The DOI matches its Crossref record; 46 of the 46 references it cites were found in a registry or an index; 5 of the 26 sources its claims cite were checked for support.
Statistical paradises and paradoxes in big data (I): Law of large populations, big data paradox, and the 2016 US presidential election
Open accessAbstract
Statisticians are increasingly posed with thought-provoking and even paradoxical questions, challenging our qualifications for entering the statistical paradises created by Big Data. By developing measures for data quality, this article suggests a framework to address such a question: “Which one should I trust more: a 1% survey with 60% response rate or a self-reported administrative dataset covering 80% of the population?” A 5-element Euler-formula-like identity shows that for any dataset of size $n$, probabilistic or not, the difference between the sample average $\overline{X}_{n}$ and the population average $\overline{X}_{N}$ is the product of three terms: (1) a data quality measure, $\rho_{{R,X}}$, the correlation between $X_{j}$ and the response/recording indicator $R_{j}$; (2) a data quantity measure, $\sqrt{(N-n)/n}$, where $N$ is the population size; and (3) a problem difficulty measure, $\sigma_{X}$, the standard deviation of $X$. This decomposition provides multiple insights: (I) Probabilistic sampling ensures high data quality by controlling $\rho_{{R,X}}$ at the level of $N^{-1/2}$; (II) When we lose this control, the impact of $N$ is no longer canceled by $\rho_{{R,X}}$, leading to a Law of Large Populations (LLP), that is, our estimation error, relative to the benchmarking rate $1/\sqrt{n}$, increases with $\sqrt{N}$; and (III) the “bigness” of such Big Data (for population inferences) should be measured by the relative size $f=n/N$, not the absolute size $n$; (IV) When combining data sources for population inferences, those relatively tiny but higher quality ones should be given far more weights than suggested by their sizes. Estimates obtained from the Cooperative Congressional Election Study (CCES) of the 2016 US presidential election suggest a $\rho_{{R,X}}\approx-0.005$ for self-reporting to vote for Donald Trump. Because of LLP, this seemingly minuscule data defect correlation implies that the simple sample proportion of the self-reported voting preference for Trump from $1\%$ of the US eligible voters, that is, $n\approx2\mbox{,}300\mbox{,}000$, has the same mean squared error as the corresponding sample proportion from a genuine simple random sample of size $n\approx400$, a $99.98\%$ reduction of sample size (and hence our confidence). The CCES data demonstrate LLP vividly: on average, the larger the state’s voter populations, the further away the actual Trump vote shares from the usual $95\%$ confidence intervals based on the sample proportions. This should remind us that, without taking data quality into account, population inferences with Big Data are subject to a Big Data Paradox: the more the data, the surer we fool ourselves.
Overview
CoChat Summary
TL;DR
Meng derives an exact identity showing that sample-mean error depends on a data defect correlation, relative sample size, and population standard deviation, yielding a Law of Large Populations and a Big Data Paradox. Applying this to 2016 US election CCES data, a tiny Trump self-reporting correlation of about −0.005 implies that data from 1% of eligible voters can have the effective SRS size of only about 400.
Key findings
- The fundamental identity decomposes the difference between sample average and population average into three terms: data quality \(\rho_{R,X}\), data quantity \(\sqrt{(N-n)/n}\), and problem difficulty \(\sigma_X\).
- For the 2016 US presidential election, CCES data suggest \(\rho_{R,X}\approx -0.005\) for self-reporting to vote for Donald Trump; because of the Law of Large Populations, a sample proportion from \(n\approx 2{,}300{,}000\) (1% of US eligible voters) has the same MSE as an SRS of \(n\approx 400\), a 99.98% reduction in sample size.
- The CCES data show that, on average, larger state voter populations had actual Trump vote shares farther from the usual 95% confidence intervals based on sample proportions; the estimated slope for Trump supporters was 0.44 (SE 0.09, \(p\approx 10^{-6}\)), versus 0.15 (SE 0.16) for Clinton supporters.
- The paper labels the resulting phenomenon the Big Data Paradox: “the more the data, the surer we fool ourselves.”
Methods
- Theoretical derivation of an exact finite-population identity and MSE decomposition, defining the data defect correlation \(\rho_{R,G}\), data defect index \(DI=E_R[\rho_{R,G}^2]\), dropout odds \(DO=(1-f)/f\), and degree of uncertainty \(DU=\sigma_G^2\); Theorem 1 equates finite design effect, \(n^{-1}\) MSE rate, and \(DI=O(N^{-1})\).
- Empirical application to the 2016 US presidential election using the Cooperative Congressional Election Study (CCES), a national stratified-sample online survey; state-level analyses compare actual vote shares with CCES estimates using raw sample averages (\(n=64{,}600\)), turnout-adjusted estimates (\(\hat n=48{,}106\)), and validated voters (\(n=35{,}829\)).
