State-of-the-Art Microeconomics (2024–2026)

A learning report on the research frontier in microeconomic theory, applied micro, behavioral economics, market design, industrial organization, labor, public, information, and experimental economics — written for a highly capable self-learner.

Executive summary

If you opened a graduate microeconomics qualifying exam in 1995, then again in 2025, the two papers would look like fields divided by a methodological revolution. Three deep currents define the present state of microeconomics:

1. The credibility revolution has matured. After Imbens and Angrist won the 2021 Nobel for the local average treatment effect (LATE) framework (Nobel Prize 2021), the next decade has been about cleaning up the staggered difference-in-differences mess. Goodman-Bacon (2021), Callaway-Sant'Anna (2021), de Chaisemartin–D'Haultfœuille (2020), and Borusyak-Jaravel-Spiess (ReStud 2024) showed that two-way fixed effects (TWFE) regressions silently weight cohort comparisons in pathological ways. The 2025 practitioner's guide by Callaway, Cunningham, Goodman-Bacon, and Sant'Anna is now the standard reference.

2. Market design left the seminar and went into operating reality. Roth's kidney exchange algorithms now run national programs; Milgrom and Wilson's 2020 Nobel for auction design (Nobel Prize 2020) recognised the 2017 FCC incentive auction that repurposed 84 MHz of TV spectrum and raised roughly $20 billion (NSF Q&A with Milgrom). Refugee resettlement has joined the catalogue of operational matching markets (Delacrétaz, Kominers, Teytelboym, AER 2023).

3. Behavioral economics has split into two halves. The "easy nudge" era is over: Ariely's 2012 PNAS paper has been retracted, Francesca Gino's lab is under HBS sanction following Data Colada investigations (Science 2023), ego depletion, power posing, and priming did not replicate. What survived — defaults, loss aversion in the loss domain, present bias, retirement auto-enrolment — are the structural features that economic theory could already absorb. Sunstein himself wrote in 2024 that "we are no longer in the era of the easy nudge" (Hoover 2024).

Around these three currents, six policy battles dominate the live frontier:

• Antitrust against Big Tech. Judge Mehta found Google liable under Sherman Act §2 in August 2024 (White & Case summary); the Amazon FTC case survived motion-to-dismiss in October 2024 (GeekWire); but Meta won its FTC case in November 2025 when Judge Boasberg found that TikTok and YouTube made the market definition untenable (NPR).
• The monopsony renaissance. Card, Manning, Azar, Marinescu, and Steinbaum have rebuilt the case that labour markets are concentrated, search-frictional, and differentiated enough that markdowns of 15–50% are credible (Azar-Marinescu AR Econ 2024).
• The FTC noncompete rule. Proposed Jan 2023, finalised April 23 2024, then enjoined by a Texas district court in August 2024 and abandoned by the FTC in September 2025 (FTC Press Release April 2024). The economic evidence — Starr, Prescott, Bishara — survives the policy.
• The Piketty-Saez-Zucman vs Auten-Splinter inequality fight. Same tax data, same income concept, opposite stories: top 1% rose from 9% to 15% (PSZ) or 8% to 9% (Auten-Splinter) of after-tax income between 1960 and 2019. The dispute hinges on the imputation of underreported business income (Brookings primer; PSZ 2024 comment; Auten-Splinter reply).
• Cash transfers and child development. The 2021 American Rescue Plan CTC cut child poverty briefly to 5.2% (National Academies). Yet the Baby's First Years RCT, designed to detect causal cash-to-child effects, found no statistically significant impacts on 4-year cognitive, executive function, language, brain, or social-emotional outcomes when comparing $333/month vs $20/month groups (NBER w33844, 2024). The discipline is wrestling with both results.
• The Census 2020 differential privacy controversy. Differential privacy was deployed at production scale for the first time at the U.S. Census, with measured effects on redistricting and voting-rights enforcement (Kenny et al. Science Advances 2021). It is the first public-economics application where mechanism design, statistics, and constitutional law meet head-on.

The throughline: microeconomics in 2024–2026 is empirically more disciplined and theoretically more humble than it was twenty years ago. The dominant style is no longer a single clever model; it is a carefully identified empirical fact combined with a sufficient-statistics formula linking it to welfare. The most important open questions involve scaling — site selection bias (List), external validity (Deaton), and how to know when a result will generalise from a Cambridge RCT to a national policy.

Research brief

This report is the third in a planned series — see cross-references at the end.

Key findings

The findings below are organised so that each can be read as a self-contained synthesis. The detailed analysis section then unpacks each in depth.

1. Mechanism design is now an applied engineering discipline

The Roth–Shapley line of work (deferred acceptance, top trading cycles, stable matching) and the Milgrom–Wilson line (multi-unit auctions, deferred-acceptance clock auctions, package bidding) are now installed in operating systems running medical, educational, and spectrum allocations. The frontier topics are maintenance (Roth 2023 NBER w31947) and robustness — what happens when the environment moves under a designed market.

2. Auction theory is reckoning with online advertising at internet scale

Google migrated Google Ad Manager from second-price to first-price in 2019 (AdExchanger) under pressure from header bidding. The theoretical justification — that second-price loses incentive compatibility when there are unobserved upstream auctions — is one of the cleaner cases of theory catching up with practice. Bid-shading then re-introduced strategic behaviour that the move was meant to eliminate, and Google's Randomized Generalized Second Price (RGSP) is the current production answer (Search Engine Land).

3. The behavioral economics canon has been re-graded

A short list of robust findings (defaults, loss aversion in losses, present bias, mental accounting in narrow framings, status-quo bias, social-norm letters at moderate effect sizes) has survived the credibility crisis. A longer list — ego depletion, power posing, priming, much of "honesty" research — has not. DellaVigna and Linos (2022) found that nudge effect sizes from in-government tests by Behavioural Insights Teams average about one-quarter of those reported in academic publications (PNAS 2022).

4. The new IO of digital platforms is fully two-sided

Rochet and Tirole's 2003/2006 foundations remain the canonical model. The 2024 frontier extends them to platform competition under multi-homing, recommender systems, and ad auctions, with Belleflamme and Peitz the standard 2024 review (CRC TR 224 DP #584, 2024). The big regulatory story is that monopolisation cases (Google search, Google ads, Amazon, Meta) are being tried on theories that explicitly invoke this literature.

5. The credibility revolution has digested its own pathologies

A wave of methodological papers (Goodman-Bacon 2021, Callaway-Sant'Anna 2021, Sun-Abraham 2021, de Chaisemartin-D'Haultfœuille 2020, Borusyak-Jaravel-Spiess 2024) showed that the workhorse two-way fixed-effects DiD regression silently does badly under staggered adoption with heterogeneous treatment effects. The field response — new estimators, imputation methods, doubly-robust group-time ATTs — is now mature and standardised.

6. Structural estimation got re-legitimised by tying itself to sufficient statistics

The Heckman-vs-Imbens debate of the 2010s has cooled. Chetty's sufficient-statistics framework gave reduced-form work the welfare claims it once lacked; BLP (Berry-Levinsohn-Pakes 1995) and its PyBLP successor (Conlon-Gortmaker JoE 2024) gave structural IO the credibility-revolution discipline it once lacked.

7. Labour economics now treats monopsony as the modal case

The Manning, Card, Krueger, and Azar-Marinescu programmes converged on a view of the labour market in which search frictions, differentiated jobs, and noncompetes create employer wage-setting power. Markdowns (wage below MPL) of 15–50% are now the consensus range (Azar-Marinescu AR Econ 2024). Cengiz, Dube, Lindner, and Zipperer's bunching estimator showed that minimum-wage increases moved jobs up the wage distribution without losing them, with confidence intervals ruling out total-employment elasticities below −0.06 (QJE 2019).

8. Public economics centres on a fight about distribution measurement

PSZ vs Auten-Splinter is the methodological battle of the decade in public finance. The Saez-Stantcheva (2014) "three elasticities" framework — labour supply, avoidance, bargaining — remains the canonical tool for top-rate analysis. Sufficient-statistics formulas pull policy work toward intuitive elasticities and away from fully parametric models.

9. Cash-transfer evidence sits in delicate tension

The 2021 ARPA CTC expansion was the largest natural experiment in US poverty policy in decades and showed historic poverty reductions (Hamilton Project). The Baby's First Years RCT, the cleanest possible cash-to-child-development causal test, returned null on cognitive outcomes at 4 years. Reconciling these requires careful thinking about what cash buys, over what horizon, in what context.

10. The state-action interface in privacy, antitrust, and labour is now where micro lives

Differential privacy at the Census, FTC rulemaking on noncompetes, antitrust against Big Tech, gig-economy worker classification, FCC spectrum auctions — these are not separate from the discipline. They are the empirical and policy laboratories in which 2024–2026 microeconomics is being argued.

Detailed analysis

The detailed analysis is organised by sub-field. Each subsection has the same structure: foundational result → frontier development → live debate → study-next pointer.

A. Market design and matching

A.1 Stable matching and deferred acceptance

Gale and Shapley's 1962 deferred acceptance (DA) algorithm is the founding result. With $n$ men and $n$ women with strict preferences over each other, DA proceeds in rounds: each unmatched man proposes to his highest-ranked woman not yet rejected him; each woman tentatively holds her favourite proposer and rejects the rest; iterate. The output is stable — no pair $(m,w)$ both prefer each other to their current match — and the proposing side gets its uniquely best stable match (Gale–Shapley 1962, AMM).

Roth's 1984 JPE paper showed that the National Resident Matching Program (NRMP), redesigned from a chaotic process to a centralised clearinghouse in the early 1950s, was actually running a variant of DA. This was a striking case of an institution evolving to mechanism-design equilibrium before the theory existed. By the early 2000s Roth, Sönmez, Ünver, Abdulkadiroğlu, and Pathak had used this same engine to redesign New York City and Boston school choice (NBER w11965).

A.2 School choice — Boston, DA, TTC

Three mechanisms compete for school-choice problems:

• Boston / Immediate Acceptance: Each round, students apply to their top remaining choice; schools admit up to capacity and rejected students fall to the next round. Easy to explain; not strategy-proof — sophisticated parents game it by lying about top choices to avoid losing a guaranteed seat at a safety school.
• Deferred Acceptance (DA): Strategy-proof for students; stable but not Pareto-efficient with respect to student preferences. The dominant choice now in many cities including Boston (after 2005 redesign), New York, and Chicago.
• Top Trading Cycles (TTC): Strategy-proof and Pareto-efficient, but not stable — students with high priority at school $s$ may be assigned to school $s'$ they prefer less than other students who got $s$.

The Boston School Committee's 2005 decision to switch from the Boston mechanism to DA is the cleanest market-design case study of how fairness and strategy-proofness compete. The argument that swung the vote was a fairness-rationale-for-strategy-proofness: when sophisticated parents game the Boston mechanism, unsophisticated parents bear the cost. Strategy-proofness equalises this strategic burden (Abdulkadiroğlu, Pathak, Roth, NBER 2006).

TTC retains adherents because Pareto efficiency is normatively attractive when priorities are exogenous (e.g. random lotteries with no fairness claim). New characterisations and weighted variants continue to appear (JET 2024).

Study next: weighted TTC, equitable TTC, and the literature on integration of priorities with multiple objectives.

A.3 Kidney exchange

The kidney-exchange algorithm (Roth, Sönmez, Ünver 2004 QJE) takes incompatible patient-donor pairs and finds cycles in a directed graph so each donor gives to a patient in another pair. Two-cycles (you donate to my patient; I donate to yours) are easiest; three-cycles add complexity but materially raise match rates; chains triggered by altruistic non-directed donors avoid the cycle constraint entirely.

The growth path:

• 2004: theoretical paper, NEPKE (New England Program for Kidney Exchange) launches
• 2007–10: chains and the National Kidney Registry scale nationally
• 2010s: global exchange becomes plausible
• 2020s: maintenance challenges — the patient-donor pool composition shifts as easier pairs match, leaving harder cases (Roth 2023, NBER w31947)

Roth's 2023 paper on "Market Design and Maintenance" argues that designed marketplaces, like designed bridges, need maintenance protocols because the underlying environment changes (here, the marginal patient becomes harder to match as easy matches are consumed, requiring algorithmic and institutional adjustment).

Live debate: whether monetary compensation should be allowed at the margin — Becker-Elias-style proposals vs the "repugnant transactions" prohibition. Roth himself has argued that some markets are repugnant for reasons economists need to take seriously (Roth, JEP 2007).

A.4 Refugee matching

Delacrétaz, Kominers, and Teytelboym (AER 2023, vol. 113(10)) formalised refugee resettlement as a matching-with-multidimensional-knapsack-constraints problem: each family has a "size" in multiple dimensions (number of children, schooling needs, healthcare needs, language services), and each community has capacity vectors that must be respected. Their mechanism integrates refugee preferences with community priorities and outperforms first-come-first-served allocation in simulations on US resettlement data.

Andersson (Economic Journal 2024) extended the literature with a constrained-priority mechanism design approach. Sweden has piloted such mechanisms in regional resettlement.

A.5 Auction design and the FCC incentive auction

Milgrom and Wilson's 2020 Nobel cited their work on auction theory and the design of auctions for new formats (Nobel Popular Information 2020). The two key innovations:

• Simultaneous Multiple Round Auction (SMRA): licences are auctioned simultaneously in rounds, with bidders able to switch across licences within a round. Introduced for the FCC's 1994–95 PCS spectrum auctions and used worldwide.
• Deferred-Acceptance Clock Auction: the engine of the 2017 FCC incentive auction. A reverse clock auction first bought back over-the-air TV spectrum from broadcasters at falling prices (each broadcaster's clock falls until they accept the current bid or drop out), then a forward clock auction sold the cleared spectrum to wireless carriers.

The 2017 incentive auction repurposed 84 MHz of spectrum and raised about $20 billion, with about $10 billion net to the US Treasury after broadcaster payments (IIC International Media 2018). Critics argued the proceeds underperformed projections, but Milgrom's response — that the auction reassigned exactly what the market valued, no more — is the right one in mechanism-design terms (Promarket 2020).

The mechanism design lesson is general: with complementarities (a wireless carrier wants a contiguous block of MHz across a region), package bidding is needed to avoid the exposure problem (a bidder who needs both A and B may win only A and overpay). The deferred-acceptance clock auction approximates an efficient outcome while preserving strategy-proofness for non-pivotal bidders.

A.6 Combinatorial and package auctions

Combinatorial auctions allow bidders to bid on bundles. The two big classes are:

• VCG auctions (Vickrey-Clarke-Groves): theoretically beautiful — truth-telling is a dominant strategy, the efficient allocation is implemented, transfers internalise externalities — but practically fragile (revenue can be negative, computational complexity is exponential, and it's vulnerable to collusion).
• Combinatorial Clock Auctions (CCA): used in UK and several EU spectrum sales; iterate ascending prices on packages while monitoring activity.

Palacios-Huerta, Parkes, and Steinberg's 2024 JEL review (JEL June 2024) is now the standard reference on "Combinatorial Auctions in Practice." Their main empirical finding: most of the welfare loss in real combinatorial auctions comes from bidders' choice of which packages to bid on, not from auction rules.

Study next: CCA's "second-stage pricing" and supplementary round mechanics; the deferred-acceptance clock auction's strategy-proofness boundary.

B. Auction theory at the frontier

B.1 First-price vs second-price — the discrete algebra

Consider two bidders, Alice and Bob, with private values $v_A$ and $v_B$ each drawn independently and uniformly from ${1, 2, 3}$ dollars.

Second-price auction: bidders submit sealed bids; the highest wins and pays the second-highest. Truth-telling is dominant. Why? Suppose $v_A = 3$. If Alice bids 3, she wins whenever Bob bids 1 or 2, paying Bob's bid. If she bid higher (say 4 — irrelevant, since 3 is max value but the logic generalises), nothing changes when she would have won. If she bid 2, she might lose to Bob's bid of 3 — but in that case Bob's value is also 3, and Alice would have paid 3 and earned 0 surplus anyway. Bidding her value can never hurt and can sometimes help. Revenue: Alice and Bob each truth-tell. The seller's revenue is the minimum of the two values. For uniform integer values on ${1,2,3}$, this expected minimum is $\mathbb{E}[\min(v_A, v_B)] = \tfrac{1\cdot5 + 2\cdot3 + 3\cdot1}{9} = \tfrac{14}{9} \approx 1.56$.

First-price auction: highest bid wins and pays own bid. Truth-telling is dominated — you must shade your bid. With two symmetric bidders and uniform values, equilibrium bidding (with continuous types) is $b(v) = v/2$. In discrete settings, you compute best-responses round by round. For finite values ${1,2,3}$, the symmetric equilibrium has each bidder bid roughly half their value. Revenue equivalence theorem (Myerson 1981, Riley-Samuelson 1981) states that in symmetric independent private-value (IPV) settings, any auction that (i) assigns the object to the bidder with the highest value and (ii) gives the lowest type zero expected surplus generates the same expected revenue. So the first-price and second-price both raise $\approx 1.56$ in our example.

