The People Were Wrong

The model said done. The voters said no. Eight years of data, all in one direction. The model was precise. The model was pointed at the wrong thing.

Cedric Atkinson

October 7, 2016. A recording surfaces of Donald Trump describing sexual assault. Within forty-eight hours, dozens of elected officials from his own party withdraw their endorsements. The editorial consensus is immediate and unanimous. The model says: done.1

He won.

December 18, 2019. The House of Representatives impeaches him for abuse of power and obstruction of Congress. The model says: damaged beyond recovery.2

His approval held.

January 6, 2021. A mob storms the United States Capitol. Congressional members shelter under desks. Five people die. The model says: permanently disqualified. No candidate survives this.3

He ran again.

Between March 2023 and August 2024, four separate grand juries returned four separate indictments. Ninety-one felony charges across four jurisdictions. The model says: unprecedented legal exposure will end his viability. No candidate in American history has faced a single felony indictment, let alone ninety-one.4

Trump's Republican primary support rose from 43 percent before the first indictment to 52 percent one week after it. By the time all four indictments had accumulated, he was polling at 59 percent. Each indictment was a new variable in the model. Each confirmed the model's assessment that Trump was fatally compromised. Each made the model's prediction more confident.5

Trump won the second time with a larger share of the vote. He carried all seven swing states and won the popular vote, something he had not done in 2016.6

The model did not fail once. The model failed repeatedly, in the same direction, for the same reason, across eight years.

The model failed on measurement, not prediction. The model was precisely measuring something. That something did not determine the outcome.

The wrong variable

The model that predicted defeat across eight years was tracking a specific set of inputs. Scandal. Character. Legal exposure. Media coverage. Expert endorsement. Approval rating. Every one of these variables was measured with increasing precision. Every one moved in the direction the model predicted. Every one said the same thing: this candidate cannot win.

The voter was tracking a different set of inputs.

In January 2020, a dozen eggs cost $1.46. By December 2024, the same dozen cost $4.15. That is a 184 percent increase. A gallon of milk rose 26 percent. A pound of white bread rose 41 percent. A pound of ground beef rose 44 percent. The 2022 increase alone, 11.4 percent for food at home, was the fastest annual rise since 1979.7

+48% Median home price increase
2019–2024
+22% Median income increase
same period
Median single-family home prices vs median household income. Source: Harvard Joint Center for Housing Studies, 2024.8

Home prices rose at more than twice the rate of income. The ratio of home price to household income reached five to one, a figure that historically sat at three to one. By 2024, a household needed to earn $106,731 per year to afford the median-priced home. The actual median household income was $83,730. The gap between what a family earned and what a home cost was $23,000 per year.9

Between 2000 and 2014, net productivity in the American economy grew 21.6 percent. Median worker compensation grew 1.8 percent. The median worker captured eight percent of the economy's productivity growth. The rest went to higher earners and capital owners.10

21.6% Net productivity growth
2000–2014
1.8% Median compensation growth
same period
Source: Economic Policy Institute. The gap is driven by rising inequality in compensation and a falling share of income going to workers relative to capital owners.10

The model tracked scandal. The voter tracked the grocery bill. The model tracked approval ratings. The voter tracked whether the paycheck bought what it used to. One set of variables was measured with increasing precision. The other set had no field in the model.

On the eve of the 2016 election, Gallup measured Trump's favorability at 36 percent. His unfavorable rating was 61 percent. He received 46.1 percent of the vote. Ten points of the electorate viewed him unfavorably and voted for him anyway. No candidate in the history of modern polling had won the presidency with numbers like these.11

The voter who disapproves of a candidate and votes for that candidate is not irrational. The voter is answering two different questions. The poll asks: do you approve of this person? The ballot asks: which direction? Approval is a character assessment. The vote is a conditions assessment. They are different variables.12

The model had accurate data. It was measuring the wrong thing.

This is the same mechanism at work in every rating system that produces confidence without accuracy. A school rating precisely measures test scores. Test scores primarily measure family income, not teaching quality. The rating is accurate. It is pointed at the wrong variable. Families pay $200,000 more for a house near the higher-rated school. They purchased a number.13

The poll is the same instrument applied to democracy. It precisely measures approval filtered through scandal exposure. The vote measures lived experience filtered through economic conditions. Both instruments are accurate. Both are pointed at the wrong thing.

