Working log behind Regression. Every query with its denominator, the report script and its unedited output, the three folds and what they absorbed, the full-text probes that were run and the ones that were rejected, the quotes spot-checked, the external sources verified, the judgement calls, and what could not be established.

Corpus-level caveats — the venue scope, the selection funnel, the provisional 2025–2026 years, the duplicate records, the posters — are on Corpus and are not restated here. Citation keys are shared with the content page and the single Bibliography; this page adds no bibliography entries of its own.

Voice here is a working log, not prose. It is for someone checking a number.

Two neighbours could have absorbed this material and both were rejected.

• Hypothesis testing (64 KB) already covers non-independence, and its own simulation measures the false-positive rate of ignoring clustering in a single test. Broadening it would have merged two different things: that page is about which test, this one is about which model, and a model has two further axes (what the outcome is, what the coefficient claims to be) that a test does not. The hypothesis-testing page also already names Statistics:Regression as the home for cluster-robust standard errors and mixed models, so the boundary was set by an earlier run.
• Pvalue corrections shares neither the population (1,025 vs 391) nor the failure mode.

Deliberate division of labour, linked rather than restated:

• The single-test clustering simulation and the ICC argument → Hypothesis testing. This page's simulation deliberately does not repeat it: part B measures a different thing (the direction of the error as a function of where the predictor varies, and the power consequence), and part A measures overdispersion, which the hypothesis-testing page does not touch.
• Multiplicity across a regression table → Pvalue corrections.
• Classifier evaluation → Website classification and Javascript.
• Committing to a specification in advance → Study preregistration.

One figure is deliberately re-derived on both pages: the crawl-versus-user-study asymmetry. It holds for modelling as it does for testing, and the parallel is the point.

Every query names its own population. regression is the denominator for the whole page.

Population                                                    Papers   Share
all extraction records                                        5,859    100.0%
inferential — any statistics tuple that is not descriptive    1,762     30.1% of 5,859
hypothesisTest — >=1 tuple with kind == 'hypothesis-test'     1,025     17.5% of 5,859
REGRESSION — >=1 tuple with kind == 'regression'                391     22.2% of 1,762
  ... regression tuples                                         530
  ... distinct method strings                                   315

Definition, in report_regression.mjs:

const hasKind = (p, k) => p.statistics.some((s) => s.kind === k);
const REG = rows.filter((p) => hasKind(p, 'regression'));
const INFERENTIAL = rows.filter((p) => p.statistics.some((s) => s.kind !== 'descriptive-only'));

This is the same INFERENTIAL definition report_hypothesis_testing.mjs and report_pvalue_corrections.mjs use, so the three pages' figures compose. Cross-checked: this script independently reports 1,762 and 1,025, matching the two sibling pages.

Checked, because a page whose whole denominator is one enum value has to know. A second query looks for a regression-shaped method string under any kind:

const REG_LIKE = /regress|ordinary least|\bOLS\b|least.squares|\bGLM\b|generali[sz]ed linear|mixed.?effect|
                  multilevel|multi-level|\blmer\b|\bglmer\b|linear model|logit|probit|\bLASSO\b|ridge|
                  cox propor|proportional hazard|cumulative link|beta regression|poisson|negative binomial/i;

papers with kind == 'regression'                391
papers with a regression-like string, any kind  380
union                                           400

Nine papers name a regression only under another kind (mostly descriptive-only for a curve fit, and one effect-size tuple naming difference-in-differences). Twenty name regression with a method string that is not a regression — those are the excluded non-regressions below. The enum holds; 391 is the right denominator, and the nine-paper gap is disclosed on the page in the identification-designs footnote.

Subset of the 391               Papers  Share    Base rate for comparison
ran a crawl                        62   15.9%    20.9% of the 1,762 inferential papers
recruited human participants      185   47.3%    40.1% of the 1,762
both                               11    2.8%
crawl, no participants             51   13.0%
neither                           155   39.6%
platforms includes 'web'          125   32.0%

const crawled = (p) => p.crawlConfig !== null || p.studyTypes.includes('automated-web-crawl');
const humanSubjects = (p) => p.participants.length > 0;

crawled is the definition in lib.mjs shared by every page on this site (1,120 papers corpus-wide). humanSubjects here is participants.length > 0 rather than the studyTypes flag, because the participants family is what carries the repeated-measures structure the page argues about.

The comparison to the base rate is load-bearing and is why it is printed. “47.3% of regression papers had participants” means nothing on its own; against 40.1% for inferential papers generally it means regression skews to user studies, and against 20.9% crawled the 15.9% means it skews away from crawls.

scripts/reg_fold.mjs. Read its header before reading any figure it produces.

statistics[].method is free text, agrees run-to-run on roughly a fifth of exact strings, and arrives here in 315 distinct spellings across 530 tuples. Ranking those strings is not merely noisy, it is the wrong operation: “mixed-effects logistic regression” is not a sibling of “logistic regression”. It is binary outcome + random effects + associational, three answers to three questions, and a single ranked list cannot show that 79.5% of papers handle dependence in no way at all because that fact is smeared across every name in the list.

Each method string is therefore assigned a value on three independent axes. First match wins within each axis; order is load-bearing.

• Axis 1, outcome family — what the response variable is.
• Axis 2, dependence — what the model does about correlated observations. Default none stated in the method name.
• Axis 3, identification — what the coefficient is claimed to be. Default associational (no design stated).

The printed residue is 0 distinct strings. That is not a claim that the fold understood everything, and the page says so. It is 0 because the last outcome rule is a catch-all:

[
  'family not stated in the method name',
  /regress|\bGLM(M|E)?s?\b|generali[sz]ed linear|mixed[-\s]?effects?|multi[-\s]?level|multilevel|
   random[-\s]?(intercept|effect|slope)|\bGEE\b|estimating equation|heckman|selection model|propensity|
   difference[-\s]?in[-\s]?difference|discontinuity|generali[sz]ed additive|\bGAM\b|elastic net|lasso|
   ridge|\bmodel\b|\bfit\b|\bfitting\b/i,
],

Everything it absorbs is reported as its own row (42 papers, 10.7%) and printed in full by the report script under THE UNSTATED-FAMILY BUCKET, PRINTED IN FULL, so the part the fold could not place stays visible instead of being folded into “continuous”. All 46 distinct strings, of which the bare-GLM group is the review finding described below:

  Ridge regression
  simple regression
  regression of inbox quantity on spambox quantity and lagged values
  time-series regression
  Forward subset selection based regression (FSSreg) with leave-one-out cross-validation
  general regression analysis
  Lasso regression
  whole-brain regression with impulsivity scores as covariate
  ridge regression with 10-fold cross-validation
  regression model
  multilevel models with country-specific random effects
  elastic net
  random-intercept multi-level regression
  weighted generalized linear regression
  multivariate stepwise regression
  generalized linear mixed model (GLMM) regression
  generalized linear multilevel model
  multiple regression
  hierarchical multiple regression
  Regression Discontinuity Design
  Heckman selection model
  multivariate regression model
  Regression of average persistence scores on time
  multivariate regression
  multiple categorical variable regressions
  regression of effective sampling frequency on zip-code, category, and sampling-date factors
  step-wise Generalised Linear Models (GLM)
  generalized linear mixed models with day-level random effects
  generalized linear mixed effects model with random intercepts and slopes
  multilevel regression models with maximum-likelihood estimation
  fuzzy regression discontinuity
  Multiple Regression Analyses (MRA)
  generalized linear regression model with log transformation
  generalized additive models
  regression analysis
  Regression of estimated treatment effects on standardized pre-treatment covariates
  Generalized Linear Mixed Models (GLMMs)
  mixed-effects model
  propensity score weighting and trimming
  staggered difference-in-differences using Callaway and Sant'Anna group-time ATTs
  multilevel regression model with maximum-likelihood estimation
  general estimating equation (GEE) with exchangeable correlation
  general estimating equation (GEE) with AR(1) autocorrelation
  generalized linear regressions
  multivariable generalized linear mixed effects (GLME) models
  fixed-effects panel regression
==============================================================================

Note what this bucket does not cost: those 42 papers still get an axis-2 and axis-3 value, because “multilevel models with country-specific random effects” states its dependence handling perfectly well even though it never says what the outcome is.

37 papers, reported separately and never counted in an outcome family:

NOT-A-REGRESSION: other multivariate method       19 papers
NOT-A-REGRESSION: model selection / diagnostic    18 papers

Examples, all real strings from kind == “regression” tuples: Akaike Information Criterion (AIC) model selection, variance inflation factor (VIF), Nagelkerke's pseudo R2, backward model selection using Bayesian Information Criterion, Structural Equation Modeling (SEM), partial least squares structural equation modelling (PLS-SEM), latent class analysis, path model / mediation analysis, Decision Tree models, log-linear Bradley-Terry (LLBT) modeling, Oaxaca-Blinder decomposition, two lines analysis, trend lines, regression test.