The proof sketch with discrete bidders:

1. Both auctions are efficient — they assign to the highest-value bidder.
2. In both, the lowest type ($v=1$) earns zero expected surplus.
3. Therefore, by a payoff-equivalence argument (the expected payment of a type-$v$ bidder is determined by the allocation rule plus the zero-rent condition), expected revenue is the same.

Study next: the full continuous proof via integration of envelope conditions. The Wikipedia presentation of revenue equivalence is a decent starting point; Krishna's Auction Theory is the standard textbook.

B.2 Generalized Second Price (GSP) and Google ads

GSP — used by Google for sponsored search from 2002 to roughly 2019 — works as follows: bidders submit bids on a keyword; the $k$th slot goes to the $k$th-highest bidder, who pays the $(k+1)$th-highest bid times a quality-adjusted score. The Edelman, Ostrovsky, Schwarz (2007) and Varian (2007) papers (Edelman et al. AER 2007) showed that GSP is not truth-telling, but has a "locally envy-free" equilibrium that matches a VCG outcome.

GSP fell out of favour at Google for online display in 2019. The reason was header bidding: publishers were running auctions before Google's auction, and the upstream price affected the optimal Google bid. In this chain of auctions, the second-price mechanism inside Google no longer had clean incentive properties. Google migrated Google Ad Manager (formerly DoubleClick) to a first-price auction in September 2019 (AdExchanger). A Digiday survey found 78% of publishers reported that first-price helped them maximise ad revenue (Adpushup).

The first-price migration immediately re-introduced strategic shading: advertisers and DSPs deployed bid shading algorithms that use historical data and ML to bid below their value but above the second-highest bid (Clickio). Prices spiked, then converged to roughly a midpoint between second-price and naive first-price equilibrium.

Randomized Generalized Second Price (RGSP) is Google's current production answer for search ads in some contexts — it randomises tie-breaking and slot allocation to reduce manipulability (Search Engine Land 2023).

The theoretical implication is that revenue equivalence fails once auctions are chained or sequential, and once bidders have asymmetric information about competitors' bids. The 2024 frontier on online advertising auctions is increasingly about information design (what the platform discloses to bidders) and the relationship to upstream and downstream auction layers.

B.3 Interdependent values and the winner's curse

Wilson's 1977 paper formalised the winner's curse: in common-value auctions (oil tract leases, IPO allocations), the winner is by definition the most optimistic bidder, and is therefore on average too optimistic. Rational bidders must "bid down" by the expected over-optimism conditional on winning.

Milgrom and Weber's 1982 Econometrica "linkage principle" showed that when values are affiliated (positively correlated), revealing more public information about value raises expected revenue, because it reduces the rent that bidders extract from informational asymmetry. This is one reason English auctions tend to dominate sealed-bid auctions empirically — the price path reveals information as it ascends.

The frontier is "interdependent values" auctions, where each bidder's value depends on others' signals — research on optimal mechanisms here is the most active corner of pure auction theory ([Athey-Levin, Bergemann et al.]).

Study next: Bergemann-Morris "robust mechanism design"; Brusco-Lopomo-Marx; auction-theoretic treatment of recommender systems.

C. Behavioral economics: triage after the replication crisis

C.1 What survived

A careful audit of the behavioral canon — from Camerer, DellaVigna, and Linos's work, the Many Labs replication projects, the Open Science Collaboration — finds the following effects robustly replicate:

• Loss aversion in the loss domain: people weight losses approximately 2x more than equivalent gains. Confirmed in many settings. Less robust in mixed gambles where the reference point is unclear.
• Default effects: organ-donor opt-in vs opt-out countries produce enormous differences in donor rates (Johnson-Goldstein 2003 Science; replicated extensively). Retirement auto-enrolment (Madrian-Shea 2001; Beshears-Choi-Laibson-Madrian) increases participation rates by 30–50 percentage points.
• Present bias / hyperbolic discounting: people consistently choose smaller-sooner over larger-later when the smaller-sooner is immediate, but reverse the choice when both rewards are delayed. Implications for retirement, exercise, savings.
• Mental accounting in narrow framings: people treat gift-card income differently from earned income; segregate house-money gains; show framing effects in consumption.
• Reciprocity and social norms in field experiments: the UK's HMRC tax letter trials by the Behavioural Insights Team (BIT) consistently increased on-time payment rates by 1–5 percentage points using injunctive-norm wording.

C.2 What did not

• Ego depletion: the original Baumeister-Vohs claim that self-control draws on a finite resource that depletes did not replicate in pre-registered multi-lab studies (Hagger et al. 2016, Perspectives on Psychological Science).
• Power posing: Carney-Cuddy-Yap 2010 Psychological Science effects did not replicate in Ranehill et al. 2015 or subsequent multi-lab efforts. Carney herself disavowed the result.
• Priming: classical social-cognition priming studies (Bargh's "elderly walking slowly" effect) failed multiple direct replications.
• Honesty research: Lisa Shu, Nina Mazar, Francesca Gino, Dan Ariely, Max Bazerman's 2012 PNAS paper claiming signature-at-top reduces dishonesty was retracted in 2021 after Data Colada showed signs of fabricated data (Data Colada). The Harvard Business School's investigation of Francesca Gino concluded she "committed research misconduct intentionally, knowingly, or recklessly," with three additional papers retracted in 2023 (Science 2024). Gino's $25 million defamation suit against the Data Colada bloggers was dismissed.

C.3 Nudges at policy scale

Stefano DellaVigna and Elizabeth Linos's 2022 PNAS paper (DellaVigna-Linos PNAS 2022) compared nudge effect sizes in in-government tests run by the US's Office of Evaluation Sciences and the UK's Behavioural Insights Team (165 trials, 24 million observations) against nudges in published academic papers. The headline:

• Academic publication-bias-corrected effect size: ~8.7% increase in target behaviour
• In-government BIT/OES trials: ~1.4% increase in target behaviour

The 6x gap reflects (i) publication bias in academic outlets (only positive results published) and (ii) the difficulty of scaling nudges that work in tightly controlled experiments to government implementations. The implication is sober but not catastrophic — nudges at 1.4% are still typically cheap enough to be cost-effective, but they are not the silver bullet some early enthusiasts portrayed.

Sunstein himself wrote in 2024 that "we are no longer in the era of the easy nudge" (Hoover Institution 2024) and called for "behavioural governance systems that adapt, learn, and respect user autonomy."

The OECD maintains an active Network of Behavioural Insights Practitioners; over 200 government units worldwide now have BI capacity (OECD 2017, updated continuously).

Live debate: are nudges politically legitimate? Sunstein argues yes, on autonomy grounds; critics argue paternalism is unavoidable; behavioral welfare economics (Bernheim, Rangel) tries to formalise welfare when revealed preferences are unreliable.

Study next: Bernheim-Rangel (2009) framework for behavioral welfare; the choice-architecture literature; recent work on boost approaches (Hertwig, Grüne-Yanoff) that build skills instead of nudging choices.

C.4 Heterogeneous treatment effects of nudges

A theme in the post-replication literature is that nudges have substantial heterogeneity in effects across populations. The same default that works for retirement saving may fail for organ donation in countries with strong religious objection. The same social-norm letter that works in suburban Britain may backfire in rural Italy. Heterogeneous treatment effects estimation (Athey-Imbens, causal forests, BCF) is now a routine companion to nudge field trials.

D. Industrial organization, platforms, and digital markets

D.1 Two-sided markets — Rochet-Tirole foundations

A two-sided (or multi-sided) market has at least two distinct user groups whose participation creates value for each other. Credit cards: cardholders are more useful when more merchants accept them; merchants pay more for terminals when more consumers carry the card. Operating systems: developers write apps where users live; users buy platforms where apps live. Dating apps, ride-share, online ads, video games, search engines — most modern digital businesses are two-sided.

Rochet and Tirole (2003 JEEA; 2006 RAND) gave the formal definition: a market is two-sided if the price structure matters beyond the price level. That is, charging side A €1 and side B €0 produces different volume and welfare than charging side A €0 and side B €1, even when total revenue is the same. This is a result of cross-side network externalities: the more cardholders, the more merchants want to participate, and vice versa. The platform's pricing problem is to allocate the burden across sides to maximise total participation.

Three classical results:

• The price on side A reflects not just side A's elasticity but side B's marginal externality on A's value.
• The side with greater elasticity often gets subsidised (sometimes to zero, sometimes negative — the "money side" and the "subsidy side").
• Competition between platforms can reduce welfare if it leads to "tipping" where one platform wins all, or if it intensifies multi-homing inefficiently.

The Belleflamme-Peitz 2024 review (CRC TR224 DP 584) is the canonical modern survey. It extends Rochet-Tirole to multi-homing (users on multiple platforms simultaneously), platform leakage (cross-side spillovers across competing platforms), and platform-mediated search.

D.2 Tipping, multi-homing, and network effects empirics

The empirical literature on network effects in digital platforms has matured. Key findings:

• Tipping: a market with strong same-side or cross-side network effects tends to converge to one platform; the conditions under which this happens are now well-understood (Caillaud-Jullien 2003; Hagiu-Wright 2015). Tipping is not automatic — multi-homing on at least one side, differentiated tastes, or platform interoperability can prevent it.
• Multi-homing: empirically, multi-homing is common on the user side of many platforms (consumers use both Lyft and Uber, both Visa and Mastercard) and rarer on the developer side (a game studio often picks PS5 or Xbox). The pricing implication, formalised in Armstrong (2006), is that the multi-homing side typically pays more.
• Identification challenge: distinguishing network effects from correlated unobservables is hard. Modern work uses (i) instruments for platform adoption (cohort-based shocks, regulatory variation); (ii) structural BLP-style demand estimation with explicit network terms; (iii) natural experiments (platform shutdowns).

D.3 Big Tech antitrust 2024–2026

This is the live policy and case-law area. The four major US actions:

United States v. Google (search) — Judge Mehta, D.D.C.

On August 5, 2024, Judge Amit Mehta found Google liable for monopolisation under Sherman Act §2 (White & Case 2024). Key factual findings:

• Google held ~90% market share in PC search and ~95% in mobile.
• Google paid Apple about $20 billion per year (revealed at trial) to be the default search engine in Safari, plus comparable amounts to Samsung, Mozilla, and others.
• These exclusive default agreements foreclosed substantial competition, denying rivals (Bing, DuckDuckGo) the scale needed to improve search quality.

The remedies decision came in September 2025: the court rejected DOJ's request to force divestiture of Chrome but ordered Google to share certain search data with rivals and prohibited some forms of exclusive default deals (CNBC December 2025). This is a meaningfully less aggressive remedy than the Microsoft 2001 settlement.

United States v. Google (ad tech) — E.D. Va.

In April 2025, a separate court found Google had monopolised parts of the digital advertising stack (publisher ad servers and ad exchanges). Remedies pending.

FTC v. Amazon — W.D. Wash., Judge Chun

Filed September 26, 2023 with 17 state AGs (FTC Press Release 2023). The complaint alleges:

• Amazon's "anti-discounting" mechanisms (downranking sellers who offer lower prices elsewhere) suppressed price competition.
• The Buy Box algorithm penalises sellers for off-Amazon discounting.
• The "Project Nessie" pricing algorithm raised prices on certain products and tested whether competitors followed.
• Tying Prime eligibility to use of Fulfillment by Amazon (FBA) extends market power into adjacent markets.

In October 2024, Judge Chun allowed most claims to proceed past Amazon's motion to dismiss (GeekWire). Trial scheduled for 2026.

FTC v. Meta — D.D.C., Judge Boasberg

Filed December 2020, alleging that Meta's 2012 Instagram and 2014 WhatsApp acquisitions were anti-competitive. Trial began April 14, 2025 (CNBC April 2025). On November 18, 2025, Judge Boasberg ruled for Meta (NPR November 2025). The decisive issue was market definition: the FTC argued for a narrow "personal social networking services" market (Facebook, Instagram, Snapchat); the court held that under modern conditions TikTok and YouTube must be included, and under the expanded market Meta lacked monopoly power.

The Meta loss is methodologically significant — it suggests courts will continue to define digital markets broadly when substitution is observed at the user behaviour level, even if the substitution is across different "kinds" of platforms.

D.4 Kill zones, killer acquisitions, and innovation

The "kill zone" thesis (Kamepalli-Rajan-Zingales 2020) argues that the threat of acquisition by a dominant platform discourages early-stage investment in companies that would otherwise compete. The intuition: VCs anticipate that any successful entrant will be bought (or worse, copied and crushed) by the incumbent, capping the upside and depressing ex-ante funding.

The empirical evidence is contested:

• Kamepalli-Rajan-Zingales found post-acquisition declines in VC funding for adjacent startups.
• AEI (2020 report) and CGO (Welcome to the kill zone?) found that overall VC activity grew dramatically through the 2010s, particularly in tech.
• Cunningham-Ederer-Ma (2021 JPE) studied "killer acquisitions" in pharmaceuticals and found 5–7% of pharma acquisitions are intended to discontinue overlapping drug projects.

The 2024 EU Digital Markets Act (DMA) and parallel UK Competition and Markets Authority interventions are partly aimed at this concern, by requiring notification of acquisitions below traditional thresholds when made by designated "gatekeepers."

D.5 Attention markets, recommender distortions, and ad markets

The digital advertising market exceeds $700 billion globally by 2025 estimate, with Google and Meta together accounting for over half. The economic literature has moved from generic "advertising-as-information" or "advertising-as-persuasion" models to:

• Recommender system distortions: engagement-maximising recommendation creates a misalignment between user welfare (satisfaction with content over time) and platform revenue (clicks now). Allcott, Gentzkow, and others have measured the welfare cost of Facebook deactivation in field experiments — finding $100+ monthly compensating valuations for staying on the platform, but also self-reported wellbeing improvements from deactivation.
• Rich-get-richer dynamics: algorithmic discovery concentrates attention on a small number of items. Empirical work on Patreon, Spotify, YouTube creator earnings shows strong Pareto distributions in earnings, often steeper than would be implied by underlying quality variance.
• Information design by platforms: platforms choose what information to surface to users (and to bidders in ad auctions). Bergemann-Bonatti work on monopoly information design is the formal frame.

Live debate: Are platform recommender systems creating welfare-reducing distortions, or simply efficiently matching attention to content? The Bursztyn-Handel-Jiménez-Durán-Roth (2024) field experiment on TikTok and Instagram found students would pay to have everyone deactivate the platform but not to deactivate themselves — a coordination failure model of attention.

E. The credibility revolution: modern causal inference

E.1 Randomized controlled trials and the LATE framework

The Imbens-Angrist 1994 Econometrica paper introduced the local average treatment effect (LATE): when an instrument $Z$ shifts treatment $D$ but not directly the outcome $Y$, the Wald ratio $\frac{\text{Cov}(Y,Z)}{\text{Cov}(D,Z)}$ identifies the average treatment effect for those who would change treatment status in response to $Z$ (the "compliers"), under the monotonicity assumption that $Z$ moves $D$ in the same direction for everyone.

The 2021 Nobel for Imbens, Angrist, and Card recognised this framework's reshaping of empirical economics. Three implications:

• Different instruments identify different LATEs (the always-takers, never-takers, compliers, defiers are distinct subpopulations).
• IV is not a tool for estimating the ATE under general heterogeneity; it estimates the LATE.
• Marginal treatment effects (Heckman-Vytlacil 2005) generalise: with a continuous instrument, you can trace out the entire distribution of treatment effects across compliance margins.

E.2 Difference-in-differences — the 2020s overhaul

The classical 2×2 DiD identity: with treatment group $T$, control group $C$, pre-period 0, and post-period 1, the DiD estimator is

$$\widehat{\delta} = (\bar{Y}{T,1} - \bar{Y}{T,0}) - (\bar{Y}{C,1} - \bar{Y}{C,0})$$

Under parallel trends (the counterfactual trend in $T$ absent treatment equals the observed trend in $C$), this identifies the average treatment effect on the treated (ATT) in the post-period.

In practice, researchers use two-way fixed effects (TWFE) regressions to handle multi-period, multi-unit, staggered adoption settings:

$$Y_{it} = \alpha_i + \gamma_t + \beta \cdot D_{it} + \epsilon_{it}$$

where $D_{it}$ is treatment status. Goodman-Bacon's 2021 JoE paper proved that this TWFE estimator is a weighted average of all possible 2×2 DiDs comparing groups treated earlier vs. later. The problem: when treatment effects evolve over time (e.g. ramping up or dynamic), "already-treated units" act as bad controls for "newly-treated units," and the weights can be negative. In pathological cases, all underlying treatment effects can be positive yet the TWFE coefficient is negative.

The new estimator menu:

• Callaway-Sant'Anna (2021 JoE): estimates group-time average treatment effects ATT(g,t) for each cohort-period pair, then aggregates with chosen weights. Uses never-treated or not-yet-treated as the comparison group, avoiding the bad-control problem. Supports outcome regression, IPW, or doubly-robust estimands.
• Sun-Abraham (2021 JoE): interaction-weighted estimator that explicitly weights cohort-relative-time effects.
• de Chaisemartin–D'Haultfœuille (2020 AER; 2023 EJ survey): estimator robust to heterogeneous effects, with the survey paper a key reference. Their 2024 paper extends to designs without never-treated stayers (2025 working paper).
• Borusyak-Jaravel-Spiess (2024 ReStud): imputation estimator. Fit unit and time fixed effects on untreated observations only; impute counterfactuals for treated obs; treatment effect = observed − imputed. Efficient under homoskedasticity. (ReStud 2024).

