{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":101,"total_is_capped":false,"direct_labels_cover":1,"predictions_cover":101,"direct_label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline (scores rank; they never assert a category)","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12","author_layer_release":"2026-06-26"},"query_hash":"4cda1ec5ea24","filters":{"venue":"Educational and Psychological Measurement"}},"results":[{"id":"W2021930444","doi":"10.1177/0013164409355698","title":"The Motivation at Work Scale: Validation Evidence in Two Languages","year":2010,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Motivation and Self-Concept in Sports","field":"Psychology","cited_by":749,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"HEC Montréal; Université du Québec à Montréal; Université de Montréal; Concordia University","funders":"","keywords":"Conceptualization; Scale (ratio); Psychology; Work (physics); Intrinsic motivation; Social psychology; Work motivation; Cognitive psychology; Computer science; Artificial intelligence; Geography; Physics","authors":[{"name":"Marylène Gagné","is_ca":true},{"name":"Jacques Forest","is_ca":true},{"name":"Marie‐Hélène Gilbert","is_ca":true},{"name":"Caroline Aubé","is_ca":true},{"name":"Estelle M. Morin","is_ca":true},{"name":"Angela Malorni","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1709997786686356,"gpt":0.4045055228999631,"spread":0.2335057442313275,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001421707,0.0001182087,0.000095246,0.00005032304,0.0001801084,0.00004700266,0.0001536707,0.00008202773,0.003370127],"category_scores_gemma":[0.0005723306,0.00007785864,0.00003147064,0.000240695,0.000134708,0.00007232493,0.00002233975,0.0002925535,0.0001583445],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005981629,"about_ca_system_score_gemma":0.00001867162,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003856379,"about_ca_topic_score_gemma":0.0001794535,"domain_scores_codex":[0.9985622,0.0001779891,0.0003025391,0.0003519855,0.0004016659,0.0002035909],"domain_scores_gemma":[0.9988819,0.0005223242,0.0001106983,0.0002639838,0.0001373845,0.00008373949],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0001014789,0.000375931,0.9524471,0.000002957899,0.000009074799,8.249946e-7,0.001316202,0.000002617651,0.001020149,0.009895257,0.01311331,0.0217151],"study_design_scores_gemma":[0.0002906848,0.00003611975,0.9847191,0.00003099235,0.000004061937,0.000009517995,0.0003363633,4.090884e-7,0.00004820761,0.001966503,0.01245823,0.00009980467],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9680256,0.0005931581,0.00003296885,0.01466586,0.001725342,0.0002884639,0.000001175324,0.00002423954,0.01464323],"genre_scores_gemma":[0.9956124,0.00004675386,0.0002249336,0.0006717569,0.0003799003,0.0001849064,0.00001260804,0.000006458795,0.002860263],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.03227201,"threshold_uncertainty_score":0.997541,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2062055237","doi":"10.1177/00131640021970871","title":"Type I Error Rate Comparisons of Post Hoc Procedures for I j Chi-Square Tables","year":2000,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":444,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Western University","funders":"","keywords":"Post hoc; Post-hoc analysis; Statistics; Pairwise comparison; Contingency table; Type I and type II errors; Homogeneity (statistics); Mathematics; Chi-square test; Null hypothesis; Monte Carlo method; Econometrics; Medicine; Dentistry","authors":[{"name":"Paul L. MacDonald","is_ca":false},{"name":"R. C. Gardner","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4095890100599183,"gpt":0.4883884670827446,"spread":0.07879945702282637,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0004558386,0.0001168146,0.000213052,0.00002178644,0.00009643747,0.00001044712,0.00008074098,0.0000518382,0.001098759],"category_scores_gemma":[0.001068668,0.00008209762,0.00004213558,0.00007171183,0.00008219629,0.00003058109,0.000006840056,0.00006826108,0.000006133817],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001678629,"about_ca_system_score_gemma":0.00005164885,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000002577913,"about_ca_topic_score_gemma":0.000008488671,"domain_scores_codex":[0.9990493,0.00007480544,0.0002718248,0.0002400084,0.0002012064,0.0001628647],"domain_scores_gemma":[0.9989524,0.0004282249,0.0000736897,0.0001174989,0.0003361742,0.00009194329],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.002024167,0.005033486,0.0007187171,0.000691299,0.0001326049,5.621497e-7,0.0006961991,0.00006690654,0.005232817,0.8630502,0.05543885,0.06691419],"study_design_scores_gemma":[0.0007871408,0.001232692,0.03205615,0.0001354691,0.0000546174,0.000006908415,0.0001555657,0.0001189724,0.0002131576,0.9480793,0.01692241,0.000237569],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7345521,0.002814109,0.2354472,0.0130172,0.0008285627,0.002399211,0.0005183006,0.00008081256,0.01034249],"genre_scores_gemma":[0.6193956,0.00005888699,0.3792747,0.0003376747,0.0001231278,0.0001200166,0.00002152025,0.00001023536,0.0006583086],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1438275,"threshold_uncertainty_score":0.9998144,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2084244869","doi":"10.1177/0013164403258450","title":"Measurement and Validity Characteristics of the Short Version of the Social and Emotional Loneliness Scale for Adults","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Health disparities and outcomes","field":"Social Sciences","cited_by":373,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of New Brunswick","funders":"","keywords":"Loneliness; Psychology; Discriminant validity; Convergent validity; UCLA Loneliness Scale; Construct validity; Concurrent validity; Scale (ratio); Clinical psychology; Test validity; Developmental psychology; Psychometrics; Social psychology; Internal consistency","authors":[{"name":"Enrico DiTommaso","is_ca":true},{"name":"Cyndi Brannen","is_ca":true},{"name":"Lisa A. Best","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2283013155929321,"gpt":0.4013158889867471,"spread":0.1730145733938151,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001127216,0.00006006375,0.0001121132,0.000009689825,0.000536169,0.0000117517,0.00009130171,0.0000575628,0.00002084429],"category_scores_gemma":[0.0003479934,0.00003454684,0.00004513334,0.0000602014,0.0003202315,0.00002671269,0.00002590089,0.0000588742,1.57348e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00007184214,"about_ca_system_score_gemma":0.0001630823,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002025778,"about_ca_topic_score_gemma":0.0002531191,"domain_scores_codex":[0.9988384,0.0001013923,0.0001989185,0.0001334105,0.0006080052,0.0001198534],"domain_scores_gemma":[0.9993839,0.00005462636,0.00007979103,0.00005704526,0.000365195,0.00005938808],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002753368,0.001332902,0.9365987,0.0005047819,0.00003272385,4.00057e-8,0.008601808,0.000001525247,0.0003453491,0.0383099,0.002365712,0.01163126],"study_design_scores_gemma":[0.0003073627,0.00003817525,0.9929418,0.0001317362,0.00001282419,4.766742e-7,0.000719969,6.01243e-7,0.00004211223,0.004337227,0.001424784,0.00004291035],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9597067,0.0001560806,0.00005372716,0.03886203,0.0006363036,0.0003429757,0.00003528514,0.000002303697,0.0002045826],"genre_scores_gemma":[0.998975,0.000080874,0.00009993433,0.0005410676,0.0002638093,0.0000201415,0.000001543254,0.000001886249,0.00001575414],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.05634316,"threshold_uncertainty_score":0.4123832,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2027051144","doi":"10.1177/00131640021970655","title":"Development and Validation of Scores on a Two-Dimensional Workplace Friendship Scale","year":2000,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Job Satisfaction and Organizational Behavior","field":"Business, Management and Accounting","cited_by":343,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Quest University Canada","funders":"","keywords":"Friendship; Nomological network; Psychology; Scale (ratio); Discriminant validity; Construct validity; Sample (material); Convergent validity; Predictive validity; Construct (python library); Social psychology; Test validity; Likert scale; External validity; Internal consistency; Replicate; Applied psychology; Psychometrics; Developmental psychology; Statistics; Mathematics; Computer science","authors":[{"name":"Ivy K. Nielsen","is_ca":true},{"name":"Steve M. Jex","is_ca":false},{"name":"Gary A. Adams","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07775349235747162,"gpt":0.2954924167858354,"spread":0.2177389244283638,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0002914167,0.0001102257,0.0001011741,0.00008261441,0.0001492014,0.00005101334,0.00005117489,0.00003780336,0.003581137],"category_scores_gemma":[0.00004231292,0.00008767795,0.00001773602,0.0001680071,0.00005491946,0.000142494,0.00001489129,0.00006830239,0.0001584974],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002487656,"about_ca_system_score_gemma":0.00002157058,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001373416,"about_ca_topic_score_gemma":0.000008552095,"domain_scores_codex":[0.9989784,0.00001324687,0.0002179362,0.0002437161,0.0004399575,0.0001067481],"domain_scores_gemma":[0.9996502,0.0000324154,0.0000653341,0.00007081317,0.0001559621,0.00002530246],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0004474447,0.002044342,0.846251,0.00006397888,0.00004085716,0.000001118111,0.0002814003,0.0002337617,0.003292841,0.0509956,0.005127225,0.0912205],"study_design_scores_gemma":[0.0005042996,0.0000219163,0.9862564,0.00004997534,0.0000135885,0.000003659892,0.00003095451,0.00000749371,0.0005796673,0.00291138,0.009492592,0.0001280629],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9904323,0.00009543337,0.000007224975,0.00238137,0.0001741527,0.0001336795,0.000001196271,0.00001900569,0.006755692],"genre_scores_gemma":[0.9971359,0.000003503564,0.001122502,0.001233583,0.0002440958,0.00002322829,0.00002760242,0.000006565314,0.0002030714],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1400055,"threshold_uncertainty_score":0.9973297,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2036831393","doi":"10.1177/0013164411409929","title":"Difference Scores From the Point of View of Reliability and Repeated-Measures ANOVA","year":2011,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":167,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia; Carleton University","funders":"","keywords":"Reliability (semiconductor); Repeated measures design; Significant difference; Statistics; Psychology; Analysis of variance; Point (geometry); Mean difference; Mathematics; Econometrics; Confidence interval","authors":[{"name":"David R. Thomas","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7554307011056853,"gpt":0.4653795213841227,"spread":0.2900511797215626,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007073482,0.0001113388,0.0002886284,0.000064427,0.00008802751,0.00002287375,0.0004211254,0.00005722675,0.000528036],"category_scores_gemma":[0.0388429,0.00005250085,0.00005942115,0.0004552616,0.0004226591,0.00004238106,0.00008314939,0.0001177334,0.000002807144],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000009741277,"about_ca_system_score_gemma":0.00003205881,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002503211,"about_ca_topic_score_gemma":0.00001019118,"domain_scores_codex":[0.9971045,0.0006320529,0.0006785488,0.0004558421,0.001002931,0.0001261573],"domain_scores_gemma":[0.991173,0.007289697,0.0003483038,0.0004926272,0.0006169453,0.00007945801],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0001130944,0.0004731831,0.8170224,0.000008307061,0.00002119778,1.818328e-7,0.0007817693,5.04177e-7,0.00194558,0.002846088,0.001250037,0.1755377],"study_design_scores_gemma":[0.0001220334,0.0001576417,0.7876037,0.00003741558,0.000008047821,0.000002672828,0.0002196643,0.000002928116,0.0002745066,0.2112522,0.0002680671,0.00005118028],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9886076,0.003997426,0.001219768,0.002071434,0.0004015245,0.000173817,0.0000154684,0.000007271472,0.003505657],"genre_scores_gemma":[0.9906312,0.0002093273,0.008859892,0.0002046245,0.00005185754,0.00001313172,8.029346e-7,0.000002249935,0.00002695076],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2084061,"threshold_uncertainty_score":0.9692534,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1763331641","doi":"10.1177/0013164415584576","title":"It Might Not Make a Big DIF","year":2015,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":130,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"York University","funders":"","keywords":"Differential item functioning; Item response theory; Statistics; Polytomous Rasch model; Computer science; Monte Carlo method; Econometrics; Type I and type II errors; Statistical hypothesis testing; Machine learning; Data mining; Artificial intelligence; Psychometrics; Mathematics","authors":[{"name":"R. Philip Chalmers","is_ca":true},{"name":"Alyssa Counsell","is_ca":true},{"name":"David B. Flora","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9118226372485674,"gpt":0.5444501445426981,"spread":0.3673724927058692,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0101387,0.0001489979,0.0002405611,0.0002087265,0.0001397812,0.0001667443,0.000480879,0.00008620465,0.0005668987],"category_scores_gemma":[0.05947007,0.00008797546,0.00007200234,0.0008806389,0.0001222944,0.0000404057,0.00007949272,0.0001640133,0.0002546447],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005420615,"about_ca_system_score_gemma":0.00008421837,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001346569,"about_ca_topic_score_gemma":0.00001142225,"domain_scores_codex":[0.9958932,0.000466154,0.000541566,0.0006236774,0.00219505,0.0002803024],"domain_scores_gemma":[0.9942334,0.00408691,0.0001782622,0.000395327,0.0007164464,0.0003896648],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0001459428,0.000893685,0.06984503,0.000003635477,0.00002580343,0.000003526481,0.0005898251,0.000008374251,0.0004786453,0.006006123,0.1942538,0.7277456],"study_design_scores_gemma":[0.0004484912,0.0002373363,0.6071104,0.00001168114,0.000005832416,0.00003237068,0.0003834005,0.00001569859,0.0000282667,0.1191553,0.2724169,0.0001543805],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5929533,0.002606433,0.00848812,0.1696946,0.009603234,0.0004496373,0.00001708687,0.00008240417,0.2161052],"genre_scores_gemma":[0.9806212,0.00002262595,0.01148,0.004778861,0.0007407291,0.00004151489,0.000001946491,0.000005069896,0.002308026],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7275912,"threshold_uncertainty_score":0.9484524,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2012610170","doi":"10.1177/0013164403260196","title":"The