The extractor scored these as regression because they appear in the same sentences as one. Counting them would inflate every family by roughly a tenth.

Each of these moves a published number, and a reasonable person could have decided otherwise.

scripts/report_regression.mjs. Reproduce with:

cd /workspace/artifacts/wiki
node scripts/report_regression.mjs            # full report, ~10 s (scans 5,869 full texts)
node scripts/report_regression.mjs --wiki     # DokuWiki tables
node scripts/report_regression.mjs --list     # the 391 papers
node scripts/report_regression.mjs --quotes 'mixed|poisson|clustered'   # evidence quotes

Real, unedited output at the time of publication:

report_regression-output.txt

==============================================================================
POPULATION
==============================================================================
corpus                                    5859
ran any inference (kind != descriptive)   1762
ran a hypothesis test                     1025
RAN A REGRESSION (kind == "regression")   391   <-- the denominator
  ... as a share of inferential papers    22.2%
  ... regression tuples                   530
  ... distinct method strings             315
regression-like string under ANY kind     380   (union with the above: 400)
 
--- What kind of study runs a regression ---
Subset of the 391             Papers  Share
----------------------------  ------  -----
ran a crawl                   62      15.9%
recruited human participants  185     47.3%
both                          11      2.8%
crawl, no participants        51      13.0%
neither                       155     39.6%
platforms includes "web"      125     32.0%
 
Base rates for comparison, so the shares above mean something:
  of the whole corpus, 1120 crawled (19.1%)
  of the whole corpus, 1357 had participants (23.2%)
  of the 1,762 inferential, 368 crawled (20.9%)
  of the 1,762 inferential, 707 had participants (40.1%)
 
--- Regression as inference vs regression as classifier ---
papers naming a regression model in tools[]        155
tools[] category of those entries  Tuples
---------------------------------  ------
ml-model-or-algorithm              148
statistics-software                11
 
papers naming one in statistics[] as well        253
overlap (same paper does both)                   31
names it ONLY as a tool, never as inference      124
 
==============================================================================
AXIS 1 — OUTCOME FAMILY (what the response variable is)
==============================================================================
papers with at least one tuple the fold read as a model:  372 of 391
Outcome family                                                        Papers  Share of 391
--------------------------------------------------------------------  ------  ------------
continuous (OLS / linear / Gaussian GLM)                              179     45.8%
binary (logistic / probit / linear probability)                       132     33.8%
ordinal (ordered logit / probit, cumulative link)                     46      11.8%
family not stated in the method name                                  42      10.7%
non-linear / curve fit (power law, log-log, polynomial, exponential)  23      5.9%
count (Poisson / negative binomial)                                   17      4.3%
continuous, non-mean estimand (quantile / robust M)                   9       2.3%
time-to-event (Cox / hazard)                                          7       1.8%
multinomial / discrete choice                                         5       1.3%
proportion (beta regression)                                          3       0.8%
 
--- Excluded: tuples the extractor scored as regression that are not models ---
Bucket                                          Papers
----------------------------------------------  ------
NOT-A-REGRESSION: other multivariate method     19
NOT-A-REGRESSION: model selection / diagnostic  18
 
==============================================================================
AXIS 2 — DEPENDENCE (what the model does about correlated observations)
==============================================================================
Dependence handling                                                            Papers  Share of 391
-----------------------------------------------------------------------------  ------  ------------
none stated in the method name                                                 311     79.5%
random effects (mixed-effects / multilevel)                                    75      19.2%
robust (heteroskedasticity-consistent) SEs only — does not address dependence  3       0.8%
fixed effects (dummies for the unit)                                           3       0.8%
GEE (population-averaged)                                                      2       0.5%
"repeated-measures", no random effect named                                    1       0.3%
cluster-robust standard errors                                                 1       0.3%
 
==============================================================================
AXIS 3 — IDENTIFICATION (what the coefficient is claimed to be)
==============================================================================
Identification strategy           Papers  Share of 391
--------------------------------  ------  ------------
associational (no design stated)  365     93.4%
difference-in-differences         3       0.8%
regression discontinuity          2       0.5%
selection model (Heckman)         2       0.5%
propensity score                  2       0.5%
interrupted time series           2       0.5%
instrumental variables / 2SLS     1       0.3%
 
==============================================================================
THE UNSTATED-FAMILY BUCKET, PRINTED IN FULL
==============================================================================
46 distinct strings, 42 papers
 
    1  Ridge regression
    1  simple regression
    1  regression of inbox quantity on spambox quantity and lagged values
    1  time-series regression
    1  Forward subset selection based regression (FSSreg) with leave-one-out cross-validation
    1  general regression analysis
    1  Lasso regression
    1  whole-brain regression with impulsivity scores as covariate
    1  ridge regression with 10-fold cross-validation
    1  regression model
    1  multilevel models with country-specific random effects
    1  elastic net
    1  random-intercept multi-level regression
    1  weighted generalized linear regression
    1  multivariate stepwise regression
    1  generalized linear mixed model (GLMM) regression
    1  generalized linear multilevel model
    1  multiple regression
    1  hierarchical multiple regression
    1  Regression Discontinuity Design
    1  Heckman selection model
    1  multivariate regression model
    1  Regression of average persistence scores on time
    1  multivariate regression
    1  multiple categorical variable regressions
    1  regression of effective sampling frequency on zip-code, category, and sampling-date factors
    1  step-wise Generalised Linear Models (GLM)
    1  generalized linear mixed models with day-level random effects
    1  generalized linear mixed effects model with random intercepts and slopes
    1  multilevel regression models with maximum-likelihood estimation
    1  fuzzy regression discontinuity
    1  Multiple Regression Analyses (MRA)
    1  generalized linear regression model with log transformation
    1  generalized additive models
    1  regression analysis
    1  Regression of estimated treatment effects on standardized pre-treatment covariates
    1  Generalized Linear Mixed Models (GLMMs)
    1  mixed-effects model
    1  propensity score weighting and trimming
    1  staggered difference-in-differences using Callaway and Sant'Anna group-time ATTs
    1  multilevel regression model with maximum-likelihood estimation
    1  general estimating equation (GEE) with exchangeable correlation
    1  general estimating equation (GEE) with AR(1) autocorrelation
    1  generalized linear regressions
    1  multivariable generalized linear mixed effects (GLME) models
    1  fixed-effects panel regression
 
==============================================================================
RESIDUE — method strings no outcome rule matched (printed in full)
==============================================================================
0 distinct strings, 0 papers (0.0% of 391)
 
 
==============================================================================
MIXED-EFFECTS / RANDOM-EFFECTS MODELS, ACROSS THE WHOLE CORPUS
==============================================================================
papers naming a random-effects model, any kind    76 of 5859
Subset                            Papers  Share of the 76
--------------------------------  ------  ---------------
recruited human participants      58      76.3%
ran a crawl                       4       5.3%
ran a crawl AND had participants  0       0.0%
neither                           14      18.4%
 
--- Every crawl paper that fits a random-effects model, in full ---
  IEEE-SP/2019/short-text-large-effect-measuring-the-impact-of-user-reviews-on-android-app-secu
      Short Text, Large Effect: Measuring the Impact of User Reviews on Android App Security & Privacy.
      [regression] Mixed-effects logistic regression | 15,835 app-update data points; random effects at application level
  IMC/2021/polls-clickbait-and-commemorative-2-bills-problematic-political-advertising-on-n
      Polls, clickbait, and commemorative $2 bills: problematic political advertising on news and media websites around the 2020 U.S. elections.
      [regression] linear mixed model analysis of variance | F(1, 744) = 0.805, n.s.
  USENIX/2023/know-your-cybercriminal-evaluating-attacker-preferences-by-measuring-profile-sal
      Know Your Cybercriminal: Evaluating Attacker Preferences by Measuring Profile Sales on an Active, Leading Criminal Market for User Impersonation at Scale
      [regression] generalized linear mixed models with day-level random effects | marginal R² = 0.264; conditional R² = 0.278
  USENIX/2024/iot-market-dynamics-an-analysis-of-device-sales-security-and-privacy-signals-and
      IoT Market Dynamics: An Analysis of Device Sales, Security and Privacy Signals, and their Interactions
      [regression] Mixed Effects Negative Binomial Model | likelihood-ratio test p < 0.001 favored device type as a random effect
 
papers saying only "repeated-measures", never naming a random effect  14
  (folding those in would take the 76 above to 90; the page publishes the strict count)
 
--- By year ---
Year  Papers
----  ------
2015  1
2016  2
2017  4
2018  2
2019  3
2020  6
2021  9
2022  12
2023  12
2024  5
2025  9
2026  11
 
--- By four-year bucket, against the regression population ---
Bucket      Regression papers  of which random effects  Share
----------  -----------------  -----------------------  -----
2010–2013   29                 0                        0.0%
2014–2017   50                 7                        14.0%
2018–2021   102                20                       19.6%
2022–2024   128                28                       21.9%
2025–2026*  82                 20                       24.4%
 