The mechanism is 250 years old

The American colonies generated significant trade revenue for the British Crown. The Treasury could calculate the economic benefit of Empire to the colonies. The model said: the colonists benefit from British rule. The data was clear.14

The colonist was measuring something the Treasury had no field for: agency, representation, whether anyone in the system heard them.

King George III addressed Parliament on October 26, 1775. He described the rebellion as the work of people who had "too successfully laboured to inflame My People in America, by gross Misrepresentations." He assured Parliament that many colonists "may still retain their Loyalty" but "the Torrent of Violence has been strong enough to compel their Acquiescence." The people were being manipulated. They did not really want what they were asking for. The model was clear. The agitators had confused them.15

Thomas Hutchinson, the former colonial governor, called the Declaration of Independence "a list of imaginary grievances" supported by "false and frivolous reasons."16

The model measured aggregate economic benefit. The colonist measured representation. The model had no category for what the colonist was measuring.

Bengal, 1943

The British administration of India tracked infrastructure output. Railways built. Roads laid. Administrative systems maintained. The model reported a functioning colonial government.17

Three million people starved to death in Bengal in 1943. India exported more than 70,000 tons of rice between January and July of that year, as the famine was killing people. British authorities stockpiled food for military use and exported quantities to forces in the Middle East. Researchers studying six major Indian famines between 1873 and 1943 found Bengal was the only one not linked to drought or crop failure. The food existed. It was directed elsewhere.18

High-ranking officials began requesting food imports through government and military channels in December 1942. Churchill's War Cabinet rejected or reduced these requests for months. Leo Amery, the Secretary of State for India, recorded in his diary that Churchill saw the famine as Indians' fault for "breeding like rabbits." Viceroy Wavell recorded a "peevish telegram" from Churchill asking why Gandhi had not died yet.19

The model measured administrative output. The citizen measured whether they could eat. The model reported a functioning system. The experience was three million dead.

Moscow, 1991

The CIA estimated Soviet gross national product at 49 percent of America's in 1990. The Soviet Union's own statistical agency, Goskomstat, estimated 58 percent. Senator Daniel Patrick Moynihan wrote in 1990 that the United States had "hugely overestimated both the size of the Soviet economy and its rate of growth" for forty years.20

CIA analysts relied on Soviet production statistics. These statistics reported what factories said they produced. They did not report what appeared on shelves. The model measured production. The citizen measured consumption. The two numbers described different realities.21

The system did not collapse because the CIA revised its estimate. The system collapsed because the variable that mattered, whether people could buy bread and shoes and soap, was not the variable being tracked.

London, 2016

The Remain campaign modeled the aggregate economic benefit of EU membership. GDP contribution. Trade flows. Financial services access. The model was clear: leaving would cost the economy billions.22

The Leave voter was measuring something else. Whether the town still had a high street. Whether local manufacturing had survived. Whether the national economic data described anything they recognized in their own lives.

Phone polls, the establishment's preferred methodology, predicted Remain would win by 2.6 percentage points. Seventy-eight percent of phone polls predicted a Remain victory. Every single final poll overstated the Remain vote share. Leave won 51.9 to 48.1. The early results from North East England showed Leave performing far better than anyone had modeled.23

Event Model's variable Voter's variable Why the model missed
American Revolution, 1776 Empire's aggregate economics Colonist's agency No field for representation
Indian Independence, 1947 Raj's infrastructure output Citizen's dignity No field for self-determination
Fall of USSR, 1991 Reported Soviet GDP Shelf inventory Measured production, not consumption
Brexit, 2016 Aggregate EU benefit Town-level decline Measured national average, not distribution
Trump, 2016 Scandal, approval, media Economic conditions Measured character, not conditions
Trump, 2024 Indictments, legal exposure Grocery bill, system trust Added more of the wrong variable

The last row is the sharpest. Between 2016 and 2024, the model did not change what it was measuring. It added more of the same kind of measurement. Legal exposure. Felony charges. January 6th. Every addition made the model more confident. The voter's variable, the one the ballot actually measures, did not change. It was the same variable in 2024 that it was in 2016. Does my life work inside this system?

The model kept refining its measurement of the thing that did not matter. The voter kept measuring the thing that did.