Analysis of Repeated Measurements with Mixed-Model Adjusted F Tests","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":130,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Manitoba","funders":"","keywords":"Type I and type II errors; Statistics; Repeated measures design; Sample size determination; Mixed model; Mathematics; Covariance; Univariate; Statistic; Homogeneity (statistics); Normality; Analysis of covariance; Heteroscedasticity; Statistical hypothesis testing; Econometrics; Multivariate statistics","authors":[{"name":"Rhonda K. Kowalchuk","is_ca":false},{"name":"H. J. Keselman","is_ca":true},{"name":"James Algina","is_ca":false},{"name":"Russell D. Wolfinger","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7953401039095214,"gpt":0.5668979603801602,"spread":0.2284421435293612,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00404919,0.0001749788,0.000440142,0.00007998074,0.0001703596,0.00002774597,0.0002301698,0.00009344631,0.00009374927],"category_scores_gemma":[0.01997148,0.0000918418,0.0001195846,0.0006472842,0.0003028479,0.00002555995,0.00002783328,0.0001711348,0.000004125243],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008847867,"about_ca_system_score_gemma":0.00009708048,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000009111795,"about_ca_topic_score_gemma":0.00005498344,"domain_scores_codex":[0.9971228,0.0003703196,0.0007395949,0.0004024204,0.001134545,0.0002303382],"domain_scores_gemma":[0.9937341,0.004680566,0.0003129091,0.0004221959,0.0006974732,0.000152818],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0029213,0.0147033,0.0481512,0.0002088747,0.01053824,0.000005610722,0.0008397246,0.002939793,0.007499366,0.8450153,0.007075713,0.06010152],"study_design_scores_gemma":[0.0009784771,0.0002980796,0.2807045,0.00006532309,0.0008793459,0.00000261399,0.00003933883,0.00006172567,0.0003004947,0.7164808,0.0000417571,0.0001475836],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8722825,0.000599512,0.1022478,0.01464621,0.0008878165,0.001465907,0.00008041185,0.0000961414,0.007693723],"genre_scores_gemma":[0.7846913,0.00002335651,0.2149148,0.0001581434,0.00005776375,0.0000819658,0.000002541168,0.000008914207,0.00006116806],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2325533,"threshold_uncertainty_score":0.9882837,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2004549974","doi":"10.1177/0013164403258402","title":"The Effect of Multicollinearity on Multilevel Modeling Parameter Estimates and Standard Errors","year":2003,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":127,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Simon Fraser University","funders":"","keywords":"Multicollinearity; Statistics; Mathematics; Multilevel model; Standard error; Covariance; Variance inflation factor; Variance (accounting); Econometrics; Correlation; Standard deviation; Linear regression","authors":[{"name":"Yann‐Yann Shieh","is_ca":false},{"name":"Rachel T. Fouladi","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3510922864786483,"gpt":0.4841609772101677,"spread":0.1330686907315194,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001351201,0.0001230735,0.0001911157,0.00001500154,0.0001656564,0.00001701284,0.00004773937,0.00004706856,0.0000211739],"category_scores_gemma":[0.009470133,0.00006421127,0.00003112334,0.00002969963,0.0001175486,0.00001848366,0.000009534802,0.0001228327,7.151593e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001915329,"about_ca_system_score_gemma":0.00001131387,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000001752265,"about_ca_topic_score_gemma":0.000001032009,"domain_scores_codex":[0.9988782,0.0002190077,0.0002330943,0.0002204474,0.0003055024,0.0001436853],"domain_scores_gemma":[0.9951861,0.004438234,0.00005968503,0.0001257482,0.0001067979,0.00008345979],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.001325989,0.001040563,0.003717189,0.0002590549,0.0001361973,0.000001297393,0.0004293642,0.0009919452,0.001291457,0.8400839,0.0004554935,0.1502676],"study_design_scores_gemma":[0.0008147741,0.001021409,0.003179143,0.00006123354,0.00004492764,0.000004790433,0.00003062017,0.01302842,0.0009046028,0.9805874,0.0001891746,0.0001334572],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5887622,0.0004608,0.4088562,0.0006720226,0.0002082711,0.0004461893,0.00001800018,0.0000119041,0.0005643793],"genre_scores_gemma":[0.7759584,0.00002116376,0.223894,0.00003747786,0.0000154165,0.00005229627,4.971135e-7,0.000005005255,0.00001578531],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1871962,"threshold_uncertainty_score":0.9988735,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2130969147","doi":"10.1177/0013164412452564","title":"The Relationship Between Root Mean Square Error of Approximation and Model Misspecification in Confirmatory Factor Analysis Models","year":2012,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":123,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Structural equation modeling; Confirmatory factor analysis; Statistics; Mathematics; Population; Cutoff; Factor analysis; Econometrics; Physics","authors":[{"name":"Victoria Savalei","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8678825713483123,"gpt":0.5198963884873063,"spread":0.3479861828610059,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007577958,0.0001086487,0.0002377927,0.0003448534,0.000199524,0.00006301185,0.0002292653,0.0000839163,0.00004083526],"category_scores_gemma":[0.01237711,0.00006407404,0.00006358908,0.001205762,0.0001245797,0.0002106039,0.00002999435,0.0001442681,0.000003308009],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004396122,"about_ca_system_score_gemma":0.00002529704,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000009985933,"about_ca_topic_score_gemma":0.000011624,"domain_scores_codex":[0.9972292,0.0005539751,0.0007073024,0.0003144112,0.00100332,0.0001918155],"domain_scores_gemma":[0.988543,0.01038811,0.000345218,0.000297356,0.0003018779,0.0001244523],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00001703137,0.0001303305,0.964588,0.000004486167,0.00002214566,9.831377e-9,0.0008410828,0.0003794606,0.00004060813,0.02462791,0.0001480322,0.009200902],"study_design_scores_gemma":[0.0001245951,0.00002085454,0.8544398,0.000006883301,0.00002262577,5.772437e-7,0.0004531077,0.003196969,0.000006812667,0.1415917,0.00006805847,0.0000679831],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9509001,0.001742189,0.04297723,0.002387522,0.0001404653,0.0002192518,0.00001285867,0.000008379014,0.001612024],"genre_scores_gemma":[0.9944417,0.00001714269,0.005340768,0.00003067745,0.00007110826,0.00003520222,0.000006359273,0.000003107759,0.00005390581],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1169638,"threshold_uncertainty_score":0.9959421,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2064379877","doi":"10.1177/0013164402238082","title":"A Monte Carlo Comparison of Item and Person Statistics Based on Item Response Theory versus Classical Test Theory","year":2002,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":121,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Western University","funders":"","keywords":"Item response theory; Classical test theory; Monte Carlo method; Statistics; Psychology; Econometrics; Psychometrics; Test theory; Item analysis; Computerized adaptive testing; Test (biology); Differential item functioning; Statistical hypothesis testing; Mathematics","authors":[{"name":"P.B. MacDonald","is_ca":false},{"name":"Sampo V. Paunonen","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7790233134897627,"gpt":0.505780321916664,"spread":0.2732429915730987,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01027207,0.0001819323,0.0003530799,0.0002524895,0.0001913847,0.00008219505,0.0002878135,0.0001049792,0.001011047],"category_scores_gemma":[0.172042,0.0001161324,0.00006305706,0.000536033,0.0003283359,0.00004402149,0.00002999475,0.0002453282,0.00002152588],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005421782,"about_ca_system_score_gemma":0.00003108215,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000002247613,"about_ca_topic_score_gemma":0.000001370805,"domain_scores_codex":[0.9951723,0.002010999,0.0005431725,0.0005771259,0.001464459,0.000231893],"domain_scores_gemma":[0.8836055,0.1152891,0.0002623309,0.0003207914,0.0003424304,0.0001797696],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.01164811,0.006034489,0.2767802,0.00003284093,0.00008277433,0.000007231965,0.002875438,0.0001539937,0.001107383,0.02980219,0.06665417,0.6048211],"study_design_scores_gemma":[0.001347528,0.002614317,0.9663729,0.00004826986,0.00002579324,0.000005199957,0.002115086,0.0057995,0.00002630321,0.01839773,0.003043179,0.0002042159],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9747216,0.002340278,0.00535482,0.007777893,0.001234807,0.0003116508,0.00009027667,0.00002997171,0.008138726],"genre_scores_gemma":[0.9894863,0.00001688989,0.009530709,0.0003684703,0.0001058357,0.00002270557,0.000001019333,0.000006495478,0.0004615909],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6895927,"threshold_uncertainty_score":0.9999022,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2065030780","doi":"10.1177/0013164404273942","title":"Understanding Parameter Invariance in Unidimensional IRT Models","year":2006,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":108,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia; University of Ottawa","funders":"","keywords":"Item response theory; Measurement invariance; Set (abstract data type); Differential item functioning; Psychology; Feature (linguistics); Property (philosophy); Econometrics; Variable (mathematics); Psychometrics; Cognitive psychology; Mathematics; Statistics; Applied mathematics; Computer science; Structural equation modeling; Confirmatory factor analysis; Mathematical analysis; Epistemology","authors":[{"name":"André Rupp","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9212118758900417,"gpt":0.4959208972683364,"spread":0.4252909786217053,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005931179,0.000134477,0.0002187688,0.0002926258,0.0001380047,0.0001140553,0.0002774851,0.00008454702,0.0004513374],"category_scores_gemma":[0.008997417,0.00008952127,0.00005106472,0.0009926866,0.0001119965,0.0001518282,0.00005021436,0.0001773529,0.00002708581],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001409192,"about_ca_system_score_gemma":0.00004424219,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004808021,"about_ca_topic_score_gemma":0.0000213187,"domain_scores_codex":[0.9968109,0.0003921403,0.0005797533,0.0005941061,0.001356022,0.000267134],"domain_scores_gemma":[0.9924677,0.006867148,0.000148665,0.0002466373,0.0001790899,0.00009078057],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001471939,0.00148447,0.252582,0.000006598603,0.00001662944,0.000007730636,0.0001493387,0.003056847,0.0009778753,0.6835403,0.04488128,0.01314969],"study_design_scores_gemma":[0.000214267,0.00004435018,0.3816177,0.00001322214,0.000001420456,0.000009310204,0.00007674439,0.0005489948,0.000004917283,0.6166646,0.000720934,0.00008355491],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.799769,0.001693163,0.09876592,0.01985271,0.001874825,0.0003820185,0.000007678052,0.00004056797,0.07761406],"genre_scores_gemma":[0.9801323,0.00001552584,0.01876162,0.0006260694,0.0001593335,0.00002558546,0.00000231156,0.000004115774,0.000273141],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1803632,"threshold_uncertainty_score":0.9993502,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2021178079","doi":"10.1177/0013164403261762","title":"The Nature and Measurement of Emotional Intelligence Abilities: Basic Dimensions and Their Relationships with Other Cognitive Ability and Personality Variables","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Emotional Intelligence and Performance","field":"Psychology","cited_by":106,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Psychology; Personality; Alexithymia; Cognition; Emotional intelligence; Developmental psychology; Big Five personality traits; Construct (python library); Congruence (geometry); Social psychology; Cognitive psychology","authors":[{"name":"Kimberly A. Barchard","is_ca":false},{"name":"A. Ralph Hakstian","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1714507792378559,"gpt":0.3521081209423749,"spread":0.180657341704519,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002901918,0.000192832,0.0001706207,0.00003936498,0.0005140873,0.00003562454,0.0000794648,0.0001436183,0.0002190113],"category_scores_gemma":[0.0005681029,0.0001091753,0.00002971789,0.0001276698,0.0009788392,0.00008266404,0.00002594439,0.0004106634,0.000003793261],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005620349,"about_ca_system_score_gemma":0.00008276197,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007413478,"about_ca_topic_score_gemma":0.0001295755,"domain_scores_codex":[0.9982433,0.0003360191,0.0003109507,0.0004686227,0.0004454566,0.0001956215],"domain_scores_gemma":[0.9979781,0.001034631,0.0001118706,0.0001649116,0.0005830452,0.0001274657],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","study_design_scores_codex":[0.0009796688,0.002474576,0.408753,0.0001752127,0.0004658292,6.830592e-7,0.0119417,0.00003043852,0.0005768703,0.5532584,0.0003158766,0.02102774],"study_design_scores_gemma":[0.0002673642,0.0004022723,0.8617002,0.0001981316,0.00002358506,0.00004518825,0.004202981,0.000007736011,0.00006823811,0.1324509,0.000488391,0.0001450413],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9783427,0.008677285,0.002299003,0.007816241,0.0002067232,0.0004390445,0.00004820019,0.00001201104,0.002158822],"genre_scores_gemma":[0.9985446,0.0002818816,0.0004315019,0.0004901984,0.00009457038,0.00009022924,0.000005876633,0.000006906545,0.00005423296],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4529472,"threshold_uncertainty_score":0.445204,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2004501768","doi":"10.1177/0013164402239317","title":"Differential Item Functioning Results May Change Depending On How An Item Is Scored: An Illustration With The Center For Epidemiologic Studies Depression Scale","year":2003,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":97,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Differential item functioning; Psychology; Item response theory; Logistic regression; Statistics; Psychometrics; Ordered logit; Scale (ratio); Clinical psychology; Center for Epidemiologic Studies Depression Scale; Persistence (discontinuity); Depressive symptoms; Mathematics; Cognition; Psychiatry","authors":[{"name":"Michaela