Dirty water and poor WASH conditions can sharply raise maternal sepsis risk.
Maternal sepsis remains a major contributor to preventable maternal mortality. Environmental conditions—especially unsafe water—play a critical role in infection risk during childbirth.
This study uses a household survey design with a binary outcome:
hospitalized = 1 if yes, 0 if no, and
insurance = 1 if covered, 0 if not.
This is a standard framework in health econometrics for estimating the effect of insurance on healthcare utilization.
Interpretation: If marginal effect = 0.08 → insurance increases hospitalization by 8 percentage points.
Why Two-Part Models Are Better
Household survey data typically contain many zeros (no hospitalization) and a skewed distribution among users.
A simple logit only models whether hospitalization occurs, ignoring intensity.
Separates access (any hospitalization) from intensity (number of visits)
Handles zero-heavy and skewed data
Uses full information instead of collapsing outcomes
Stata Example
* Part 1: probability of any hospitalization
logit hospitalized i.insurance age i.sex i.education ///
ln_income chronic_illness i.rural
* Part 2: intensity (only if hospitalized)
glm n_admissions i.insurance age i.sex i.education ///
ln_income chronic_illness i.rural if hospitalized==1, ///
family(gamma) link(log)
Interpretation
Part 1: Insurance increases likelihood of hospitalization
Part 2: Insurance affects number of admissions or length of stay
Policy Insight (Kenya Context)
In Kenya and similar settings, insurance schemes often:
Increase access to care (more people hospitalized)
Increase intensity of care among users
Two-part models capture both effects, while simple logit only captures the first.
A strong econometric model in health research combines a clear causal or predictive question with an appropriate regression structure and data design. Below is a practical framework with examples applicable in Stata or any statistical software.
1. Common Econometric Models for Health
Health economics typically focuses on:
Health outcomes: hospitalization, mortality, disease prevalence
Building Scalable, Transparent Econometric Workflows in Stata SE
Building Scalable, Transparent Econometric Workflows in Stata SE
In modern econometrics, the challenge is no longer just estimation—it’s scale, reproducibility, and credibility. When working with millions of observations and policy-relevant questions, your Stata workflow must be both computationally efficient and fully transparent.
Large-Scale Data Management and Cleaning
Handling large datasets in Stata SE requires careful attention to memory and execution speed. A simple but powerful habit is using compress immediately after loading data. This reduces storage requirements without altering values.
Stata’s frames (introduced in version 16) allow you to keep multiple datasets in memory simultaneously, avoiding repeated saves and merges.
Automation becomes critical at scale. Regular expressions (regexm, regexs) help clean messy string data such as IDs or survey responses. For faster aggregation and joins, the ftools package significantly improves performance.
Validation is essential. Use assert statements to enforce assumptions:
Income must be positive
Dates must fall within valid ranges
Pair this with datasignature to detect unintended data changes across sessions.
Advanced Econometric Modeling
With a robust data pipeline, you can move beyond basic OLS into more realistic models.
High-dimensional fixed effects:reghdfe
Treatment effects:teffects
Instrumental variables:ivreg2
Dynamic panels:xtabond2
These tools enable rigorous causal inference and efficient estimation even with large datasets.
Reproducibility and Transparency
Your code is part of your evidence. A well-structured project should include:
The Knowledge Paradox: When Does Sharing Become Theft? | Zacharia Nyambu
The Knowledge Paradox: When Does Sharing Become Theft?
March 14, 2026 | By Zacharia Nyambu
The “knowledge paradox” in AI and finance is about a structural shift: the same openness that once leveled the playing field now fuels models that centralize informational, financial, and computational power in a few hands.
In finance and financial engineering, that shift collides with copyright,
data-protection, and market‑abuse rules in ways that make “sharing vs theft”
not just an ethical debate, but a legal and economic fault line.
From Commons To Collateral: How Finance Uses “Open” Data
For decades, open data and open research infrastructures have been justified as public goods that lower information asymmetries in markets.
In practice, financial institutions and quantitative funds now treat open datasets—academic working papers, GitHub code, open‑access journals, public filings, and Creative Commons‑licensed content—as raw material for proprietary alpha generation and risk models.
Three examples in finance and financial engineering
Asset pricing research: Open macro and firm‑level datasets feed factor models and ML pipelines that underpin commercial “smart beta” and multi‑factor products, while the models and parameters are fully proprietary.
Alternative data: Web‑scraped reviews, job postings, satellite feeds, and social media are harvested—often under ambiguous licenses—to build credit, sentiment, and now AI‑driven trading signals.
Retail analytics and credit scoring: “Consent” to share data in apps or platforms often cascades into data brokers and lenders, who treat that data as a monetizable asset, not a shared commons.
Legally, much of this sits in a grey zone between lawful re‑use and potential copyright
or database‑right infringement, depending on jurisdiction and on whether scraping respects
contractual and technical access restrictions. Economically, it creates an inversion:
public and open resources become informational collateral for private balance‑sheet gains,
reinforcing the knowledge paradox that this Creative Commons session surfaces.
When Does Sharing Become Legal “Theft” In AI Training?
The law does not recognize “theft of openness” as such; it talks in terms of copyright
infringement, breach of contract, database rights, trade secrets, unfair competition and,
in finance, market abuse and consumer‑protection norms. But recent AI‑training cases begin
to sketch a legal answer to “when does sharing become theft?” that is directly relevant to
financial and quantitative use‑cases.