- Binary-outcome analysis uses Hoeffding–Fréchet bounds, estimates state-level \(\hat\rho_N\), and regresses \(\log|Z_{n,N}|\) on \(\log N\); extensions cover weighted estimators, stratified sampling, Monte Carlo/Quasi-Monte Carlo, and privacy-enhancing adjustments.
Limitations
- The paper states that \(\rho_{R,G}\) cannot be estimated from the sample itself; it requires post-election or external information, and the empirical estimates rely on retrospective CCES data and validated-voter subsamples.
- The paper states that response bias is assumed negligible, that CCES stratification specifications were unavailable, that validation via matching is subject to errors, and that weighting is “a mess”; it also notes that defining d.d.i. for estimands other than sample/population averages remains open.
- Evident from the text: the empirical application is limited to one election and one survey source, so using the estimated \(\rho_N\) for future elections requires borrowing information or prior assumptions, as the paper itself proposes.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 7 Not valid · 17 Not checked · 95 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“But it is rarely possible to do so when the weights themselves are subject to errors, typically much larger than O(N -1 ) for large N [see Kim and Kim (2007), Kim and Riddles (2012)].”
What the sources say
- Not checked — no full text of the sourcecites Kim and Kim (2007) source
- Partly supportedcites Kim and Riddles (2012)Experimental — accuracy not yet measured
Our copy of Kim and Riddles (2012) builds its whole theory on weights that are themselves estimated and erroneous: Section 1 (eq. (2)) assumes the response-model parameter estimator satisfies (φ̂ − φ0) = Op(n^{-1/2}), and Sections 2 and 4 (eqs. (3), (9), (14), (17), (24)–(26)) carry that weight error through the linearization and variance formulas, with the residual design terms bounded only at order O(n N^{-1}) (eq. (25)). That supports the sub-claim that the weights are subject to errors of a rate far coarser than N^{-1}. But no passage states the comparison with O(N^{-1}), nor the conclusion that it is 'rarely possible' to drive such a term down to N^{-1}; the paper's own negligibility statements are of the form op(1) / op(V) (e.g. Section 3, eq. (14) region: '(| ) = (1) x h i p E o γ', and Section 4: HT { ( , ) } = (1) N N N V n V o η Φ R), i.e. relative orders, not an N^{-1} target. The Introduction also credits the rate/asymptotics work to another paper — 'Kim and Kim (2007) used a Taylor expansion to obtain the asymptotic mean and variance of PSA estimators' — so part of the cited proposition originates there rather than here. (LLM, conf=0.62)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“But none of them suggests the remote possibility of reducing ρ R,G to the order of N -1 / 2 , especially for large N .”
What the sources say
- Not matchedcites Chen et al. (2006)
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Could not verifycites Gelman (2007) sourceExperimental — accuracy not yet measured
Our copy of Gelman (2007) does discuss weight modification in the terms the claim gestures at — 'truncation of weights' among the arbitrary choices in creating practical weights (Practical limitations of weighting) and hierarchical poststratification as 'approximately equivalent to a shrinkage of weights' (Hierarchical regression, eq. 13) — but nowhere does it define or analyse a correlation ρ_{R,G}, nor state anything about reducing such a quantity at rate N^{-1/2} or for large N (searched: ρ/correlation between regression and poststratified estimators, trimming/power shrinkage of weights, effective sample size, order 1/sqrt(n)). The variance material that exists (3.3 Properties of the Model-Based Poststratified Estimates; 1.3 competing methods/Table 3) concerns standard errors, not a correlation's convergence rate. The rate claim therefore cannot be checked against this source; nothing in our copy contradicts it either. (LLM, conf=0.65)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“This quantitative finding provides numerical evidence to the general belief that there was serious underreporting by Trump's supporters [see e.g., Cohn (2017), Gelman and Azari (2017)].”
What the sources say
- Not matchedcites Cohn (2017)
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Partly supportedcites Gelman and Azari (2017) sourceExperimental — accuracy not yet measured
In Section 5 ('Survey Nonresponse Bias is a Thing') Gelman and Azari do give a quantitative anchor — 'The final polls were off by about 2 percentage points, suggesting that, even at the end, Trump supporters were responding at a lower rate than Clinton supporters, most notably in certain key swing states' — so a numerical indication of shortfall in reaching Trump supporters is in the source. However, the same section is explicitly skeptical of the underreporting/'shy Trump voter' explanation (Trump outperformed most in North Dakota, West Virginia, and Republican Senate races fit differential nonresponse or turnout rather than reluctance to admit support), and Sections 3 and the Abstract ('An error of 2 percentage points is no great embarrassment'; the error was historically expected) characterise the magnitude as small, not 'serious'. The citing sentence therefore takes the 2-point figure as numerical evidence for serious underreporting, a reading this source disputes and downplays. (LLM, conf=0.78)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Estimates obtained from the Cooperative Congressional Election Study (CCES) of the 2016 US presidential election suggest a ρR,X ≈-0 . 005 for self-reporting to vote for Donald Trump.”
hedgedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“This negative impact is particularly pronounced when the (relative) variation in the weights is high, as measured by CV 2 W , and the sampling rate f is high.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
55 more claims
Create a free account to see all 123 claims and to verify the 3 references still open.