N. Gelin","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8833118557082188,"gpt":0.5341469955629262,"spread":0.3491648601452926,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008030978,0.0002288073,0.0003077641,0.0001629654,0.000825972,0.000238538,0.0003107138,0.000107634,0.0001045223],"category_scores_gemma":[0.02549889,0.0001035856,0.00007400031,0.0004474425,0.0001470396,0.0002792774,0.00002766658,0.0001995674,0.000004381514],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005999839,"about_ca_system_score_gemma":0.00001894698,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000005045115,"about_ca_topic_score_gemma":0.00002111606,"domain_scores_codex":[0.9958541,0.001344573,0.0004612504,0.0009558714,0.001051645,0.0003324981],"domain_scores_gemma":[0.9886667,0.009783723,0.0004050175,0.0004350494,0.0005479865,0.000161582],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.002155089,0.002204346,0.6345406,0.00002769171,0.0001319167,0.000001456404,0.003105217,0.00009719392,0.001259532,0.005647318,0.02736664,0.323463],"study_design_scores_gemma":[0.001065587,0.00171708,0.9748963,0.00008042205,0.00002373629,0.00002320208,0.003854853,0.0002288187,0.000163909,0.01097805,0.006736757,0.0002312727],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9712744,0.0005340384,0.01011754,0.01473618,0.001604813,0.0007173943,0.00003037283,0.00003429111,0.0009510112],"genre_scores_gemma":[0.9908785,0.00004851168,0.006738887,0.001222352,0.0006059188,0.0002733947,0.00002005527,0.000008017185,0.0002043438],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3403558,"threshold_uncertainty_score":0.9827098,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2033678552","doi":"10.1177/0013164409344548","title":"The Impact of Outliers on Cronbach’s Coefficient Alpha Estimate of Reliability: Ordinal/Rating Scale Item Responses","year":2009,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":96,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Cronbach's alpha; Outlier; Categorical variable; Ordinal data; Statistics; Likert scale; Mathematics; Econometrics; Reliability (semiconductor); Sample (material); Rating scale; Psychology; Psychometrics","authors":[{"name":"Yan Liu","is_ca":true},{"name":"Amery D. Wu","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2803599964946322,"gpt":0.5147372816258765,"spread":0.2343772851312443,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001861384,0.0001356447,0.0002386892,0.00003129429,0.0001520009,0.0000141897,0.0001181076,0.00004766347,0.00004545667],"category_scores_gemma":[0.005239347,0.00007257509,0.00009197636,0.0001034151,0.0001931917,0.00002276175,0.00001290661,0.0001253084,0.000001294407],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00007259304,"about_ca_system_score_gemma":0.00005520847,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000005282257,"about_ca_topic_score_gemma":8.142163e-7,"domain_scores_codex":[0.9984006,0.0002535229,0.0004457395,0.0002517816,0.0004595848,0.0001887253],"domain_scores_gemma":[0.9967936,0.002356022,0.0001993764,0.0002483189,0.0003024874,0.0001001532],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.005410805,0.01360208,0.008583158,0.0002133838,0.0001743443,0.000002326962,0.002058135,0.002530069,0.02141706,0.586468,0.01035732,0.3491833],"study_design_scores_gemma":[0.0004055288,0.002408241,0.2346339,0.0001219988,0.000026809,0.000005432453,0.0001252846,0.0006029804,0.0005968607,0.7608211,0.0001130606,0.0001388108],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9291537,0.0002474397,0.06422069,0.003109749,0.0001536955,0.0003985462,0.00003493697,0.00001553999,0.002665735],"genre_scores_gemma":[0.9200715,0.00001794376,0.07970042,0.00005768016,0.00003568509,0.00001822031,0.000001221698,0.000004904099,0.00009240677],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3490445,"threshold_uncertainty_score":0.6272366,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1501970797","doi":"10.1177/0013164415593777","title":"Testing Mediation in Structural Equation Modeling","year":2015,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Attachment and Relationship Dynamics","field":"Psychology","cited_by":93,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa; Carleton University","funders":"","keywords":"Bootstrapping (finance); Structural equation modeling; Monte Carlo method; Latent variable; Mediation; Econometrics; Contrast (vision); Statistical hypothesis testing; Statistics; Computer science; Latent variable model; Mathematics; Artificial intelligence","authors":[{"name":"Craig Leth‐Steensen","is_ca":true},{"name":"Elena Gallitto","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4340302585753368,"gpt":0.4648977677929242,"spread":0.03086750921758735,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007914322,0.00009550372,0.00008710481,0.00008402816,0.00006727328,0.00002059227,0.00006608482,0.00009612281,0.0001520782],"category_scores_gemma":[0.0003600032,0.0000765965,0.00001610267,0.0001948364,0.00003001993,0.00007805286,0.00001173272,0.000157786,0.00005765739],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001241239,"about_ca_system_score_gemma":0.00004041168,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000463321,"about_ca_topic_score_gemma":0.00001354858,"domain_scores_codex":[0.9988624,0.0001189237,0.0002596558,0.0002506426,0.0003484621,0.0001599223],"domain_scores_gemma":[0.9994659,0.00009408989,0.00006344831,0.00009361228,0.000179371,0.0001036072],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00008031313,0.0003679759,0.9222157,0.000005542098,0.00001890683,0.000001229066,0.00336591,0.004582977,0.00009138951,0.05688288,0.0004493702,0.0119378],"study_design_scores_gemma":[0.0007680241,0.0001228707,0.8682479,0.00001589211,0.000007687935,0.00001068015,0.0007219403,0.0494392,4.01628e-7,0.08046315,0.00005865296,0.0001435822],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9720743,0.0006000636,0.004081113,0.003000506,0.001307381,0.0002159744,0.000001479741,0.00002633042,0.01869287],"genre_scores_gemma":[0.9970842,0.000001786939,0.002241686,0.0001504579,0.0003012338,0.00005985066,0.00002835191,0.000005400768,0.0001270384],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.05396777,"threshold_uncertainty_score":0.3123515,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2758319537","doi":"10.1177/0013164417719111","title":"Recommendations on the Sample Sizes for Multilevel Latent Class Models","year":2017,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Bayesian Methods and Mixture Models","field":"Computer Science","cited_by":93,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University","funders":"","keywords":"Latent class model; Sample size determination; Statistics; Sample (material); Rule of thumb; Nested set model; Computer science; Model selection; Econometrics; Data mining; Mathematics; Algorithm","authors":[{"name":"Jungkyu Park","is_ca":true},{"name":"Hsiu‐Ting Yu","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4657980968857349,"gpt":0.4110614061339838,"spread":0.05473669075175114,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009958328,0.0001037531,0.00009412987,0.00001959352,0.0008100831,0.0002198267,0.0005848045,0.00004648841,0.00005520288],"category_scores_gemma":[0.0004660402,0.00005898724,0.0000568991,0.00002178265,0.00006113881,0.0001328984,0.00005893615,0.0001007465,0.000006570171],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002751907,"about_ca_system_score_gemma":0.000030499,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000009636351,"about_ca_topic_score_gemma":0.00000373865,"domain_scores_codex":[0.9990324,0.00008753361,0.0001496927,0.0003368294,0.0002325653,0.0001609124],"domain_scores_gemma":[0.9986386,0.0005082146,0.00009706865,0.000518464,0.0001559582,0.0000817043],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000009863841,0.0002523706,0.00003456768,0.000002297084,0.00001163451,4.116742e-8,0.0001206943,0.000006344914,0.0000720485,0.8193682,0.01352469,0.1665972],"study_design_scores_gemma":[0.0002696629,0.0001298434,0.02560972,0.00002363389,0.000005034115,0.000001867537,0.00000506956,0.008538443,0.00006374146,0.951153,0.01408371,0.0001162132],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.0009032149,0.00006229416,0.8258319,0.1686484,0.000681794,0.00029396,0.00001538592,0.00001434169,0.003548719],"genre_scores_gemma":[0.6427089,0.00003468111,0.3535258,0.003058024,0.0001857965,0.000269798,0.000003273268,0.000003671835,0.0002101295],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6418056,"threshold_uncertainty_score":0.6230586,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2036299295","doi":"10.1177/0013164405275668","title":"A Comparison of Four Methods for Detecting Differential Item Functioning in Ordered Response Items","year":2005,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":92,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia; University of Ottawa","funders":"","keywords":"Differential item functioning; Statistics; Psychology; Logistic regression; Item response theory; Type I and type II errors; Linear discriminant analysis; Ordinal regression; Skewness; Ordered logit; Discriminant function analysis; Sample (material); Psychometrics; Mathematics","authors":[{"name":"Elizabeth Kristjansson","is_ca":true},{"name":"Richard Aylesworth","is_ca":true},{"name":"Ian McDowell","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8267139799800711,"gpt":0.6035828191741548,"spread":0.2231311608059163,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01660389,0.000134141,0.0003899913,0.0004020942,0.0001638882,0.00007330627,0.0002766927,0.00009501773,0.000329475],"category_scores_gemma":[0.123021,0.00009296528,0.00009579861,0.0008843424,0.00006671863,0.00007808873,0.00004428082,0.0001724039,0.00000419145],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006617919,"about_ca_system_score_gemma":0.00004288567,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000008133952,"about_ca_topic_score_gemma":0.00001028462,"domain_scores_codex":[0.9961089,0.001427582,0.0009758526,0.0005150642,0.0007266143,0.0002459678],"domain_scores_gemma":[0.9650272,0.03378924,0.0003946032,0.0002350469,0.000459236,0.0000946554],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.001067661,0.0006333358,0.1681289,0.000008109153,0.00001717727,7.877512e-8,0.0005126914,0.0001314781,0.02222869,0.0003845797,0.0007858325,0.8061015],"study_design_scores_gemma":[0.0008067818,0.0004340136,0.9704582,0.00003370296,0.000008449601,0.000006885484,0.0007167471,0.002786707,0.0007629854,0.01714009,0.006710089,0.0001353615],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7998247,0.0007578594,0.1947046,0.003047766,0.0008755152,0.0002670674,0.000002289899,0.00001288854,0.0005073022],"genre_scores_gemma":[0.8085319,0.000003805697,0.1909927,0.00008508631,0.0002025916,0.00007145528,7.896008e-7,0.000004391317,0.0001073229],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8059661,"threshold_uncertainty_score":0.8843662,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1977904348","doi":"10.1177/0013164404264840","title":"Testing the Structure of Psychological Empowerment: Does Gender Make a Difference?","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Community Health and Development","field":"Health Professions","cited_by":83,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Western University; Université de Montréal","funders":"","keywords":"Psychology; Confirmatory factor analysis; Competence (human resources); Social psychology; Construct validity; Structural equation modeling; Developmental psychology; Group analysis; Empowerment; Psychometrics; Statistics; Mathematics","authors":[{"name":"Jean‐Sébastien Boudrias","is_ca":true},{"name":"Patrick Gaudreau","is_ca":true},{"name":"Heather K. Spence Laschinger","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3919868797244075,"gpt":0.4792874805514811,"spread":0.08730060082707353,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007906764,0.000165069,0.0002049448,0.00004471909,0.0008341648,0.000009321045,0.0002569772,0.0001516428,0.0009567236],"category_scores_gemma":[0.0006690194,0.00007733965,0.00003783006,0.0002026048,0.0001622166,0.00001747914,0.00008459518,0.0006567539,0.00002286265],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001313585,"about_ca_system_score_gemma":0.0002265206,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007052096,"about_ca_topic_score_gemma":0.00007389175,"domain_scores_codex":[0.9978554,0.0003924437,0.0005375526,0.000298293,0.0005429169,0.0003733558],"domain_scores_gemma":[0.9984539,0.0005176175,0.0001944385,0.0003152557,0.000328108,0.0001906696],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0005968903,0.004882672,0.7922832,0.0006015426,0.0001170764,0.000003216969,0.02164433,0.00001388855,0.02435311,0.09631622,0.005098606,0.0540893],"study_design_scores_gemma":[0.0006449467,0.0001436791,0.8605259,0.00009731046,0.000007586149,0.000007181342,0.0008939397,3.447592e-7,0.00001720644,0.1358724,0.001692901,0.00009654409],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.974505,0.0003799218,0.00007403993,0.01711805,0.001057738,0.0007502025,0.00001756765,0.00002602865,0.006071475],"genre_scores_gemma":[0.9913296,0.00003989005,0.002503928,0.005689784,0.0002620502,0.00008457841,0.0000103777,0.00000620319,0.00007357757],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.06824276,"threshold_uncertainty_score":0.9999565,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3011484504","doi":"10.1177/0013164420911136","title":"Evaluating the Performances of Missing Data Handling Methods in Ability Estimation From Sparse Data","year":2020,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":79,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Missing data; Imputation (statistics); Statistics; Cart; Data set; Regression; Mean squared error; Mathematics; Computer science; Regression analysis; Data mining; Engineering","authors":[{"name":"Jiaying Xiao","is_ca":true},{"name":"Okan