The 2025 Callaway-Cunningham-Goodman-Bacon-Sant'Anna practitioner's guide is the standard reference now.

Live debate: pre-trend testing. The Roth (2022) AER:Insights paper showed that conventional pre-trend tests have low power against meaningful violations, and the Rambachan-Roth (2023 ReStud) honest inference framework allows researchers to bound treatment effects under "approximately parallel trends" rather than assume parallelism exactly.

E.3 Event studies

An event study generalises DiD by tracking dynamic effects $\beta_k$ for periods $k$ relative to treatment onset. The classical specification:

$$Y_{it} = \alpha_i + \gamma_t + \sum_{k \neq -1} \beta_k \cdot 1{t - g_i = k} + \epsilon_{it}$$

where $g_i$ is the treatment-start period for unit $i$ and $k=-1$ is the omitted period. Under staggered adoption with heterogeneous effects, the same negative-weights problem afflicts these estimates — sometimes called the "spurious pre-trend" phenomenon, where pre-treatment leads appear non-zero just because of TWFE contamination.

The fix is the same: use Callaway-Sant'Anna, BJS imputation, or de Chaisemartin–D'Haultfœuille. Most stata users now run did_imputation (Borusyak) or csdid (Callaway-Sant'Anna).

E.4 Regression discontinuity

RD identifies treatment effects from the discontinuity in conditional expectations at a cutoff. If treatment $D = 1{X \geq c}$ for some running variable $X$ with cutoff $c$, then under continuity of the conditional outcome means $E[Y(0)|X]$ and $E[Y(1)|X]$ at $c$:

$$\tau = \lim_{x \downarrow c} E[Y|X=x] - \lim_{x \uparrow c} E[Y|X=x]$$

Algebraic intuition: compare students just above and just below an admissions threshold; compare counties just above and just below a vote-share cutoff for an electoral result; compare drinkers just above and just below 21.

The 2010s Calonico-Cattaneo-Titiunik (2014 Econometrica) rdrobust package made bias-corrected, robust inference standard. Their 2024 Cambridge textbook (Cattaneo-Idrobo-Titiunik) is the practitioner's bible. The frontier:

• Bunching estimators: when running variables are manipulable (e.g. self-reported income at notches in tax schedules), the bunching mass at the threshold identifies the elasticity of behavioural response (Saez 2010; Kleven-Waseem 2013).
• Treatment effect heterogeneity in RD (Calonico-Cattaneo-Farrell-Palomba-Titiunik 2025): conditional ATEs as functions of covariates.

Study next: the local-linear vs local-quadratic bias tradeoff; covariate adjustment in RD; geographic RD; the "RD as quasi-experiment" reading.

E.5 Synthetic control

Abadie-Diamond-Hainmueller (2010 JASA; 2015 AJPS) developed synthetic control as a credible alternative to DiD when one unit (e.g. one state, country, school) receives treatment and many other units are available as comparisons. The method constructs a weighted average of comparison units whose pre-period outcome trajectory best matches the treated unit; treatment effect = post-period (treated − synthetic).

The 2020s extensions:

• Augmented Synthetic Control (Ben-Michael-Feller-Rothstein 2021): combines synthetic control weights with an outcome model to address bias from poor pre-treatment fit.
• Penalized Synthetic Control (Abadie-L'Hour 2021 JASA): smooths between one-to-one matching and SCM weights.
• Spatially augmented SCM (2024): incorporates spatial autocorrelation in outcome.

Study next: the matrix completion synthesis (Athey-Bayati-Doudchenko-Imbens-Khosravi 2021); generalized synthetic control (Xu 2017); synthetic DiD (Arkhangelsky et al. 2021).

E.6 Instrumental variables — beyond LATE

Weak-instrument concerns (Stock-Yogo, Andrews-Stock-Sun) and the "many weak instruments" problem (Bekker, Hansen-Hausman-Newey) reshape applied IV. The current best practice: report the Effective F-statistic (Olea-Pflueger), use Anderson-Rubin confidence intervals when the first stage is weak, and disclose first-stage robustness.

Local IV / marginal treatment effects (Heckman-Vytlacil 2005, 2007) extend LATE to identify the full distribution of treatment effects, given a continuous instrument and the monotonicity assumption.

Shift-share / Bartik IVs: Goldsmith-Pinkham-Sorkin-Swift (2020 AER) showed that the validity of a Bartik instrument rests on the exogeneity of the shares, not the shifts. Adão-Kolesár-Morales (2019 QJE) provided inference adjustments when both shares and shifts contribute identifying variation.

F. Structural estimation revival

F.1 Sufficient statistics

Raj Chetty's 2009 AR Econ article (Annual Reviews) crystallised the sufficient statistics approach. The premise: for many welfare questions, the policy-relevant quantity is a function of a small number of estimable behavioural elasticities, not the full structural primitives of a model. You can estimate the elasticities credibly with reduced-form methods, plug them into the formula, and get a policy answer.

Examples:

• Optimal income tax (Saez 2001 ReStud): at the top, the optimal marginal rate $t^$ is a function of the elasticity of taxable income $e$ and the Pareto parameter of the income distribution $a$. The formula $t^ = \frac{1}{1 + a \cdot e}$ (when the social value of marginal top consumption is zero) is sufficient — you don't need to model anyone's utility function.
• Optimal social insurance (Chetty 2006): the welfare-maximising UI replacement rate depends on the elasticity of unemployment duration with respect to benefits and the coefficient of relative risk aversion.
• Behavioural welfare (Mullainathan-Schwartzstein-Congdon): when consumers are biased, welfare formulas adjust by adding correction terms involving the gap between revealed and "true" preferences.

The sufficient-statistics approach is the bridge between the credibility revolution and structural economics. It accepts that you cannot identify deep parameters, but argues that for some welfare questions you don't need to. Kleven (2020 AR Econ "Sufficient Statistics Revisited") is the modern critical assessment.

F.2 Dynamic discrete choice — Rust and successors

John Rust's 1987 Econometrica paper on bus engine replacement (Harold Zurcher, the maintenance manager of the Madison, WI bus company) launched dynamic discrete choice estimation. The setup: each period the agent chooses whether to replace (incurring a fixed cost) or keep maintaining (with rising costs as the engine ages); the optimal stopping rule is characterised by a Bellman equation, which the econometrician inverts to recover the cost and discount parameters.

The 2024 frontier:

• Hotz-Miller-style two-step estimators for high-dimensional state spaces
• Machine-learning-aided value function approximation (deep DDC)
• Applications to firm decisions, marriage and divorce, occupation choice, retirement timing.

F.3 BLP and demand estimation for differentiated products

Berry, Levinsohn, and Pakes (1995 Econometrica) — universally "BLP" — estimates demand for differentiated products using market-share data, instruments for endogenous prices, and a random-coefficients logit model. The output is a system of elasticities that lets the analyst simulate counterfactual mergers, tax changes, or product entry.

The PyBLP package by Conlon and Gortmaker has become the de facto standard. Their 2024 Journal of Econometrics paper (Conlon-Gortmaker JoE 2024) extends best practices for incorporating micro moments — individual-level survey or admin data — into BLP estimation, substantially improving finite-sample performance.

Applications: merger simulation (the FTC's work in the FTC v. Sysco/US Foods case, the Whole Foods case), retail demand, pharmaceutical pricing, and increasingly digital platform behavior. The 2024 frontier extends BLP to two-sided markets (Lee 2013; Fan-Yang 2020) and to dynamic settings.

F.4 The Heckman-Imbens debate (truce)

Through the 2000s and 2010s, Heckman (structural) and Imbens (reduced-form) disagreed publicly about empirical method. Heckman's complaint: LATE is "policy-irrelevant" because policy interventions don't typically operate on the same compliance margin as historical instruments. Imbens' complaint: structural models impose untestable parametric restrictions and obscure identification.

The 2020s truce: the disciplines have converged. Modern structural papers (BLP, dynamic discrete choice) report identification strategies explicitly; modern reduced-form papers (sufficient statistics, marginal treatment effects, identification at infinity) close some of the welfare-extrapolation gap. The PhD curriculum increasingly teaches both.

G. Labour economics

G.1 The monopsony renaissance

For decades, the textbook labour market was competitive: many employers, each facing perfectly elastic labour supply, paying wage = marginal product. Card and Krueger's 1994 AER minimum-wage paper ("Myth and Measurement") challenged this empirically — minimum wage increases in New Jersey did not destroy fast-food jobs. The theoretical apparatus to explain why was monopsony.

The 2020s reformulation has three pillars:

1. Search frictions (Mortensen, Pissarides, Diamond — 2010 Nobel): workers don't see all job openings instantly; firms post wages knowing workers arrive stochastically. The labour supply curve to a firm has finite elasticity.
2. Job differentiation (Card-Cardoso-Heining-Kline; Lamadon-Mogstad-Setzler): jobs differ in non-wage amenities (commute, hours, culture); workers have heterogeneous preferences over these.
3. Concentration (Azar-Marinescu-Steinbaum, Benmelech-Bergman-Kim): in many local labour markets, a small number of employers account for most posted vacancies. Higher concentration → more wage suppression.

The Azar-Marinescu 2024 Annual Review of Economics (AR Econ 2024) synthesises the evidence. The headline numbers:

• Markdown estimates (wage as fraction of MRPL): 50–85%, implying wages would rise 15–50% if monopsony power were eliminated.
• Concentration is highest in rural areas, low-skill occupations, and certain professions (nursing, K-12 teaching).
• Concentration measures correlate negatively with wages at the local-occupation level (Azar-Marinescu-Steinbaum 2022 JHR).

G.2 Minimum wage: the Cengiz-Dube-Lindner-Zipperer bunching estimator

The Cengiz-Dube-Lindner-Zipperer (2019 QJE) bunching estimator addressed a measurement problem: when a minimum wage rises, jobs paying below the new floor "disappear" — either through job loss (bad) or through workers being paid more (good). Conventional employment regressions can't distinguish these.

The bunching approach uses the distribution of wages directly. Before the increase, you can see jobs paying below the new minimum. After the increase, those jobs should either (i) disappear (jobs lost) or (ii) move to the new minimum or just above (jobs reorganised). By counting the excess mass at the new minimum vs the missing mass below it, you can decompose effects.

The headline finding: the excess mass approximately equals the missing mass. Total low-wage employment is roughly unchanged after typical state minimum wage increases. Confidence intervals rule out total employment elasticities below −0.06. Effect estimates extend through Dube's 2024 NBER review ("Minimum Wages in the 21st Century" NBER w32878).

Nuances:

• Effects in tradable sectors are more negative (consistent with elastic labour demand).
• Effects in more-concentrated markets are less negative (consistent with monopsony — minimum wage moves toward competitive wage).
• Very high minimum wages (Seattle 2015–17 at $15) may have non-negligible negative effects ([UW evaluation 2017; offset by Reich et al.]).

G.3 Noncompetes and the FTC rule

The economic literature on noncompetes (Starr, Prescott, Bishara; Lipsitz-Starr; Marx-Strumsky-Fleming) found:

• Roughly 18% of US workers are bound by noncompete agreements, including many low-wage workers (security guards, hair stylists) where the firm-specific human capital story is implausible.
• States with stricter enforceability of noncompetes show lower wages, less mobility, and reduced startup formation in affected industries (Starr-Prescott-Bishara 2021 JLE).
• The 2008 Michigan policy change to enforce noncompetes is associated with reduced inventor mobility and worse innovation outcomes (Marx-Strumsky-Fleming 2009).

The FTC's April 23 2024 rule (3-2 vote) would have banned new noncompetes and voided most existing ones nationwide (FTC Final Rule 2024). The FTC's economic analysis projected:

• +$524/year wages for the average worker
• 8,500 additional new firms annually
• 17,000–29,000 additional patents per year over 10 years
• $194B in healthcare cost savings over 10 years (noncompetes restrict mobility of healthcare workers, raising prices)

On August 20 2024, Judge Ada Brown (N.D. Tex.) issued a nationwide injunction holding the FTC lacked statutory authority to issue substantive competition rules. The FTC dismissed its appeal in September 2025 (DLA Piper). The economic evidence stands; the federal regulatory tool does not, leaving the field to state-by-state action (California, North Dakota, Oklahoma already ban; Minnesota, Colorado tightening).

G.4 Automation, AI, and the task framework

Acemoglu and Restrepo's task framework (2018 AER; 2019 JEP; 2020 AER) replaces the old skill-biased technical change story with one based on tasks. Production requires bundles of tasks; technology automates some tasks and creates new ones. The labour share rises when new tasks outpace automation; it falls when automation outpaces.

Their 2024 NBER paper "The Simple Macroeconomics of AI" (Acemoglu 2024) applies the framework to AI. Their estimate: AI is unlikely to cause large wage declines or productivity surges within a decade. The hard task to automate is not the technical capability but the integration into production processes.

David Autor's complementary work (Autor MIT lectures 2024) documents the end of barbell polarization: the 2000s pattern in which middle-wage jobs declined relative to both high-wage (cognitive) and low-wage (manual service) jobs has shifted to a one-sided pattern favouring well-compensated, high-training jobs. STEM employment share grew from 6.5% in 2010 to nearly 10% in 2024.

The early micro evidence on generative AI in the workplace (Brynjolfsson-Li-Raymond 2023 on call centre agents; Noy-Zhang 2023 on writing tasks; Peng-Kalliamvakou-Cihon-Demirer 2023 on Copilot for coders) shows productivity gains of 14–56%, with bigger gains for lower-skill workers (compression, not amplification of inequality). Whether this scales is the central open question.

G.5 Gig work and the classification debate

The gig economy raises a foundational question: are Uber drivers employees or independent contractors? The economic stakes:

• Independent contractors fund their own benefits (no employer share of payroll taxes, no health insurance, no unemployment insurance, no workers' compensation).
• In Massachusetts alone, Uber and Lyft avoided ~$47 million in employer-side payroll taxes in 2023 by classifying drivers as contractors (Slate 2024).
• Driver earnings (after expenses): 2024 weekly average ~$513 (Uber) and ~$318 (Lyft) — Uber driver hourly equivalent is roughly the 10th percentile of W-2 workers, below the legal minimum wage in 13 of 20 major US cities.

The 2024–2025 regulatory landscape:

• California Supreme Court upheld Proposition 22 (which exempts gig drivers from AB5's employee classification) in July 2024.
• US Department of Labor finalised an FLSA classification rule (effective March 2025) that returned to a more worker-friendly multi-factor test.
• The EU passed a Platform Workers Directive in 2024 establishing a presumption of employment for platform workers under certain conditions.

The economic question: are the contracting savings to platforms passed to consumers (lower prices, more service) or captured as platform rents? The Berkeley Labor Center 2024 study and Reich-Parrott work suggest most of the cost shift accrues to drivers rather than consumers — drivers earn below minimum wage on net because the surplus from contractor classification doesn't trickle down.

H. Public economics

H.1 Optimal income taxation

The Mirrleesian framework (Mirrlees 1971 ReStud; Saez 2001 ReStud) characterises the optimal nonlinear income tax. The Saez (2001) formula for the optimal top rate is:

$$t^* = \frac{1 - g}{1 - g + a \cdot e}$$

where $g$ is the social marginal welfare weight on top earners (often set to 0), $a$ is the Pareto parameter of the top income distribution, and $e$ is the elasticity of taxable income (ETI). With $g = 0$, this simplifies to $t^* = \frac{1}{1 + a \cdot e}$.

Saez-Slemrod-Giertz (2012 JEL) survey the ETI. The plausible range is $e \in [0.2, 0.5]$. With $a \approx 1.5$ for the US, this yields optimal top rates between roughly 60% and 80%.

Piketty-Saez-Stantcheva (2014 AEJ:Policy) extended this to three elasticities:

• Labour supply elasticity: people work less when taxed more.
• Avoidance elasticity: people reclassify income, defer realisation, shift to tax-favoured forms.
• Bargaining elasticity: top earners bargain harder for compensation when their take-home is higher.

The bargaining channel is zero-sum in aggregate — what one CEO bargains away is captured by other shareholders or workers. So the optimal top rate is higher when the bargaining elasticity is larger.

Study next: the calculus version of the Mirrleesian optimal tax problem (uses calculus of variations); dynamic optimal taxation (Stantcheva 2017 Econometrica); optimal capital taxation (Saez-Stantcheva 2018 JPubE).

H.2 The Piketty-Saez-Zucman vs Auten-Splinter inequality debate

This is the most consequential measurement dispute in modern public finance.

Both sides agree on:

• The data source (IRS tax returns)
• The income concept (national income — labour, capital, transfers minus taxes)
• The basic objective: measure the share of national income going to the top 1%, 0.1%, etc.