- Probably not a substantial claim. 22
- It may state two separate claims as one. 45
- The quote only partly states this claim. 3
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“A previously formulated statistical framework 1 permits us to exactly decompose the total error of a survey estimate into three components:”
From Unrepresentative big surveys significantly overestimated US vaccine uptake (2021) · cites this paper as [1]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“This decomposition is why our comparison of the surveys is consistent with the 'Law of Large Populations' 1 (estimation error increases with N ), even though all three surveys ultimately target the same US adult population.”
From Unrepresentative big surveys significantly overestimated US vaccine uptake (2021) · cites this paper as [1]
The cited source says what the claim attributes to it.
In the citing paper
“Data quality will depend on the interplay of sample sizes and various sample biases—for a detailed and formal treatment of these issues, see Meng (2018).”
From Social Data: Biases, Methodological Pitfalls, and Ethical Boundaries (2019) · cites this paper as Meng, 2018
The cited source says what the claim attributes to it.
In the citing paper
“However, counter to common intuition 12 , larger sample sizes alone do not ensure lower error. Instead, small biases are compounded as sample size increases 1 .”
From Unrepresentative big surveys significantly overestimated US vaccine uptake (2021) · cites this paper as [1]
The cited source says what the claim attributes to it.
In the citing paper
“This is thev Big Data Paradox: 'the bigger the data, the surer we fool ourselves' 1 when we fail to account for bias in data collection.”
From Unrepresentative big surveys significantly overestimated US vaccine uptake (2021) · cites this paper as [1]
1 more claim about this paper
Create a free account to see every claim other papers make about this one.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4. Applications to binary outcome and the 2016 US general election. · p. 23
“= 1 /( 200 √ 0 . 99 )”
§ 4. Applications to binary outcome and the 2016 US general election. · p. 23
“= -0 . 00502”
Citations
Cites
RUBIN, D. B. (1976) · Biometrika · doi:10.1093/biomet/63.3.581 · Cited by 9,925
A generalization of sampling without replacement from a finite universe
Horvitz, D. G., & Thompson, D. J. (1952) · Journal of the American Statistical Association · doi:10.1080/01621459.1952.10483446 · Cited by 4,293 · cited 1×
“If πj is known for those observed Gj , then a well-known weighting scheme is to set Wj ∝ π−1j , which leads to the Horvitz–Thompson estimator [Horvitz and Thompson (1952)].”
BackgroundA Generalization of Sampling Without Replacement From a Finite Universe
Horvitz, D. G., & Thompson, D. J. (1952) · Journal of the American Statistical Association · doi:10.2307/2280784 · Cited by 3,011
EFRON, B., & HINKLEY, D. V. (1978) · Biometrika · doi:10.1093/biomet/65.3.457 · Cited by 952
Donoho, D. (2017) · Journal of Computational and Graphical Statistics · doi:10.1080/10618600.2017.1384734 · Cited by 695
Struggles with Survey Weighting and Regression Modeling
Gelman, A. (2007) · Statistical Science · doi:10.1214/088342306000000691 · Cited by 604
Metropolized independent sampling with comparisons to rejection sampling and importance sampling
Liu, J. S. (1996) · Statistics and Computing · doi:10.1007/bf00162521 · Cited by 507
(1951) · Medical Entomology and Zoology · Cited by 426
Validation: What Big Data Reveal About Survey Misreporting and the Real Electorate
Ansolabehere, S., & Hersh, E. (2012) · Political Analysis · doi:10.1093/pan/mps023 · Cited by 416
(2000) · Technometrics · doi:10.1080/00401706.2000.10486038 · Cited by 412
Research works printed but not linked: 7
- Chen, C., Duan, N., Meng, X.-L. and Alegria, M. (2006). Power-shrinkage and trimming: Two ways to mitigate excessive weights. In Proceedings of the Survey Research Methods Section of the American Statistical Association 2839–2846.
- Cohn, N. (2017). Election review: Why crucial state polls turned out to be wrong. The New York Times, June 1 st.