Bulut","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9728668887097844,"gpt":0.6848813766788585,"spread":0.2879855120309259,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03422786,0.0001089755,0.0002737855,0.00006850577,0.0001648603,0.0001177388,0.001761088,0.00004917249,0.0003891205],"category_scores_gemma":[0.1934306,0.00005919431,0.00002389971,0.000880895,0.0001455662,0.0002906318,0.0004866393,0.0002080863,0.000006295177],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001825044,"about_ca_system_score_gemma":0.0000888725,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007103551,"about_ca_topic_score_gemma":0.000005508873,"domain_scores_codex":[0.995159,0.001683264,0.0008591573,0.0008548279,0.001289184,0.0001545813],"domain_scores_gemma":[0.9725879,0.02504437,0.0004881371,0.001478358,0.0002893704,0.0001118864],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00006719087,0.0001416847,0.05570584,0.000007987333,0.00001092363,1.398034e-7,0.0004956385,0.0003927793,0.002193673,0.00007479744,0.0007508764,0.9401585],"study_design_scores_gemma":[0.0002449466,0.000121897,0.7066829,0.00003916008,0.00001452093,0.000001844986,0.0004920312,0.2314029,0.00005521453,0.06028666,0.0005731066,0.0000848261],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7741733,0.003859561,0.1887154,0.03122724,0.0007960083,0.0003196622,0.0000668081,0.00001452045,0.0008275712],"genre_scores_gemma":[0.58206,0.00001880249,0.4174232,0.0003223959,0.0001410507,0.000005126597,0.00002582852,0.000001923946,0.000001632145],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9400737,"threshold_uncertainty_score":0.9944656,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1981571713","doi":"10.1177/0013164403258407","title":"Balancing the Need for Reliability and Time Efficiency: Short Forms of the Wechsler Adult Intelligence Scale-III","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Cognitive Abilities and Testing","field":"Psychology","cited_by":72,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Windsor","funders":"","keywords":"Wechsler Adult Intelligence Scale; Wechsler Preschool and Primary Scale of Intelligence; Psychology; Reliability (semiconductor); Short Forms; Scale (ratio); Wechsler Intelligence Scale for Children; Intelligence quotient; Statistics; Psychometrics; Raw score; Developmental psychology; Clinical psychology; Mathematics; Cognition; Psychiatry; Cartography; Raw data","authors":[{"name":"Sharon L. E. Jeyakumar","is_ca":true},{"name":"Erin M. Warriner","is_ca":true},{"name":"Vaishali V. Raval","is_ca":true},{"name":"Saadia Ahmad","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.08200673823526182,"gpt":0.3386034828471236,"spread":0.2565967446118618,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010669,0.0001140534,0.0001308883,0.00001681408,0.0002319741,0.0000169489,0.0001602945,0.00005855015,0.0002135668],"category_scores_gemma":[0.0006421967,0.00005636213,0.00006780606,0.0001275707,0.000361743,0.00002676826,0.00004041258,0.0001340811,0.00000579807],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004305633,"about_ca_system_score_gemma":0.00003832548,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000511337,"about_ca_topic_score_gemma":0.00001211009,"domain_scores_codex":[0.9988655,0.00007023907,0.0002890961,0.0003236742,0.0002556055,0.0001958295],"domain_scores_gemma":[0.9988277,0.0004678395,0.0000681529,0.0002275365,0.0003539043,0.00005489106],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","study_design_scores_codex":[0.001512311,0.008830819,0.2923724,0.0002894314,0.0002847259,7.881883e-7,0.02808273,0.000234739,0.008331126,0.3736137,0.004243255,0.2822039],"study_design_scores_gemma":[0.0005083358,0.0004386819,0.8235789,0.0001023106,0.00003627206,0.00002380581,0.001981443,0.00003847077,0.000403493,0.1723334,0.0003970274,0.00015783],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.979445,0.0002989559,0.001657528,0.01037148,0.0004229417,0.0007152491,0.00001196961,0.00001163596,0.007065257],"genre_scores_gemma":[0.9984432,0.00000611828,0.000341211,0.0007837841,0.000114983,0.0001677365,0.000002137885,0.000004997873,0.0001358141],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5312065,"threshold_uncertainty_score":0.2338406,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1977836768","doi":"10.1177/0013164402239324","title":"Psychometric Properties Of The French Canadian Version Of The State-Trait Anxiety Inventory For Children","year":2003,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Child and Adolescent Psychosocial and Emotional Development","field":"Psychology","cited_by":70,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Institut universitaire en santé mentale de Montréal","funders":"","keywords":"State-Trait Anxiety Inventory; Psychology; Anxiety; Psychometrics; Trait anxiety; Trait; Clinical psychology; Developmental psychology; Internal consistency; Psychiatry","authors":[{"name":"Lyse Turgeon","is_ca":true},{"name":"Elise Chartrand","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.07757421657594854,"gpt":0.276566351871343,"spread":0.1989921352953945,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004707737,0.0001192007,0.0001296564,0.00006748577,0.0002201687,0.000008372325,0.0002324982,0.00006992266,0.0004169337],"category_scores_gemma":[0.0001155907,0.00006266927,0.0001329244,0.0003394243,0.0001641145,0.00001980582,0.00001308154,0.0001183556,0.000005461169],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006935687,"about_ca_system_score_gemma":0.0001864167,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001651973,"about_ca_topic_score_gemma":0.0008265441,"domain_scores_codex":[0.9987738,0.0001182415,0.0002830225,0.0002464716,0.0003764406,0.000202075],"domain_scores_gemma":[0.9993719,0.00001659759,0.0001264203,0.0001885009,0.0001931183,0.0001034176],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00009077661,0.002411121,0.8671469,0.00008283286,0.0002050076,2.043393e-8,0.001825296,0.000005185561,0.001608098,0.01616655,0.1005455,0.009912643],"study_design_scores_gemma":[0.0005009168,0.00005625038,0.9774894,0.00009418323,0.00001424249,0.000004017526,0.00007021254,1.276605e-7,0.0002651149,0.005942801,0.0154739,0.00008880837],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9850994,0.001508673,0.000008431532,0.004555902,0.002474578,0.0006298014,0.00006364176,0.000003771298,0.00565579],"genre_scores_gemma":[0.9980035,0.00003736374,0.00004439047,0.0007386136,0.00009501472,0.0000440879,0.00000544905,0.000005462672,0.001026082],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1103425,"threshold_uncertainty_score":0.4565131,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1995196170","doi":"10.1177/0013164403256358","title":"Approximate Confidence Intervals for Effect Sizes","year":2003,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":70,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Manitoba","funders":"","keywords":"Statistics; Confidence interval; Interval (graph theory); Contrast (vision); Coverage probability; Mathematics; Population; Credible interval; Tolerance interval; Econometrics; Demography; Computer science; Combinatorics","authors":[{"name":"James Algina","is_ca":false},{"name":"H. J. Keselman","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7704275307910567,"gpt":0.5976349069448019,"spread":0.1727926238462548,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006547533,0.0001611382,0.0003679041,0.0000246575,0.00009761367,0.0000377462,0.0001284556,0.0001002052,0.000828165],"category_scores_gemma":[0.1121442,0.000107719,0.0001025942,0.00007100306,0.0001431431,0.00002580918,0.00001389064,0.0001277271,0.00001926902],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003449664,"about_ca_system_score_gemma":0.00002661991,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":6.379288e-7,"about_ca_topic_score_gemma":3.831762e-7,"domain_scores_codex":[0.9977139,0.000780175,0.0004945093,0.0004031129,0.0003821647,0.0002262019],"domain_scores_gemma":[0.975208,0.02409479,0.0001313311,0.0002069461,0.0002114082,0.0001475343],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001605331,0.0006803941,0.0005580263,0.00014899,0.00005155542,3.36728e-7,0.00004318759,1.277427e-7,0.0009075246,0.9791241,0.009429845,0.008895325],"study_design_scores_gemma":[0.0008250854,0.0006506283,0.003207789,0.0000730859,0.0000441678,0.000007808377,0.00001210289,0.000004600516,0.001185393,0.9916546,0.002190242,0.000144546],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.1229184,0.001259465,0.8090975,0.01256959,0.00685794,0.005344098,0.0001017702,0.0001887871,0.04166248],"genre_scores_gemma":[0.4791962,0.00001795851,0.5191544,0.0005466006,0.0002475076,0.0005943283,9.807452e-7,0.00001240821,0.0002296453],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.3562778,"threshold_uncertainty_score":0.9067826,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2158599033","doi":"10.1177/0013164407313368","title":"Factorial Invariance of the Academic Amotivation Inventory (AAI) Across Gender and Grade in a Sample of Canadian High School Students","year":2008,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Motivation and Self-Concept in Sports","field":"Psychology","cited_by":65,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Ottawa; Université du Québec en Outaouais","funders":"","keywords":"Amotivation; Psychology; Measurement invariance; Metric (unit); Sample (material); Internal consistency; Factorial; Construct validity; Confirmatory factor analysis; Developmental psychology; Test validity; Psychometrics; Social psychology; Structural equation modeling; Statistics; Mathematics; Intrinsic motivation","authors":[{"name":"Isabelle Green‐Demers","is_ca":true},{"name":"Lisa Legault","is_ca":true},{"name":"Daniel Pelletier","is_ca":true},{"name":"Luc G. Pelletier","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2520589651374427,"gpt":0.3812125439226223,"spread":0.1291535787851796,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005904109,0.0001107056,0.0001655488,0.0001178865,0.0001035651,0.00000584622,0.0001919695,0.0001512511,0.001033473],"category_scores_gemma":[0.0003390339,0.00008145763,0.0000305412,0.0003332029,0.000210972,0.00005255689,0.00002557443,0.0002575804,0.000003947376],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008985195,"about_ca_system_score_gemma":0.0001151029,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01233363,"about_ca_topic_score_gemma":0.001727895,"domain_scores_codex":[0.9984673,0.0001678906,0.0004065469,0.0002777365,0.0004852402,0.0001952536],"domain_scores_gemma":[0.9992865,0.00008904954,0.0001773475,0.0001731235,0.0001240502,0.0001499441],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00003913377,0.0004299157,0.9831203,0.000007981651,0.00002129634,3.983284e-7,0.005049207,0.000002415868,0.0002130542,0.007949858,0.002978842,0.0001876151],"study_design_scores_gemma":[0.0007625921,0.00005910245,0.991302,0.000024852,0.000004153621,0.000005234649,0.0004675948,1.598379e-7,0.00002838193,0.005045071,0.00222148,0.00007938869],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9962611,0.0002688811,0.0000208559,0.0009414493,0.001335537,0.0003348314,0.00003398113,0.000003981811,0.0007994347],"genre_scores_gemma":[0.9988304,0.00006747055,0.00009768832,0.0006165267,0.0001907897,0.0000525157,0.00001323582,0.000005247715,0.0001261183],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.01060573,"threshold_uncertainty_score":0.9998797,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1978985895","doi":"10.1177/0013164409344498","title":"Testing the Second-Order Factor Structure and Measurement Equivalence of the Wong and Law Emotional Intelligence Scale Across Gender and Ethnicity","year":2009,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Emotional Intelligence and Performance","field":"Psychology","cited_by":65,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Marriott International (Canada)","funders":"","keywords":"Equivalence (formal languages); Psychology; Ethnic group; Emotional intelligence; Scale (ratio); Social psychology; Context (archaeology); Mathematics; Law; Political science","authors":[{"name":"Daniel S. Whitman","is_ca":false},{"name":"David La Rooy","is_ca":true},{"name":"Chockalingam Viswesvaran","is_ca":false},{"name":"Eyran Kraus","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3337503138767035,"gpt":0.4099216136148018,"spread":0.07617129973809833,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008299434,0.0001928343,0.0001608337,0.00001958864,0.0004350641,0.00005110424,0.0001916952,0.0001143142,0.0007099021],"category_scores_gemma":[0.000149972,0.0001083509,0.00003149033,0.0001704259,0.000634068,0.00009207976,0.00006047039,0.000330014,0.000003356774],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003095352,"about_ca_system_score_gemma":0.00003356196,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003554484,"about_ca_topic_score_gemma":0.0000551295,"domain_scores_codex":[0.9982008,0.0001606922,0.0003314035,0.0004373132,0.0006088564,0.0002609429],"domain_scores_gemma":[0.9989169,0.000192293,0.0001368626,0.0002172364,0.000431179,0.0001055744],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0004214112,0.002678202,0.5313779,0.000347801,0.0003258176,0.000002003692,0.02441176,0.00007246901,0.05627546,0.1743285,0.002041282,0.2077174],"study_design_scores_gemma":[0.0001259574,0.0001957244,0.9622928,0.00006706847,0.00001129179,0.00007413684,0.0004246533,0.00001965675,0.0005597667,0.0357689,0.0003247345,0.0001353052],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9896703,0.00363701,0.0002467562,0.00336461,0.0004311968,0.0003039627,0.0000489842,0.000009314278,0.002287824],"genre_scores_gemma":[0.9971108,0.00006396321,0.000402384,0.002118187,0.00017804,0.00001523049,0.000001747384,0.000005049619,0.0001045577],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4309149,"threshold_uncertainty_score":0.777293,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2143769021","doi":"10.1177/0013164406296976","title":"The Impact of Outliers on Cronbach's Coefficient Alpha Estimate of Reliability: Visual Analogue Scales","year":2007,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":63,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Cronbach's alpha; Outlier; Statistics; Sample (material); Population; Reliability (semiconductor); Statistic; Context (archaeology); Mathematics; Econometrics; Psychometrics; Demography; Geography; Chemistry","authors":[{"name":"Yan Liu","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2458833365065614,"gpt":0.5254813432007601,"spread":0.2795980066941987,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002581417,0.0001298856,0.0002243327,0.00003638637,0.0001143068,0.000008474264,0.0001075958,0.00005522149,0.0000431854],"category_scores_gemma":[0.002524885,0.0000683578,0.00009598085,0.0001059753,0.0003381782,0.00001664167,0.00001854261,0.00011994,0.000001747409],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00007629671,"about_ca_system_score_gemma":0.00003547875,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001055611,"about_ca_topic_score_gemma":0.000005029783,"domain_scores_codex":[0.9984846,0.0001255393,0.0004587228,0.0002354667,0.0004819173,0.0002137595],"domain_scores_gemma":[0.9971873,0.002023011,0.000178036,0.0001923088,0.0002961921,0.0001231575],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.001976905,0.01101406,0.009969966,0.000155805,0.0002244305,0.00000206087,0.0006473643,0.00123419,0.005878607,0.7504942,0.004144181,0.2142582],"study_design_scores_gemma":[0.0004030046,0.001523878,0.3433439,0.00006309153,0.00002372365,0.00000386083,0.0001127603,0.000344616,0.0005379723,0.6533895,0.0001226109,0.000131072],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7237012,0.0001962302,0.2728417,0.0007107306,0.0002106186,0.0003200495,0.00003100732,0.00001043339,0.00197805],"genre_scores_gemma":[0.9517711,0.00001595622,0.04808679,0.00003195209,0.00004566477,0.00001306913,0.000002730885,0.000006157929,0.00002661119],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3333739,"threshold_uncertainty_score":0.3022706,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2160937459","doi":"10.1177/0013164406288161","title":"Confidence