They disagree by 1.2 percentage points on the change in the top 1% share between 1979 and 2019, accounting for about 30% of the gap in trends. PSZ find the top 1% share rose from 9% in 1960 to 15% in 2019. Auten-Splinter find a rise from 8% to 9%.

The technical hinge: how to allocate underreported income (income on national accounts but not on tax returns) to households.

• PSZ allocate underreported income proportional to reported income. Since top earners have more pass-through business income, which is where most underreporting happens, this assignment concentrates underreported income at the top, inflating top shares.
• Auten-Splinter use audit-based corrections, which suggest underreporting is more uniformly distributed across the income distribution.

The debate has been carried out across multiple papers (PSZ 2024 Tax Notes; Auten-Splinter 2024 reply; NBER w33678 2025). The Tax Policy Center and Brookings have written primers attempting to mediate (Tax Policy Center).

There is no resolution. The honest summary: inequality has risen, but how much depends on methodological choices that reasonable people disagree about. Headline numbers in media reports should always specify whose series they cite.

H.3 EITC and Child Tax Credit

The Earned Income Tax Credit (EITC) is the largest US anti-poverty program for working families. Empirical findings (Hoynes, Romich, Eissa-Liebman, Bhargava-Manoli):

• A $1,000 increase in EITC raises single-mother employment by ~7.3 percentage points (Eissa-Liebman 1996; Hoynes-Patel 2018).
• The EITC reduces poverty by 9.4 percentage points for affected families.
• The EITC has long-run effects on children: birth outcomes (Hoynes-Miller-Simon 2015), educational attainment (Manoli-Turner 2018), adult earnings (Bastian-Michelmore 2018).

The 2021 American Rescue Plan Child Tax Credit (CTC) expansion was the largest single anti-poverty intervention in decades:

• Maximum credit rose to $3,600/child under 6, $3,000/child age 6–17.
• Made fully refundable (no minimum earnings requirement).
• Distributed monthly July–December 2021.

Outcomes:

• Child poverty fell to 5.2% in 2021 (from 9.7% in 2020) — the lowest on record.
• The expansion lifted >2 million children out of poverty (National Academies report).
• Largest reductions: single-parent households, large families, families with parents with less education or part-time work.

When the expansion expired at end-2021, child poverty doubled in 2022. This was an unintentional "natural experiment" confirming the causal direction.

Live debate: did the expansion reduce parental labour supply? Bastian (2024) finds modest negative effects; Corinth-Meyer-Stadnicki-Wu (2022) initially projected larger ones; Roll-Hamilton find essentially no effect in surveys. The political question whether to restore the expansion remains active.

H.4 Cash transfers and child development — Baby's First Years

The Baby's First Years RCT (designed by Greg Duncan, Lisa Gennetian, Katherine Magnuson, Kimberly Noble, Hirokazu Yoshikawa, Sarah Halpern-Meekin, Nathan Fox) recruited 1,000 mothers with newborns from postpartum wards in four US cities (NYC, New Orleans, Omaha, Twin Cities) in 2018–19. All had incomes below the federal poverty line. Half were randomly assigned to receive $333/month unconditional cash for the first several years of the child's life; half received $20/month as a placebo control.

Pre-registered primary outcomes:

• Cognitive: language, executive function
• Brain: high-frequency EEG activity
• Social-emotional: behavioural assessment

Year-one finding (Troller-Renfree et al. 2022 PNAS): infant brain activity in the $333 group showed more high-frequency (alpha, beta, gamma) activity, consistent with patterns associated with later cognitive development. This was hailed as the first causal evidence that cash transfers shape infant brain development.

Year-four finding (Troller-Renfree et al. 2024 NBER w33844): on four pre-registered primary outcomes (language, executive function, social-emotional problems, high-frequency brain activity) and three secondary outcomes, no statistically significant impacts of cash transfers were found. Some secondary subgroup analyses found language effects, but the main pre-registered tests are null.

Possible reconciliations:

• Infant brain differences faded by age 4 because the environment around the children equalised.
• Effects are real but smaller than the trial was powered to detect.
• Mothers in the control group adjusted other behaviours, narrowing the realised resource gap.
• COVID-19 shock affected outcomes in ways unanticipated by the pre-registration.

This is one of the cleanest tests of "money causes child development." The honest interpretation is that unconditional cash to poor mothers is not a magic developmental intervention at the dose tested. It does materially reduce hardship, and the macro-poverty evidence (from CTC, EITC) suggests it does improve long-run child outcomes through earnings and education effects — but the immediate, in-childhood causal pathway is harder to detect than expected.

H.5 Early childhood education and Perry Preschool at 50

The Perry Preschool Project (Ypsilanti, MI, 1962–67) randomised 123 low-income Black children into a high-quality preschool + weekly home-visit program or a control group. Follow-ups have tracked participants for 50+ years.

Heckman, García, Baulos NBER w32972 (2024) "Perry Preschool at 50" (NBER 2024) reaffirm:

• Treatment effects on educational attainment, earnings, criminal behaviour persist into participants' 50s.
• Estimated annual social rate of return: 7–10%.
• Significant intergenerational effects — children of Perry participants show improved education, health, and earnings.
• Common criticisms (IQ fadeout, small sample, generalisability) are largely addressed when proper small-sample inference is used.

Skeptical takes (Bond-Lang, Whitehurst, Hechinger Report) argue the small sample, the historical context (1960s Michigan), and the unique program intensity limit generalisation to modern universal pre-K. The Boston Pre-K and Tulsa Pre-K studies (Gormley, Weiland-Yoshikawa) suggest meaningful short-run gains; long-run effects from large-scale public preschool (Tennessee's TN-VPK; Head Start) are more mixed.

H.6 Charter schools

Angrist-Pathak-Walters (2013 AEJ:Applied) "Explaining Charter School Effectiveness" (AEJ:Applied 2013) used Massachusetts charter school admissions lotteries to identify causal effects. Urban charters in Boston produced large achievement gains; non-urban charters were essentially ineffective.

The Boston-specific result: attendance at oversubscribed charter schools raises low-income student test scores by approximately one-third of a standard deviation per year — enough to close the Black-white test score gap after a few years.

The "No Excuses" model (KIPP, Match, Roxbury Prep, Excel Academies) — strict discipline, extended school day, intensive tutoring, high expectations — emerges as the active ingredient. Cohodes-Setren-Walters (2021 AEJ:Applied) studied Boston's scaling — when charters doubled enrollment, effectiveness did not dilute. This is uncommon in education research; most successful programs fade at scale.

Live debate: extrapolation from oversubscribed urban charters to the broader charter sector is contested. CREDO national charter studies (Hanushek, Raymond) find modest average effects with high variance — many charters perform like local district schools, some much better, some worse.

I. Information economics, signaling, and privacy

I.1 Akerlof, Spence, Stiglitz — the 2001 trinity

Three foundational papers structure the field:

Akerlof (1970 QJE) "The Market for Lemons": With asymmetric information about quality, markets unravel. Used-car buyers can't distinguish good cars ("peaches") from bad ("lemons"). They pay a price equal to the average quality. Sellers of peaches refuse to sell at the average. Average quality drops. Buyers' willingness to pay drops. More peaches exit. The market may collapse entirely.

Spence (1973 QJE) "Job Market Signaling": An informed sender (the worker, who knows her productivity) chooses a costly signal (education); an uninformed receiver (employer) infers productivity from the signal. In a separating equilibrium, high-productivity workers acquire more education not because it raises productivity, but because the cost of education is lower for them, allowing them to signal type without lying. Implication: the private return to education may exceed its productivity contribution.

Rothschild-Stiglitz (1976 QJE) "Equilibrium in Competitive Insurance Markets" / Stiglitz screening: When the uninformed party moves first (offering a menu of contracts), the informed party self-selects. In insurance, low-risk types choose contracts with high deductibles; high-risk types choose full coverage. Pooling equilibria don't exist; separating equilibria may not exist either.

The trinity won the 2001 Nobel. These ideas are now embedded in every applied microeconomics question involving asymmetric information — credit markets, insurance, labour markets, education policy, regulation.

The Lemons algebra in tidy form. Suppose a continuum of cars; quality $q$ uniform on $[0, 2]$. Sellers value a car at $q$ (their reservation). Buyers value a car at $1.5q$ (they get more out of it). With full information, all trade happens (any $q$, buyer pays between $q$ and $1.5q$). With asymmetric info, buyers only see the average quality among traded cars. Suppose all $q \leq q^$ sell. The average quality is $q^/2$. Buyers pay $1.5 \cdot q^/2 = 0.75 q^$. Sellers sell only if $0.75 q^* \geq q$, i.e. $q \leq 0.75 q^$. So $q^ = 0.75 q^$ requires $q^ = 0$ — no trade happens. The market collapses entirely under uniform distribution. With other distributions and partial overlap, partial trade can survive but is below efficient.

I.2 Information design and Bayesian persuasion

Kamenica-Gentzkow (2011 AER) Bayesian Persuasion reversed the signaling problem. Now the sender chooses what information to provide before the receiver acts, committing in advance to an information-disclosure rule. A prosecutor with discretion over which evidence to investigate; a search platform deciding what to surface; a salesperson deciding what to demonstrate.

Key concept: the concavification of the sender's value function over the receiver's posterior beliefs. The sender benefits from optimal information revelation when the sender's expected value is not concave in beliefs — when there are particular beliefs the sender wants to induce.

Recent advances (2024):

• Curello-Sinander 2024 "Comparative Statics of Persuasion" — how the optimal persuasion strategy varies with primitives
• Hedlund 2024 MSS "Signaling through Bayesian persuasion" — combining signaling and information design
• Multiple-receiver, dynamic, and computational extensions

Bergemann-Morris (2019 JEL) survey the broader information-design literature, including the related concept of informationally robust mechanism design — mechanisms whose properties hold under any feasible distribution of beliefs.

Live applications: data brokers, ad targeting, regulatory disclosure mandates, school-quality rating systems, restaurant inspection grades.

I.3 Differential privacy and the 2020 Census controversy

Differential privacy (DP), introduced by Dwork-McSherry-Nissim-Smith (2006), provides formal mathematical guarantees that any individual's data has bounded influence on published statistics. A mechanism is $\epsilon$-differentially private if changing one person's data changes the probability of any output by at most a factor $e^\epsilon$.

The US Census Bureau adopted DP for the 2020 Census, replacing the older "swapping" disclosure-avoidance method. The new TopDown algorithm injects calibrated noise into all published statistics, with the privacy budget $\epsilon$ allocated across geographies and tables.

The controversy:

• Quantitative biases: Kenny et al. (Science Advances 2021) showed that the noise systematically reduces racial heterogeneity across small geographies (the DP algorithm pushes counts toward state-level averages). This affects Voting Rights Act §2 enforcement, which depends on small-area minority population counts.
• Redistricting: simulation evidence from Harvard's Algorithmic Redistricting Lab (Imai et al. DAS) found DP noise levels would meaningfully change redistricting outcomes in many states.
• Legal challenge: Alabama filed a federal lawsuit in March 2021 to block DP's application to redistricting; the case was eventually dismissed.

The economic-and-statistical tension: privacy mechanism design must trade off privacy ($\epsilon$ smaller → more private) against accuracy ($\epsilon$ larger → more accurate). The Census Bureau's chosen budgets reflect their reading of legal disclosure obligations; researchers argued the budgets sacrificed too much accuracy for the marginal privacy gain.

This is the first major economic-policy application of differential privacy. It will not be the last — IRS administrative data, healthcare data, education data, all face similar pressures.

I.4 Price discrimination via personal data

Modern personalised pricing — different prices to different customers based on observed data — sits between classical 3rd-degree price discrimination (segments) and 1st-degree (personalised). The welfare implications are theoretically ambiguous: personalisation can increase or decrease consumer surplus depending on the elasticity structure.

Empirical work has mostly found limited deployment of true personalised pricing in retail (Hannak et al. 2014; Mikians et al. 2012), but extensive personalisation of offers, recommendations, and product positioning. The European Union's 2018 GDPR and 2022 Digital Services Act impose constraints on personalisation that vary by jurisdiction.

The 2024 frontier: data markets, the value of consumer data as an economic asset, the platform-firm-consumer triangle in data sharing. Acquisti, Taylor, Wagman (JEL 2016) is the standard reference; the question of how data should be priced and traded remains open.

J. Experimental economics

J.1 Lab, lab-in-field, field

Three modalities form the experimental hierarchy:

• Lab: students in computer labs play games for stakes. High internal validity; questionable external.
• Lab-in-field: experiments with subjects in their actual context (farmers in Bihar, traders in fish markets, soldiers). Bridges the lab/field gap.
• Field: natural-environment experiments, sometimes invisible to subjects. High external validity; often lower internal control.

Harrison-List (2004 JEL) classified field experiments into artefactual, framed, natural. The Levitt-List (2009 JEL) updates and characterises three phases of economic field experimentation. List's 2024 Sociological Methods & Research paper ("Field Experiments: Here Today Gone Tomorrow?") argues the methodology has matured but is at risk of being replaced by AI simulation.

J.2 Large-scale field experiments and the voltage effect

John List's "voltage effect" (List 2022) names a systematic problem in scaling: programs that work in pilots often fail at scale. The five reasons List identifies:

• False positives in pilot evaluation
• Wrong-population scaling (the pilot subjects weren't representative)
• Spillover non-replication (positive spillovers in pilots evaporate in dense scaling)
• Cost economies that don't materialise
• Compromised fidelity (the scaled version drops the active ingredient)

List's 2024 Nature paper "Optimally generate policy-based evidence before scaling" proposes a sequential experimentation protocol designed to detect voltage drops before scaling.

J.3 LLMs as research subjects — homo silicus

John Horton's 2023 working paper "Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?" (Horton, Filippas, Manning 2023; SSRN 4413859) launched the homo silicus research program. The premise: LLMs trained on human-generated text encode human-like preferences, beliefs, and behaviours. They can be queried as simulated economic agents.

Replicated results in LLM agents:

• Endowment effects (Kahneman-Knetsch-Thaler 1991)
• Dictator/Ultimatum game choices (Charness-Rabin 2002)
• Loss aversion
• Status quo bias (Samuelson-Zeckhauser 1988)

Limitations:

• LLMs may reflect reported rather than revealed preferences (a survey-response bias).
• Results vary across model families and versions.
• Distribution of outputs differs from human distribution (less heterogeneity).

The 2024 frontier: using LLMs to pilot experiments before running with humans, to generate hypotheses by varying conditions cheaply, and to simulate counterfactual policies in agent-based models. The skeptical position (Oprea 2024b) is that LLM behaviour is unstable enough to make causal inference fragile.

K. Development economics

K.1 The RCT critique (Deaton vs Banerjee)

The 2019 Nobel for Banerjee, Duflo, and Kremer recognised the RCT-driven approach to development economics — running field experiments on micro-policy questions (deworming, microfinance, agricultural extension, conditional cash transfers) and aggregating findings.

Angus Deaton's critique (most fully in his 2010 JEL "Instruments, Randomization, and Learning about Development"): RCTs are not the unique standard of credibility they're presented as. They estimate effects for the population sampled, in the context studied, at the moment of intervention. Generalising requires structural understanding, not just more RCTs. Site-selection bias (places that allow RCTs may differ systematically from places that don't) and external validity weakness are systemic.

Banerjee's reply: RCTs force researchers to confront data. Imperfect external validity is better than no internal validity. And the cumulative body of RCT evidence — meta-analysed, mechanism-tested, replicated — has produced policy-relevant knowledge that observational economics couldn't.

The current synthesis: RCTs are a powerful tool; the discipline benefits from triangulating with structural estimation, natural experiments, and qualitative work. Pre-registration, multi-site trials, and explicit external-validity discussion are now standard practice. The J-PAL and IPA networks have institutionalised this.

K.2 Cash transfer evidence

Unconditional cash transfers (UCTs) and conditional cash transfers (CCTs) are the most-studied development interventions. Stylised findings (2024 meta-analyses, e.g. IPR WP 2024 Bayesian meta-analysis):

• UCTs increase household consumption, food security, child schooling, asset accumulation.
• Effects on adult labour supply are small to zero (the "lazy poor" hypothesis is rejected).
• Effects on long-run outcomes (adult productivity, child human capital) are positive but smaller than enthusiasts hoped.

The big innovations:

• GiveDirectly's randomised large-scale UCTs in Kenya and Uganda.
• Egger et al. (2022 Econometrica) measured general-equilibrium spillover effects of large UCTs — cash recipients spend in local markets, raising non-recipient incomes too. The total multiplier was large (~2.5x).
• Banerjee-Duflo-Goldberg-Karlan-Osei-Pariente-Shapiro-Thuysbaert-Udry (2015 Science) "Graduation" programs combining cash, assets, training, and savings showed persistent effects.

K.3 Microfinance after the hype

Microfinance was originally pitched (Yunus's Grameen Bank) as poverty-eradicating. Early observational evidence was glowing. The Banerjee-Duflo et al. multi-site RCT (Karlan-Zinman 2010 Science; Banerjee-Duflo-Glennerster-Kinnan 2015 AEJ:Applied; Crepon-Devoto-Duflo-Pariente 2015 AEJ:Applied) found modest effects on entrepreneurship and consumption smoothing, but no transformational poverty reduction.