Recent U.S. decisions such as Bartz v. Anthropic and Kadrey v. Meta—part of a first wave of AI‑training litigation—apply the four‑factor fair‑use test in 17 U.S.C. §107 to large‑scale ingestion of copyrighted works. Courts there distinguished between:
Transformative learning: Using lawfully obtained works to train a model that does not substitute for those works, and whose outputs are not substantially similar or market‑replacing, which courts have tended to treat as fair use in the U.S. context.
Substitutional copying: Using works to build a system that effectively competes with, or reproduces, the market function of the original, which courts have signaled is much less likely to qualify as fair use.
One federal analysis framed the emerging principle this way: “transformation protects learning;
substitution invites liability,” tying legality to whether AI training or outputs erode the original
work’s market. For financial and legal databases—think proprietary datasets like Westlaw in
Thomson Reuters v. Ross Intelligence or high‑value paywalled datasets used in quantitative
finance—copying for a competing product is more likely to be seen as infringing than as acceptable
text‑and‑data mining.
For finance professionals, that means:
Using open or lawfully licensed data to train risk models, pricing engines, and robo‑advisors is more defensible when outputs do not reproduce the source content and do not undercut the rights‑holder’s core product.
Building AI tools that approximate or replace a subscription data vendor using that vendor’s own content for training crosses the line from “sharing” into probable infringement under current U.S. precedent.
Financial Regulation: Data As Market Power, Not Just IP
Beyond copyright, financial regulation treats information asymmetry and data concentration as
core systemic‑risk and market‑fairness issues. Open data used to be a counterweight to incumbents’
informational advantages, but AI flips that logic: firms with the capital to train large models on
open resources can reinforce their lead rather than democratize access.
Three legal and regulatory levers
Market abuse and unfair practices: Misuse of non‑public data can breach insider‑trading and market‑manipulation prohibitions, while mass appropriation of “open” data that violates terms of use can trigger unfair‑competition or consumer‑protection scrutiny.
Open banking and data portability: Frameworks that force banks to share customer data via APIs aim to empower consumers and foster competition, but they also require strict governance around consent, security, and secondary uses such as AI training for credit models.
Algorithmic accountability: Regulators increasingly expect transparency about data provenance, explainability around model decisions, and evidence that models do not encode discriminatory bias or unfair outcomes.
In effect, financial law reframes the knowledge paradox as a question of who holds informational
advantage and who bears the risk. If open data trains proprietary credit or trading models that
entrench incumbents and amplify systemic risk, regulators may respond with data‑governance,
model‑risk, and competition‑law interventions.
Designing Contracts And Licenses For Financial Engineering
If we accept that “sharing becomes theft” when open contributions are systematically turned into proprietary financial edge without regard to contributors’ rights or expectations, then a core solution is contractual and licensing innovation. Creative Commons has shown how standardized licenses can embed norms into legal code; similar moves are emerging around AI and finance.
Key contractual tools and design choices
AI‑restricted licenses: Terms that permit human re‑use but restrict training of commercial AI or require separate paid licenses, especially in high‑value financial contexts.
Data‑scraping codes of conduct: Standards that set out acceptable scraping practices, require documentation of data provenance, and distinguish between non‑profit research and leveraged commercial re‑use.
Revenue‑sharing and data trusts: Data trusts or cooperatives that negotiate licenses with financial firms and share downstream value with contributors.
API‑first access: Controlled APIs that restrict bulk extraction for model training while enabling legitimate research and transactional access.
From a financial‑engineering perspective, training data becomes an intangible asset with pricing,
legal, and governance constraints that must be modeled alongside capital, liquidity, and risk.
Open vs Closed Data Practices In Finance
Dimension
“Pure Open” Practice
Guard‑railed Open Practice
Closed / Proprietary Practice
Access to datasets
Unrestricted download and scraping; attribution only
Open for human and non‑AI use; separate license for AI training
Paywalled, contract‑bound, API‑gated
AI training use
Implicitly allowed unless terms forbid
Explicitly licensed with conditions, fees, or purpose limits
Prohibited absent negotiated license
Value capture
Value concentrated in those with compute and capital
Shared via revenue‑sharing or negotiated AI licenses
Concentrated in rights‑holder and direct clients
Legal risk (copyright/IP)
High ambiguity for commercial AI use
Lower, because scope and terms are clear
Lower, but possible antitrust scrutiny
Impact on financial markets
Can widen informational gaps
More balanced; contributors participate in value
Stronger incumbency advantages
Keeping Knowledge Open Without Fueling Extraction
The Creative Commons panel at SXSW asks how to keep knowledge open without facilitating
exploitation at scale, precisely when AI makes extraction cheap and proprietary capture
highly profitable. In finance and financial engineering, a workable answer likely blends
legal rules, contract design, technical controls, and community norms.
Four practical directions
Specify AI uses up front: Choose licenses that clearly permit or restrict AI training, and state expectations around commercial re‑use.
Build transparent data‑lineage into models: Log which datasets and licenses feed each model so compliance can audit for violations.[web:10]
Advocate for sector‑specific TDM exceptions: Allow socially beneficial research while imposing duties of non‑substitution, non‑discrimination, and reasonable revenue‑sharing.[web:10]
Align incentives with fiduciary and ESG duties: Make “not stealing the commons” part of responsible investment and risk management.
The paradox becomes a design question: how do we structure contracts, incentives, and constraints
so that open knowledge remains a shared input to market innovation, instead of an unpriced subsidy
to whoever has the biggest model and the lowest cost of capital?
Read This Page As:
A primer on how AI is reshaping the economics of open data in finance.
A quick legal guide to when AI training crosses from “sharing” into potential infringement.
A starting point for quants, lawyers, and policymakers designing fairer data and model practices.