Interval Coverage for Cohen's Effect Size Statistic","year":2006,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":62,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Manitoba","funders":"","keywords":"Statistics; Confidence interval; Percentile; Statistic; Mathematics; Sample size determination; Coverage probability; Population; Robust confidence intervals; Truncated mean; Econometrics; Medicine","authors":[{"name":"James Algina","is_ca":false},{"name":"H. J. Keselman","is_ca":true},{"name":"Randall D. Penfield","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2409555901132229,"gpt":0.467946665240317,"spread":0.2269910751270941,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008694425,0.0001524737,0.0002339806,0.00001644695,0.0001128298,0.00003122111,0.00008309371,0.00005401506,0.0002835433],"category_scores_gemma":[0.003321554,0.0001093976,0.00005282411,0.0000415813,0.00009378977,0.00003696384,0.00001392421,0.00009579843,0.000007064588],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005319736,"about_ca_system_score_gemma":0.00002102066,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000008709344,"about_ca_topic_score_gemma":0.000004415031,"domain_scores_codex":[0.9987074,0.0001460001,0.0002881111,0.0003304374,0.0003171143,0.000211004],"domain_scores_gemma":[0.9950171,0.004502312,0.00008584469,0.0001261754,0.0001767567,0.00009186548],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001677613,0.000484753,0.0001975011,0.0001027106,0.00001794579,0.000001108807,0.00002967684,0.000004513136,0.001266797,0.9768467,0.01086278,0.01001777],"study_design_scores_gemma":[0.0006562854,0.0005407867,0.01636171,0.00004589925,0.00003018201,0.000008258628,0.000006584885,0.0000516735,0.0000991089,0.9799551,0.002097334,0.0001471147],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.02507291,0.0001775345,0.969767,0.00133628,0.0004196357,0.0005986587,0.00008356532,0.00002724539,0.002517208],"genre_scores_gemma":[0.7541516,0.000006364191,0.2445773,0.0002679806,0.0002210328,0.0002706592,0.000008859582,0.000009474693,0.0004866369],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.7290788,"threshold_uncertainty_score":0.4461104,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2309234137","doi":"10.1177/0013164415596421","title":"Improving the Factor Structure of Psychological Scales","year":2015,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Child and Adolescent Psychosocial and Emotional Development","field":"Psychology","cited_by":58,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Likert scale; Psychology; Scale (ratio); Confirmatory factor analysis; Exploratory factor analysis; Acquiescence; Guttman scale; Psychometrics; Social psychology; Statistics; Clinical psychology; Structural equation modeling; Developmental psychology; Mathematics","authors":[{"name":"Xijuan Zhang","is_ca":true},{"name":"Victoria Savalei","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1701546745592259,"gpt":0.3575145933061743,"spread":0.1873599187469484,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0003507639,0.0001812745,0.0001879974,0.00004192416,0.0001397613,0.00002742423,0.0002418323,0.000125126,0.001260848],"category_scores_gemma":[0.0001313029,0.0001006904,0.0000786173,0.0001618974,0.0002358825,0.00003809835,0.0000369101,0.0002317773,0.00005666243],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004279183,"about_ca_system_score_gemma":0.00003417223,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002150044,"about_ca_topic_score_gemma":0.000007595842,"domain_scores_codex":[0.9982392,0.0001542081,0.0003521957,0.0004109164,0.0005958862,0.0002476072],"domain_scores_gemma":[0.999136,0.00007285228,0.0001362955,0.0002131785,0.0002379217,0.0002037009],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.001495394,0.01145882,0.2676418,0.0001208306,0.0004132916,0.000007168411,0.007830977,0.000009526429,0.02467436,0.2598835,0.1944737,0.2319905],"study_design_scores_gemma":[0.0006670816,0.0002517314,0.9736795,0.00002905002,0.0000104059,0.00003313701,0.0003406731,8.976442e-7,0.00004746363,0.01914757,0.005630061,0.0001623892],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9787708,0.001758141,0.0001257801,0.009052657,0.003038447,0.0002936782,0.00003848308,0.00002497309,0.006897032],"genre_scores_gemma":[0.9972616,0.00002040328,0.0003050652,0.001291339,0.0006922423,0.00001718119,0.00001344941,0.000007246496,0.0003914345],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7060377,"threshold_uncertainty_score":0.9996521,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2568498297","doi":"10.1177/0013164416685599","title":"Retrofitting Diagnostic Classification Models to Responses From IRT-Based Assessment Forms","year":2017,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":55,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Retrofitting; Item response theory; Computer science; Process (computing); Test (biology); Data mining; Engineering; Psychometrics; Mathematics; Statistics","authors":[{"name":"Ren Liu","is_ca":false},{"name":"Anne Corinne Huggins‐Manley","is_ca":false},{"name":"Okan Bulut","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8783454710157835,"gpt":0.5694813197426731,"spread":0.3088641512731104,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008993718,0.0001813124,0.0002833432,0.0002246848,0.000899116,0.0006676288,0.0009590128,0.0001014814,0.0004272539],"category_scores_gemma":[0.126592,0.000117094,0.00008495786,0.0003042845,0.0001244911,0.0002398118,0.0001161724,0.0001997486,0.00006889915],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001152383,"about_ca_system_score_gemma":0.0001124176,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006475648,"about_ca_topic_score_gemma":0.00001110335,"domain_scores_codex":[0.9957938,0.0004965072,0.0006600329,0.0008731709,0.001882763,0.0002936843],"domain_scores_gemma":[0.9773806,0.02017188,0.0004968618,0.001054152,0.0006110707,0.0002853962],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002351562,0.0008246595,0.6559843,0.000005435209,0.00003165763,0.000002794897,0.0002034041,0.0001819361,0.002983168,0.02102429,0.01181897,0.3067042],"study_design_scores_gemma":[0.0002782426,0.0001662484,0.7916595,0.00003754191,0.000007388584,0.000001231027,0.000177543,0.000883112,0.00003165029,0.2050371,0.001583425,0.0001369729],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8676457,0.0002523679,0.07320807,0.03922619,0.001484805,0.0003982214,0.00003173963,0.00003583509,0.01771703],"genre_scores_gemma":[0.9380841,0.00001545155,0.06003863,0.001214104,0.0003000869,0.0001395614,0.000006445977,0.000006980553,0.0001946045],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3065673,"threshold_uncertainty_score":0.8807651,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2081705014","doi":"10.1177/0013164414527449","title":"Seeing Perfectly Fitting Factor Models That Are Causally Misspecified","year":2014,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":53,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Factor (programming language); Factor analysis; Causal model; Econometrics; Causal structure; Psychology; Statistics; Mathematics; Computer science","authors":[{"name":"Leslie A. Hayduk","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8554756584236994,"gpt":0.4901287910866036,"spread":0.3653468673370959,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007614276,0.0002047761,0.0003317494,0.0002088554,0.000375578,0.0002580219,0.000505332,0.0001092446,0.0009531891],"category_scores_gemma":[0.03187465,0.0001308369,0.000101174,0.0005551794,0.0000892616,0.0001615242,0.00008705482,0.0002429453,0.00007422789],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005173477,"about_ca_system_score_gemma":0.0000290475,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000009931778,"about_ca_topic_score_gemma":0.000003583861,"domain_scores_codex":[0.9957297,0.0006968732,0.0005377262,0.0008098857,0.001836822,0.0003889878],"domain_scores_gemma":[0.9899974,0.008510029,0.0003494757,0.0004280913,0.0004725409,0.0002424994],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0001107646,0.0008389912,0.4263179,0.00002280814,0.00005716894,0.000002777104,0.00097331,0.0003278076,0.002614633,0.01678118,0.02120681,0.5307458],"study_design_scores_gemma":[0.0002805086,0.0001288329,0.842547,0.00003679071,0.000005389289,0.00001527145,0.0003734984,0.0005717279,0.00005039683,0.145068,0.01072212,0.000200478],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8962564,0.001363949,0.04224876,0.0104976,0.00210697,0.000258757,0.000007877064,0.00007308759,0.04718663],"genre_scores_gemma":[0.9855183,0.00002930051,0.012297,0.001083683,0.0005058317,0.00002485243,0.000001503027,0.000008488586,0.0005310408],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5305454,"threshold_uncertainty_score":0.9999601,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2941810097","doi":"10.1177/0013164419844305","title":"A Measurement Is a Choice and Stevens’ Scales of Measurement Do Not Help Make It: A Response to Chalmers","year":2019,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":51,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Relevance (law); Confusion; Reliability (semiconductor); Task (project management); Psychology; Test (biology); Epistemology; Measure (data warehouse); Computer science; Philosophy; Psychoanalysis; Data mining; Economics","authors":[{"name":"Bruno D. Zumbo","is_ca":true},{"name":"Edward Kroc","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4019735413709748,"gpt":0.46501596832292,"spread":0.06304242695194523,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005523759,0.0003008524,0.000480223,0.0001147498,0.0001075355,0.00003501314,0.000173653,0.0001094452,0.0002988053],"category_scores_gemma":[0.003506663,0.000231719,0.00008609946,0.0001730451,0.0001133517,0.00004859041,0.00007331889,0.0001873463,0.00002444335],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002109827,"about_ca_system_score_gemma":0.00009537757,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000009675781,"about_ca_topic_score_gemma":0.00001849763,"domain_scores_codex":[0.9957165,0.0004668012,0.0006232193,0.0007218259,0.002102521,0.0003692013],"domain_scores_gemma":[0.9974847,0.0005743544,0.0001871699,0.0004018653,0.000986611,0.000365338],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","study_design_scores_codex":[0.01111135,0.01167906,0.01095894,0.00236599,0.001066696,0.000008094775,0.008864738,0.0000381653,0.4204154,0.2220055,0.02666836,0.2848177],"study_design_scores_gemma":[0.005245067,0.004507681,0.5077362,0.002174842,0.0002478012,0.00005365024,0.001456012,0.00007509607,0.004758426,0.4365331,0.03566463,0.001547515],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8654582,0.001962929,0.07454339,0.05039191,0.0009387875,0.002899448,0.0001212043,0.00005439824,0.003629755],"genre_scores_gemma":[0.9303914,0.0000391546,0.06733705,0.001751187,0.00009553604,0.0001781807,7.38716e-7,0.00002047163,0.0001863399],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4967772,"threshold_uncertainty_score":0.9449225,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2139940744","doi":"10.1177/0013164404268668","title":"Reliability Generalization of Responses by Care Providers to the Center for Epidemiologic Studies-Depression Scale","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Mental Health Treatment and Access","field":"Psychology","cited_by":51,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Simon Fraser University","funders":"","keywords":"Center for Epidemiologic Studies Depression Scale; Reliability (semiconductor); Scale (ratio); Generalization; Psychology; Clinical psychology; Depression (economics); Psychometrics; Item response theory; Gerontology; Psychiatry; Depressive symptoms; Medicine; Cognition; Mathematics","authors":[{"name":"Norm O’Rourke","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2445451191917063,"gpt":0.4761799233039707,"spread":0.2316348041122644,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008813435,0.0001288465,0.0001846294,0.00003010373,0.000251915,0.000008746177,0.0001325611,0.00006985499,0.0001000421],"category_scores_gemma":[0.0003456555,0.00006964962,0.00005563045,0.00009853625,0.00009609323,0.00003605084,0.00002415344,0.00006259627,0.00001304212],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001477939,"about_ca_system_score_gemma":0.0000196415,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006352242,"about_ca_topic_score_gemma":0.00002243392,"domain_scores_codex":[0.9985415,0.000270477,0.0003515256,0.0004092107,0.0002231624,0.0002041283],"domain_scores_gemma":[0.9990596,0.0002456497,0.0001191711,0.0002207024,0.0002627935,0.00009207868],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.007668562,0.01133728,0.4135198,0.0006840693,0.0003246607,9.911049e-7,0.02387781,0.0002964864,0.005635023,0.009240372,0.3674271,0.1599879],"study_design_scores_gemma":[0.002425653,0.001620486,0.9526126,0.0001519735,0.00004336711,0.000005594438,0.003172787,9.076867e-7,0.0009034702,0.01294178,0.02590424,0.000217169],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.961713,0.007648647,0.0005497359,0.02653277,0.00109276,0.001799592,0.0001005794,0.00001580422,0.0005471782],"genre_scores_gemma":[0.9948219,0.00008373947,0.001175342,0.002379652,0.0001527696,0.001100034,0.00006612928,0.000005966046,0.0002144785],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5390928,"threshold_uncertainty_score":0.2840229,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1973081171","doi":"10.1177/0013164408322017","title":"A Cross-Cultural Examination of the Psychometric Properties of Responses to the Achievement Goal Questionnaire","year":2008,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Education, Achievement, and Giftedness","field":"Psychology","cited_by":51,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Simon Fraser University","funders":"","keywords":"Psychology; Confirmatory factor analysis; Structural equation modeling; Measurement invariance; Metric (unit); Psychometrics; Test validity; Academic achievement; Developmental psychology; Social psychology; Statistics; Mathematics","authors":[{"name":"Kou Murayama","is_ca":false},{"name":"Mingming