can gain a wedge via political connection that produces static gains but dynamic losses via weaker entry/reallocation of resources. Naturally, automated hiring can mirror this wedge through scaled selection control. By enabling incumbents to process massive applicant pools cheaply, this standardization may filter out unconventional candidates who are oftentimes key to novel recombination whereas entrants may face a higher eligibility threshold to compete for talent. As such, if these hiring tools reshape selection at scale, they may become a new institutional bottleneck for innovation and mobility, especially in concentrated industries.

To study this, I propose to test whether firm level adoption of automated hiring tools act as a shock to hiring capacity and selection structure to observe if before versus after adoption, there is an increase in creative destruction or incumbent entrenchment. To do so, I propose using job postings and hiring workflow signals as evidence of an asynchronous interview request/pre-employment test along with outcomes via hiring volumes (job openings filled), time to fill (posting duration) and employee composition (education,geography, and prior employment). By then treating adoption of the tools as a shock, I will apply a Difference-in-Difference analysis with the treatment group as the first to adopt automated hiring tools while the control group are those which are similar in industry size, and growth which do not. By comparing treatments against control firms before and after adoption we can find an estimate of whether adoption shifts hiring growth, talent composition, and performance. Further then by splitting firms by “incumbent” versus “entrant” via age, market power, and baseline growth rate, we can test whether the same hiring tool can have opposite effects. By analyzing post shock effect on hiring rate, homogeneity, and share in total hiring volume, we can then conclude how these automated hiring tools affect innovation.

[csardinas]

Carolina Sardinas

Research Question

Do countries with stronger incumbent-protecting regulation in network industries exhibit systematically more defensive patenting behavior in network technologies relative to competitive technologies?

Motivation/Background

This question builds from the broad idea that barriers to entry and incumbent protection can shape innovation incentives, but instead of focusing on firm-level outcomes, it tests whether cross-country regulatory environments predict measurable differences in how patents are used—particularly whether patenting becomes more defensive (heavy citing of prior art with weaker subsequent impact) in sectors where incumbents are structurally protected.
This perspective is closely related to Baslandze (2021), who shows that large incumbent firms increasingly rely on non-productive strategies, including defensive patenting, to preserve market power and slow creative destruction. The paper documents that such patents tend to receive fewer forward citations and serve primarily to block entry rather than advance technology. This memo extends that logic from firms to countries by testing whether incumbent-protecting regulation in network industries is associated with systematically more defensive patenting in the technologies used in those sectors.

Research Design

I construct a country-level panel by combining (i) OECD PMR “yes/no” regulatory questions from the Econwide PMR database and (ii) country-by-technology patent summary statistics from the USPTO PatentsView bulk tables. First, I classify PMR questions into two baskets: network sectors (electricity, natural gas, e-communications, rail/air/water/road transport) and competitive sectors (retail and selected professional services plus digital markets). For each country, I compute PMR restrictiveness indices as the mean share of “yes” answers within each basket and define the main policy variable as ΔPMR = PMR_network − PMR_competitive. Second, I classify technologies using CPC section letters. Using inventor location, I assign each patent to a primary inventor country (ISO2). For each country × CPC section, I summarize citation behavior and construct a defensive patenting measure Defensive = (avg backward citations)/(avg forward citations + 1), where backward citations proxy the intensity of prior-art referencing and forward citations proxy impact. I then compute country-level sector tilts: Defensive_network and Defensive_competitive, and the main outcome ΔDefensive = Defensive_network − Defensive_competitive. The main empirical test is a cross-country regression and/or rank correlation of ΔDefensive on ΔPMR, with robustness checks using levels (PMR_network, PMR_competitive) and alternative CPC mappings (e.g., telecom-heavy subsets within H/G). The identification is descriptive/associational, but the within-country differencing (network minus competitive) helps control for slow-moving country traits (legal origin, development level) that affect patenting overall.

Visualizations

Note: Within each country, network sectors often have a different restrictiveness level than competitive sectors; the slope size visualizes “incumbent protection tilt” (ΔPMR).

[bheim]

Do patent filings predict market cap increases among tech companies?

Question: Do past patent filings by US tech companies predict future changes to their market cap? The larger question we want to get at is whether patent filings are strong signals for innovation within an economy.