- Fienberg, S. E. (1996). Applying statistical concepts and approaches in academic administration. In Education in a Research University 65–82. Stanford Univ. Press, Stanford.
- Fuller, W. A. (2011). Sampling Statistics Wiley, New York.
- Kish, L. (1965). Survey Sampling. Wiley, New York.
- Mehrhoff, J. (2016). Executive summary: Meng, X.-L. (2014), “A trio of inference problems that could win you a Nobel prize in statistics (if you help fund it)”. Conference handout.
- Meng, X.-L. (2018). Statistical paradises and paradoxes in big data (II): Multi-resolution inference, Simpson’s paradox, and individualized treatments. Preprint.
Cited by
Social Data: Biases, Methodological Pitfalls, and Ethical Boundaries
Olteanu, A., Castillo, C., Diaz, F., & Kıcıman, E. (2019) · Frontiers in Big Data · doi:10.3389/fdata.2019.00013 · Cited by 763
“Data quality will depend on the interplay of sample sizes and various sample biases— for a detailed and formal treatment of these issues, see Meng (2018).”
BackgroundSocial Data: Biases, Methodological Pitfalls, and Ethical Boundaries
Olteanu, A., Castillo, C., Diaz, F., & Kiciman, E. (2016) · doi:10.2139/ssrn.2886526 · Cited by 763
Unrepresentative big surveys significantly overestimated US vaccine uptake
Bradley, V. C., Kuriwaki, S., Isakov, M., Sejdinovic, D., Meng, X.-L., & Flaxman, S. (2021) · Nature · doi:10.1038/s41586-021-04198-4 · Cited by 248
“While the data defect correlation ddc is not directly observed, COVID-19 surveys present a rare case in which it can be deduced because all other terms in Equation (1) are known (see “Calculation and interpretation of ddc” in the Methods for an in-depth explanation).”
BackgroundMethodResultInfluentialSelf-citationThe case for formal methodology in scientific reform
Devezer, B., Navarro, D. J., Vandekerckhove, J., & Ozge Buzbas, E. (2021) · Royal Society Open Science · doi:10.1098/rsos.200805 · Cited by 188
“Example: statistical methods assume probabilistic sampling and do not make any claims in a non-probabilistic sampling framework [43].”
The case for formal methodology in scientific reform
Devezer, B., Navarro, D. J., Vandekerckhove, J., & Buzbas, E. O. (2020) · doi:10.1101/2020.04.26.048306 · Cited by 195
Metadata
- Authors
- Meng, X.-L.
- Affiliations
- Harvard University (US)
- Corresponding author
- Meng, Xiao-Li
- ORCID
- Meng, Xiao-Li
- Venue
- The Annals of Applied Statistics
- Publisher
- Institute of Mathematical Statistics
- Published
- 2018-06-01
- Volume
- 12
- Issue
- 2
- Pages
- 685-726
- ISSN
- 1932-6157 (ISSN-L)
- 1941-7330
- OpenAlex
- W2883811944
- Semantic Scholar
- 8b58f608261132a950a5e83ec00aa3b3836ccab7
- Cited by
- 316
- References
- 46
- Citations by year
- 2026: 29 · 2025: 49 · 2024: 44 · 2023: 58 · 2022: 46 · 2021: 40 · 2020: 29 · 2019: 17 · 2018: 4
- FWCI
- 18.88
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- The Annals of Applied Statistics (publishedVersion)
- Project Euclid (Cornell University) (submittedVersion)
- Funders
- National Science Foundation
- John Templeton Foundation
- University of Chicago
- Arizona State University
- Rice University
- University of Pittsburgh
- University of Cincinnati
- McGill University
- Colby College
- University of Cambridge
- University of Oxford
- Institut Carnot Santé Animale
- Harvard Data Science Initiative, Harvard University
- Topics
- Physical Sciences › Mathematics › Statistics and Probability › Survey Sampling and Estimation Techniques
- Social Sciences › Sociology and Political Science › Survey Methodology and Nonresponse
- Social Sciences › Decision Sciences › Information Systems and Management › Big Data Technologies and Applications
- Keywords
- big data, data quality, population inference, data combination, self-reported data, 2016 US presidential election, Cooperative Congressional Election Study, mean squared error, survey response rates, administrative data
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Statistical paradises and paradoxes in big data (I): Law of large populations, big data paradox, and the 2016 US presidential election. https://app.cochat.ai/papers/10.1214/18-aoas1161sf
- The paper
- Meng, X.-L. (2018). Statistical paradises and paradoxes in big data (I): Law of large populations, big data paradox, and the 2016 US presidential election. The Annals of Applied Statistics. https://doi.org/10.1214/18-aoas1161sf
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.