The 2024 meta-analysis literature finds:

• Positive effects on women's empowerment, financial inclusion, asset accumulation
• Mixed-to-null effects on income, business profits, poverty headcount
• Publication bias inflates the apparent positive effects

The honest summary: microfinance is a useful financial-inclusion tool, not a transformational anti-poverty intervention. The development field has moved on to graduation programs, cash transfers, savings devices, and digital financial inclusion.

K.4 Behavioural interventions in low- and middle-income settings

Behavioural-economics interventions in LMIC contexts have been more robust than the lab nudges that failed replication:

• SMS reminders for medication adherence (HIV, TB)
• Commitment savings devices (Ashraf-Karlan-Yin 2006 QJE)
• Default enrollment in agricultural insurance
• Information provision about returns to schooling (Jensen 2010 QJE; Nguyen 2008)

These are often best understood not as nudges to overcome bias but as removing genuine information or cost frictions.

L. Inequality and intergenerational mobility

L.1 The Chetty Opportunity Atlas

Chetty, Friedman, Hendren, Jones, Porter (2018 NBER) "The Opportunity Atlas" linked 20 million children's outcomes to the Census tract in which they grew up. The Atlas is publicly available at opportunityatlas.org. Headline findings:

• Large variation across neighbourhoods, even within counties, in children's adult earnings.
• The "land of opportunity" varies enormously: top neighbourhoods predict 30+ percentile points higher adult income than bottom ones for kids from similar families.
• Race-specific effects: Black children — particularly Black boys — have markedly worse mobility, even when comparing across children of the same family income who grow up on the same block (Chetty-Hendren-Jones-Porter 2020 QJE).

The 2024 Brookings paper "The changing landscape of economic opportunity by race and class in America" updated the Atlas. Findings:

• For low-income children, the Black-white mobility gap narrowed between cohorts born in 1978 and 1992.
• For Hispanic children, intergenerational mobility is high — they are catching up to white income levels generationally.
• The Black-white gap remains; the Hispanic-white gap is closing.

L.2 Derenoncourt et al. on the racial wealth gap

Derenoncourt-Kim-Kuhn-Schularick (2024 QJE "Wealth of Two Nations: The U.S. Racial Wealth Gap, 1860–2020", QJE 2024) built a 160-year time series of US Black and white wealth. The findings reshape the conversation:

• In 1860, the white-to-Black wealth ratio was 56:1 (driven primarily by slavery and the absence of Black wealth ownership).
• From 1860 through 1950, the ratio narrowed sharply, driven by emancipation, the rise of Black property ownership, and early-20th-century Black economic mobility.
• From 1950 through the early 1980s, the ratio stayed roughly flat — convergence stalled.
• Since the 1980s, the ratio has widened. Capital gains on equity and housing accrued disproportionately to white households; Black households held a higher share of wealth in lower-yielding assets.

The take-away: initial-condition disadvantage (slavery's stripping of capital from Black Americans) plus century-long capital-market exclusion plus modern asset-distribution differences combine to produce a wealth gap whose convergence has stalled. The mechanical "if Black incomes were 80% of white incomes" simulation under-explains the gap because the gap is in wealth, which compounds over generations.

L.3 The Great Gatsby Curve revisited

Krueger's 2012 "Great Gatsby Curve" cross-country plot showed that countries with higher inequality (Gini) tend to have lower intergenerational income mobility. The 2020s replication and extension work (Corak; Chetty et al.; Bratberg et al.) confirms the basic empirical pattern but qualifies the interpretation:

• The relationship is robust across measurement choices.
• Within countries, mobility varies enormously across regions (US: Salt Lake City high; Charlotte and Atlanta low).
• The mechanism — whether high inequality causes low mobility through compounding human-capital investments, neighbourhood effects, or political-economy capture — remains debated.

M. Climate and environmental microeconomics

M.1 Carbon pricing — cap-and-trade vs Pigouvian tax

The theoretical equivalence of a Pigouvian tax and a cap-and-trade system (Weitzman 1974 ReStud) holds under certainty. With uncertainty over abatement costs, Weitzman showed that price instruments (taxes) are preferred when marginal damage is relatively flat, and quantity instruments (caps) are preferred when marginal damage is relatively steep. Climate change has uncertain enough damage curves that the choice depends on parameterisation.

In practice:

• The EU Emissions Trading System (ETS) — the world's largest carbon market — operates a cap-and-trade across power and heavy industry. Phase 4 (2021–30) tightened the cap and introduced the Market Stability Reserve.
• California's cap-and-trade program (launched 2013) covers ~80% of state GHGs. May 2024 auction settlement: $38.35/tonne, floor $24.04 (EIA).
• RGGI (Regional Greenhouse Gas Initiative) covers Northeast US power sector since 2009.
• Sweden's carbon tax (since 1991) is among the highest in the world at €120+/tonne; widely cited as evidence that a carbon tax can be politically durable.

Robert Stavins's 2019 "Carbon Pricing 100" updates and his cap-and-trade lessons paper (REEP 2017) are the standard references.

M.2 Social cost of carbon

The social cost of carbon (SCC) is the present-value damage caused by emitting one additional tonne of CO₂. EPA's 2023 update (using the Rennert-Errickson-Prest-Rennels et al. framework) raised the central SCC estimate to ~$190/tonne at a 2% discount rate, up from previous estimates around $50/tonne. The increase reflects:

• Lower discount rates (Ramsey-derived, accepting that damages compound over centuries)
• Updated damage functions incorporating tipping points, biodiversity, mortality
• Faster-than-expected climate impacts

This SCC is consequential: it enters cost-benefit analyses of nearly every federal regulation. The 2025 Trump administration EPA is expected to revise downward, restoring the practical debate.

M.3 Behavioural interventions for energy use

Allcott-Mullainathan (2010 Science) and the OPower experiments showed that home energy reports comparing your usage to your neighbours reliably reduce electricity consumption by 1–3%. The effects:

• Persist for years (with diminishing returns)
• Are cost-effective at displacing demand-side efficiency programs
• Have larger effects on bigger users (heterogeneous treatment effects)
• Don't backfire on low users (no boomerang on the "right" side)

This is among the most-robust behavioural-policy findings — it scales, it persists, and it transfers across utilities. It's a counter-example to the broader nudge-disappointment narrative.

N. Crypto and on-chain mechanism design

N.1 Automated Market Makers and constant-function market makers

Uniswap (2018) introduced the constant-product AMM: a pool holds reserves $(x, y)$ of two tokens; trades preserve $x \cdot y = k$. To buy $\Delta y$, you pay $\Delta x = \frac{x \cdot \Delta y}{y - \Delta y}$. Price slippage is built in; liquidity providers earn fees but bear impermanent loss (the opportunity cost vs. holding the tokens).

The economics literature on AMMs (Angeris-Chitra-Evans-Cao; Capponi-Jia; Lehar-Parlour) treats them as a special class of mechanism design problems:

• Optimal fee design tradeoffs LP revenue against trader welfare
• Concentrated liquidity (Uniswap v3) trades range customisation against gas costs
• Multi-asset and curve-shape variants (Balancer, Curve) optimise for stable-asset pairs

N.2 Maximum extractable value (MEV)

MEV is the value extractable by reordering, including, or excluding transactions in a block. Sandwich attacks (front-running a swap to push the price, letting the swap fill at the bad price, then exiting) are the canonical example. By 2024, total MEV extraction on Ethereum was approximately $1.1 billion annually (vs. $550M in 2021) (ESMA 2025). Sandwich attacks account for ~66% of MEV, arbitrage ~33%, liquidations <1%.

Proposer-Builder Separation (PBS), implemented via MEV-Boost (Flashbots), is the production response. Validators no longer build their own blocks; specialised builders construct blocks (extracting MEV optimally) and pay validators via auction. By 2025, >90% of Ethereum validators use MEV-Boost. The 2024 frontier:

• In-protocol PBS (built into Ethereum consensus rather than as middleware)
• Fair-ordering protocols (Aequitas) that try to enforce order-fairness rather than ratify the highest bidder
• Application-level MEV capture and redistribution — RediSwap (2024) is one design that returns sandwich profits to traders/LPs

N.3 Token-based mechanism design and governance

DAOs (decentralised autonomous organisations) deploy governance tokens that grant voting rights over protocol parameters (fee levels, reserve requirements, treasury spending). The economics literature studies:

• Voter participation problems (most DAOs have low turnout)
• Whale governance — token holdings are concentrated, so a few addresses dominate votes
• Vote buying via secondary markets ("CowSwap for governance votes")
• Quadratic voting (Posner-Weyl) implementations (Gitcoin Grants for retroactive funding)

The 2024 frontier is "futarchy" experiments (decisions made by prediction-market-determined criteria) and AI-mediated governance.

O. Antitrust and competition policy live debates 2024–2026

Beyond the Big Tech cases discussed in section D.3, several other live debates frame the 2024–2026 antitrust landscape:

O.1 The "neo-Brandeisian" turn vs the consumer welfare standard

The dominant US antitrust frame since the late 1970s — the Robert Bork "consumer welfare standard" — focuses on whether mergers and conduct raise prices to consumers. The "neo-Brandeisian" school (Lina Khan, Tim Wu, Barry Lynn, Sandeep Vaheesan) argues this is too narrow: antitrust should also consider effects on competition itself, labour markets, political power, innovation, and small business viability.

The FTC under Khan (2021–25) reflected the neo-Brandeisian turn. The Trump 2.0 administration (2025–) has pivoted back to more traditional consumer welfare framing in priorities, though many cases continue.

O.2 The 2023 Horizontal Merger Guidelines

The DOJ-FTC 2023 Merger Guidelines replaced the 2010 version. Key changes:

• Lower concentration thresholds for "presumptively anti-competitive" mergers (HHI > 1800 with $\Delta$HHI > 100)
• Explicit attention to vertical mergers, labour markets, and potential competition
• Less weight on efficiencies defences

Critics argue the new guidelines exceed statutory mandate; defenders argue they update for an actual concentration trend.

O.3 Live cases beyond Big Tech

• Visa antitrust complaint (DOJ Sept 2024) over debit card market
• Live Nation/Ticketmaster antitrust case (DOJ May 2024)
• JetBlue/Spirit blocked merger (early 2024)
• Kroger/Albertsons blocked merger (Dec 2024)
• Pharma generic suppression cases (ongoing)

P. Health, education, family economics — additional notes

P.1 Healthcare cost growth

US healthcare spending continues at ~17% of GDP, roughly double most peer countries. The Cutler-Skinner work, the Garber-Skinner JEP work, and the Finkelstein-Gentzkow-Williams (2016) work on geographic variation all point to the same conclusion: most variation is supply-side (provider practice styles, technology adoption, hospital pricing), not patient demand-side or population health.

The 2024 Inflation Reduction Act's Medicare drug-price negotiation provisions launched the first-ever Medicare negotiation cycle. The drugs subject to first-round negotiation are taking 38–79% price reductions when negotiated prices take effect in 2026.

P.2 Deaths of despair

Case-Deaton's Deaths of Despair and the Future of Capitalism (Princeton 2020) named the rising mortality from drug overdoses, alcohol-related deaths, and suicides in non-college-educated white Americans aged 45–54. The pattern extended through the 2010s and was supercharged by COVID-19 and the fentanyl epidemic.

By 2018, US deaths from these causes totaled ~158,000 (up from ~65,000 in 1995). US life expectancy declined to 76.4 years in 2021 — the steepest two-year drop since World War II.

2024 updates:

• Friedman-Hansen (UCLA 2024) extended the framework to Black Americans, where similar trends in overdose deaths now match or exceed white rates.
• Native American midlife mortality from deaths of despair: 241.7 per 100,000 in 2022 (2.36x the white rate).
• Fentanyl overdoses peaked in 2022 and have begun to decline modestly in 2023–24.

The economic interpretation: deaths of despair correlate strongly with declining real wages, declining marriage rates, and weakening community institutions in affected groups. The proximate cause is opioid availability; the deeper cause is structural economic dislocation.

Mathematical foundations

This section presents the algebra you need to operate fluently with the literature above. Each subsection ends with a "study next" callout for the calculus or measure-theory version.

M1. Consumer choice with a budget line and indifference logic

A consumer with income $I$ buys quantities $(x_1, x_2)$ of two goods at prices $(p_1, p_2)$. The budget constraint is

$$p_1 x_1 + p_2 x_2 \leq I$$

with equality if the consumer spends all income (which she does if she prefers more of either good).

Without calculus, characterise optimal choice by the indifference logic. A consumer's preference over bundles is captured by indifference curves — sets of bundles among which she is indifferent. At an interior optimum:

1. She is on the budget line (not below — that wastes income).
2. The indifference curve through her chosen point is tangent to the budget line — they touch but don't cross.

Tangency means the slopes match. The slope of the budget line is $-p_1/p_2$ (the price ratio with a sign for the trade-off). The slope of the indifference curve is the marginal rate of substitution: how much $x_2$ the consumer would give up for one more $x_1$ to stay equally happy. At optimum:

$$MRS_{1,2} = \frac{p_1}{p_2}$$

That is: the consumer's willingness to substitute equals the market's required substitution rate. If $MRS > p_1/p_2$, she values good 1 more than the market does and should buy more of it; if $MRS < p_1/p_2$, less.

Worked example: Suppose $I = $10$, $p_1 = $2$, $p_2 = $1$. Budget line: $2x_1 + x_2 = 10$, so $x_2 = 10 - 2x_1$. If preferences are symmetric and the consumer always prefers a balanced bundle of equal quantity ($MRS = 1$ at $x_1 = x_2$), the optimum has $MRS = 2$ (must equal $p_1/p_2 = 2$). With symmetric preferences this happens where $x_2 = 2x_1$. Combined with budget: $2x_1 + 2x_1 = 10$, so $x_1 = 2.5$, $x_2 = 5$.

Study next: utility maximisation via Lagrangian, first-order conditions, the Marshallian and Hicksian demand functions, the Slutsky equation.

M2. Producer surplus, consumer surplus, geometrically and algebraically

The market demand curve $P = a - bQ$ (with positive $a$ and $b$) and supply curve $P = c + dQ$ (with positive $c$ and $d$, $c < a$) intersect at equilibrium:

$$Q^* = \frac{a-c}{b+d}, \qquad P^* = \frac{a d + b c}{b + d}$$

Verify: $P^* = a - b Q^$ and $P^ = c + d Q^*$ both give the same value.

Consumer surplus (CS) is the area below the demand curve and above the equilibrium price, between $Q = 0$ and $Q = Q^*$. With linear demand, this is a triangle:

$$CS = \tfrac12 \cdot (a - P^) \cdot Q^$$

Producer surplus (PS) is the area above the supply curve and below the equilibrium price:

$$PS = \tfrac12 \cdot (P^* - c) \cdot Q^*$$

Total surplus $TS = CS + PS$ is the gain from trade. A tax of $t$ per unit drives a wedge: buyers pay $P^* + t$, sellers receive $P^* - 0$ (or however it splits — depends on relative elasticities). The new quantity $Q'$ satisfies $P^(buyer) - P^(seller) = t$. Tax revenue is $t Q'$. Deadweight loss is a triangle with height $t$ and base $(Q^* - Q')$.

Worked numerical: $a = 10$, $b = 1$, $c = 2$, $d = 1$. Then $Q^* = 4$, $P^* = 6$. $CS = \tfrac12 \cdot 4 \cdot 4 = 8$. $PS = \tfrac12 \cdot 4 \cdot 4 = 8$. $TS = 16$. A $t = 2$ tax: buyers pay 7, sellers receive 5, $Q' = 3$. Revenue $= 6$. DWL $= \tfrac12 \cdot 2 \cdot 1 = 1$.

M3. Nash equilibrium in finite normal-form games

A 2-player game is a normal form: each player has a finite set of strategies; payoffs are functions of joint strategies.

The classic Prisoner's Dilemma (PD):

A Nash equilibrium is a strategy pair $(s_1, s_2)$ such that neither player can improve by unilaterally deviating. In PD, $(\text{Defect}, \text{Defect})$ is the unique NE — given the other defects, defecting yields 1 vs. 0 for cooperating. The Pareto-dominant outcome $(\text{Coop}, \text{Coop})$ — with payoffs $(3,3)$ — is not a NE.

The Coordination Game:

Two pure NE: $(\text{L}, \text{L})$ and $(\text{R}, \text{R})$. The former is payoff-dominant (yields more for both); $(\text{L}, \text{L})$ is risk-dominant if mistakes are likely.

Mixed strategies arise when there's no pure NE. In Matching Pennies:

No pure NE. Mixed NE: each player plays Heads with probability $\tfrac12$.

Computing mixed NE algebraically: suppose Player 2 plays Heads with probability $q$ and Tails with $1-q$. Player 1's payoff to Heads = $q \cdot 1 + (1-q) \cdot (-1) = 2q - 1$. Player 1's payoff to Tails = $q \cdot (-1) + (1-q) \cdot 1 = 1 - 2q$. Player 1 is indifferent when these are equal: $2q - 1 = 1 - 2q$, so $q = \tfrac12$. By symmetry, Player 1 plays Heads with $p = \tfrac12$.