The policy produced both numbers

The aggregate model was not lying. GDP grew. Unemployment fell. The stock market reached record highs. These numbers were real.

The grocery bill rose 26 percent. Home prices outpaced income two to one. The median worker captured eight cents of every dollar the economy produced in additional productivity. Trade deals that raised GDP closed factories. Low interest rates that inflated asset prices priced out every family that did not already own a home. These numbers were also real.

Between 2000 and 2024, real GDP grew 66 percent. Over the same period, manufacturing employment fell 27 percent. The economy added $9.4 trillion in output. It lost 4.6 million factory jobs. Between 2000 and 2007 alone, trade with China eliminated between 1.5 and 2 million of those jobs.24

In Macomb County, Michigan, manufacturing employment collapsed from 106,000 jobs to 52,000 in a single decade. A 52 percent decline. Median household income fell from $72,000 to $53,000. The county voted for Obama twice. Then Trump three times.25

In Erie County, Pennsylvania, the pattern was the same. A blue-collar county, whiter and poorer than the state or the country. Obama won it twice. Trump took it in 2016 by fewer than 2,000 votes, lost it in 2020 by fewer than 1,500, and won it back in 2024. Manufacturing is one of the county's three largest employment sectors. The national GDP number did not describe anything they recognized.26

The trade deal raised GDP. The trade deal closed the factory. The same policy produced both numbers. The model tracked one. The voter experienced both.

Low interest rates grew household wealth for homeowners. The same low interest rates priced out every family that had not yet purchased a home. The model measured the wealth effect. The voter measured whether they could afford a house.

The establishment is not being contradicted by ignorant voters. The establishment is being contradicted by the consequences of its own policies, measured at a level its model does not track.

The aggregate number says the economy grew. The local number says the grocery bill doubled. Both are accurate. One is tracked. One is lived. The gap between them is not a prediction error. The gap is what the vote is measuring.43

Precise instruments, wrong variable

The polls did not fail technically. After 2016, pollsters introduced education weighting to correct for the undersampling of non-college white voters. They improved likely-voter screens. They adjusted for social desirability bias. The methodology became more sophisticated.27

The 2020 polls were worse. They missed by a larger margin than 2016, still in the same direction. The 2024 polls improved, missing by 2.4 percentage points nationally and 2.6 in swing states. In every swing state, Trump outperformed his final polling number by approximately three points. The error was smaller. The direction was the same.28

Three consecutive election cycles. Three rounds of methodology refinements. The same directional miss every time.

The precision of the instrument does not matter if the instrument is pointed at the wrong variable.

The poll measures who answers. Six percent answer. The poll asks them a character question: do you approve? Do you find the candidate favorable?

The ballot asks a different question. Does my life work? Is it better than it was? Can I afford what I could afford four years ago?

In 1997, 36 percent of people called by pollsters picked up the phone. By 2018, 6 percent did. The polls are not surveying a representative sample of the electorate. They are surveying the 6 percent who agree to be surveyed, then weighting the results to represent the other 94 percent. Whether the 6 percent who answer resemble the 94 percent who do not is the question the model cannot test.29

The voter who disapproves in the poll and votes for the candidate anyway is not hiding. The voter is a person whose actual variable, conditions, does not match the variable the poll is asking about, character. They do not answer the poll because the poll is not asking their question. Their absence from the data is the data.30

The academic literature on economic voting has documented this for decades. Douglas Hibbs built a model that predicts presidential election outcomes using two variables: per capita real income growth and military fatalities. Zero character variables. Zero scandal variables. It outperformed twenty-two alternative models. In the 2024 exit polls, two-thirds of voters described the economy as "bad." Seventy-three percent said they were angry or dissatisfied with the country's direction. The variable that predicted the outcome was the one the polling model was not tracking.31

When the establishment listened

In 1944, the consensus model for postwar Germany was punishment. Treasury Secretary Henry Morgenthau proposed transforming Germany into a pastoral agricultural economy. Strip the Ruhr of industrial capacity. Deindustrialize the country. Herbert Hoover investigated the proposal and concluded it would result in the starvation of up to 25 million Germans.32

The model said: punish. The local reality said: a destroyed continent sliding toward communism. The establishment updated the model.