==============================================================================
THE 62 CRAWL PAPERS THAT RAN A REGRESSION
==============================================================================
denominator: 62 papers that ran a crawl AND a regression
  of the 1,120 crawl papers in the corpus, that is 5.5%
 
--- Outcome family, of the 62 ---
Outcome family                                                        Papers  Share of 62
--------------------------------------------------------------------  ------  -----------
continuous (OLS / linear / Gaussian GLM)                              28      45.2%
binary (logistic / probit / linear probability)                       20      32.3%
count (Poisson / negative binomial)                                   7       11.3%
non-linear / curve fit (power law, log-log, polynomial, exponential)  4       6.5%
family not stated in the method name                                  4       6.5%
time-to-event (Cox / hazard)                                          3       4.8%
ordinal (ordered logit / probit, cumulative link)                     2       3.2%
continuous, non-mean estimand (quantile / robust M)                   1       1.6%
multinomial / discrete choice                                         1       1.6%
proportion (beta regression)                                          1       1.6%
 
--- Dependence handling, of the 62 ---
Dependence handling                          Papers  Share of 62
-------------------------------------------  ------  -----------
none stated in the method name               54      87.1%
random effects (mixed-effects / multilevel)  4       6.5%
"repeated-measures", no random effect named  1       1.6%
cluster-robust standard errors               1       1.6%
fixed effects (dummies for the unit)         1       1.6%
 
--- By four-year bucket ---
Bucket      Crawl papers in corpus  of which ran a regression  Share
----------  ----------------------  -------------------------  -----
2010–2013   102                     8                          7.8%
2014–2017   167                     4                          2.4%
2018–2021   308                     16                         5.2%
2022–2024   345                     21                         6.1%
2025–2026*  198                     13                         6.6%
 
--- All 62, listed ---
  2010  CCS/2010/dissecting-one-click-frauds
        power-law fit
  2011  CCS/2011/fashion-crimes-trending-term-exploitation-on-the-web
        logit regression ; linear regression
  2011  IMC/2011/understanding-website-complexity-measurements-metrics-and-implications
        LASSO linear regression
  2011  USENIX/2011/measuring-and-analyzing-search-redirection-attacks-in-the-illicit-online-prescri
        Cox proportional hazard model
  2011  USENIX/2011/show-me-the-money-characterizing-spam-advertised-revenue
        least-squares linear fit
  2012  IMC/2012/new-kid-on-the-block-exploring-the-google-social-graph
        simple statistical linear regression in the log-log scale
  2013  WWW/2013/bitsquatting-exploiting-bit-flips-for-fun-or-profit
        linear regression
  2013  WWW/2013/traveling-the-silk-road-a-measurement-analysis-of-a-large-anonymous-online-marke
        linear regression
  2015  CCS/2015/an-empirical-study-of-web-vulnerability-discovery-ecosystems
        linear regression
  2017  NDSS/2017/automated-analysis-of-privacy-requirements-for-mobile-apps
        linear regression ; polynomial regression ; binary logistic regression
  2017  PETS/2017/topics-of-controversy-an-empirical-analysis-of-web-censorship-lists
        Cox proportional hazard model
  2017  USENIX/2017/characterizing-the-nature-and-dynamics-of-tor-exit-blocking
        linear regression
  2018  CCS/2018/fraud-de-anonymization-for-fun-and-profit
        regularized logistic regression
  2018  CCS/2018/predicting-impending-exposure-to-malicious-content-from-user-behavior
        logistic regression ; backward model selection using Bayesian Information Criterion
  2019  CCS/2019/oh-the-places-youve-been-user-reactions-to-longitudinal-transparency-about-third
        repeated-measures ordinal logistic regression
  2019  IEEE-SP/2019/short-text-large-effect-measuring-the-impact-of-user-reviews-on-android-app-secu
        Mixed-effects logistic regression
  2019  NDSS/2019/quantity-vs-quality-evaluating-user-interest-profiles-using-ad-preference-managers
        negative binomial regression
  2019  USENIX/2019/towards-the-detection-of-inconsistencies-in-public-security-vulnerability-report
        linear regression
  2019  USENIX/2019/understanding-ios-based-crowdturfing-through-hidden-ui-analysis
        linear forecast regression
  2019  WWW/2019/before-and-after-gdpr-the-changes-in-third-party-presence-at-public-and-private
        linear regression
  2019  WWW/2019/estimating-the-total-volume-of-queries-to-google
        Nonlinear Least Square (NLS) regression
  2019  WWW/2019/web-experience-in-mobile-networks-lessons-from-two-million-page-visits
        multiple-linear regression with standard step-wise selection and P-value filtering
  2020  CCS/2020/impersonation-as-a-service-characterizing-the-emerging-criminal-infrastructure-f
        linear regression ; logistic regression
  2020  IMC/2020/accept-the-risk-and-continue-measuring-the-long-tail-of-government-https-adoptio
        linear regressions
  2020  USENIX/2020/from-needs-to-actions-to-secure-apps-the-effect-of-requirements-and-developer-pr
        Decision Tree models
  2020  WWW/2020/the-pod-people-understanding-manipulation-of-social-media-popularity-via-recipro
        linear regression
  2021  IMC/2021/polls-clickbait-and-commemorative-2-bills-problematic-political-advertising-on-n
        linear mixed model analysis of variance
  2021  PETS/2021/privacy-preference-signals-past-present-and-future
        logistic regression
  2022  CCS/2022/a-run-a-day-wont-keep-the-hacker-away-inference-attacks-on-endpoint-privacy-zone
        logistic regression
  2022  IMC/2022/causal-impact-of-android-go-on-mobile-web-performance
        ordinary least squares (OLS) regression ; Heckman selection model ; probit regression
  2022  PETS/2022/checking-websites-gdpr-consent-compliance-for-marketing-emails
        logistic regression
  2022  PETS/2022/how-can-and-would-people-protect-from-online-tracking
        binary logistic regression
  2022  WWW/2022/left-or-right-a-peek-into-the-political-biases-in-email-spam-filtering-algorithm
        logistic regression with lasso regularization
  2022  WWW/2022/leveraging-googles-publisher-specific-ids-to-detect-website-administration
        linear regression ; linear regression
  2022  WWW/2022/moral-emotions-shape-the-virality-of-covid-19-misinformation-on-social-media
        negative binomial regression
  2022  WWW/2022/the-impact-of-twitter-labels-on-misinformation-spread-and-user-engagement-lesson
        robust linear regression ; multinomial regression ; beta regression
  2022  IEEE-SP/2022/robbery-on-devops-understanding-and-mitigating-illicit-cryptomining-on-continuou
        linear regression
  2023  IMC/2023/evolving-bots-the-new-generation-of-comment-bots-and-their-underlying-scam-campa
        ordinary least squares linear regression ; multiple categorical variable regressions
  2023  IEEE-SP/2023/sok-decentralized-finance-defi-attacks
        Structural Equation Modeling (SEM) ; ordinary least squares
  2023  PETS/2023/comparing-large-scale-privacy-and-security-notifications
        logistic regression
  2023  WWW/2023/a-method-to-assess-and-explain-disparate-impact-in-online-retailing
        regression of effective sampling frequency on zip-code, category, and sampling-date factors ; generalized least squares with errors clustered at the product level
  2023  WWW/2023/misbehavior-and-account-suspension-in-an-online-financial-communication-platform
        logistic regression
  2023  USENIX/2023/know-your-cybercriminal-evaluating-attacker-preferences-by-measuring-profile-sal
        generalized linear mixed models with day-level random effects
  2024  IMC/2024/of-choices-and-control-a-comparative-analysis-of-government-hosting
        Ordinary Least Squares (OLS) regression
  2024  CCS/2024/characterizing-and-mitigating-phishing-attacks-at-cctld-scale
        linear regression
  2024  USENIX/2024/does-online-anonymous-market-vendor-reputation-matter
        Cox proportional-hazards regression
  2024  USENIX/2024/iot-market-dynamics-an-analysis-of-device-sales-security-and-privacy-signals-and
        Mixed Effects Negative Binomial Model
  2024  PETS/2024/what-does-it-mean-to-be-creepy-responses-to-visualizations-of-personal-browsing
        ordinal logistic regression
  2024  WWW/2024/social-media-discourses-on-interracial-intimacy-tracking-racism-and-sexism-throu
        multivariate linear probability regression with short-video ID fixed effects
  2025  IMC/2025/learning-as-to-organization-mappings-with-borges
        linear regression
  2025  NDSS/2025/attributing-open-source-contributions-is-critical-but-difficult-a-systematic-analysis-of-github-practices-and-their-impact-on-software-supply-chain-security
        linear regression
  2025  PETS/2025/tracker-installations-are-not-created-equal-understanding-tracker-configuration
        Logistic Regression
  2025  PETS/2025/more-and-scammier-ads-the-perils-of-youtubes-ad-privacy-settings
        Poisson regression ; negative binomial regression
  2025  USENIX/2025/assessing-the-aftermath-the-effects-of-a-global-takedown-against-ddos-for-hire-s
        negative binomial regression for interrupted time series
  2025  WWW/2025/causal-insights-into-parlers-content-moderation-shift-effects-on-toxicity-and-fa
        Difference-in-Differences (DiD) with Ordinary Least Squares (OLS)
  2025  PETS/2025/sheeps-clothing-wolfish-intent-automated-detection-and-evaluation-of-problematic
        mediation analysis
  2026  PETS/2026/because-i-didnt-touch-these-and-even-dont-know-why-i-should-to-change-these-why
        negative binomial regression (NB2)
  2026  PETS/2026/the-role-of-online-forums-in-developer-understanding-of-privacy-law-a-reddit-cas
        LASSO logistic regression
  2026  PETS/2026/overcoming-language-barriers-multilingual-analysis-of-the-2023-swiss-privacy-law
        difference-in-differences using Ordinary Least Squares (OLS) ; logistic difference-in-differences regression
  2026  NDSS/2026/repairing-trust-in-domain-name-disputes-practices-insights-from-a-quarter-centurys-worth-of-squabbles
        logistic regression
  2026  WWW/2026/moral-outrage-shapes-commitments-beyond-attention-multimodal-moral-emotions-on-y
        negative binomial regression
  2026  USENIX/2026/chameleon-channels-measuring-youtube-accounts-repurposed-for-deception-and-profi
        Design-based Supervised Learning logistic regression
 