Study next: best-response correspondences in larger games; sequential games and subgame-perfect equilibrium; Bayesian games (incomplete information); evolutionary stable strategies.

M4. Auction algebra (discrete bidders)

We already did the discrete revenue-equivalence example in section B.1. To summarise the algebra:

Two bidders, values in ${1, 2, 3}$ uniformly.

Second-price auction: bid your value. Expected revenue = $E[\min(v_1, v_2)] = \frac{14}{9}$.

First-price auction: equilibrium bid $b(v)$ in discrete settings is determined by indifference. In the continuous symmetric IPV case with uniform $v$ on $[0,1]$, $b(v) = v/2$. The discrete analog gives essentially the same expected revenue under revenue equivalence.

Revenue equivalence in one line: any auction that (a) assigns the good efficiently, (b) gives the lowest type zero expected surplus, yields the same expected revenue under symmetric IPV.

Why it fails empirically: bidders' values are not IPV (common value components), bidders are not symmetric, the auction is part of a larger market (header bidding), there is asymmetric information about competitors' bids.

Study next: continuous-type revenue equivalence proof via envelope conditions; Myerson optimal mechanism with virtual values; auction design with risk aversion and budget constraints.

M5. Akerlof Lemons algebra

Recap: continuum of cars; quality $q$ uniform on $[0, 2]$. Sellers reservation = $q$. Buyer valuation = $1.5q$. Buyers see only the average $q$ among traded cars.

Suppose all cars with $q \leq q^$ trade. Then average traded quality is $q^/2$. Buyers pay $1.5 \cdot q^/2 = 0.75 q^$ for any car. Sellers trade iff $0.75 q^* \geq q$, i.e. $q \leq 0.75 q^$. So we need $q^ = 0.75 q^$, only satisfied at $q^ = 0$.

The market collapses entirely. Pure adverse selection. With other distributions, partial trade survives but never reaches the efficient outcome.

The fix is signaling (sellers offer warranties; high-quality cars carry credible signals) or screening (buyers offer menus of price/inspection contracts) or third-party certification (Carfax, dealer reputations, Lemons-law statutes).

M6. Bayesian updating with discrete priors

Bayes' rule: $P(H|E) = \frac{P(E|H) P(H)}{P(E)}$, where $P(E) = \sum_H P(E|H) P(H)$.

Worked example: A doctor knows that a disease has prior probability $P(D) = 0.01$ in the population. A test has true positive rate $P(+|D) = 0.99$ and false positive rate $P(+|\neg D) = 0.05$. A patient tests positive. What is $P(D|+)$?

$$P(D|+) = \frac{P(+|D) P(D)}{P(+|D) P(D) + P(+|\neg D) P(\neg D)} = \frac{0.99 \cdot 0.01}{0.99 \cdot 0.01 + 0.05 \cdot 0.99} = \frac{0.0099}{0.0594} \approx 0.167$$

Only 17% chance the patient has the disease — counterintuitive base-rate result.

This is the engine behind Bayesian persuasion, Bayesian games, and information design. The concavification trick in Bayesian persuasion follows directly: the sender chooses what posterior beliefs to induce, subject to martingale (the prior is the expected posterior).

M7. DiD 2x2 algebraic identity

The 2x2 DiD estimator with two periods (pre, post) and two groups (treated $T$, control $C$):

$$\widehat{\delta}{DiD} = (\bar{Y}{T,\text{post}} - \bar{Y}{T,\text{pre}}) - (\bar{Y}{C,\text{post}} - \bar{Y}_{C,\text{pre}})$$

Under parallel trends — $E[Y(0){T,\text{post}} - Y(0){T,\text{pre}}] = E[Y(0){C,\text{post}} - Y(0){C,\text{pre}}]$ — this identifies the ATT in the post-period.

Worked example: A state $T$ raises minimum wage in 2023. Pre-period (2022): $\bar{Y}_T = 5$ jobs lost per 1000 workers; $\bar{Y}_C = 4$ jobs lost. Post-period (2024): $\bar{Y}_T = 6$ jobs lost; $\bar{Y}_C = 7$ jobs lost. DiD = $(6-5) - (7-4) = 1 - 3 = -2$. The minimum wage reduced job loss by 2 per 1000 vs. the counterfactual.

Why staggered TWFE fails: with multiple treatment cohorts and dynamic effects, the TWFE coefficient is a weighted sum of all $2 \times 2$ comparisons — including comparisons where the "control" is an already-treated unit, in which case the comparison subtracts away the late-treated unit's evolving treatment effect. If treatment effects ramp up over time, this comparison is biased negative.

M8. Event study discrete setup

In an event-study, you index time by $k$ relative to treatment onset. With clean control (untreated units), the estimator is:

$$\widehat{\beta}k = (\bar{Y}{T,k} - \bar{Y}{T,-1}) - (\bar{Y}{C,k} - \bar{Y}_{C,-1})$$

for periods $k$ relative to the omitted $k = -1$ pre-period.

Modern best practice (Borusyak-Jaravel-Spiess):

1. Estimate unit and period FE using only untreated observations.
2. Impute $\hat{Y}_{it}(0)$ for treated obs using these FE.
3. Treatment effect per obs: $Y_{it} - \hat{Y}_{it}(0)$.
4. Aggregate by event-time $k$.

M9. RDD as a discontinuity-in-conditional-mean estimator

For a sharp RD at cutoff $c$ on running variable $X$:

$$\hat{\tau} = \hat{m}{+}(c) - \hat{m}{-}(c)$$

where $\hat{m}{+}(c)$ is the conditional mean just above $c$, and $\hat{m}{-}(c)$ just below. In practice, you fit a local linear regression on each side and read off the intercepts at $c$.

Algebraic local-linear analog: take observations within bandwidth $h$ on each side. Fit $Y = a + b(X - c) + \epsilon$ on each side separately. The above-intercept $a^+$ and below-intercept $a^-$ at $X = c$ give $\hat\tau = a^+ - a^-$.

Modern best practice (Calonico-Cattaneo-Titiunik): use bias-corrected confidence intervals that account for the asymptotic bias of local-linear regression near the boundary.

M10. Sufficient statistics intuition with discrete examples

Chetty's framework: a welfare formula $W = f(\theta_1, \theta_2, \ldots, \theta_K)$ where $\theta_i$ are reduced-form elasticities you can estimate.

Example: optimal commodity tax on labour. With linear demand, the deadweight loss of a small tax $t$ is approximately $\frac{1}{2} t^2 \epsilon \cdot Q^* / P^$, where $\epsilon$ is the elasticity of labour supply. To get welfare implications of changing $t$, you need only know $\epsilon$ and the equilibrium $(P^, Q^*)$ — not the underlying utility function.

Example: optimal UI replacement rate (Baily-Chetty formula). Letting $b$ be the unemployment benefit and $w$ be wage, the optimal replacement rate $r = b/w$ satisfies:

$$r^* = \frac{r}{1-r} \cdot \frac{\text{coefficient of relative risk aversion}}{\text{elasticity of unemployment duration with respect to benefits}}$$

You don't need to model search, leisure preferences, or job offers; you need the two elasticities and the risk aversion parameter.

M11. Optimal income tax Mirrleesian intuition via elasticity

The Saez (2001) formula for the optimal top marginal tax rate:

$$t^* = \frac{1 - g}{1 - g + a \cdot e}$$

with $g$ the social welfare weight on top incomes (often 0), $a$ the Pareto parameter of the top income distribution (around 1.5 for US), and $e$ the elasticity of taxable income (estimates between 0.2 and 0.5).

With $g = 0$ and $a = 1.5$:

• If $e = 0.2$, $t^* = 1/(1 + 0.3) = 77%$
• If $e = 0.5$, $t^* = 1/(1 + 0.75) = 57%$

That range — 57% to 77% — is what defines the policy debate. The 1980 US top federal rate was 70%; today's top federal rate is 37% (plus state, plus Medicare). The framework suggests US top rates are below the Saez optimum.

Study next: the full Mirrleesian derivation via the Lagrangian on the incentive-compatibility constraint and the social welfare functional. Saez (2001) is the canonical paper; Diamond-Saez (2011 JEP) is the policy-applied summary.

Risks, limitations, and open questions

Methodological frontier risks

1. The DiD methodology is unfinished. The Callaway-Sant'Anna, BJS, and de Chaisemartin estimators agree on staggered-adoption problems but differ in efficiency, robustness to violations of parallel trends, and finite-sample behaviour. The 2025 practitioner's guide (Callaway-Cunningham-Goodman-Bacon-Sant'Anna) is a guide, not a settled consensus.

2. External validity of RCTs remains unresolved. List's voltage effect identifies pathologies; the discipline has not produced a reliable forecasting tool for which RCT findings will scale.

3. Sufficient statistics formulas hide assumptions. The elegance of the Baily-Chetty UI formula or the Saez top-tax formula obscures the assumption that the underlying environment is locally well-approximated by the model the formula is derived from. Behavioural agents, general equilibrium spillovers, dynamic learning all complicate the picture.

4. AI/ML in causal inference: methods like double machine learning, causal forests, and BCF promise heterogeneous treatment effect estimation. But the assumptions (overlap, no unmeasured confounding) are the same as ever; the methods don't manufacture credibility.

Substantive open questions

1. What is the cause of stagnant intergenerational mobility? Chetty's neighbourhoods variation tells us where mobility is high or low. The mechanism (peer effects, school quality, network access, family stability) remains contested.

2. How much of inequality is real vs. measurement? PSZ vs. Auten-Splinter is unsettled and may stay so for years.

3. Will the noncompete economic evidence outlast the failed federal rule? State-by-state action is happening but uneven. Will a future FTC re-attempt with explicit statutory authority?

4. What replaces Roe v. Wade's effect on labour supply? Dobbs (2022) has triggered a new cohort of natural experiments on the labour-supply effects of abortion access, with early evidence still emerging.

5. Will antitrust law catch up with platform economics? The Meta loss (Nov 2025) suggests courts will define digital markets broadly when behaviour supports it. The DMA in Europe may produce different jurisprudence.

6. Are nudges legitimate after the credibility crisis? DellaVigna-Linos quantified the gap between academic effect sizes and real-world impact; the political/ethical question of when nudges are appropriate is not closed.

7. Will AI compress or expand wage inequality? Early evidence (Brynjolfsson-Li-Raymond on call centres; Noy-Zhang on writing tasks) suggests compression in some settings; Autor 2024 suggests the broader pattern favours high-skill workers. The next five years will be decisive.

8. Can cash transfers improve child development causally? Baby's First Years' null at age 4 is one data point; replication and longer follow-up are essential before drawing strong conclusions.

9. What is the right discount rate for climate damages? The Ramsey (1928) approach favoured by Stern (2007) gives low rates and high SCC; the descriptive market-rate approach (Nordhaus) gives higher rates and lower SCC. EPA 2023's 2% pick lands closer to Stern.

10. Can on-chain mechanism design solve MEV? PBS contained the harms; in-protocol PBS and fair-ordering proposals may go further. Or MEV may just migrate to different layers.

Risks to the field's authority

• Replication crisis spillover: behavioural economics has had retractions; macro and finance have faced similar challenges. If applied micro produces more high-profile failures, public trust may erode.
• Politicisation: when inequality measurement, antitrust, climate policy, gig-economy classification are decided by partisan courts, the discipline's reputation as neutral expert advisor erodes.
• AI generated research: the next 3–5 years will see massive amounts of AI-assisted empirical work. Peer review and disclosure norms need to adapt.

Recommendations for further study

For the self-learner aiming to operate at the frontier:

Year 1: foundations

• Textbooks: Varian's Intermediate Microeconomics (algebra), then Microeconomic Analysis (calculus). Mas-Colell-Whinston-Green for the canonical graduate text. For econometrics: Stock-Watson, then Wooldridge.
• Empirical methods: Angrist-Pischke Mostly Harmless Econometrics (the classic intro to causal inference), then Mastering 'Metrics (gentler). Cunningham Causal Inference: The Mixtape (free online, programming-heavy).
• Game theory: Tadelis Game Theory: An Introduction (algebra-level).

Year 2: applied micro and methods

• Industrial organization: Tirole The Theory of Industrial Organization; Belleflamme-Peitz Industrial Organization: Markets and Strategies.
• Public economics: Salanie The Economics of Taxation; Hindriks-Myles Intermediate Public Economics.
• Behavioural economics: Thaler-Sunstein Nudge; Kahneman Thinking, Fast and Slow; Camerer Behavioral Game Theory.
• Market design: Roth Who Gets What and Why (popular); Vohra Mechanism Design: A Linear Programming Approach (technical).

Year 3: frontier methods and primary literature

• Modern causal inference: read Callaway-Cunningham-Goodman-Bacon-Sant'Anna 2025 practitioner's guide end-to-end. Read Roth-Sant'Anna-Bilinski-Poe 2023 JoE DiD survey. Read Cattaneo-Idrobo-Titiunik 2024 RDD book.
• NBER working papers in your area of interest. Subscribe to the weekly NBER digest.
• JEP (Journal of Economic Perspectives) is the best long-form audience-friendly source.

Resources for staying current

• Marginal Revolution (Cowen-Tabarrok) — daily; opinionated but widely-read
• A Fine Theorem (Kevin Bryan) — slower; technical deep-dives
• VoxEU (CEPR) — policy-focused short pieces
• Promarket (Stigler Center) — competition and antitrust
• Development Impact (World Bank) — development economics
• J-PAL and IPA evaluation databases — RCT findings
• NBER reporter, JEP, Annual Reviews of Economics — survey articles

For deeper math

• Mas-Colell-Whinston-Green for measure-theoretic micro
• Krishna Auction Theory for continuous types
• Stokey-Lucas Recursive Methods in Economic Dynamics for dynamic optimisation
• Wasserman All of Statistics and Hansen Econometrics for econometric theory
• For machine learning in econ: Athey-Imbens 2017 JEP; Mullainathan-Spiess 2017 JEP; Chernozhukov et al. on double machine learning

Source inventory

Foundational classics (cited as underpinning)

• Akerlof, G. (1970). "The Market for Lemons: Quality Uncertainty and the Market Mechanism." QJE 84(3). Sfu mirror
• Arrow, K. & Debreu, G. (1954). "Existence of an Equilibrium for a Competitive Economy." Econometrica.
• Gale, D. & Shapley, L. (1962). "College Admissions and the Stability of Marriage." American Mathematical Monthly.
• Kahneman, D. & Tversky, A. (1979). "Prospect Theory: An Analysis of Decision under Risk." Econometrica.
• Mirrlees, J. (1971). "An Exploration in the Theory of Optimum Income Taxation." ReStud.
• Roth, A. & Sotomayor, M. (1990). Two-Sided Matching. Cambridge University Press.
• Rothschild, M. & Stiglitz, J. (1976). "Equilibrium in Competitive Insurance Markets." QJE.
• Spence, M. (1973). "Job Market Signaling." QJE.
• Tirole, J. (1988). The Theory of Industrial Organization. MIT Press.
• Vickrey, W. (1961). "Counterspeculation, Auctions, and Competitive Sealed Tenders." JF.
• Weitzman, M. (1974). "Prices vs. Quantities." ReStud.

Market design and matching

• Abdulkadiroğlu, A., Pathak, P., & Roth, A. (2006). "Changing the Boston School Choice Mechanism." NBER w11965
• Delacrétaz, D., Kominers, S., & Teytelboym, A. (2023). "Matching Mechanisms for Refugee Resettlement." AER 113(10)
• Roth, A. (2023). "Market Design and Maintenance." NBER w31947
• Roth, A., Sönmez, T., & Ünver, M. (2004). "Kidney Exchange." QJE

Auctions

• Edelman, B., Ostrovsky, M., & Schwarz, M. (2007). "Internet Advertising and the Generalized Second-Price Auction." AER.
• Milgrom, P. & Weber, R. (1982). "A Theory of Auctions and Competitive Bidding." Econometrica.
• Myerson, R. (1981). "Optimal Auction Design." Mathematics of Operations Research.
• Palacios-Huerta, I., Parkes, D., & Steinberg, R. (2024). "Combinatorial Auctions in Practice." JEL.
• NSF Q&A with Milgrom on 2017 incentive auction
• Nobel Popular Information 2020 on Milgrom-Wilson
• AdExchanger on Google's first-price transition
• Search Engine Land on RGSP

Behavioral economics

• DellaVigna, S. & Linos, E. (2022). "RCTs to Scale: Comprehensive Evidence from Two Nudge Units." PNAS
• Data Colada blog
• Science 2024 on Gino research misconduct finding
• Hoover 2024 on Sunstein
• Horton, J., Filippas, A., & Manning, B. (2023). "Large Language Models as Simulated Economic Agents." SSRN 4413859

Industrial organization and antitrust

• Belleflamme, P. & Peitz, M. (2024). "The Economic Theory of Two-Sided Platforms." CRC TR224 DP 584
• Rochet, J.-C. & Tirole, J. (2003). "Platform Competition in Two-Sided Markets." JEEA.
• Rochet, J.-C. & Tirole, J. (2006). "Two-Sided Markets: A Progress Report." RAND.
• White & Case on Mehta Google ruling
• FTC v. Amazon press release
• GeekWire on Chun motion to dismiss ruling
• NPR on Meta FTC ruling November 2025
• Kamepalli, S., Rajan, R., & Zingales, L. (2020). "Kill Zone." NBER w27146.