Methods: I returned to my best friend, the google patent database, to investigate this question. I also used yfinance to pull market cap data from two tech companies. My intention was to run the regression on a couple of companies to see if I could identify a strong relationship. If I could, I’d dive deeper into this analysis in future weeks.

The model I ran was a fairly naive model that was looking to see if patents were predictive of changes in market cap:

$$ \Delta M_t = α₀ + α₁ Patents_{t-k} + α₂ SelfCit_{t-k} + α₃ ExternalCit_{t-k} + ε_{t} $$

where variables are defined as follows:

Delta M_t → the change in market cap period for the company between period t and t-1

Patents_{t-k} → the number of patents filed by the company since k periods ago (I varied this in my analysis from 1 to 3 periods)

SelfCit_{t-k} → the number of self citations (e.g., IBM citing their own patent) that the patents filed in the the last t-k periods received

ExternalCit_{t-k} → the number of citations from other filers that the patents filed in the last t-k periods received.

I chose two public companies to analyze for this memo. The first was IBM — the biggest patent filer in recent years. The second was Intuit — a tech company that files some patents but not too many. Since the sample size is so small, I won’t be indexing much on the p-value associated with the coefficients.

Results: Initially, I thought that patent filings had a near zero effect on the market cap of the companies. However, inspired by [Hall, Jaffe, and Trajtenberg (2005)](https://www.jstor.org/stable/1593752?seq=1), I took a look to see if self-citation played a role, and it indeed does appear to do so. For both IBM and Intuit, self citations on patents in recent periods predicted growth in market cap, and this was even significant for Intuit (impressive for such a scarce sample).

While I could share the regression results as my image, I don’t think they’re so interesting, so instead I’ll show a scatter plot for Inuit. These results are promising in linking market cap toward self citations. I think an interesting direction to go in the future will be incorporating the idea of surprise as a predictor for these patents and see if that leads to an increase in market cap or if we see (at least) the same strong degree of variance in impact from surprising patents as we see in terms of the number of citations surprising patents get.

[ShiyangLai]

Shiyang Lai

Research Question.
Do LLMs redirect the selected development pathways of scientific knowledge by changing which assembly trajectories—among many plausible recombination routes—become institutionally validated, standardized, and widely reused?

Research Design.
This project treats scientific progress not only as producing new “objects” (papers, methods, pipelines) but as selecting paths: sequences of incremental joins through which a community converges on particular methodological lineages. Assembly Theory (AT) motivates this by emphasizing that selection is visible when complex assemblies arise through repeated reproduction—but the deeper implication is pathway-level: diffusion concentrates on certain assembly routes while alternatives remain latent. I ask whether LLMs shift this pathway selection by (i) making some joins easier to propose and articulate, (ii) increasing the local branching factor of plausible next-steps, and (iii) changing what becomes legible and persuasive to reviewers, collaborators, and adopters.

Empirically, I study a bounded domain with recognizable modular “pipelines” (e.g., computational biology workflows, quantitative social science methods). I construct a corpus of papers (2016–2025) and model knowledge development as a method lineage graph: nodes are method modules (e.g., specific dataset–model–evaluation bundles), and directed edges represent development steps (a paper that modifies, extends, or recombines prior modules). Edges are inferred from citation contexts plus structured signals (shared code dependencies, dataset reuse, protocol references). The central outcome is pathway concentration: whether diffusion accumulates along a narrower set of lineages (path dependence) or spreads across many alternative trajectories (exploration).

Identification leverages the post–late 2022 LLM shock in an event-study / difference-in-differences design. I compare subfields with higher vs lower LLM exposure (proxied by code/writing intensity, availability of modular components, and openness of tooling) and estimate changes in: (1) the branching factor of proposed development steps (new edges per lineage), (2) the selection gradient from proposed steps to adopted steps (edge survival into follow-on work), and (3) standardization speed (time from module introduction to becoming a common dependency/citation anchor). A key distinction is between candidate edges (suggested joins, visible in paper prose and released code) and selected edges (joins that later papers actually build upon). If LLMs mainly expand ideation, we should see higher branching but unchanged pathway concentration; if they redirect selection, we should observe systematic re-weighting of which lineages dominate downstream work.