==============================================================================
COUNT AND PROPORTION OUTCOMES OVER TIME
==============================================================================
count-model papers (any kind)       24
beta-regression papers (any kind)   3
Bucket      Regression papers  count models  Share of regression papers
----------  -----------------  ------------  --------------------------
2010–2013   29                 0             0.0%
2014–2017   50                 0             0.0%
2018–2021   102                6             5.9%
2022–2024   128                5             3.9%
2025–2026*  82                 6             7.3%
 
--- Every count-model paper, in full ---
  2011  IEEE-SP/2011/homealone-co-residency-detection-in-the-cloud-via-side-channel-analysis
        [hypothesis-test] Poisson-binomial probability distribution with thresholds Tc and Td
  2012  CCS/2012/manufacturing-compromise-the-emergence-of-exploit-as-a-service
        [descriptive-only] Poisson process model
  2016  CCS/2016/mems-gyroscopes-as-physical-unclonable-functions
        [descriptive-only] Poisson distribution fit
  2016  PETS/2016/listening-to-whispers-of-ripple-linking-wallets-and-deanonymizing-transactions-i
        [descriptive-only] Poisson distribution
  2017  NDSS/2017/a-large-scale-analysis-of-the-mnemonic-password-advice
        [descriptive-only] negative binomial models
  2019  CCS/2019/privacy-aspects-and-subliminal-channels-in-zcash
        [descriptive-only] Poisson distribution
  2019  IMC/2019/booting-the-booters-evaluating-the-effects-of-police-interventions-in-the-market
        [regression] negative binomial regression
  2019  NDSS/2019/quantity-vs-quality-evaluating-user-interest-profiles-using-ad-preference-managers
        [regression] negative binomial regression
  2019  USENIX/2019/cognitive-triaging-of-phishing-attacks
        [regression] Poisson regression
  2020  IMC/2020/turning-up-the-dial-the-evolution-of-a-cybercrime-market-through-set-up-stable-a
        [regression] Zero-inflated Poisson regression
  2020  USENIX/2020/understanding-security-mistakes-developers-make-qualitative-analysis-from-build
        [regression] Poisson regression
  2021  USENIX/2021/now-im-a-bit-angry-individuals-awareness-perception-and-responses-to-data-breach
        [regression] quasi-Poisson regression
  2022  USENIX/2022/seeing-the-forest-for-the-trees-understanding-security-hazards-in-the-3gpp-ecosy
        [regression] Negative Binomial Distribution
  2022  WWW/2022/effective-messaging-on-social-media-what-makes-online-content-go-viral
        [regression] Zero-Inflated Negative Binomial Regression Model (ZINBR)
  2022  WWW/2022/moral-emotions-shape-the-virality-of-covid-19-misinformation-on-social-media
        [regression] negative binomial regression
  2022  WWW/2022/hate-speech-in-the-political-discourse-on-social-media-disparities-across-partie
        [regression] beta regression
  2022  WWW/2022/the-impact-of-twitter-labels-on-misinformation-spread-and-user-engagement-lesson
        [regression] beta regression
  2023  USENIX/2023/mixed-signals-analyzing-ground-truth-data-on-the-users-and-economics-of-a-bitcoi
        [hypothesis-test] Poisson binomial distribution
  2023  WWW/2023/combining-worker-factors-for-heterogeneous-crowd-task-assignment
        [regression] Beta Regression
  2024  USENIX/2024/iot-market-dynamics-an-analysis-of-device-sales-security-and-privacy-signals-and
        [regression] Mixed Effects Negative Binomial Model
  2024  IEEE-SP/2024/patchy-performance-uncovering-the-vulnerability-management-practices-of-iot-cent
        [regression] negative binomial generalized linear model with log link
  2025  IMC/2025/on-youtube-search-api-use-in-research
        [regression] Poisson regression objective in a gradient boosting model
  2025  PETS/2025/more-and-scammier-ads-the-perils-of-youtubes-ad-privacy-settings
        [regression] Poisson regression
        [regression] negative binomial regression
  2025  USENIX/2025/assessing-the-aftermath-the-effects-of-a-global-takedown-against-ddos-for-hire-s
        [regression] negative binomial regression for interrupted time series
  2026  PETS/2026/because-i-didnt-touch-these-and-even-dont-know-why-i-should-to-change-these-why
        [regression] negative binomial regression (NB2)
  2026  WWW/2026/consensus-stability-of-community-notes-on-x
        [regression] Poisson regression with robust standard errors
  2026  WWW/2026/moral-outrage-shapes-commitments-beyond-attention-multimodal-moral-emotions-on-y
        [regression] negative binomial regression
 
==============================================================================
CAUSAL IDENTIFICATION DESIGNS OVER TIME
==============================================================================
papers naming a causal-identification design, any kind   11 of 5859
Bucket      Regression papers  causal design  Share
----------  -----------------  -------------  -----
2010–2013   29                 0              0.0%
2014–2017   50                 0              0.0%
2018–2021   102                0              0.0%
2022–2024   128                5              3.9%
2025–2026*  82                 5              6.1%
 
--- Every one of them, in full ---
  2022  CCS/2022/empirical-analysis-of-eip-1559-transaction-fees-waiting-times-and-consensus-secu
        [regression] Regression Discontinuity Design
  2022  IMC/2022/causal-impact-of-android-go-on-mobile-web-performance
        [regression] Heckman selection model
  2023  USENIX/2023/glowing-in-the-dark-uncovering-ipv6-address-discovery-and-scanning-strategies-in
        [effect-size] difference-in-differences
  2023  WWW/2023/automated-content-moderation-increases-adherence-to-community-guidelines
        [regression] fuzzy regression discontinuity
        [regression] two-stage least squares regression
  2023  WWW/2023/hidden-indicators-of-collective-intelligence-in-crowdfunding
        [regression] two-stage Heckman probit model
  2023  WWW/2023/longitudinal-assessment-of-reference-quality-on-wikipedia
        [regression] logistic regression for propensity scores
  2025  IEEE-SP/2025/understanding-users-security-and-privacy-concerns-and-attitudes-towards-conversa
        [regression] Ordinary Least Squares interrupted time-series regression
  2025  USENIX/2025/assessing-the-aftermath-the-effects-of-a-global-takedown-against-ddos-for-hire-s
        [regression] negative binomial regression for interrupted time series
  2025  WWW/2025/causal-insights-into-parlers-content-moderation-shift-effects-on-toxicity-and-fa
        [regression] Difference-in-Differences (DiD) with Ordinary Least Squares (OLS)
  2026  WWW/2026/community-fact-checks-do-not-break-follower-loyalty
        [regression] propensity score weighting and trimming
        [regression] staggered difference-in-differences using Callaway and Sant'Anna group-time ATTs
  2026  PETS/2026/overcoming-language-barriers-multilingual-analysis-of-the-2023-swiss-privacy-law
        [regression] difference-in-differences using Ordinary Least Squares (OLS)
        [regression] logistic difference-in-differences regression
 