The Marshall Plan invested $13.3 billion, equivalent to roughly $137 billion in today's dollars, in European reconstruction. Five percent of the American economy at the time. Between 1948 and 1953, Western European GDP grew more than 30 percent. Industrial production increased 35 percent. It was the fastest period of growth in European history.33

The model changed because the establishment treated the contradiction between its punishment framework and the reality on the ground as information, not as evidence that the Europeans were ungrateful or that the locals did not understand their own situation.

In April 1961, President Kennedy authorized an invasion of Cuba at the Bay of Pigs. The decision was supported by a near-unanimous consensus among his advisors. The operation was a disaster.34

Thirteen months later, the Soviet Union placed nuclear missiles in Cuba. Kennedy convened the Executive Committee of the National Security Council. He structured the process differently. He deliberately cultivated dissent. Robert Kennedy challenged his colleagues to consider the moral dimension of a military strike, equating a first strike on Cuba with Japan's attack on Pearl Harbor. The Joint Chiefs and a majority of the committee favored an attack. Kennedy chose a blockade.35

Had he accepted the recommendation that the consensus supported, the historical analysis concludes he would have precipitated a nuclear war. Soviet tactical nuclear weapons were already on the island, a fact unknown to American intelligence at the time. Moscow had ordered they not be used, but lacked the technical controls to prevent it.36

The people did not change between April 1961 and October 1962. The same advisors sat in the same rooms. What changed was the response to contradiction. At the Bay of Pigs, consensus was treated as confidence. During the missile crisis, consensus was treated as a signal that something might be missing from the room.

Sub-Saharan Africa, 2003

Twenty-five million people in sub-Saharan Africa were living with HIV. Antiretroviral treatment existed but cost more than $12,000 per patient per year. The model measured complexity and concluded the problem exceeded the capacity of intervention.37

On January 28, 2003, George W. Bush announced the President's Emergency Plan for AIDS Relief in his State of the Union address. Fifteen billion dollars over five years. The largest commitment by any nation to combat a single disease in history. Generic production brought treatment costs below $300 per patient per year. Then below $45.38

The program has saved an estimated 25 million lives. HIV-related deaths in PEPFAR-supported countries fell roughly 59 percent. It was reauthorized under Obama, maintained under Trump, continued under Biden. Four administrations. Both parties. Twenty-two years. The medicine was always there. The model had measured complexity. The contradiction was a single fact: the treatment exists.39

Medellín, 2004

In 1991, Medellín recorded 6,809 homicides. A rate of 416 per 100,000. The highest of any city in the world. The model measured crime and deployed resources against crime.40

Sergio Fajardo, a mathematician, won the mayor's office in 2004. He measured something the model did not track. Not the crime rate. Whether anyone was investing in the neighborhoods where the crime rate was highest. He built library-parks in the most dangerous comunas. Not police stations. Libraries. The city built a gondola system connecting hillside slums to the metro below. It built outdoor escalators up the steep slopes of Comuna 13, where residents climbed the equivalent of a 28-story building to reach their homes.41

The homicide rate fell 90 percent from its peak. In 2013, Medellín won the Wall Street Journal's Innovative City of the Year award, beating New York and Tel Aviv. The transformation had many causes. The decline of the Escobar cartel. National security reforms. Fajardo did not cause the entire decline. He measured the citizen's variable. Libraries instead of police stations. Cable cars instead of containment.42

When the establishment treats the contradiction as information, the outcome changes. When it treats the contradiction as proof the people are wrong, the gap widens.

The gap widened from 2016 to 2024.

The measurement and the outcome

October 2016. The model said done. The voters said no.

December 2019. The model said damaged. The voters said no.

January 2021. The model said finished. The voters said no.

2023. 2024. Ninety-one felony charges across four jurisdictions. The model said: unprecedented. Unrecoverable. The variables are overwhelming.

The voter measured something the model did not track. Eggs up 184 percent. Homes at five times income. Productivity up 21.6 percent, wages up 1.8 percent. The trade deal that raised GDP and closed the factory. The interest rate that built wealth for owners and priced out everyone else.

The model added variables between 2016 and 2024. Legal exposure. Felony charges. More of the same kind of measurement. The voter's variable did not change. It was the same question in 2024 that it was in 2016.

Does my life work inside this system?

The question was never whether the data was accurate. The data was always accurate. The question was whether the data was measuring what determined the outcome.