Zhou","is_ca":true},{"name":"John C. Nesbit","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.205160770575988,"gpt":0.388235285087074,"spread":0.183074514511086,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001463423,0.0001632694,0.0001695229,0.0001273477,0.0003704241,0.00002200075,0.0003790891,0.00008390309,0.0002941336],"category_scores_gemma":[0.0007109512,0.00008186716,0.00008516163,0.0006852693,0.0005204353,0.00007673888,0.00004068916,0.0001585316,0.00002074457],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006853663,"about_ca_system_score_gemma":0.00008502373,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001148104,"about_ca_topic_score_gemma":0.00002240474,"domain_scores_codex":[0.9976435,0.0005340815,0.0004869731,0.0003452769,0.0007895458,0.0002005868],"domain_scores_gemma":[0.9983391,0.00008795855,0.0002338967,0.0004126438,0.0008450917,0.00008130188],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0003985594,0.00373581,0.9535444,0.00004746519,0.0001020013,1.551321e-7,0.01273877,0.000008192408,0.004892571,0.01018132,0.00504431,0.009306384],"study_design_scores_gemma":[0.0003942657,0.0003322137,0.9945287,0.00007726993,0.00001846142,0.00001530522,0.0008621926,4.372204e-7,0.0005495354,0.0002929777,0.002817721,0.0001109705],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9816932,0.001833923,0.0000399061,0.01225028,0.001836094,0.00063468,0.00001261269,0.00001152363,0.00168776],"genre_scores_gemma":[0.9960143,0.0000535422,0.00009289777,0.0006220079,0.0002836016,0.00031848,0.000004144483,0.000007609095,0.002603419],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.04098419,"threshold_uncertainty_score":0.3338446,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1966169084","doi":"10.1177/0013164410387382","title":"Invariance of the Measurement Model Underlying the Wechsler Adult Intelligence Scale-IV in the United States and Canada","year":2010,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Cognitive Abilities and Testing","field":"Psychology","cited_by":48,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Calgary; BC Mental Health & Substance Use Services; University of British Columbia","funders":"American Psychological Association; American Educational Research Association","keywords":"Psychology; Wechsler Adult Intelligence Scale; Measurement invariance; Normative; Generality; Construct validity; Intelligence quotient; Latent variable; Wechsler Preschool and Primary Scale of Intelligence; Psychometrics; Cognition; Developmental psychology; Construct (python library); Scale (ratio); Set (abstract data type); Latent variable model; Wechsler Intelligence Scale for Children; Sample (material); Statistics; Confirmatory factor analysis; Structural equation modeling; Mathematics","authors":[{"name":"Stephen C. Bowden","is_ca":false},{"name":"Donald H. Saklofske","is_ca":true},{"name":"Lawrence G. Weiss","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2390830827106653,"gpt":0.3553275337429329,"spread":0.1162444510322676,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001511904,0.000122785,0.000103624,0.00003167726,0.0002479476,0.00003891837,0.0002939835,0.00004850691,0.000173989],"category_scores_gemma":[0.0005319834,0.00005951138,0.00002605338,0.0002185588,0.0003242094,0.0000228429,0.00003978294,0.000401422,0.000001311351],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005675619,"about_ca_system_score_gemma":0.0001566524,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.1464472,"about_ca_topic_score_gemma":0.4040575,"domain_scores_codex":[0.9984238,0.0002274023,0.0002855479,0.0002724339,0.0005904518,0.000200389],"domain_scores_gemma":[0.9986559,0.0004926046,0.0001016086,0.0002750482,0.0004278204,0.00004701815],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","study_design_scores_codex":[0.0006582119,0.005203539,0.3533269,0.0002074449,0.0003449172,0.000004922586,0.06324825,0.001566081,0.005865741,0.4152454,0.02607961,0.128249],"study_design_scores_gemma":[0.0002501798,0.00007134879,0.9264714,0.00007472785,0.00002060168,0.00001857093,0.007652985,0.0007759113,0.00003005518,0.06307649,0.001417563,0.0001401721],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.967531,0.0002175338,0.0002211367,0.02696557,0.0004927019,0.000394957,0.00001110486,0.000004278883,0.004161716],"genre_scores_gemma":[0.9956884,0.0000289992,0.0001287081,0.003926845,0.00004840506,0.0001235425,0.000002601,0.000004815519,0.00004762504],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5731445,"threshold_uncertainty_score":0.8592367,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2026405046","doi":"10.1177/0013164404268675","title":"Effect Sizes and their Intervals: The Two-Level Repeated Measures Case","year":2005,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":45,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Manitoba","funders":"","keywords":"Statistics; Mathematics; Confidence interval; Estimator; Sample size determination; Variance (accounting); Coverage probability; Population; Econometrics; Demography","authors":[{"name":"James Algina","is_ca":false},{"name":"H. J. Keselman","is_ca":true},{"name":"Randall D. Penfield","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7635696793982186,"gpt":0.5823470726920558,"spread":0.1812226067061629,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009215028,0.0002309397,0.0003752982,0.00003012026,0.0002311203,0.00007041399,0.0001659142,0.00009811774,0.0004451975],"category_scores_gemma":[0.04252147,0.0001111132,0.00008988933,0.0001030727,0.0003096977,0.00003838418,0.00006129594,0.0002982629,0.00001920919],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004680136,"about_ca_system_score_gemma":0.00002138285,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001061805,"about_ca_topic_score_gemma":0.000020458,"domain_scores_codex":[0.9965042,0.001837054,0.0005730296,0.0004490913,0.00040603,0.0002305348],"domain_scores_gemma":[0.9755256,0.02363487,0.0001442804,0.0003198123,0.0002045845,0.0001708981],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0004990446,0.001440519,0.001090198,0.00009404421,0.0002644234,0.00001872789,0.0009300114,0.000001460859,0.001368083,0.06585635,0.01180534,0.9166318],"study_design_scores_gemma":[0.001900279,0.0009764268,0.01459187,0.0001448617,0.0001341725,0.0009421635,0.000184582,0.00006311847,0.001122121,0.9750037,0.004587272,0.0003494593],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9256414,0.002512825,0.02113685,0.04099717,0.001483913,0.001663851,0.0000829389,0.0001234186,0.006357629],"genre_scores_gemma":[0.9143822,0.00004792983,0.08349995,0.001008495,0.0007572703,0.0001558736,7.655659e-7,0.00001306507,0.0001344213],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9162824,"threshold_uncertainty_score":0.9655438,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2016558976","doi":"10.1177/0013164402062004013","title":"Stability of the Reliability of LibQual+™ Scores a Reliability Generalization Meta-Analysis Study","year":2002,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Digital Mental Health Interventions","field":"Psychology","cited_by":45,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Generalization; Reliability (semiconductor); Meta-analysis; Psychology; Statistics; Scale (ratio); Mathematics; Medicine; Geography","authors":[{"name":"Bruce Thompson","is_ca":false},{"name":"Colleen Cook","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4337064491574427,"gpt":0.4399387765720935,"spread":0.006232327414650796,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002201724,0.000163852,0.000540058,0.00007677715,0.0001021852,0.00001253438,0.0002814966,0.00007287463,0.008617758],"category_scores_gemma":[0.0005889367,0.00009706376,0.0006324793,0.0007456363,0.0003214891,0.00006920828,0.00006208979,0.0001366201,0.00001554408],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008059701,"about_ca_system_score_gemma":0.00001700259,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000236008,"about_ca_topic_score_gemma":0.00008507298,"domain_scores_codex":[0.9964108,0.001089398,0.001019795,0.0005567564,0.0007422581,0.0001809776],"domain_scores_gemma":[0.9978687,0.0002509938,0.0003935605,0.0008692311,0.000516929,0.0001006154],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002298488,0.06331404,0.9027196,0.0001626098,0.01270794,2.572921e-7,0.003478438,0.00008963816,0.0001203156,0.01198088,0.003461395,0.001735083],"study_design_scores_gemma":[0.0003002196,0.0007704425,0.983462,0.000005026858,0.006372178,0.000001022336,0.0002464231,0.00001308535,0.0001241614,0.008482116,0.0001258084,0.00009753046],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9874513,0.001010478,0.00009498042,0.005884586,0.000587856,0.001035669,0.00008823896,0.00001333911,0.003833518],"genre_scores_gemma":[0.9989361,0.00000274503,0.0002064576,0.0001502658,0.00002990715,0.0003151556,0.000005490655,0.00000558545,0.0003482641],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.08074243,"threshold_uncertainty_score":0.9922885,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2168109053","doi":"10.1177/0013164413492419","title":"Effectiveness of Combining Statistical Tests and Effect Sizes When Using Logistic Discriminant Function Regression to Detect Differential Item Functioning for Polytomous Items","year":2013,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":43,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Statistics; Differential item functioning; Polytomous Rasch model; Discriminant function analysis; Sample size determination; Mathematics; Logistic regression; Linear discriminant analysis; Psychology; Statistical hypothesis testing; Econometrics; Psychometrics; Item response theory","authors":[{"name":"Juana Gómez‐Benito","is_ca":false},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5320350509278347,"gpt":0.4801596014508845,"spread":0.05187544947695016,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004952921,0.000215017,0.00045911,0.0002551084,0.0003139445,0.0001738374,0.0001722506,0.00009801979,0.0001842049],"category_scores_gemma":[0.05126907,0.0001256745,0.00006770079,0.000361326,0.0001445648,0.0001099424,0.00008164125,0.0001327765,0.000005932791],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005789775,"about_ca_system_score_gemma":0.0000274422,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007661951,"about_ca_topic_score_gemma":0.000002067983,"domain_scores_codex":[0.9965783,0.000995175,0.0005762501,0.0006816829,0.0008808359,0.0002877587],"domain_scores_gemma":[0.9648725,0.03396068,0.0002546351,0.0002200931,0.0004835558,0.0002085083],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.002156497,0.000564463,0.3081382,0.0003364558,0.00009710169,9.382959e-7,0.0002681679,0.0001083749,0.1888231,0.002678513,0.001617451,0.4952107],"study_design_scores_gemma":[0.0007151644,0.002334883,0.9229598,0.0002352608,0.00004670861,0.00001558789,0.0001078809,0.0007051837,0.0003875064,0.07228404,0.00004673122,0.0001612185],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7450351,0.0003489867,0.2524864,0.0001798402,0.001149279,0.0006821795,0.000009389906,0.00001487993,0.00009390678],"genre_scores_gemma":[0.9764791,0.000004587695,0.02311096,0.00004264078,0.0001386311,0.0001887726,0.000005661194,0.000009192694,0.00002052812],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6148217,"threshold_uncertainty_score":0.9567225,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1975494033","doi":"10.1177/0013164413487738","title":"False Positives in Multiple Regression","year":2013,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":42,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Type I and type II errors; Statistics; Regression; False positive paradox; Regression analysis; Econometrics; Observational error; Regression diagnostic; Nominal level; Inflation (cosmology); Computer science; Linear regression; False positives and false negatives; Mathematics; Polynomial regression; Confidence interval","authors":[{"name":"Benjamin R. Shear","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.450798048209222,"gpt":0.4850311437518074,"spread":0.03423309554258541,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003928839,0.00009539343,0.000130848,0.00003182792,0.00005754854,0.00001594623,0.0000551281,0.00004780065,0.0004158877],"category_scores_gemma":[0.001398457,0.00006095699,0.00002124748,0.00005725665,0.00005635683,0.00006018958,0.00001666317,0.0001071893,0.00002443259],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003225187,"about_ca_system_score_gemma":0.000009238373,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001078703,"about_ca_topic_score_gemma":0.000004853232,"domain_scores_codex":[0.999048,0.000122609,0.0002004043,0.0002518154,0.0002283687,0.0001488262],"domain_scores_gemma":[0.9990852,0.0005633185,0.00004286361,0.0001078371,0.0001035328,0.00009728966],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001147826,0.003868335,0.01513179,0.00007721171,0.00002512522,0.000003341016,0.0007792232,0.000004551204,0.02063109,0.7100284,0.01132927,0.2380069],"study_design_scores_gemma":[0.0002612132,0.0000704745,0.1881197,0.00005502917,0.000002491864,0.000003283601,0.00006265477,0.00005150238,0.0001012813,0.8109105,0.000278363,0.00008343883],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8638424,0.0006613889,0.1163965,0.0107698,0.0003743314,0.0007774924,0.000008923082,0.00003505592,0.007134133],"genre_scores_gemma":[0.8089048,0.00001930528,0.1904628,0.0002370197,0.00005855909,0.0001403084,0.000001673928,0.000004450955,0.000171165],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2379235,"threshold_uncertainty_score":0.4553679,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2037678252","doi":"10.1177/00131640021970998","title":"Testing Repeated Measures Hypotheses When Covariance Matrices are Heterogeneous: Revisiting the Robustness of the Welch-James Test Again","year":2000,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":42,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Manitoba","funders":"","keywords":"Statistics; Bootstrapping (finance); Estimator; Mathematics; Covariance; Robustness (evolution); Type I and type II errors; Sample size determination; Repeated measures design; Covariance matrix; Statistical hypothesis testing; Analysis of covariance; Econometrics","authors":[{"name":"H. J. Keselman","is_ca":true},{"name":"James Algina","is_ca":false},{"name":"Rand R. Wilcox","is_ca":false},{"name":"Rhonda