==============================================================================
REPORTING GAPS
==============================================================================
papers whose regression tuples name no model the fold could place   19  (4.9% of 391)
    WWW/2012/a-dual-mode-user-interface-for-accessing-3d-content-on-the-world-wide-web :: analysis of variance (ANOVA)
    WWW/2012/investigating-the-distribution-of-password-choices :: least-squares line
    IMC/2016/an-empirical-analysis-of-a-large-scale-mobile-cloud-storage-service :: maximum likelihood estimation
    IMC/2017/connected-cars-in-cellular-network-a-measurement-study :: trend lines
    WWW/2017/some-recipes-can-do-more-than-spoil-your-appetite-analyzing-the-security-and-pri :: Nagelkerke's pseudo R2
    WWW/2019/how-serendipity-improves-user-satisfaction-with-recommendations-a-large-scale-us :: path analysis
    IEEE-SP/2020/security-update-labels-establishing-economic-incentives-for-security-patching-of :: latent class analysis
    PETS/2020/explaining-the-technology-use-behavior-of-privacy-enhancing-technologies-the-cas :: partial least squares structural equation modelling (PLS-SEM)
    USENIX/2020/from-needs-to-actions-to-secure-apps-the-effect-of-requirements-and-developer-pr :: Decision Tree models
    PETS/2022/personal-information-inference-from-voice-recordings-user-awareness-and-privacy :: regression analyses based on the Akaike information criterion
    PETS/2023/investigating-privacy-decision-making-processes-among-nigerian-men-and-women :: Partial Least Squares structural equation modelling (PLS-SEM)
    USENIX/2023/bug-hunters-perspectives-on-the-challenges-and-benefits-of-the-bug-bounty-ecosys :: log-linear Bradley-Terry (LLBT) modeling
    WWW/2023/not-seen-not-heard-in-the-digital-world-measuring-privacy-practices-in-childrens :: regression test
    PETS/2024/internet-users-willingness-to-disclose-biometric-data-for-continuous-online-acco :: Partial Least Squares based Structural Equation Modeling (PLS-SEM)
    PETS/2024/simply-tell-me-how-on-trustworthiness-and-technology-acceptance-of-attribute-bas :: covariance-based structural equation modeling with WLSMV estimation
    WWW/2024/global-news-synchrony-and-diversity-during-the-start-of-the-covid-19-pandemic :: regression models with VIF- and AIC-based feature/model selection
    PETS/2025/privacy-perceptions-and-behaviors-towards-targeted-advertising-on-social-media-a :: path model / mediation analysis
    WWW/2026/how-graphs-can-help-you-stay-informed-in-an-evolving-world :: regression line
    PETS/2025/sheeps-clothing-wolfish-intent-automated-detection-and-evaluation-of-problematic :: mediation analysis
 
papers with at least one bare "regression"/"regression analysis"    99  (25.3% of 391)
 
papers with a reported value on >=1 regression tuple (detail != null)  313  (80.1% of 391)
 
What the `detail` field carries, of the 391 (a FLOOR, not a measurement — see the page):
Detail mentions        Papers  Share of 391
---------------------  ------  ------------
an R² or pseudo-R²     31      7.9%
a coefficient value    130     33.2%
a p-value              76      19.4%
a confidence interval  23      5.9%
a sample size          14      3.6%
 
==============================================================================
FULL-TEXT PROBES (lower bounds — a paper may phrase it differently)
==============================================================================
searched 5869 paper.cols.txt files, whitespace collapsed
 
Probe                             Papers matching
--------------------------------  ---------------
cluster-robust standard errors    4
the words "random effect"         61
"random intercept"                29
generalised estimating equations  2
"overdispersion"                  6
"zero-inflated"                   4
a variance inflation factor       27
"linear probability model"        2
"marginal effect"                 20
"odds ratio"                      100
"interrupted time series"         5
"regression discontinuity"        6
"difference-in-differences"       11
 
--- Proximity probe: a random-effects phrase within 120 chars of a site/domain/page word ---
  IEEE-SP/2010/a-practical-attack-to-de-anonymize-social-network-users  (1 windows)
      directory is not restricted. As a result, everyone can download it from the web. The directory itself is organized in a multilevel hierarchical collection of alphabetically ordered lists that provide pointers to individual web pages to make it convenient for a
  PETS/2015/an-automated-approach-for-complementing-ad-blockers-blacklists  (1 windows)
      dern web traffic. In Proc. IMC '11, pages 295-312, 2011. [23] T. Karagiannis, K. Papagiannaki, and M. Faloutsos. Blinc: multilevel traffic classification in the dark. SIGCOMM Comput. Commun. Rev., 35(4):229-240, 2005. [24] H. Kim, K. Claffy, M. Fomenkov, D. Ba
  IMC/2016/sneaking-past-the-firewall-quantifying-the-unexpected-traffic-on-major-tcp-and-u  (1 windows)
      LOBECOM'04. IEEE, volume 3, pages 1532-1538. IEEE, 2004. [12] T. Karagiannis, K. Papagiannaki, and M. Faloutsos. BLINC: Multilevel Traffic Classification in the Dark. In ACM SIGCOMM Computer Communication Review, volume 35, pages 229-240. ACM, 2005. [13] G. Ma
  PETS/2016/do-not-track-me-sometimes-users-contextual-preferences-for-web-tracking  (1 windows)
      he use of several classifiers (including Asymmetric AdaBoost [42], Support Vector Machines [18], and Generalized Linear Mixed Effects Regression [10]) to predict a user's comfort with tracking of specific page visits based on properties of the web page and tha
  NDSS/2017/catching-worms-trojan-horses-and-pups-unsupervised-detection-of-silent-delivery  (1 windows)
      imization problem to detect locksteps, which makes it highly sensitive to the choice of seed domains and FastGreedy [7] Multilevel [12] the times provided. Furthermore, this serial implementation of Number of Communities 6919 6439 Average #nodes/community 21 2
  IMC/2018/multilevel-mda-lite-paris-traceroute  (1 windows)
       the previous section. Fakeroute is available as free open-source software at the URL mentioned at the end of Sec. 1. 4 MULTILEVEL ROUTE TRACING The third principal contribution of this paper, after the MDA-Lite and Fakeroute of the previous sections, is IPv4 
  USENIX/2021/a-large-scale-study-of-user-behavior-expectations-and-engagement-with-android-pe  (1 windows)
      e as a fixed effect and the par 1 Some of the explanations were actually permission requests by web ticipant and app as random effects. The trained model shows pages in a browser a significant difference between expecting and not expecting 810 30th USENIX Secu
  IMC/2022/what-factors-affect-targeting-and-bids-in-online-advertising-a-field-measurement  (9 windows)
      anatory variables), as well as the website the ads appeared on, the bidder, the individual, and the category of the ad (random effects). We selected our model using the top-down method suggested by Zuur et al. [49]: we started with a full specified model, incl
  PETS/2022/analyzing-the-feasibility-and-generalizability-of-fingerprinting-internet-of-thi  (1 windows)
      r and Communications Security (CCS), page 263-274, 2014. [25] T. Karagiannis, K. Papagiannaki, and M. Faloutsos. Blinc: Multilevel traffic classification in the dark. In Proceedings of the 2005 Conference on Applications, Technologies, Architectures, and Proto
  PETS/2022/if-this-context-then-that-concern-exploring-users-concerns-with-ifttt-applets  (1 windows)
      Conference on Human Factors in Computing Systems, pages 2003-2012. ACM, 2009. [39] Gerhard Tutz and Wolfgang Hennevogl. Random effects in ordinal regression models. Computational Statistics & Data Analysis, 22(5):537-557, 1996. [40] Blase Ur, Elyse McManus, Me
  CCS/2024/the-illusion-of-randomness-an-empirical-analysis-of-address-space-layout-randomi  (3 windows)
      e, which is 64 bits or 32 bits according to the hardware architecture. However, the practical limit is lower because of Multilevel Paging and memory pages. Multilevel Paging is a technique that translates virtual memory addresses into physical ones. The idea i
  IEEE-SP/2024/to-auth-or-not-to-auth-a-comparative-analysis-of-the-pre-and-post-login-security  (1 windows)
      urs and did not attempt to visit intentionally invalid URLs which may have different security configurations [48]), and random effects might affect both states differently. We tried to minimize all such errors by starting the crawl in both states for each site
  WWW/2024/are-adversarial-phishing-webpages-a-threat-in-reality-understanding-the-users-pe  (1 windows)
      * worryingly, we found that their ability to recognize phishing web-Table 2: Webpage Classification Analysis - Logistic mixed-effects repages is much worse; intriguingly, however, it appears that our gression model: we predict whether a website is classified c
  CCS/2025/layered-overlapping-and-inconsistent-a-large-scale-analysis-of-the-multiple-priv  (1 windows)
      ivacy policies, identifying persistent clarity issues despite regulatory efforts, and finding that regulation has had a mixed effect on privacy policy transparency [8]. For example, Chen et al.'s [15] analysis of the privacy policies of 95 popular websites fou
  IEEE-SP/2025/restricting-the-link-effects-of-focused-attention-and-time-delay-on-phishing-war  (1 windows)
       pre-registration), including participants' interactions with false positives and different phishing URL manipulations, mixed-effect regressions on predictive factors for clicking on different types of link, and qualitative analysis of open-text responses. 61 
  IEEE-SP/2025/understanding-the-efficacy-of-phishing-training-in-practice  (1 windows)
       Security (SOUPS), pages 339-357, August 2021. [15] Andrew Gelman and Jennifer Hill. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press, December 2006. [16] William J. Gordon, Adam Wright, Robert J. Glynn, Jigar Kadak
 
  16 papers of 5869
 
--- Cluster-robust matches, listed, because the page names them individually ---
  WWW/2023/a-method-to-assess-and-explain-disparate-impact-in-online-retailing
  PETS/2026/overcoming-language-barriers-multilingual-analysis-of-the-2023-swiss-privacy-law
  WWW/2026/community-fact-checks-do-not-break-follower-loyalty
  WWW/2026/consensus-stability-of-community-notes-on-x

scripts/counts_and_repeats.py, published verbatim on the page as a <file python> block. Synthetic data, fixed seed 20260819, numpy 2.4.6, statsmodels 0.14.6. Nothing in it is a measurement claim.