Causal inference methods

• Borusyak, K., Jaravel, X., & Spiess, J. (2024). "Revisiting Event-Study Designs: Robust and Efficient Estimation." ReStud 91(6) 3253-3285
• Callaway, B., Cunningham, S., Goodman-Bacon, A., & Sant'Anna, P. (2025). "Difference-in-Differences Designs: A Practitioner's Guide." arXiv 2503.13323
• Callaway, B. & Sant'Anna, P. (2021). "Difference-in-Differences with Multiple Time Periods." JoE.
• de Chaisemartin, C. & D'Haultfœuille, X. (2020). "Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects." AER.
• Goodman-Bacon, A. (2021). "Difference-in-Differences with Variation in Treatment Timing." JoE.
• Imbens, G. & Angrist, J. (1994). "Identification and Estimation of Local Average Treatment Effects." Econometrica.
• Sun, L. & Abraham, S. (2021). "Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects." JoE.
• Calonico, S., Cattaneo, M., & Titiunik, R. (2014). "Robust Nonparametric Confidence Intervals for Regression-Discontinuity Designs." Econometrica.
• Cattaneo, M., Idrobo, N., & Titiunik, R. (2024). A Practical Introduction to Regression Discontinuity Designs: Extensions. Cambridge
• Abadie, A. (2021). "Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects." JEL.
• Conlon, C. & Gortmaker, J. (2024). "Incorporating Micro Data into Differentiated Products Demand Estimation with PyBLP." JoE

Labor economics

• Azar, J. & Marinescu, I. (2024). "Monopsony Power in the Labor Market: From Theory to Policy." AR Econ
• Cengiz, D., Dube, A., Lindner, A., & Zipperer, B. (2019). "The Effect of Minimum Wages on Low-Wage Jobs." QJE. NBER w25434
• Dube, A. (2024). "Minimum Wages in the 21st Century." NBER w32878
• FTC Noncompete Rule Final Press Release April 2024
• Acemoglu, D. (2024). "The Simple Macroeconomics of AI." NBER w32487
• Acemoglu, D. & Restrepo, P. (2022). "Tasks, Automation, and the Rise in U.S. Wage Inequality." Econometrica.
• Berkeley Labor Center 2024 on gig driver pay
• EPI on Uber wages

Public economics, inequality, transfers

• Auten, G. & Splinter, D. (2024). "Income Inequality in the United States: Using Tax Data to Measure Long-term Trends." JPE.
• Auten, G. & Splinter, D. (2024). "Reply to Piketty, Saez, and Zucman." PDF
• Piketty, T., Saez, E., & Zucman, G. (2018). "Distributional National Accounts: Methods and Estimates for the United States." QJE.
• Piketty, T., Saez, E., & Zucman, G. (2024). "Income Inequality in the United States: A Comment." Zucman.eu
• Saez, E. (2001). "Using Elasticities to Derive Optimal Income Tax Rates." ReStud.
• Saez, E., Slemrod, J., & Giertz, S. (2012). "The Elasticity of Taxable Income with Respect to Marginal Tax Rates." JEL.
• Piketty, T., Saez, E., & Stantcheva, S. (2014). "Optimal Taxation of Top Labor Incomes: A Tale of Three Elasticities." AEJ:Policy.
• Chetty, R. (2009). "Sufficient Statistics for Welfare Analysis." AR Econ
• Hamilton Project on CTC anti-poverty effects
• National Academies on 2021 CTC
• Brookings primer on inequality measurement debate
• Troller-Renfree, S. et al. (2024). "The Effect of a Monthly Unconditional Cash Transfer on Children's Development at Four Years of Age." NBER w33844
• Heckman, J., García, J., & Baulos, A. (2024). "Perry Preschool at 50." NBER w32972

Information economics and privacy

• Bergemann, D. & Morris, S. (2019). "Information Design: A Unified Perspective." JEL.
• Kamenica, E. & Gentzkow, M. (2011). "Bayesian Persuasion." AER, NBER w15540
• Acquisti, A., Taylor, C., & Wagman, L. (2016). "The Economics of Privacy." JEL.
• Kenny, C. et al. (2021). "The use of differential privacy for census data and its impact on redistricting." Science Advances

Inequality and mobility

• Chetty, R., Friedman, J., Hendren, N., Jones, M., & Porter, S. (2018). "The Opportunity Atlas." opportunityatlas.org
• Chetty, R., Hendren, N., Jones, M., & Porter, S. (2020). "Race and Economic Opportunity in the United States." QJE.
• Derenoncourt, E., Kim, C., Kuhn, M., & Schularick, M. (2024). "Wealth of Two Nations." QJE
• Brookings 2024 Chetty et al. mobility update

Climate

• Stavins, R. (2017). "Lessons Learned from Three Decades of Experience with Cap-and-Trade." REEP.
• EIA on California cap-and-trade May 2024 auction
• Rennert, K. et al. (2022). "Comprehensive evidence implies a higher social cost of CO2." Nature.

Crypto and on-chain

• Angeris, G. et al. (various). "Constant function market makers."
• ESMA 2025 on MEV implications for crypto markets
• "Mechanism Design for Automated Market Makers." arXiv 2402.09357
• "RediSwap: MEV Redistribution Mechanism for CFMMs." arXiv 2410.18434

Development

• Banerjee, A. & Duflo, E. (2019). Good Economics for Hard Times. PublicAffairs.
• Deaton, A. (2010). "Instruments, Randomization, and Learning about Development." JEL.
• Egger, D., Haushofer, J., Miguel, E., Niehaus, P., & Walker, M. (2022). "General Equilibrium Effects of Cash Transfers." Econometrica.

Health and deaths of despair

• Case, A. & Deaton, A. (2020). Deaths of Despair and the Future of Capitalism. Princeton University Press.
• Friedman, J. & Hansen, H. (2024). Various papers on extension to Black and Native American mortality patterns.

Charter schools

• Angrist, J., Pathak, P., & Walters, C. (2013). "Explaining Charter School Effectiveness." AEJ:Applied.
• Cohodes, S., Setren, E., & Walters, C. (2021). "Can Successful Schools Replicate? Scaling Up Boston's Charter School Sector." AEJ:Applied.

Glossary appendix

A working vocabulary for the field. Terms are listed alphabetically.

Adverse selection — Pre-contract asymmetric information leading to a worse pool of contracting parties; the Akerlof Lemons setup.

Affiliated values — Bidders' values are positively correlated; a generalisation of common values used in Milgrom-Weber linkage principle.

ATE (average treatment effect) — $E[Y(1) - Y(0)]$ over the population.

ATT (average treatment effect on the treated) — $E[Y(1) - Y(0) | D=1]$ for those who actually received treatment.

Augmented synthetic control method (ASCM) — Ben-Michael-Feller-Rothstein 2021 method combining synthetic control with outcome regression for bias correction.

Bayesian persuasion — Kamenica-Gentzkow framework where a sender commits to an information disclosure rule to influence a Bayesian receiver.

Behavioural Insights Team (BIT) — UK government unit applying behavioural economics to policy; founded 2010.

BLP — Berry-Levinsohn-Pakes; the canonical structural model for differentiated-products demand estimation.

Bunching estimator — Saez 2010; Kleven-Waseem 2013; identifies behavioural elasticities from the mass of observations at notches/kinks in budget sets.

Callaway-Sant'Anna — 2021 JoE estimator for staggered DiD using group-time ATTs.

Cap-and-trade — Quantity-based pollution regulation: total emissions capped, allowances tradable.

Causal forest — Athey-Wager machine-learning estimator for heterogeneous treatment effects.

CCA (combinatorial clock auction) — Auction format used in UK, EU spectrum sales; iterates ascending clock prices on packages.

Consumer surplus — Area between demand curve and price, integrated over quantity.

Constant function market maker (CFMM) — A market-making protocol (Uniswap, Balancer, Curve) that prices trades to keep a function of reserves constant.

Continuous-type mechanism — A mechanism designed for agents with values drawn from continuous distributions; requires calculus to solve.

Cremer-McLean — 1985, 1988 papers showing that with correlated types, the seller can extract all surplus.

DA (deferred acceptance) — Gale-Shapley 1962 algorithm; produces stable matchings; strategy-proof for the proposing side.

Deaths of despair — Case-Deaton's term for deaths from drug overdose, alcohol abuse, suicide.

Default effect — Tendency for people to remain at the option pre-selected for them.

Differential privacy — Dwork et al.; mathematical framework guaranteeing bounded influence of any individual on published statistics.

DiD (difference-in-differences) — Estimator that subtracts a control's pre-post change from a treated unit's pre-post change to identify ATT under parallel trends.

Direct mechanism — Mechanism where each agent reports a type and outcome is determined by the type profile.

Dominant strategy — A strategy that is best regardless of others' strategies; truth-telling in second-price auctions.

ETI (elasticity of taxable income) — Percent change in taxable income for 1% change in retention rate; used in optimal-tax formulas.

EITC (Earned Income Tax Credit) — US wage subsidy for low-income working families.

Endogeneity — When a regressor is correlated with the error term, biasing OLS.

Event study — Empirical design tracking outcome differences relative to treatment onset.

Externality — Cost or benefit imposed on third parties; can be positive or negative.

Field experiment — Experiment conducted in a real-world (natural) context, sometimes invisible to subjects.

First-price auction — Highest bid wins, pays own bid; not truth-telling.

Fixed effects (FE) — Regression terms absorbing time-invariant unit characteristics (unit FE) or unit-invariant time characteristics (time FE).

Game theory — The study of strategic interaction; the math of multi-agent decision-making.

Generalized second-price (GSP) auction — Sponsored-search auction format; not truth-telling but has equilibria matching VCG outcomes.

Gini coefficient — Inequality summary statistic; 0 = perfect equality, 1 = perfect inequality.

Great Gatsby Curve — Krueger 2012; cross-country negative correlation between inequality and intergenerational mobility.

Hedonic regression — Regression of price on product attributes; used in CPI construction, real estate.

Heterogeneous treatment effects — Treatment effects that vary across individuals or contexts.

HHI (Herfindahl-Hirschman Index) — Concentration measure: sum of squared market shares (in percent).

Hicksian demand — Compensated demand; holds utility constant while varying prices.

Hidden action (moral hazard) — Post-contract asymmetric information about effort or behaviour.

Imputation estimator — Borusyak-Jaravel-Spiess; impute counterfactuals using untreated observations.

Incentive compatibility (IC) — Mechanism property: truth-telling is in each agent's interest.

Indifference curve — Set of bundles among which a consumer is indifferent.

Individual rationality (IR) — Mechanism property: each agent prefers participating to opting out.

Information design — Sender's optimal disclosure strategy given receiver's Bayesian inference; Bayesian persuasion.

Instrument variable (IV) — Variable correlated with treatment, uncorrelated with outcome shocks, that identifies causal effects.

Intergenerational elasticity (IGE) — Slope of child's income on parent's income; measure of mobility.

JTPA (Job Training Partnership Act) — US training program; subject of classic RCT evaluation by Heckman et al.

Killer acquisition — Acquisition intended to discontinue a competing product.

Kill zone — Hypothesised effect where incumbent platform's acquisition threat discourages startup investment.

Kidney exchange — Mechanism design for matching incompatible kidney donor-recipient pairs.

Land of opportunity — High-mobility geographic areas in Chetty's Opportunity Atlas.

LATE (local average treatment effect) — Treatment effect for compliers in an IV setting; Imbens-Angrist 1994.

Lemons — Akerlof's bad-quality cars in his 1970 market collapse model.

Linear quadratic model — Standard tractable framework for dynamic optimisation in macroeconomics.

Linkage principle — Milgrom-Weber 1982; public information disclosure can increase auction revenue.

Local-linear regression — Nonparametric estimation using a linear model fit within a bandwidth.

Loss aversion — Asymmetric weighting of losses (heavier) vs equivalent gains.

Many weak instruments — IV setup where each instrument is weak but many are available; requires specific inference adjustments.

Marginal treatment effect (MTE) — Heckman-Vytlacil; treatment effect at a particular point in the distribution of unobserved heterogeneity.

Market design — Engineering branch of microeconomics; designing rules of allocation mechanisms.

Matching with contracts — Hatfield-Milgrom 2005; generalisation of two-sided matching to richer terms.

Mental accounting — Thaler 1985; people segment money mentally rather than treating it as fungible.

Mechanism design — Theory of designing rules of interaction to achieve desired outcomes given strategic agents.

MEV (maximum extractable value) — Profit extractable through transaction ordering by block builders on a blockchain.

Minimum wage — Legal floor on hourly wage.

Monopsony — Single buyer; in labour markets, employer wage-setting power above competitive equilibrium.

Multi-homing — Users participating on multiple competing platforms simultaneously.

Mirrlees — 1971 optimal income tax model; foundational.

Nash equilibrium — Strategy profile where no player can improve by unilateral deviation.

No Excuses model — Charter school pedagogical approach: strict discipline, extended day, high expectations.

Noncompete clause — Employment contract restricting employee's ability to work for competitors after leaving.

Nudge — Choice architecture intervention that steers behaviour without restricting options.

Optimal mechanism — Myerson 1981; revenue-maximising mechanism in an auction setting.

Parallel trends — DiD identifying assumption: counterfactual trends would have been the same in treated and control units.

Pareto efficiency — Allocation where no one can be made better off without making someone worse off.

Pareto parameter — Tail thickness parameter of the Pareto distribution; ~1.5 for US income top tail.

PBS (proposer-builder separation) — Ethereum architecture separating block proposal from block construction.

Personalised pricing — Setting different prices to different customers based on observed data.

Perry Preschool — 1962–67 RCT of high-quality preschool for low-income Black children; long-run follow-ups.

Pigouvian tax — Tax equal to the marginal external cost of the activity; corrects externalities.

Platform — A two-sided market intermediary connecting groups whose interactions create network value.

Power posing — Discredited behavioural finding that body posture affects hormones; failed replications.

Present bias — Disproportionate weighting of immediate vs delayed payoffs; hyperbolic discounting.

PSZ — Piketty-Saez-Zucman; collaboration on US inequality and capital accounts.

Quantile regression — Estimation of conditional quantiles (median, 75th percentile) rather than mean.

Quasi-experiment — Natural variation that approximates experimental conditions without random assignment.

RD (regression discontinuity) — Causal identification via discontinuity in conditional means at a treatment threshold.

RCT (randomised controlled trial) — Experiment with random assignment to treatment vs control.

Replication crisis — Discovery in the 2010s that many published findings in psychology and behavioural economics failed to replicate.

Revealed preference — Inferring preferences from observed choices.

Revelation principle — Any equilibrium of any mechanism can be replicated by a direct mechanism in which truth-telling is the equilibrium.

Revenue equivalence — Symmetric IPV auctions with efficient allocation and zero rent at the lowest type yield same expected revenue.

RGGI (Regional Greenhouse Gas Initiative) — Northeast US cap-and-trade for power-sector CO2 since 2009.

RGSP (Randomized Generalized Second Price) — Google's randomised sponsored-search auction variant.

Risk dominance vs payoff dominance — Two equilibrium-selection criteria in coordination games.

Roth-Sotomayor — Canonical textbook on two-sided matching theory.

Saez — Emmanuel Saez; key empirical and theoretical contributor to optimal taxation.

Sample selection — Heckman 1979; selection on outcomes generates bias unless modelled.

Sandwich attack — MEV strategy: front-run a user's trade, then back-run to extract value.

Search frictions — Costs of finding trading partners; foundational to modern labour markets.

Second-price auction — Highest bid wins, pays second-highest bid; truth-telling dominant.

Selection effect — Difference between treated and untreated groups due to selection rather than treatment.

Self-control / time inconsistency — Strotz 1955; preferences over consumption that change with the passage of time.

Sherman Act §2 — US antitrust statute prohibiting monopolisation; Google found liable Aug 2024.

Signaling — Spence 1973; informed sender uses costly action to communicate type.

Sufficient statistics — Chetty 2009; welfare formulas expressed in estimable elasticities.

Synthetic control — Abadie; comparison unit constructed as weighted average of controls matching pre-period trajectory.

Spectrum auction — FCC auctions for wireless licences; foundational case for auction design.

Tipping — Network market converging to single platform.

Top trading cycles (TTC) — Pareto-efficient matching mechanism; strategy-proof but not stable.

Two-sided market — Market with at least two distinct user groups whose participation creates value for each other.

TWFE (two-way fixed effects) — Regression with unit and time fixed effects; broken under heterogeneous staggered DiD.

Vickrey-Clarke-Groves (VCG) — Efficient mechanism with dominant-strategy truth-telling; theoretically beautiful, practically fragile.