K. Kowa","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.350129595397415,"gpt":0.4037648988950279,"spread":0.05363530349761286,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001431265,0.0001782306,0.0002501656,0.00001762849,0.0003342387,0.00004224678,0.0002866162,0.00006098898,0.000285762],"category_scores_gemma":[0.008231862,0.00008961707,0.0000682247,0.0001817952,0.0002023959,0.00004297098,0.00003142161,0.0001848818,0.000002748031],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003552088,"about_ca_system_score_gemma":0.00003292793,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000009335567,"about_ca_topic_score_gemma":0.000002798591,"domain_scores_codex":[0.9980108,0.0003966494,0.0004580317,0.0003430554,0.0005717261,0.0002197104],"domain_scores_gemma":[0.9957351,0.003236631,0.0002815359,0.0003371146,0.0003408614,0.00006876465],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002868681,0.002400733,0.01689238,0.0006359305,0.0002638006,0.000006850803,0.001441211,0.006287569,0.007275415,0.03516737,0.005929096,0.9234128],"study_design_scores_gemma":[0.0006527101,0.0002350421,0.1347924,0.001485501,0.000172451,0.0001264147,0.0004099645,0.001326623,0.0006565979,0.8536454,0.005946451,0.0005504214],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7686144,0.01114655,0.1542087,0.04855612,0.001051968,0.002852426,0.0002103761,0.0002080441,0.01315142],"genre_scores_gemma":[0.8393068,0.00008264666,0.1593441,0.0004449133,0.000429121,0.00009191246,0.000001003366,0.00001635114,0.0002832187],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9228624,"threshold_uncertainty_score":0.9854902,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2066906964","doi":"10.1177/0013164409344520","title":"Athletes’ Perceptions of Coaching Competency Scale II-High School Teams","year":2009,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Sport Psychology and Performance","field":"Psychology","cited_by":40,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Psychology; Coaching; Athletes; Confirmatory factor analysis; Applied psychology; Construct validity; Scale (ratio); Construct (python library); Exploratory factor analysis; Structural equation modeling; Clinical psychology; Psychometrics; Statistics; Physical therapy; Mathematics; Medicine","authors":[{"name":"Nicholas D. Myers","is_ca":false},{"name":"Melissa A. Chase","is_ca":false},{"name":"Mark R. Beauchamp","is_ca":true},{"name":"Ben Jackson","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.06080880170417884,"gpt":0.3496966861001845,"spread":0.2888878843960057,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008450949,0.0001954031,0.0002566729,0.0001054195,0.0003290474,0.00001212443,0.0002300721,0.0001796219,0.01239816],"category_scores_gemma":[0.00005278204,0.0001584862,0.00008750377,0.0001974006,0.000197752,0.00007780943,0.00001837759,0.0003562672,0.0002959475],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004256546,"about_ca_system_score_gemma":0.00004608814,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002877492,"about_ca_topic_score_gemma":0.000005915041,"domain_scores_codex":[0.9982503,0.000145334,0.0004422684,0.0004919781,0.0003628676,0.0003072421],"domain_scores_gemma":[0.99907,0.00005346708,0.0001453105,0.0003592109,0.0001698923,0.0002020567],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0006337305,0.01287565,0.6838999,0.0000274972,0.0001777577,0.000004491978,0.004131122,0.00001444934,0.01146838,0.1169395,0.05882403,0.1110034],"study_design_scores_gemma":[0.0006598158,0.0009014199,0.9803407,0.00004629745,0.00002263891,0.00004805349,0.0003412337,0.000001216968,0.00002011461,0.01463057,0.002802306,0.0001855973],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9297983,0.001033331,0.0002256909,0.007180746,0.001347832,0.0002465655,0.00001379601,0.00003967436,0.06011407],"genre_scores_gemma":[0.9940315,0.00007251061,0.001729163,0.002238658,0.0004747781,0.00005774004,0.00002814361,0.000007316208,0.001360156],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2964408,"threshold_uncertainty_score":0.9885046,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2019602618","doi":"10.1177/0013164404267277","title":"Validity Studies Psychometric Properties of Scores on the French and Korean Versions of the Hexaco Personality Inventory","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Personality Traits and Psychology","field":"Psychology","cited_by":39,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université du Québec à Montréal; Brock University; University of Calgary; Concordia University","funders":"Agence Nationale de la Recherche","keywords":"Psychology; Personality; Internal consistency; Psychometrics; Personality Assessment Inventory; California Psychological Inventory; Personality test; Big Five personality traits; Scale (ratio); Test validity; Social psychology; Clinical psychology","authors":[{"name":"Kathleen Boies","is_ca":true},{"name":"Tae-Yong Yoo","is_ca":false},{"name":"Annik Ebacher","is_ca":true},{"name":"Kibeom Lee","is_ca":true},{"name":"Michael C. Ashton","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4726173115767044,"gpt":0.3908928201723001,"spread":0.08172449140440435,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001263054,0.0001777006,0.0002582366,0.00008457927,0.0002408067,0.000009953235,0.0002757802,0.00009679774,0.0003245499],"category_scores_gemma":[0.0004235137,0.00008557979,0.0001122854,0.000342815,0.001140204,0.00002717821,0.00004401842,0.0002518834,0.000008426831],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005110757,"about_ca_system_score_gemma":0.00005020135,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002166226,"about_ca_topic_score_gemma":0.00004381854,"domain_scores_codex":[0.9980837,0.0004311726,0.0003506542,0.0003902766,0.0005417259,0.0002025069],"domain_scores_gemma":[0.9989244,0.0002176913,0.0001900221,0.0003477542,0.0002393871,0.00008076279],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0007347097,0.01079362,0.6704273,0.0002745453,0.001002021,0.00000115341,0.02600091,0.000007040623,0.006559938,0.2385717,0.04073629,0.004890692],"study_design_scores_gemma":[0.000626752,0.0004830228,0.9812106,0.000133292,0.00004077879,0.00001338445,0.001161911,1.346116e-7,0.0001847003,0.01514756,0.0008916423,0.0001062509],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9630446,0.005078255,0.000004052918,0.02610469,0.001070688,0.0003262729,0.00002719678,0.000007755762,0.004336515],"genre_scores_gemma":[0.9977415,0.0001557394,0.0000405993,0.00166724,0.0001334028,0.00007207543,0.000001430744,0.000005923177,0.0001820634],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3107832,"threshold_uncertainty_score":0.4201127,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2912880345","doi":"10.1177/0013164419829855","title":"Item Response Tree Models to Investigate Acquiescence and Extreme Response Styles in Likert-Type Rating Scales","year":2019,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":39,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Acquiescence; Psychology; Likert scale; Rating scale; Item response theory; Social psychology; Explanatory model; Scale (ratio); Psychometrics; Statistics; Developmental psychology; Mathematics","authors":[{"name":"Minjeong Park","is_ca":true},{"name":"Amery D. Wu","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7658315465227954,"gpt":0.4754891438004715,"spread":0.2903424027223239,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02036774,0.0002083406,0.00034326,0.0005136818,0.0001489678,0.0002020068,0.0004364091,0.0001059247,0.0001967638],"category_scores_gemma":[0.07539219,0.0001422982,0.00004151578,0.001543644,0.0001454443,0.0002125412,0.0001460738,0.0002049378,0.00007136424],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00007438178,"about_ca_system_score_gemma":0.00009923058,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001711011,"about_ca_topic_score_gemma":0.00001807516,"domain_scores_codex":[0.9945995,0.002099877,0.00070328,0.0009236594,0.001321693,0.0003520384],"domain_scores_gemma":[0.9777385,0.02090774,0.0001833791,0.0004444535,0.00041189,0.0003140045],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.01225237,0.0007334612,0.7199343,0.00002208412,0.00002427036,0.000009036084,0.002987203,0.0004888008,0.1205586,0.004461916,0.008088537,0.1304395],"study_design_scores_gemma":[0.0004295992,0.0005287902,0.9520442,0.00009241828,0.000002373401,0.00001733903,0.001022192,0.0004757549,0.00008869696,0.04367717,0.001418799,0.0002026428],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9806734,0.001268786,0.0004874561,0.01522847,0.0005318762,0.0003795493,0.000003696863,0.00002380219,0.001402962],"genre_scores_gemma":[0.9776695,0.00003598116,0.02005758,0.001459073,0.00007116534,0.00003437386,7.394071e-7,0.000007588339,0.0006639539],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2321099,"threshold_uncertainty_score":0.9323962,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2014421030","doi":"10.1177/00131640021970970","title":"Structural Equation Models and the Regression Bias for Measuring Correlates of Change","year":2000,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":38,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Lakehead University; University of Manitoba","funders":"","keywords":"Structural equation modeling; Skewness; Analysis of covariance; Regression analysis; Regression; Statistics; Psychology; Linear regression; Econometrics; Regression toward the mean; Covariance; Mathematics","authors":[{"name":"Robert A. Cribbie","is_ca":true},{"name":"John Jamieson","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9410559504515326,"gpt":0.523874003364629,"spread":0.4171819470869036,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00713073,0.0001011049,0.0002122605,0.00008780471,0.0002218011,0.0000618308,0.0002089225,0.00005688857,0.0002938679],"category_scores_gemma":[0.01246389,0.00004570737,0.00005997557,0.0003169197,0.000187932,0.0001229041,0.00002347123,0.00008117962,0.00000245883],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001303161,"about_ca_system_score_gemma":0.00001235771,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001904707,"about_ca_topic_score_gemma":0.000001281146,"domain_scores_codex":[0.9978947,0.0003524174,0.0004349313,0.0003300046,0.000851439,0.0001364877],"domain_scores_gemma":[0.9924899,0.0067143,0.0002093113,0.0001891081,0.0003355541,0.00006179332],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0005845542,0.00009645323,0.03796306,0.00001183528,0.00002080679,7.384963e-8,0.001214339,0.0001545024,0.00008549957,0.02557369,0.0006474599,0.9336478],"study_design_scores_gemma":[0.0008185556,0.000115325,0.4363746,0.00003717657,0.00001068648,0.000007799115,0.0001739329,0.007264998,0.00001527742,0.5547156,0.0003920105,0.00007397938],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9835448,0.004270598,0.00257571,0.006476539,0.0005218126,0.0004954164,0.000006035543,0.000009898579,0.002099243],"genre_scores_gemma":[0.9945387,0.0001798955,0.004705471,0.0002048575,0.0001440502,0.00008938443,0.000001774152,0.000003023221,0.000132882],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9335737,"threshold_uncertainty_score":0.9958546,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1981221874","doi":"10.1177/0013164403261051","title":"A Note on How to Quantify and Report Whether Irt Parameter Invariance Holds: When Pearson Correlations are Not Enough","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":37,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Item response theory; Measurement invariance; Pearson product-moment correlation coefficient; Mathematics; Econometrics; Statistics; Scale (ratio); Measure (data warehouse); Correlation; Psychology; Psychometrics; Confirmatory factor analysis; Computer science; Structural equation modeling; Data mining","authors":[{"name":"André Rupp","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7338274295775041,"gpt":0.4936132193884701,"spread":0.240214210189034,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005228688,0.0001955677,0.0002957988,0.0002365564,0.000276062,0.0002883537,0.0002997609,0.0001248259,0.0002089257],"category_scores_gemma":[0.06219046,0.0001287343,0.00006779595,0.0006376781,0.0001126881,0.0001269226,0.00006224519,0.0002604339,0.0000945006],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00007410339,"about_ca_system_score_gemma":0.00005632257,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002900305,"about_ca_topic_score_gemma":0.00001274242,"domain_scores_codex":[0.99652,0.0003189657,0.0004921851,0.0009272744,0.001465384,0.0002762033],"domain_scores_gemma":[0.9934185,0.005073769,0.0003184405,0.0005053644,0.0004233784,0.0002605563],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.001060827,0.005557962,0.4999641,0.00004088665,0.0001702833,0.000121387,0.007542775,0.0009446129,0.005681368,0.06386925,0.1283634,0.2866831],"study_design_scores_gemma":[0.0003620776,0.0003259489,0.8711762,0.00006578492,0.00000814466,0.00009021415,0.0002225405,0.00001019364,0.00003351469,0.1050626,0.02245886,0.0001839153],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7591046,0.0004419774,0.03144462,0.201983,0.001913612,0.000511174,0.00001601098,0.00004126283,0.004543674],"genre_scores_gemma":[0.9336081,0.0000206525,0.0611322,0.00398329,0.0002699507,0.00007005443,0.000002542926,0.000007836528,0.0009053679],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3712122,"threshold_uncertainty_score":0.9457091,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2140588557","doi":"10.1177/0013164414548894","title":"A Cautionary Note on the Use of the Vale and Maurelli Method to Generate Multivariate, Nonnormal Data for Simulation Purposes","year":2014,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Statistical Methods and Bayesian Inference","field":"Mathematics","cited_by":37,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Kurtosis; Skewness; Statistics; Univariate; Mathematics; Monte Carlo method; Population; Variable (mathematics); Multivariate statistics; Sample (material); Econometrics; Algorithm","authors":[{"name":"Oscar L. Olvera Astivia","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.6134090036945616,"gpt":0.4939607930004505,"spread":0.1194482106941111,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001550841,0.00009820388,0.0001254735,0.00001806992,0.0001708144,0.0000327052,0.0001726051,0.00004599249,0.00005013205],"category_scores_gemma":[0.005167484,0.00004778763,0.00002369552,0.00007563266,0.00006455286,0.00003861812,0.00006599741,0.00008105265,0.000001525921],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001313718,"about_ca_system_score_gemma":0.0000216897,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001254846,"about_ca_topic_score_gemma":0.000003285183,"domain_scores_codex":[0.9987095,0.00042235,0.0002201699,0.0002695693,0.0002701409,0.000108297],"domain_scores_gemma":[0.9942276,0.005105071,0.00008770207,0.0003439042,0.0001777089,0.00005793713],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001560996,0.0005353406,0.0003859081,0.00004547654,0.00003158688,4.469248e-8,0.0001792796,0.0002863086,0.002047007,0.9247472,0.005565681,0.06602012],"study_design_scores_gemma":[0.0003279559,0.0003491435,0.1174316,0.00007564614,0.00006160278,0.000003163609,0.00001340621,0.03590698,0.0002853294,0.8327518,0.01264376,0.0001495434],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.01226457,0.00001625825,0.9757903,0.0108731,0.0002045924,0.0005862925,0.000137359,0.000005663516,0.0001218209],"genre_scores_gemma":[0.3402989,0.000002189721,0.6583948,0.001050191,0.0001297226,0.00006700528,0.000006656954,0.00000471309,0.00004575176],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.3280344,"threshold_uncertainty_score":0.6186335,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2154633345","doi":"10.1177/0013164403258393","title":"Impact