What it does and does not establish:

• Part A measures the false-positive rate of Poisson regression on negative-binomial counts with variance/mean = 13, under a true null. It establishes that the estimator is anti-conservative at that dispersion. It does not establish that real tracker counts have that dispersion — nobody in this corpus reports one, which is why it is an open question on the page.
• Part B measures the false-positive rate and power of a naive logistic regression against a site-clustered one, at a latent-scale ICC of 0.41. Again, the ICC is chosen to be plausible; no paper reports a real one.
• The ICC is on the latent (logit) scale, computed as σ²/(σ² + π²/3). That is the standard convention for a logistic random-intercept model and it is not the same as an observed-scale ICC; the number is not comparable with an ICC reported for a continuous outcome.
• Part B's random-number stream is shared across its sub-experiments, so the B1 two-wave row appears twice with slightly different values (7.8% in the first block, 8.6% in the wave-scaling block). That is Monte-Carlo noise at 500 iterations, not a bug, and both are printed rather than one being suppressed. At 500 iterations the standard error of a 5% rate is about 1.0 percentage point.

Real, unedited output:

counts_and_repeats-output.txt

Two things that change a regression's answer in a web measurement.
 
Both run on SYNTHETIC data with a fixed seed, because the point is a property
of the estimator rather than a fact about the web. Nothing here is a
measurement claim; every number is reproducible by running the file.
 
  A. TRACKER COUNTS ARE OVERDISPERSED. "Number of third parties on a site" is a
     count with variance many times its mean. A Poisson regression assumes
     variance == mean, so it reports standard errors that are far too small and
     rejects a true null far more often than its nominal alpha. This part
     measures that rate, and shows the fixes: a negative-binomial model, a
     quasi-Poisson scale correction, and heteroskedasticity-robust standard
     errors on the Poisson fit.
 
     The negative-binomial row is fitted TWICE on purpose. `sm.families.
     NegativeBinomial(alpha=...)` takes the dispersion as a FIXED input, so
     handing it the true simulated value is an oracle an analyst does not have.
     The second row estimates the dispersion jointly with the coefficients,
     which is what you actually do, and is the number to read.
 
  B. REPEATED CRAWLS OF THE SAME SITE CUT BOTH WAYS. When the same sites are
     crawled twice, the rows are not independent, and what that does to your
     test depends on WHERE the predictor varies:
 
       B1. A SITE-LEVEL predictor (category, rank bucket, CMP vendor) is
           constant within a site. Non-independence then inflates the
           false-positive rate, exactly as on Statistics:Hypothesis testing.
 
       B2. A WITHIN-SITE predictor (before/after, treatment/control applied to
           the same site) varies inside the site. Non-independence then makes
           the naive test CONSERVATIVE: the standard error is too LARGE and you
           lose power. Clustering on the site is still the right thing to do,
           but the direction of the error is the opposite one, and "clustering
           always inflates false positives" is the wrong rule to carry around.
 
     Both are measured against cluster-robust standard errors clustered on the
     site, which is the cheapest correct analysis and the one the corpus's own
     panel study uses.
 
seed = 20260819; numpy 2.4.6; statsmodels 0.14.6
 
A. OVERDISPERSED COUNTS  (500 iterations, 1000 sites each)
   third parties per site ~ NegBin(mean=12, var=156), variance/mean = 13.0
   two groups drawn from that SAME distribution -> every rejection is false
 
   model                                         false positives   (nominal 5%)
   Poisson GLM                                            60.4%
   Poisson GLM + robust (HC0) SE                           5.2%
   quasi-Poisson (Pearson scale)                           5.0%
   negative binomial GLM, alpha KNOWN (oracle)             5.0%
   negative binomial, alpha ESTIMATED                      5.0%
 
B. THE SAME SITES, CRAWLED TWICE  (500 iterations, 500 sites x 2 waves)
   site random intercept SD = 1.5, latent-scale ICC = 0.41
 
   B1  site-level predictor (constant within a site), TRUE NULL
      naive logistic, rows treated as independent    7.8% false positives   mean SE 0.135
      same model, SEs clustered on the site          5.2% false positives   mean SE 0.152
   B2  within-site predictor (wave 1 vs wave 2), TRUE NULL
      naive logistic, rows treated as independent    3.0% false positives   mean SE 0.135
      same model, SEs clustered on the site          4.8% false positives   mean SE 0.114
 
   B1 again, as the panel gets longer (site-level predictor, TRUE NULL)
       crawls per site     rows     naive   clustered
                     2     1000     8.6%       4.4%
                     4     2000    11.4%       3.6%
                     8     4000    25.2%       6.2%
                    12     6000    32.4%       4.2%
 
   B3  the same within-site predictor with a REAL effect (log-odds 0.4)
      naive logistic                                61.6% power
      SEs clustered on the site                     75.2% power
 
Nothing above is a measurement. It is what these estimators do to data 
shaped like a crawl's.

Every probe is a lower bound. They exist as an independent check on the tuple-based counts, computed from the same paper.cols.txt the extractor read (5,869 files).

Whitespace is collapsed before matching, and this was a real bug in the first draft. multilevel\nmixed-effects in WWW/2026 consensus-stability-of-community-notes-on-x is a line break falling inside a two-word phrase; the un-normalised probe missed it. Normalising (-\n joined, \s+ collapsed) moved four published counts: “random effect” 59→61, “random intercept” 27→29, “odds ratio” 99→100, “linear probability model” 1→2. The cluster-robust count did not move. The naive version of this probe would have published “one paper in 5,869 uses a linear probability model” when there are two.

Doing it by holding all 5,869 normalised texts in memory exhausts the V8 heap (OOM at ~4 GB). The script streams one file at a time and tests all thirteen probes per file.

The proximity probe, added during review. A keyword probe cannot tell you what a random effect is:

const RE   = /random (intercepts?|effects?|slopes?)|mixed[- ]effects?|multilevel|\(1 ?\| ?\w+\)/gi;
const UNIT = /\b(website|web site|websites|site|sites|domain|domains|page|pages|url|urls)\b/i;
// a hit is any RE match with a UNIT word inside the surrounding 260 characters

16 papers of 5,869. Fifteen are bibliography entries (“Blinc: multilevel traffic classification”), unrelated senses (“Multilevel Paging”, “multilevel hierarchical collection”), or a random effect for something else with a site word nearby. The sixteenth is Zeng et al. [1Zeng, Eric; McAmis, Rachel; Kohno, Tadayoshi; Roesner, Franziska (2022): "What Factors Affect Targeting and Bids in Online Advertising? A Field Measurement Study", in: Proceedings of the ACM Internet Measurement Conference, pp. 210-229. (DOI)] (IMC 2022), who fit “random intercepts for website, participant, bidder, and ad category”.

This probe exists because the first draft of the page was wrong. It claimed that no paper in the corpus models the website as a random effect. The keyword probe could not have found Zeng et al., because nothing in their sentence is a phrase a keyword probe would carry. The page now says one paper does, and that it is a field study with participants rather than a repeated crawl — which is the claim the evidence actually supports. All 16 hits are printed by the report script so a reader can dismiss the fifteen themselves.

Probe hits read individually, because a keyword hit is not a use. The four zero-inflated matches: IMC/2020 turning-up-the-dial and WWW/2022 effective-messaging fit one (Zero-inflated Poisson regression, Zero-Inflated Negative Binomial Regression Model (ZINBR)); WWW/2019 modeling-item-specific-temporal-dynamics and USENIX/2024 iot-market-dynamics have the phrase only in their bibliographies (a zero-inflated Poisson paper, and the glmmTMB package paper). The second of those is the only crawl paper in the corpus with the phrase anywhere in it. A first draft of the page said “four papers mention zero-inflation and none is a crawl”, which was wrong on both halves.