Contribution.
Rather than treating LLMs as uniformly accelerating science, this study isolates a more structural mechanism: LLMs may change which roads get paved. By measuring how diffusion reallocates across competing development pathways, the project links AI assistance to path dependence, standardization, and the institutional selection of scientific trajectories—offering a quantitative account of how “what gets built next” shifts when recombination and articulation are partially automated.

[HenryPinner7]

Research Question: As clean-tech manufacturers scale, do they reallocate technical labor away from product innovation and toward regulatory/compliance functions, and is that associated with slower innovative performance (e.g., fewer breakthrough patents, narrower novelty, etc.)?

Background and Motivation: This question fits the “productive vs. protective” tension emphasized throughout this week’s readings. Akcigit et al. (“Connecting to Power”) show that firms with political connections can grow faster without corresponding increases in productivity, suggesting that resources can be used to protect incumbency rather than expand frontier innovation. I want to test a potentially analogous mechanism in clean tech: instead of political ties, I use workforce composition as a marker of whether firms are leaning into productive innovation (R&D/product engineering) versus protective capacity (regulatory navigation/compliance).

Research Design: To answer this question, I would build a firm-year panel of clean-tech manufacturers and integrators (e.g., solar modules/inverters, batteries, power electronics, etc.) from roughly 2010-2025. My first step would be to compile a large sample of firms using industry lists and public sources, then match each firm to LinkedIn company pages to approximate annual workforce composition. Using a transparent keyword dictionary (and LLM-assisted title classification), I would group employee job titles into three buckets: (1) innovation/product engineering roles (R&D, product design, materials, firmware/controls), (2) compliance/regulatory roles (regulatory affairs, certification, safety standards, environmental compliance/QA systems), and (3) operations/execution roles (manufacturing, supply chain, production). The main explanatory variable would be the compliance share (# compliance employees divided by the total across these core categories), and scale would be measured by total headcount and year-over-year growth.

To measure innovation performance, I would match firms to patent assignees (using USPTO or similar) and construct outcomes like patents per year, citation-weighted patents per innovation employee, field-normalized forward citations (as a proxy for impact), and an indicator for “breakthrough” patents (e.g., top 10% citations within a technology class-year). I would then estimate firm and year fixed-effects regressions relating scaling to compliance share, and compliance share (lagged by one year) to subsequent innovation outcomes (treating this as correlational evidence, causality could come in a future memo). As a potential mechanism check, I would test whether higher compliance share is associated with lower patent originality/breadth.

Figure: The simulated figures illustrate the basic pattern I would test in the data. The top panel shows that as clean-tech firms scale (larger headcount), a higher share of their core roles shifts toward compliance work. The bottom panel shows that higher compliance share is associated with fewer “breakthrough” innovations (lower breakthrough patent share). In this way, this research memo aims to see if it is possible to look at a firm’s employee breakdown as a way to measure some of the anti-competitive habits we have discussed in class.

[isocitrate]

Justin Chen

Question

Do equity markets price firm-specific regulatory risk differently than sector-wide regulatory risk, and can this differential response identify the private value of political connections in the AI industry?

Motivation

The political economy of Schumpeterian destruction suggests that incumbents facing creative destruction invest in political influence to erect barriers to entry. However, measuring the returns to such investments is empirically challenging: firms that lobby heavily may do so precisely because they face greater regulatory exposure. Fisman's (2001) event study of political connections in Indonesia provides a useful template whereby one can exploit variation in whether regulatory events target specific firms versus the entire sector to identify how markets price political risk in big tech.

Methodology

A difference-in-differences event study would compare stock price responses of targeted versus non-targeted firms around firm-specific regulatory events (DOJ/FTC antitrust suits against individual companies), contrasted with sector-wide events (legislative proposals, leadership appointments, EU regulations).

Identification strategy: When the DOJ files suit against Google specifically, Google faces direct regulatory costs while Amazon and Microsoft experience only indirect effects. This isolates the firm-specific component of regulatory risk from sector-wide effects. While lobbying levels are endogenous, the timing of DOJ/FTC enforcement actions provides plausibly exogenous variation.