King George measured trade surplus. The colonist measured representation. Churchill measured administrative output. The Indian measured whether they could eat. The CIA measured production. The Soviet citizen measured what was on the shelf. The Remain campaign measured GDP. The Leave voter measured the high street.

The model measured scandal. The voter measured the grocery bill.

Two numbers. Same reality. One is tracked. One is lived.

The people were not wrong. The model was precise. The model was pointed at the wrong thing.

New pieces when they're ready. Nothing else.

Sources

  1. The Access Hollywood tape was published by the Washington Post on October 7, 2016. Within days, dozens of Republican officials withdrew endorsements or called for Trump to drop out of the race.
  2. U.S. House of Representatives voted to impeach on December 18, 2019, on charges of abuse of power and obstruction of Congress. The Senate acquitted on February 5, 2020.
  3. January 6, 2021 attack on the U.S. Capitol. Five deaths confirmed. Congressional members evacuated or sheltered in place.
  4. Four indictments: Manhattan DA (March 2023, 34 counts), federal classified documents (June 2023, 40 counts), federal January 6th (August 2023, 4 counts), Fulton County GA (August 2023, 13 counts). Total: 91 felony charges.
  5. Republican primary polling data: RealClearPolitics average showed 43% before first indictment (March 2023), rising to 52% within one week (April 7, 2023). FiveThirtyEight recorded a 12-point climb to 58.7% after all indictments accumulated. Yahoo News/YouGov poll post-first indictment: lead over DeSantis grew from 8 points to 26 points.
  6. 2024 presidential election results. Trump won all seven swing states: Arizona, Georgia, Michigan, Nevada, North Carolina, Pennsylvania, Wisconsin. Trump won the popular vote with approximately 49.9% (he lost it in 2016 with 46.1%). His vote share increased by nearly 4 points between elections.
  7. Bureau of Labor Statistics, Average Price Data. Eggs (FRED series APU0000708111): $1.46/dozen (Jan 2020) to $4.15 (Dec 2024), +184%. Whole milk (APU0000709112): $3.25/gal to $4.10, +26%. White bread (APU0000702111): $1.35/lb to $1.91, +41%. Ground beef (APU0000703112): $3.89/lb to $5.61, +44%. Food-at-home CPI: +11.4% in 2022 (all food +9.9%), fastest since 1979.
  8. Harvard Joint Center for Housing Studies, 2024. Median single-family home prices rose 48% between 2019 and 2024, more than double the 22% increase in median income.
  9. National Association of Realtors and Federal Reserve data. Home price-to-income ratio reached approximately 5:1 in 2024, up from historical norms of 3:1. Income required to afford median home at prevailing mortgage rates: $106,731. Median household income: $83,730.
  10. Economic Policy Institute, "The Productivity-Pay Gap" and "Understanding the Historic Divergence Between Productivity and a Typical Worker's Pay." Net productivity growth 2000-2014: 21.6%. Median worker compensation growth: 1.8%. More than 80% of the divergence driven by rising inequality in compensation and falling labor share of income.
  11. Gallup final favorability survey (November 2-5, 2016): 36% favorable, 61% unfavorable. Lowest election-eve favorability of any major-party candidate since Gallup began measuring in 1956. Actual vote share: 46.1%. The gap between favorability and vote share was 10 percentage points. Source: Gallup, "Trump and Clinton Finish With Historically Poor Images," November 8, 2016.
  12. The distinction between character assessment and conditions assessment is documented across the economic voting literature. The poll's categories (approve/disapprove, favorable/unfavorable) are character variables. The ballot is a directional choice about conditions. See footnote 31.
  13. The school rating mechanism is documented in "The School Was Rated" on this site. Sandra Black (Federal Reserve Bank of New York, 1999) found parents paid roughly 2% more per 5% increase in test scores using a boundary-discontinuity design. Brookings Institution: homes near high-scoring schools sell for $205,000 more on average. See /school.
  14. British colonial trade data. The colonies were a significant source of revenue and raw materials for the British Empire. Navigation Acts directed colonial trade through British ports.
  15. King George III, speech to Parliament, October 26, 1775. Full text available at ushistory.org.