Probes that were run and rejected:

• clustered SE (case-insensitive) — matched 8 papers, of which four were clustered sequences, clustered security events and similar. Prefix matching on a two-letter abbreviation is not a probe. Replaced with the explicit list above.
• clustered at the unanchored — matched “usage is clustered at the low end” in PETS/2026/personal-data-flows. Anchored to a unit noun (site, website, domain, user, note, product, participant), which removed it.
• cox|hazard|survival for time-to-event — matched 135 papers on “hazard” and “survival” in unrelated security senses. Not used; the axis-1 fold's \bcox\b|proportional[-\s]?hazard is what the page reports.
• Shell grep -rl was not used for any published probe. Some paper.cols.txt files are detected as binary and are silently skipped; the node probe reads every file with an explicit latin1 encoding. See the corpus page for the general form of this trap.

Disagreement with Hypothesis testing, which this run did not resolve. That page reports a probe for clustered standard errors finding three papers; this page's probe finds four, adding [2Chuai, Yuwei; Lenzini, Gabriele; Pröllochs, Nicolas (2026): "Consensus Stability of Community Notes on X", in: Proceedings of the ACM Web Conference, pp. 8885-8896. (DOI)]. Two things differ and both favour the higher count: this probe is anchored on a unit noun (clustered at the (site|website|domain|user|note|product|participant)) where the sibling's is not, and this one collapses whitespace first. The two pages also state different full-text denominators — 5,855 files there against 5,869 here — which was not chased. Neither page has been changed to match the other, because re-deriving the sibling's figure means re-running the sibling's own probe, which belongs to that page's audit trail and not to this one. Recorded so that whoever refreshes statistics:hypothesis_testing next knows this is waiting. A reader can reach both numbers in two clicks, which is exactly why it is written down rather than left.

Disagreement between the probe and the fold, disclosed on the page. The probe finds 4 clustering papers; axis 2 finds 1. The gap is Chuai et al. [2Chuai, Yuwei; Lenzini, Gabriele; Pröllochs, Nicolas (2026): "Consensus Stability of Community Notes on X", in: Proceedings of the ACM Web Conference, pp. 8885-8896. (DOI)] and Bobek et al. [3Bobek, Michelle; Pröllochs, Nicolas (2026): "Community Fact-Checks Do Not Break Follower Loyalty", in: Proceedings of the ACM Web Conference. (DOI)], whose regression tuples say “robust standard errors” while the clustering statement lives in a different sentence (in one case in a descriptive-only tuple), plus Nenadić et al. [4Nenadić, Luka; Rodriguez, David; Calandrino, Joseph A. (2026): "Overcoming Language Barriers: Multilingual Analysis of the 2023 Swiss Privacy Law's Impact", Proceedings on Privacy Enhancing Technologies 2026(4):703-723. (DOI)], whose clustering is stated in prose rather than in the method name. The fold reads method strings; the probe reads the paper. Where they disagree the probe is right and the page uses the probe's figure for the “four papers cluster” claim while reporting the fold's 1 in the fold's own table.

node scripts/quote_check.mjs –statistics regression –show 400 against data/fulltext/<year>/<venue>/<slug>/paper.cols.txt:

530 quotes checked: 267 exact, 170 partial (>=60% of 5-word windows), 93 below threshold,
0 with no full text on disk.

Below-threshold is not “unsupported”. The five-word-window test is harsh on a fifteen-word quote (eleven windows), and a dropped citation marker or a two-column splice moves several windows at once. Example: USENIX/2020 understanding-security-mistakes scored below threshold because the extractor's quote omits the reference [15, 67-106] that sits inside the sentence in the PDF. The 93 were not individually chased.

No published quote failed. The one flagged row ([9Votipka, Daniel; Fulton, Kelsey R.; Parker, James; Hou, Matthew; Mazurek, Michelle L.; Hicks, Michael (2020): "Understanding security mistakes developers make: Qualitative analysis from Build It, Break It, Fix It", in: Proceedings of the USENIX Security Symposium. (Link)]) is a dropped citation marker inside an otherwise exact sentence.

Every external reference was verified against a primary record on 2026-08-19, by fetching, not by recall.

Rejected sources. No blog post, vendor page, SEO listicle or secondary summary was used for any claim on this page. Two categories were considered and rejected:

• Textbook chapters found by search (e.g. a Gujarati chapter on count models that Crossref returned alongside Ver Hoef & Boveng). Rejected in favour of a primary journal article that makes the specific comparison the page makes.
• Preprint DOIs where a published version exists (Callaway & Sant'Anna, Shmueli, Barr et al. all have OSF/SSRN records). Rejected on principle: the page cites what a reader will be asked to cite.

28 entries appended before the closing </bibtex> in Bibliography across four saves, taking it from 402 to 430:

• 13 corpus papers generated by node scripts/bibgen.mjs <venue>/<year>/<slug> …, so titles and DOIs are publisher metadata rather than model recall.
• 15 methodological references hand-written from the Crossref records in the table above.
• 5 generated entries were dropped because the paper was already in the bibliography under the identical key and DOI: mai2025_more, zeng2021_polls, butkiewicz2011_website, becerrilarreola2023_method, bobek2026_community. The page cites the existing entries.
• Duplicate check was three-way — key string, DOI, and normalised title — across the whole live bibliography, not a key comparison alone. A key-only check has passed a real duplicate on this wiki before.
• PETS and USENIX index records carry no authors, so scripts/fetch_authors.py was run for the six affected slugs. Five resolved from the landing page. One did not — PETS/2025 tracker-installations-are-not-created-equal — and its author list was read off petsymposium.org/popets/2025/popets-2025-0151.php by hand and written into out/authors.json. That is a hand edit and is recorded here because nothing else records it.

Render verification after saving. The bibtex4dw plugin serves a cached parse, so the very first render showed 7 of the then-30 references with no warning at all. ?purge=true on Bibliography and then on the citing page fixes it, and it has to be redone after every bibliography save.

Final state, verified post-purge on the published pages:

The content page's 37th marker is LePochat2019_tranco inside the template's /* … */ comment block, which is not rendered — every page on this wiki carries it. Four internal links render red (artifacts, design:automated_measurements, design:user_studies, statistics:biases); all four are pages the wiki already promises from start and from Hypothesis testing, so they are planned pages rather than broken links. Both cross-page anchors were checked against the sibling's rendered id attributes.

• The real intra-class correlation of any web-measurement outcome. No paper in the corpus reports one. The simulation's 0.41 is chosen, not measured, and the page says so. This is the single figure that would make the page's central argument quantitative rather than illustrative.
• The real overdispersion of a tracker count. Same problem: six papers in 5,869 use the word “overdispersion” and none reports a variance-to-mean ratio for a crawl outcome.
• Whether papers that do not name a dependence treatment nonetheless applied one. The extraction records method strings; the full-text probes are the check and they agree with the tuple counts, but neither can prove a negative.
• Whether the detail field measures reporting practice. It does not, and the page says so in a footnote. detail is a capped free-text field the extractor filled opportunistically, so “23 of 391 papers record a confidence interval” is a floor on the extraction, not a rate in the literature. A figure that was considered and not published: “only 5.9% of regression papers report a confidence interval”. It would have been the page's most quotable number and it is not supportable.
• How many of the 132 binary-outcome papers report marginal effects rather than odds ratios. The full-text probe gives 20 and 99 corpus-wide, but not restricted to the 391 and not restricted to the paper's own model, so the page reports it as a corpus-wide ratio and does not claim more.
• Whether the ten causal-design papers are a turn or a coincidence. Half sit in the provisional 2025–2026 years.
• Four pre-existing duplicate pairs in Bibliography, surfaced by this run's whole-bibliography normalised-title scan and not fixed here because none is cited by this page and each belongs to another page: lerner2016internet/lerner2016_internet, bouhoula2024automated/bouhoula2024_automated, fouad2022my/fouad2022_cookie, bottger2025_regional/boettger2025_regional. Recorded so the next run does not have to rediscover them.

• Re-run report_regression.mjs after the next corpus refresh and diff against report_regression-output.txt. Every figure on the page comes from it.
• If anyone measures an ICC or an overdispersion for a real crawl outcome, the simulation's parameters should be replaced with it and the page's “chosen, not measured” caveats removed.
• Reconcile the cluster-robust count with Hypothesis testing (three there, four here) and the full-text denominator (5,855 there, 5,869 here). Belongs to whoever refreshes that page.
• Merge the four duplicate bibliography pairs listed above, on whichever page owns them.
• quote_check.mjs's five-word-window threshold is harsh on short quotes; a length-aware threshold would cut the 93 below-threshold figure without loosening the test. Affects every page on this site, not just this one.

Four reviewers, all told explicitly that the author's context might not be exhaustive and all handed the page text, the report script, its real output, the simulation and these notes. The three focused ones ran in parallel first; the generic one ran last, after their findings were applied.