Welfare economics — Analysis of social welfare implications of policies.

Winner's curse — Wilson 1977; the auction winner is the most optimistic bidder, hence on average too optimistic.

Zero-rated — Pricing $0 on one side of a two-sided market; common in platforms.

End of report.

Appendix: extended discussion of selected frontier topics

A. The 2024–2026 antitrust trilogy in greater depth

The Google, Amazon, and Meta cases — taken together — constitute the most important set of antitrust decisions in a generation. They are worth understanding in detail because the economic theories they ratified or rejected will shape the next two decades of digital platform regulation.

A.1 Google search — what the court ratified

Judge Mehta's opinion in United States v. Google (D.D.C. Aug 5, 2024) accepted several economic propositions that had been argued in academic literature but not previously embraced by US courts:

1. Defaults matter in digital markets. The trial established that Apple Safari's default search setting determined where ~50% of mobile search queries went. Google paid Apple approximately $20 billion per year for that default position. The court accepted that paying for defaults at this scale is not "competition on the merits" but rather a foreclosure mechanism — Google's rivals (Bing, DuckDuckGo) could not match the payment, so they were excluded from the default position regardless of product quality.

2. Scale effects and the search-quality flywheel. The court accepted Google's own internal evidence that search quality improves with query volume (more queries → more data → better results → more users → more queries). This is a classic learning-curve / network effect on the user side. The implication: once a search engine has scale, it has structural advantage that newer entrants cannot overcome without preferential distribution channels — which Google had foreclosed.

3. Specialized vertical search markets. The court found that Google held monopoly power in general search and in general search text advertisements (the slightly different market where Google's ads compete for placement). This dual-market finding mattered for damages calculations and remedies design.

4. Indirect distribution as exclusion. Mehta accepted that excluding rivals from default positions, not direct exclusion from the search market, was the violation. This is conceptually similar to the Microsoft (2001) treatment of operating-system tying — both cases involved excluding rivals from contested distribution chokepoints.

The remedies decision in September 2025 was milder than DOJ requested. The court rejected divestiture of Chrome (DOJ's headline request). It ordered Google to share certain syndication and search data with rivals and prohibited certain exclusive default-distribution payments. The reasoning: divestiture is appropriate only when less drastic remedies cannot restore competition. The court judged data-sharing sufficient to give rivals scale opportunities while preserving Google's economic structure.

Economically, this is a "structural-conduct hybrid" remedy. It mirrors the EU DMA's approach of mandating interoperability and data portability rather than breaking up firms. Whether it actually restores competition is the empirical question that will dominate the next five years.

A.2 Amazon — the unsettled case

The FTC v Amazon complaint (Sept 2023) advances a theory of monopolisation that is more aggressive than the Google search theory. The core allegations:

1. Anti-discounting punishment. Amazon's algorithm "demotes" sellers from the Buy Box (the default purchase button) when those sellers offer their products at lower prices on competing sites. This pressures sellers to price-match Amazon, raising prices across the broader retail market.

2. Project Nessie pricing algorithm. An internal Amazon algorithm that systematically tested how rivals responded to Amazon's price increases. If rivals followed Amazon's price up, Amazon kept the higher price; if they didn't, Amazon backed down. The FTC argues this is conscious coordination to raise prices.

3. Tying Prime to FBA. To qualify products for the "Prime" badge (which drives a substantial sales boost), sellers must use Amazon's Fulfillment by Amazon (FBA) logistics service. The FTC argues this is an unlawful tie, extending Amazon's marketplace power into the fulfillment market.

Judge Chun's October 2024 ruling on Amazon's motion to dismiss let most of these theories through to trial. The trial is scheduled for 2026 and will be decided under the consumer-welfare standard, with the FTC needing to show consumer harm from each practice.

Economically, the Amazon case tests the Brandeisian-leaning theory of platform power — that a platform can be too powerful for non-pricing reasons (information control, terms of service, ecosystem lock-in) even if prices appear competitive. If the FTC wins, it will substantially expand antitrust scope. If it loses, it will narrow it back to the post-Bork consumer-welfare frame.

A.3 Meta — what the FTC loss teaches

The November 2025 Meta verdict reveals a structural constraint on antitrust against digital platforms: market definition. The FTC's case rested on a narrow "Personal Social Networking Services" market (Facebook, Instagram, Snapchat). Judge Boasberg held that this definition was untenable because user behaviour shows substantial substitution between social-networking services and short-form video (TikTok, YouTube Shorts).

The economic principle at work is the hypothetical monopolist test (the SSNIP test in the 2010 and 2023 Merger Guidelines): would a hypothetical monopolist of the candidate market be able to profitably raise prices by 5% above competitive levels? In a digital market with $0 user pricing, the test pivots to substitution of attention and engagement. When TikTok captures a substantial share of user attention historically allocated to Facebook/Instagram, the court ruled, TikTok must be in the same antitrust market.

This is a methodological problem for digital antitrust. If "the market" is defined by what users can substitute to, almost any digital service is in the same market — gaming, messaging, video, news. Market definition becomes a binding constraint that lets dominant firms argue their dominance is illusory.

A possible response (some neo-Brandeisians' position) is to abandon the market-definition step in some cases, focusing instead on conduct — exclusive dealing, predatory acquisitions, refusal to interoperate — without needing to first define a market. The 2023 Merger Guidelines move slightly in this direction. Whether courts will follow remains to be seen.

B. The credibility revolution applied to platform policy

B.1 Causal estimation in digital markets

A key methodological challenge: many digital-economics questions (does Facebook deactivation improve wellbeing? do recommender systems polarise users? do gig drivers prefer flexibility to employment?) are hard to answer without RCTs because observational data is plagued by selection.

The 2024 frontier of credible empirical platform research:

• Allcott-Braghieri-Eichmeyer-Gentzkow 2020 "The Welfare Effects of Social Media": RCT in which participants were paid to deactivate Facebook before the 2018 midterm elections. Deactivation increased subjective wellbeing, reduced political polarisation, and reduced news knowledge. Most participants who tried it valued not having Facebook at $100+/month after the experiment.

• Bursztyn-Handel-Jiménez-Durán-Roth 2024 on TikTok and Instagram: students would pay to coordinate everyone off the platform but not to deactivate themselves alone. This is a coordination failure / unraveling model — the platform is welfare-reducing in equilibrium because of strategic complementarities in usage.

• Mosquera et al. 2020 RCT on Facebook deactivation in Spain: lower platform use during the experiment was associated with less stress, lower anxiety, and worse short-term news knowledge.

• Levy 2021 AER on Facebook's effect on news consumption and political views: a field experiment showed that exposure to outlets with opposing political views (vs. one's own preferred outlets) did not increase polarisation, contradicting some echo-chamber claims.

The collective body of evidence is that social media use is welfare-reducing for substantial subgroups, that exposure to opposing views doesn't reliably reduce polarisation, but that the welfare cost of platforms is genuine and not just rhetorical. Whether this implies regulation is a normative question; the empirical inputs are increasingly credible.

B.2 Platform recommender system effects on welfare

The 2024 empirical literature on recommender systems is split:

• Concentration effects: recommender systems concentrate attention on a small set of popular items, generating a heavier-tailed earnings distribution than would be expected from underlying quality variance alone.
• Discovery effects: in some platforms (Spotify, TikTok), recommenders surface content users would not otherwise have found, expanding niche access.
• Polarisation effects: contested. Some evidence (Bakshy-Messing-Adamic 2015 Science on Facebook) suggests recommenders amplify pre-existing preferences. Other evidence (Allcott-Gentzkow various) suggests echo-chamber effects are smaller than public discourse implies.

The economic model that frames this is attention as a scarce resource. Each user has a finite attention budget; platforms compete to capture it; consumers (and policymakers) cannot easily observe the opportunity cost of attention spent. This is the central economics of the "attention economy" and the 2024 frontier of platform welfare analysis.

C. The gig economy as a worker-classification natural experiment

The gig economy provides a rich set of natural experiments on labour market regulation. California's Assembly Bill 5 (AB5), passed September 2019, reclassified most gig workers as employees. Proposition 22, passed November 2020, then partly reversed AB5 for app-based delivery and rideshare drivers, treating them as independent contractors with some benefit floors. The California Supreme Court upheld Prop 22 in July 2024.

The economic evidence:

• Cook-Diamond-Hall-List-Oyer 2021 ReStud "The Gender Earnings Gap in the Gig Economy": studied Uber drivers, found that male drivers earn ~7% more per hour due to (i) experience, (ii) preferences over locations and times, (iii) driving speed. No discrimination; gender pay gaps in flexible-hours platforms can persist for non-discriminatory reasons.
• Hall-Krueger 2018 ILR Review on Uber driver earnings: earnings vary substantially by city; net of expenses, hourly earnings are often below typical alternatives for low-credentialed workers.
• Berger-Chen-Frey 2018, Chen-Sheldon 2016 on supply elasticity: gig drivers' supply responds substantially to surge pricing, suggesting genuine intertemporal substitution.

The classification question — employee vs contractor — has profound economic implications:

• Employees receive employer-sponsored health insurance, employer payroll tax payment, workers' comp, unemployment insurance, minimum wage and overtime protection, paid leave.
• Contractors receive flexibility, no employer-side tax burden on the company, no employment protections.

Most gig drivers, in surveys, prefer contractor status for the flexibility — though this preference may reflect lack of awareness of foregone protections. The DOL's March 2025 FLSA rule re-tightened the classification test, while California's Prop 22 carve-out remains. The legal patchwork makes it hard to learn from clean cross-state comparisons.

The economic conclusion: gig classification has substantial distributional consequences. The savings to platforms from classifying workers as contractors are real and large; the question is whether they accrue to consumers (lower prices) or shareholders (rents). Most evidence suggests the latter.

D. The differential privacy debate beyond the Census

The 2020 Census differential privacy controversy was the first salvo. The broader question — how to make administrative data useful for research and policy while protecting individual privacy — is a major frontier.

The economic stakes:

• Administrative data (IRS tax returns, Medicare claims, K-12 student records, criminal records) is enormously valuable for research. Chetty-Saez-Friedman-Hendren-Jones built modern intergenerational mobility research on linked IRS-Census data.
• Privacy concerns are real. With sufficient auxiliary data, even "anonymised" datasets can be re-identified (Sweeney 1997 famously re-identified the Massachusetts governor from "anonymised" health records).

The differential privacy framework provides formal guarantees. But there are tradeoffs:

• Noisy statistics may be biased in ways correlated with policy-relevant variables (the Census case for racial heterogeneity).
• Federated learning and other privacy-preserving computation protocols allow analysis without raw data exposure but at the cost of computational and statistical complexity.
• Synthetic data (machine-generated data with similar statistical properties) preserves aggregate patterns but may not preserve causal relationships needed for treatment-effect estimation.

The IRS administrative data Chetty et al. have used illustrates one model: data resides at the IRS, researchers run code on the data without seeing individual records, output is disclosure-reviewed. This works for small, vetted research projects but doesn't scale to a broader research community.

The economic question (formalised in Acemoglu-Makhdoumi-Malekian-Ozdaglar 2022 AER:I): when one person's data reveals information about others (because of correlations), the social value of each person's data exceeds their private value. Without coordination, individuals under-supply data to platforms; with coordination, privacy may be undersupplied if individuals don't internalise externalities. The 2024 frontier formalises this as a mechanism-design problem on data markets.

E. AI and the future of micro

The penetration of generative AI into economic research and policy is reshaping micro in three ways:

1. AI as research subject. Horton's homo silicus program is one strand. Others use LLMs to simulate consumer choice (Brand-Israeli-Ngwe 2023), to play game-theoretic games at scale, to generate synthetic survey responses for pilot studies. The methodological challenge: LLMs reproduce patterns from training data, which may include biased or non-representative human responses.

2. AI as research method. Machine learning has entered econometrics through double machine learning (Chernozhukov et al. 2018), causal forests (Wager-Athey 2018), and BCF (Hahn-Murray-Carvalho 2020). These methods enable heterogeneous treatment effect estimation at scale. The discipline is now standardising how to report and validate these approaches.

3. AI as labour market shock. Brynjolfsson-Li-Raymond 2023 on call centres; Noy-Zhang 2023 on writing tasks; Peng-Kalliamvakou-Cihon-Demirer 2023 on coding — early evidence shows productivity gains of 14-56% from generative AI assistance, with bigger gains for lower-skill workers. This pattern (compression of skill gaps) is opposite to the SBTC pattern that explained 1980-2010 wage inequality dynamics. Whether the compression scales beyond the studied contexts is the open question.

The 2025-2030 research frontier will likely see AI-augmented research at scale, with both productivity gains and new credibility challenges. Pre-registration, replication, and audit norms need to adapt to a world where empirical results can be produced rapidly with AI assistance and where adversarial use of AI tools may game peer review.

F. The frontier of climate microeconomics

Beyond the cap-and-trade vs Pigouvian-tax debate, the 2024-2026 frontier of climate microeconomics includes:

• Border carbon adjustments (CBAM): The EU's Carbon Border Adjustment Mechanism (transition 2023, full implementation 2026) imposes carbon prices on imports based on their embodied carbon. The economic literature on CBAMs (Cosbey, Felder et al.) is developing rapidly.
• Subsidies-as-policy: The US Inflation Reduction Act of 2022 represents a shift from carbon-pricing to massive clean-energy subsidies (~$370B over 10 years). The economic-efficiency case for subsidies vs taxes is contested, with the Stern-Stiglitz position favoring subsidies under political-economy constraints.
• Climate adaptation economics: Heat, sea-level rise, agricultural disruption. The Burke-Hsiang-Miguel (2015 Nature) macroeconomic damage function is the standard; micro-level work on residential adaptation, agricultural choice, labour productivity in heat is active.
• Negative emissions and carbon removal: New markets are forming for direct-air-capture, biochar, enhanced weathering. Mechanism design questions parallel the 1990s SO2 cap-and-trade market.
• Distributional effects: Carbon pricing is regressive in its incidence (poor households spend more of income on energy). Climate damage is regressive in incidence (poor households more exposed to heat, flooding, displacement). Policy design balancing these is the live distributive-justice question.

G. The Roe-Dobbs natural experiment

The June 2022 Dobbs v. Jackson Women's Health decision overturned Roe v. Wade, returning abortion regulation to states. This generated a natural experiment on the labour-supply, fertility, marriage, and educational effects of abortion access — variables previously studied only through interstate variation pre-Dobbs.

Early evidence (2023-2025):

• Fertility: state-level fertility rates show measurable upticks in trigger-law states (Texas, Mississippi, others) — 1-3% increases in birth rates relative to non-trigger states.
• Labour force participation: women in trigger-law states show small declines in labour force participation; effects are larger for younger and lower-educated women.
• Education: enrollment and completion patterns showing early signs of disruption; longer follow-up needed.

This is one of the cleanest natural experiments on the economic effects of reproductive policy ever available. The full evidence will take a decade to mature, but early results are consistent with the cross-state pre-Dobbs literature (Levine, Donohue-Levitt) on the substantial economic effects of abortion access.

H. The persistent puzzle of charter school scaling

Boston's charter sector is unusual: it expanded substantially without losing effectiveness. Most successful educational interventions fade or fail when scaled. The Boston exception is worth examining for what it teaches about scalability.

Possible explanations:

• Selection of effective school leaders: Boston's charter sector recruited educators with strong track records and gave them autonomy.
• Pedagogical coherence: the "No Excuses" model is well-defined, easy to monitor, and ports between schools.
• Strong local accountability: Boston has aggressive student-level data tracking, lottery-based admissions, and clear standards for performance.
• Structural constraints absent: many barriers to scaling (teacher union contracts, district facility constraints, transportation funding) were limited in Boston's charter context.

The hard question: can the Boston model be exported to cities without these conditions? Replications in Newark, New Orleans (post-Katrina), Washington DC show some success but with mixed magnitudes. The honest synthesis: charter effectiveness is highly context-dependent, and scaling has narrow operational tolerances.

I. The frontier of inequality measurement

Beyond PSZ vs Auten-Splinter, several methodological questions in inequality measurement are active:

• Top-coding and survey limitations: CPS, SCF, and other surveys top-code high incomes for privacy, biasing top shares downward. Linking surveys to administrative data (as in DINA accounts) addresses this but raises privacy concerns (back to differential privacy).
• Wealth measurement: the Saez-Zucman (2016) "capitalisation method" infers wealth from capital income, which assumes constant returns across wealth levels. Alternative approaches (estate-tax-based, survey-based) give different totals.
• Lifetime vs annual inequality: people experience income shocks; annual inequality overstates lifetime inequality because some bottom incomes are transitory. Kopczuk-Saez-Song (2010) show lifetime inequality has risen less than annual.
• Consumption inequality: typically rises less than income inequality, perhaps because of credit access. Aguiar-Bils (2015) and the Meyer-Sullivan series show consumption inequality has been more stable.

These methodological choices have substantial policy implications. The headline number for "is inequality rising" is sensitive to the methodological choice in ways that can be lost in popular discussion.

Research conducted May 2026 based on web sources, NBER working papers, and primary economic literature.