  16. Thomas Hutchinson, "Strictures upon the Declaration of the Congress at Philadelphia," 1776. Hutchinson served as the last civilian royal governor of Massachusetts Bay Province.
  17. British Raj infrastructure claims were a standard justification for colonial rule. Railways, telegraph systems, and administrative structures were cited as evidence of the benefit of British governance.
  18. Bengal famine of 1943. Death toll: approximately 3 million (Britannica, PMC). India exported 70,000+ tons of rice January-July 1943. Researchers found Bengal was the only major Indian famine (1873-1943) not linked to drought or crop failure. Al Jazeera, CNN reporting on Geophysical Research Letters study (2019).
  19. Amery diary: Churchill attributed the famine to Indians "breeding like rabbits" and resisted requests for food imports. The "why hasn't Gandhi died yet" remark comes from Wavell's diary (July 1944), paraphrasing a Churchill telegram. Some historians argue the telegram concerned Gandhi's recovery from illness, not the famine directly. The broader pattern of War Cabinet resistance to food imports is well documented regardless of the precise context of individual remarks. Sources: Amery's published diaries; Wavell's diary; Mukerjee, "Churchill's Secret War" (2010). Disputed by Hillsdale Churchill Project and Andrew Roberts.
  20. CIA estimate of Soviet GNP: 49% of U.S. GNP for 1990. Goskomstat estimate: 58% for 1987. Senator Moynihan quote from 1990. Sources: Texas National Security Review, "Assessing Soviet Economic Performance During the Cold War" (2018); GAO report NSIAD-91-274.
  21. CIA analysts relied on Soviet production statistics that were "almost surely inflated intentionally, both by Soviet leaders for propaganda purposes and by Soviet producers who had incentives to misreport." GAO report, CIA reading room documents.
  22. Brexit referendum: Remain campaign's economic forecasts projected GDP losses from leaving the EU. HM Treasury published analysis in April 2016 projecting significant economic costs.
  23. Brexit polling: phone polls predicted Remain +2.6%, online polls predicted Leave +1.2%. 78% of phone polls predicted Remain victory. Result: Leave 51.9%, Remain 48.1%. Every final poll overstated the Remain vote share. British Polling Council post-referendum analysis.
  24. Real GDP (FRED series GDPC1): $14.2 trillion (Q4 2000) to $23.6 trillion (Q4 2024), +66%. Manufacturing employment (FRED series MANEMP): 17.3 million (Jan 2000) to 12.7 million (Dec 2024), -27%. China trade shock: 1.5-2 million manufacturing jobs lost 2000-2007 (Autor, Dorn, and Hanson, "The China Shock," 2016). EPI: 3.7 million jobs lost 2001-2018.
  25. Macomb County, MI. Manufacturing jobs: 106,415 (2000) to 51,526 (2010), -52%. Median household income: $71,797 (1999) to $52,978 (2013), -26%. Voted Obama 2008, Obama 2012, Trump 2016, 2020, 2024. Sources: Crain's Detroit Business, BLS QCEW, Census.
  26. Erie County, PA. "Deeply blue-collar county that is whiter and poorer than both the commonwealth and the country overall." Manufacturing among the three largest employment sectors (18,951 employed, 2023). Obama 2008, Obama 2012, Trump 2016 (by <2,000 votes), Biden 2020 (by <1,500 votes), Trump 2024. Sources: NPR, Census QuickFacts.
  27. Post-2016 polling methodology changes: education weighting introduced, likely-voter screen improvements, social desirability bias adjustments. Despite changes, 2020 polls were worse (larger miss than 2016). 2024 polls improved but still missed in same direction. Source: NBC News analysis, Sabato's Crystal Ball.
  28. 2024 polling errors: national miss of 2.4 percentage points, swing state miss of 2.6 points. In every swing state, Trump outperformed by approximately 3 points. Pennsylvania: NYT/Siena final 48-48, actual 50.4-48.7. Michigan: NYT/Siena final 47-47, actual 49.7-48.3. Sources: NBC News, UCR analysis, U.S. News.
  29. Pew Research Center: polling response rates declined from 36% (1997) to 9% (2012) to 6% (2018). Source: Pew Research, "Response Rates in Telephone Surveys Have Resumed Their Decline" (February 27, 2019).
  30. The voter's absence from the data as data: FiveThirtyEight (2020) identified the issue as differential nonresponse, not voter dishonesty. People with low institutional trust correlate with both Trump support and survey nonparticipation. The "shy voter" framing puts the blame on the voter. The "nonresponse bias" framing puts the blame on the instrument.