Three further defects were found by re-reading the page against the data rather than by any reviewer. They are listed because a review log that only records what reviewers caught overstates what the review layer is worth.

Plus a set of consistency slips a fresh read caught: a heading that said “Three remedies” over a four-row table, a stale 33 where the fold now gives 42, an unsourced date on the sandwich estimator, and an odds ratio count that had not been updated with the whitespace-normalised probe.

Handed both pages, both scripts, both outputs, and the live rendered HTML, with no checklist beyond “whatever the focused three were not looking for”. It was the most productive of the four.

It also disclosed that both source files changed under it mid-review. They did — the author was applying the self-review fixes at the time — and the disclosure was right to make.

The generic reviewer was worth more than the three focused ones combined, and by a wide margin: it produced the page's only substantive content gaps, the cross-page contradiction, and four self-contradictions on this page that a focused reviewer had no reason to look for. If a future run has budget for one reviewer, make it this one.

The three focused reviewers between them produced two defects that changed published figures (the \blinear\b fold bug and the bibliography slug title), two that changed a published claim (the negative-binomial oracle and the truncated table), five reference-currency additions, and one wrong finding (item 1 of the citations review, and item 5's overreach).

The three most consequential errors on the page — the site-random-effect claim, the zero-inflation claim, and the fold-versus-probe disagreement about clustering crawl papers — were not found by any reviewer. All three were false negatives of the same kind: a claim that nothing in the corpus does X, resting on a probe that could not have found it if it did. That is the failure mode to design the next review prompt around, and it is the argument for not treating the review layer as the last line.

• Regression — the page these notes are behind.
• Corpus — corpus-level provenance: the venue scope, the selection funnel, the provisional years.
• hypothesis_testing — the sibling page's log; the population definitions are shared and cross-checked against it.
• Bibliography — the single shared bibliography.

[1]

Zeng, Eric; McAmis, Rachel; Kohno, Tadayoshi; Roesner, Franziska (2022): "What Factors Affect Targeting and Bids in Online Advertising? A Field Measurement Study", in: Proceedings of the ACM Internet Measurement Conference, pp. 210-229. (DOI)

[2]

Chuai, Yuwei; Lenzini, Gabriele; Pröllochs, Nicolas (2026): "Consensus Stability of Community Notes on X", in: Proceedings of the ACM Web Conference, pp. 8885-8896. (DOI)

[3]

Bobek, Michelle; Pröllochs, Nicolas (2026): "Community Fact-Checks Do Not Break Follower Loyalty", in: Proceedings of the ACM Web Conference. (DOI)

[4]

Nenadić, Luka; Rodriguez, David; Calandrino, Joseph A. (2026): "Overcoming Language Barriers: Multilingual Analysis of the 2023 Swiss Privacy Law's Impact", Proceedings on Privacy Enhancing Technologies 2026(4):703-723. (DOI)

[5]

Mai, Cat; Coelho, Bruno; Kieserman, Julia; Matsumoto, Lexie; Spinelli, Kyle; Yang, Eric; Andreou, Athanasios; Greenstadt, Rachel; Lauinger, Tobias; McCoy, Damon (2025): "More and Scammier Ads: The Perils of YouTube's Ad Privacy Settings", in: Proceedings on Privacy Enhancing Technologies. (DOI)

[6]

Kieserman, Julia B.; Andreou, Athanasios; Geeng, Chris; Lauinger, Tobias; McCoy, Damon (2025): "Tracker Installations Are Not Created Equal: Understanding Tracker Configuration of Form Data Collection", in: Proceedings on Privacy Enhancing Technologies, pp. 679-695. (DOI)

[7]

Yuan, Ying; Hao, Qingying; Apruzzese, Giovanni; Conti, Mauro; Wang, Gang (2024): ""Are Adversarial Phishing Webpages a Threat in Reality?" Understanding the Users' Perception of Adversarial Webpages", in: Proceedings of the ACM Web Conference, pp. 1712-1723. (DOI)

[8]

Collier, Ben; Thomas, Daniel R.; Clayton, Richard; Hutchings, Alice (2019): "Booting the Booters: Evaluating the Effects of Police Interventions in the Market for Denial-of-Service Attacks", in: Proceedings of the ACM Internet Measurement Conference, pp. 50-64. (DOI)

[9]

Votipka, Daniel; Fulton, Kelsey R.; Parker, James; Hou, Matthew; Mazurek, Michelle L.; Hicks, Michael (2020): "Understanding security mistakes developers make: Qualitative analysis from Build It, Break It, Fix It", in: Proceedings of the USENIX Security Symposium. (Link)

[10]

Vu, Anh V.; Collier, Ben; Thomas, Daniel R.; Kristoff, John; Clayton, Richard; Hutchings, Alice (2025): "Assessing the Aftermath: the Effects of a Global Takedown against DDoS-for-hire Services", in: Proceedings of the USENIX Security Symposium. (Link)

[11]

Abdullah, Muhammad; Qazi, Zafar Ayyub; Qazi, Ihsan Ayyub (2022): "Causal impact of Android go on mobile web performance", in: Proceedings of the ACM Internet Measurement Conference, pp. 113-129. (DOI)

[12]

Butkiewicz, Michael; Madhyastha, Harsha V.; Sekar, Vyas (2011): "Understanding website complexity: measurements, metrics, and implications", in: Proceedings of the ACM Internet Measurement Conference. (DOI)

[13]

Gelman, Andrew; Hill, Jennifer (2007): "Data Analysis Using Regression and Multilevel/Hierarchical Models". Cambridge University Press. (DOI)

[14]

Bates, Douglas; Mächler, Martin; Bolker, Ben; Walker, Steve (2015): "Fitting Linear Mixed-Effects Models Using lme4", Journal of Statistical Software 67(1):1-48. (DOI)

[15]

Barr, Dale J.; Levy, Roger; Scheepers, Christoph; Tily, Harry J. (2013): "Random Effects Structure for Confirmatory Hypothesis Testing: Keep It Maximal", Journal of Memory and Language 68(3):255-278. (DOI)

[16]

Mood, Carina (2010): "Logistic Regression: Why We Cannot Do What We Think We Can Do, and What We Can Do About It", European Sociological Review 26(1):67-82. (DOI)

[17]

Ver Hoef, Jay M.; Boveng, Peter L. (2007): "Quasi-Poisson vs. Negative Binomial Regression: How Should We Model Overdispersed Count Data?", Ecology 88(11):2766-2772. (DOI)

[18]

Ferrari, Silvia; Cribari-Neto, Francisco (2004): "Beta Regression for Modelling Rates and Proportions", Journal of Applied Statistics 31(7):799-815. (DOI)

[19]

Callaway, Brantly; Sant'Anna, Pedro H. C. (2021): "Difference-in-Differences with Multiple Time Periods", Journal of Econometrics 225(2):200-230. (DOI)

[20]

Moineddin, Rahim; Matheson, Flora I.; Glazier, Richard H. (2007): "A Simulation Study of Sample Size for Multilevel Logistic Regression Models", BMC Medical Research Methodology 7:34. (DOI)

[21]

Shmueli, Galit (2010): "To Explain or to Predict?", Statistical Science 25(3):289-310. (DOI)

[22]

Vittinghoff, Eric; McCulloch, Charles E. (2007): "Relaxing the Rule of Ten Events per Variable in Logistic and Cox Regression", American Journal of Epidemiology 165(6):710-718. (DOI)

[23]

Angrist, Joshua D.; Pischke, Jörn-Steffen (2009): "Mostly Harmless Econometrics: An Empiricist's Companion". Princeton University Press. (DOI)

[24]

Colin Cameron, A.; Miller, Douglas L. (2015): "A Practitioner’s Guide to Cluster-Robust Inference", Journal of Human Resources 50(2):317-372. (DOI)

[25]

Gelman, Andrew; Hill, Jennifer; Vehtari, Aki (2020): "Regression and Other Stories". Cambridge University Press. (DOI)

[26]

Matuschek, Hannes; Kliegl, Reinhold; Vasishth, Shravan; Baayen, Harald; Bates, Douglas (2017): "Balancing Type I Error and Power in Linear Mixed Models", Journal of Memory and Language 94:305-315. (DOI)

[27]

MacKinnon, James G.; Nielsen, Morten Ørregaard; Webb, Matthew D. (2023): "Cluster-Robust Inference: A Guide to Empirical Practice", Journal of Econometrics 232(2):272-299. (DOI)

[28]

Huang, Francis L. (2026): "When Cluster-Robust Inferences Fail", Educational and Psychological Measurement 86(3):579-601. Published online 2025-12-19; issue dated 2026-06 (DOI)

[31]

Becerril-Arreola, Rafael (2023): "A Method to Assess and Explain Disparate Impact in Online Retailing", in: Proceedings of the ACM Web Conference. (DOI)