Empirical specification:
$$CAR_{i,t} = \alpha + \beta_1 \cdot Targeted_{i,t} + \beta_2 \cdot EventType_t + \gamma \cdot X_{i,t} + \epsilon_{it}$$

Where $Targeted_{i,t}$ indicates whether firm $i$ is the specific target of event $t$, and $EventType$ captures heterogeneity across antitrust, AI regulation, and leadership events.

Preliminary results

Using daily stock returns from Yahoo Finance (2020-2024) and 12 major political events, the cumulative abnormal returns are calculated using a market model with SPY as the benchmark. The figure below presents results for firm-specific events (DOJ/FTC suits against Google, Meta, Amazon, and Microsoft).

The results show that: targeted firms experience -0.92%pp CAR at day +5, while non-targeted peers experience +1.03%pp CAR, a difference of almost 2 %pp. Note that non-targeted firms show positive abnormal returns, suggesting markets may interpret competitor-specific regulatory actions as beneficial for rivals (reduced competitive pressure) or as signals that sector-wide regulation is less likely.

Further analysis reveals substantial heterogeneity by event type: antitrust events show high-lobbying firms outperforming by +0.65 %pp, while AI regulation and leadership events show the opposite pattern. This heterogeneity suggests that different regulatory mechanisms may operate via different channels.

Implications

These preliminary findings suggest that (1) firm-specific regulatory targeting produces measurable market effects distinct from sector-wide sentiment, (2) the "competitor benefit" from rival targeting may partially offset Schumpeterian destruction concerns, and (3) simple lobbying expenditure may be a poor proxy for political influence, requiring more granular measures of firm-regulator relationships.

[tepiv1]

Ved Jethani

Research Question
Does the 'revolving door' between government research agencies and the pharmaceutical industry create a preferential velocity for Phase 1 trials?

Motivation
As Professor Akcigit discussed using empirical data for Italy, hiring in-office politicians helps existing firms survive, lowers productivity, and removes an institutional wedge that might otherwise exist. Overall, their impact is causally negative on innovativeness by acting as friction for new firms. In the US pharmaceutical industry (and many more worldwide), the phenomena of the revolving door is observed. While simultaneously holding office and being employed at a high level is not permitted, many people go from private industry to government positions and vice versa. In particular, studies show that between 2004 and 2020, 15% of employees from the Department of Health and Human Services were privately employed prior to appointment and at the end of their appointment, 32% exited to private firms. Last week, I aimed to analyze the impact of budget cuts for the NIH on the economy, one impact being the lost productivity from researchers. The revolving door theory suggests that many of these researchers who the NIH can’t fund will enter industry. Compared to Italian politicians, the researchers are not government and private employees simultaneously, so they should not have similar leverage, but it is interesting to see whether they have some level of “institutional information” which aids the firm in going to market at a quicker speed.

Research Design
This study collects data of NIH researchers who exited industry to small startups. Ideally, they would hold high level roles ( CSO or head of research) at the startup as to suggest their relative importance. Then, the velocity metric is measured by Velocityyears = Date of phase 1 initiation - Company Founding date and comparing a control set with no ex-NIH researcher in order to identify a difference. Compiling NIH researchers is simple but figuring out whether they left the NIH and to where requires manual searching. Due to time constraints, I compiled more notable cases for the sake of visualization. Company founding dates are pulled through SEC filings and Phase 1 trials are filed with the government publicly.

Discussion
While not statistically significant (perhaps with more data), ex-NIH median velocity is lower. Causality is also much harder to prove here. Unlike the case of Italy, it is unclear whether this is inherently bad. While scientists may leave the NIH for a number of reasons, private firms may be more efficient in developing life-saving technologies and provide a better array of resources to the scientist. In essence, the NIH research might fund de-risk technology enabling new startups, rather than connections of the researchers reducing a wedge that exists for other competitors. However, when separating ex-NIH directors who joined/founded startups (n=5), median velocity was 3.7 years, which may indicate the power of their individual network. In the future, it would be interesting to analyze the success rate (reaching market) of ex-NIH research startups vs the industry baseline, to see whether there is a higher probability of innovativeness from the group.