  31. Douglas Hibbs, "Bread and Peace Voting in U.S. Presidential Elections," Public Choice, 2000. Two variables (per capita real income growth, military fatalities) outperformed 22 alternative models. 2024 exit polls: CBS News, two-thirds described economy as "bad." 73% angry or dissatisfied with country's direction. See also Lewis-Beck and Stegmaier, "Economic Determinants of Electoral Outcomes," Annual Review of Political Science, 2000.
  32. Morgenthau Plan: proposed by Treasury Secretary Henry Morgenthau Jr. in 1944. Advocated transforming Germany into an agricultural economy. Herbert Hoover's investigation concluded it would result in up to 25 million German deaths from starvation.
  33. Marshall Plan: $13.3 billion (approximately $137 billion in 2025 dollars), equal to 5% of U.S. GDP. Enacted April 3, 1948. Western European GDP grew 30%+ from 1948-1953. Industrial production increased 35%. Described as the fastest period of growth in European history. Sources: EH.net, Wikipedia, CFR.
  34. Bay of Pigs invasion: April 17-19, 1961. Kennedy authorized the CIA-planned operation. Near-unanimous consensus among advisors. The operation failed within three days.
  35. Cuban Missile Crisis ExComm: convened October 16, 1962. Kennedy structured the committee to force dissent. Robert Kennedy challenged the consensus for a military strike. Joint Chiefs and majority of ExComm favored attack. Kennedy chose blockade. Sources: JFK Library, State Department Office of the Historian, Time.
  36. Soviet tactical nuclear weapons on Cuba: confirmed at a 1992 conference in Havana. Their presence was unknown to American intelligence during the crisis. Moscow explicitly ordered they not be used without Kremlin authorization, but lacked technical negative controls to prevent unauthorized use. Soviet field commanders had already mated warheads to delivery systems without Moscow's knowledge.
  37. UNAIDS estimated 25 million people living with HIV in sub-Saharan Africa by 2003. US antiretroviral treatment cost exceeded $12,000 per patient per year. The consensus assessment was that treatment in Africa was logistically impossible at scale.
  38. PEPFAR: announced January 28, 2003, State of the Union address. $15 billion authorized over five years. Cumulative funding through 2024: $120 billion. Generic ARV costs fell from $1,053/patient/year (2005) to $339 (2011) to under $45 (current). Sources: White House archives, GAO-13-345, PEPFAR annual reports, Clinton Health Access Initiative.
  39. PEPFAR: 25 million lives saved (US government estimate, 2023; updated to 26 million by 2024). AIDS-related deaths declined approximately 59% in PEPFAR-supported countries; globally, HIV deaths fell approximately 70% from the 2004 peak. Mother-to-child transmission declined over 80%. Reauthorized: Tom Lantos and Henry J. Hyde Act (2008, $48 billion). Maintained under Trump and Biden administrations. Sources: State Department, NEJM (2018), USIP (2023).
  40. Medellín 1991 homicides: 6,809 (rate of 416 per 100,000). Source: Medellín municipal records. The city held the highest murder rate of any city in the world. Pablo Escobar was killed in December 1993; the cartel's decline contributed significantly to the reduction in violence alongside subsequent policy interventions.
  41. Sergio Fajardo: mathematician and academic, elected 2003, took office January 2004. Library-parks (Parques Biblioteca) built in comunas with highest homicide rates, including Santo Domingo Savio (architect: Giancarlo Mazzanti, opened 2007). Metrocable gondola system planned and built under predecessor Luis Pérez Gutiérrez, inaugurated mid-2004 under Fajardo, who expanded the system. Comuna 13 outdoor escalators: 384 meters, installed 2011. Sources: Americas Quarterly, Rapid Transition Alliance, Urban Land Institute.
  42. Medellín homicide rate declined approximately 90% from 1991 peak by 2007. Wall Street Journal/Citi/Urban Land Institute "Innovative City of the Year" 2013, selected over New York and Tel Aviv. Sources: Wall Street Journal, Business Wire (March 1, 2013), ColombiaOne (2024, lowest rate in 82 years).
  43. The divergence is documented in the Economic Policy Institute's work, including "The Productivity-Pay Gap" and "Wage Stagnation in Nine Charts." See also the broader economic conditions data in footnotes 7-10 and 24.