of Test Design, Item Quality, and Item Bank Size on the Psychometric Properties of Computer-Based Credentialing Examinations","year":2004,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":36,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ministry of Education and Child Care","funders":"","keywords":"Credentialing; Computerized adaptive testing; Item bank; Test (biology); Quality (philosophy); Computer science; Consistency (knowledge bases); Test design; Item response theory; Psychology; Psychometrics; Applied psychology; Statistics; Reliability engineering; Artificial intelligence; Medical education; Test method; Medicine; Mathematics; Clinical psychology","authors":[{"name":"Dehui Xing","is_ca":true},{"name":"Ronald K. Hambleton","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8330425941791564,"gpt":0.5172334143779688,"spread":0.3158091798011876,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008970078,0.0001673623,0.000334348,0.000400802,0.0001984456,0.0001037613,0.0004178979,0.00007152487,0.0001873679],"category_scores_gemma":[0.1097481,0.00008266083,0.000111165,0.002107907,0.0002775501,0.00006438365,0.00004419868,0.0001426403,0.000003959107],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005952946,"about_ca_system_score_gemma":0.0001185307,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005090276,"about_ca_topic_score_gemma":0.000001421721,"domain_scores_codex":[0.996316,0.0007359767,0.0008326067,0.0004527099,0.001474914,0.0001877484],"domain_scores_gemma":[0.9602904,0.03767937,0.0005363902,0.0003807231,0.001000899,0.000112242],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0004194061,0.008656241,0.7963242,0.00008407261,0.0001895068,7.959198e-7,0.0009826015,0.002358467,0.04658718,0.008919579,0.003256941,0.132221],"study_design_scores_gemma":[0.0005210896,0.001110561,0.9796048,0.00009821734,0.00001004679,0.0000048983,0.0001108185,0.0001033368,0.0007538317,0.01755325,0.00002258937,0.0001064941],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.975822,0.0008044172,0.01803228,0.004136453,0.0003508073,0.000358375,0.00001169164,0.00001201578,0.0004719102],"genre_scores_gemma":[0.9853181,0.00002704813,0.01426061,0.0002388922,0.0001006227,0.00003023563,7.114274e-7,0.000004948118,0.00001884103],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1832807,"threshold_uncertainty_score":0.8977509,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2083403230","doi":"10.1177/0013164406292030","title":"Confidence Intervals for an Effect Size Measure in Multiple Linear Regression","year":2007,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Methods and Models","field":"Mathematics","cited_by":34,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Manitoba","funders":"","keywords":"Statistics; Confidence interval; Coverage probability; Mathematics; CDF-based nonparametric confidence interval; Percentile; Linear regression; Robust confidence intervals; Sample size determination; Regression analysis; Tolerance interval; Measure (data warehouse); Confidence distribution; Regression; Computer science","authors":[{"name":"James Algina","is_ca":false},{"name":"H. J. Keselman","is_ca":true},{"name":"Randall D. Penfield","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4771766274732445,"gpt":0.5316271763859817,"spread":0.05445054891273715,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004365105,0.0001530602,0.0002443375,0.0000358352,0.0000876762,0.00001299789,0.0001054801,0.00009776549,0.00006317302],"category_scores_gemma":[0.01070344,0.000100813,0.00004755397,0.00007321099,0.00007033067,0.00006328088,0.0000146898,0.0001542514,0.00000251344],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000577083,"about_ca_system_score_gemma":0.00001802116,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000005759754,"about_ca_topic_score_gemma":0.00003666292,"domain_scores_codex":[0.9984366,0.0002080261,0.0003580518,0.0003987257,0.0003520951,0.0002464442],"domain_scores_gemma":[0.9952188,0.004145924,0.00008744358,0.0001786764,0.0002037157,0.000165434],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.004291633,0.005780488,0.01003163,0.0005049775,0.00006687089,0.000008018605,0.001487094,0.00001582546,0.0553905,0.570478,0.003516058,0.3484289],"study_design_scores_gemma":[0.001352946,0.001168664,0.04502387,0.0002512332,0.00001649788,0.000008614476,0.00006742975,0.0001928944,0.001531212,0.9495791,0.0005954161,0.0002121106],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1812214,0.0002502282,0.8154432,0.0009054958,0.0004136634,0.0008528099,0.00001321451,0.00002662443,0.0008733869],"genre_scores_gemma":[0.7266189,0.000006213244,0.2728433,0.0001994011,0.0001565656,0.0001066305,0.000002891816,0.000008266777,0.00005783197],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.5453975,"threshold_uncertainty_score":0.9976298,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2154413108","doi":"10.1177/0013164406299132","title":"Correction for Attenuation With Biased Reliability Estimates and Correlated Errors in Populations and Samples","year":2007,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Statistical Methods and Bayesian Inference","field":"Mathematics","cited_by":34,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Statistics; Reliability (semiconductor); Monte Carlo method; Correction for attenuation; Attenuation; Sample size determination; Observational error; Mathematics; Population; Sampling error; Sample (material); Sampling (signal processing); Random variable; Econometrics; Computer science; Physics; Demography; Optics","authors":[{"name":"Donald W. Zimmerman","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3426256692564635,"gpt":0.449982568421525,"spread":0.1073568991650615,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001153325,0.00009227342,0.000128526,0.00005103781,0.0001047751,0.00002111631,0.00002143118,0.00005986179,0.00003251789],"category_scores_gemma":[0.00307251,0.00006474728,0.000009619359,0.0001021601,0.0001004159,0.00003934249,0.000005862543,0.00008391107,3.442668e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003740915,"about_ca_system_score_gemma":0.00001593702,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000372942,"about_ca_topic_score_gemma":0.0001125837,"domain_scores_codex":[0.9991913,0.00005225444,0.0002332689,0.0002464007,0.0001548555,0.0001218725],"domain_scores_gemma":[0.9980823,0.001562024,0.00006791364,0.00006666782,0.0001475503,0.00007351834],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0005829203,0.001117521,0.6447486,0.0001191925,0.00001985389,6.825098e-7,0.0006150044,0.000005742834,0.0006340893,0.2465374,0.001321086,0.1042979],"study_design_scores_gemma":[0.000272099,0.0001594885,0.6173151,0.00005016556,0.00001372561,0.000007054945,0.00006510629,0.0003568377,0.00001793266,0.3816535,0.00002713074,0.00006189423],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5564217,0.00006575381,0.4413445,0.001347724,0.0002286007,0.0003803177,0.000008810413,0.00001393134,0.000188695],"genre_scores_gemma":[0.6704177,0.00000485158,0.3294379,0.00004949926,0.00002514301,0.00004240906,0.000008864227,0.000003352443,0.00001026508],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1351161,"threshold_uncertainty_score":0.3678304,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2406198090","doi":"10.1177/0013164416651116","title":"Multidimensional Extension of Multiple Indicators Multiple Causes Models to Detect DIF","year":2016,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Advanced Statistical Modeling Techniques","field":"Computer Science","cited_by":33,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Differential item functioning; Item response theory; Type I and type II errors; Statistics; Statistical power; Sample size determination; Context (archaeology); Correlation; Mathematics; Econometrics; Psychometrics","authors":[{"name":"Soo In Lee","is_ca":false},{"name":"Okan Bulut","is_ca":true},{"name":"Youngsuk Suh","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1697549661461686,"gpt":0.3494013428645477,"spread":0.1796463767183791,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004203982,0.0001562215,0.0001826585,0.0001323008,0.00008321844,0.00001085518,0.0003039197,0.00006772664,0.00002668421],"category_scores_gemma":[0.0009513221,0.00009513635,0.00004293161,0.0001980709,0.00009752422,0.0001534738,0.000145389,0.00007656603,0.00001598087],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006688434,"about_ca_system_score_gemma":0.00004463078,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001364663,"about_ca_topic_score_gemma":0.000005961357,"domain_scores_codex":[0.99817,0.00008375914,0.0003303967,0.000568315,0.0006170517,0.000230491],"domain_scores_gemma":[0.998426,0.0005811085,0.00009263586,0.00036811,0.000300301,0.0002318204],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002899514,0.001884964,0.007200102,0.00003142278,0.00005308492,0.000004887745,0.0004981621,0.0005207825,0.3589428,0.2103523,0.003715921,0.4165056],"study_design_scores_gemma":[0.00151245,0.001173214,0.296593,0.0003716574,0.00001420242,0.00002299648,0.00001674906,0.006679452,0.03542852,0.655746,0.001728309,0.0007135021],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1332045,0.0001273368,0.8630716,0.002984921,0.0001973869,0.0002557783,0.00001385509,0.00007582259,0.00006878404],"genre_scores_gemma":[0.6457766,0.00001885306,0.3537998,0.0002945873,0.00002857285,0.00006215786,7.977504e-7,0.000004629731,0.00001399233],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.5125722,"threshold_uncertainty_score":0.3879548,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2166456938","doi":"10.1177/0013164406292089","title":"A Test of the Age-Based Measurement Invariance and Temporal Stability of Antonovsky's Sense of Coherence Scale","year":2007,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Health, psychology, and well-being","field":"Health Professions","cited_by":33,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of British Columbia","funders":"","keywords":"Measurement invariance; Structural equation modeling; Psychology; Scale (ratio); Stability (learning theory); Population; Test (biology); Statistics; Developmental psychology; Mathematics; Demography; Confirmatory factor analysis; Geography; Computer science","authors":[{"name":"Chris G. Richardson","is_ca":true},{"name":"Pamela A. Ratner","is_ca":true},{"name":"Bruno D. Zumbo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2646435922938717,"gpt":0.4371949212201265,"spread":0.1725513289262547,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007325362,0.0001416442,0.0003219963,0.00005077264,0.0002294554,0.000002485968,0.0001387678,0.0001747196,0.0002025525],"category_scores_gemma":[0.001193516,0.00009278931,0.00005425926,0.0002305012,0.0005190454,0.00002917937,0.00003414217,0.000364093,0.000002316341],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009786906,"about_ca_system_score_gemma":0.0003881897,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003208688,"about_ca_topic_score_gemma":0.0003974776,"domain_scores_codex":[0.9971014,0.0006212026,0.0009131567,0.0003464271,0.0007194176,0.0002984068],"domain_scores_gemma":[0.9974397,0.0008137036,0.0004604878,0.0003568181,0.0007875617,0.0001417328],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002551334,0.001692868,0.9577019,0.0004888366,0.00001669136,4.42172e-7,0.001605653,4.795347e-7,0.03283965,0.0008583998,0.0006665252,0.003873423],"study_design_scores_gemma":[0.0008469713,0.0003336221,0.9913112,0.0003709533,0.00001608191,0.000001873393,0.0004648121,0.00000225957,0.002332404,0.003442588,0.0007895829,0.00008768457],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9869743,0.000477889,0.0003853055,0.005089591,0.0005767117,0.0008859834,0.00002248955,0.000008186495,0.005579518],"genre_scores_gemma":[0.9974154,0.000028811,0.001300082,0.00108627,0.0001004485,0.00004035313,0.000002428306,0.000006068803,0.00002018089],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.03360927,"threshold_uncertainty_score":0.3783838,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4388140417","doi":"10.1177/00131644231202949","title":"Comparing RMSEA-Based Indices for Assessing Measurement Invariance in Confirmatory Factor Models","year":2023,"lang":"en","type":"article","venue":"Educational and Psychological Measurement","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":33,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"York University","funders":"","keywords":"Structural equation modeling; Confirmatory factor analysis; Mathematics; Goodness of fit; Statistics","authors":[{"name":"Nataly Beribisky","is_ca":true},{"name":"Gregory R. Hancock","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9282427079429465,"gpt":0.5543654165441689,"spread":0.3738772913987776,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01399356,0.0001904563,0.0003754273,0.0005678244,0.0002572484,0.0002998393,0.0004941294,0.0001022524,0.000138931],"category_scores_gemma":[0.02272173,0.0001359468,0.00008562994,0.001448076,0.00009759735,0.0002431214,0.00005408397,0.0001958715,0.00002928321],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001476238,"about_ca_system_score_gemma":0.0001702269,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001956253,"about_ca_topic_score_gemma":0.00001836127,"domain_scores_codex":[0.9953379,0.0004935041,0.0007831517,0.0007400192,0.002236641,0.0004088095],"domain_scores_gemma":[0.9921908,0.006378666,0.0003024663,0.0003072849,0.0006512982,0.0001695066],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002431028,0.001465138,0.8307145,0.0001055601,0.00006048222,0.000004614496,0.0009188183,0.006134728,0.00345903,0.01227559,0.01768203,0.1269364],"study_design_scores_gemma":[0.0007700779,0.00007291818,0.8826404,0.00008748665,0.000004462303,0.000001722252,0.0003716276,0.008063969,0.00004756506,0.1056735,0.002079491,0.0001868211],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9347901,0.001086542,0.04833745,0.007699773,0.001940648,0.0007530217,0.00001469975,0.00009510481,0.005282701],"genre_scores_gemma":[0.9898903,0.00001405574,0.009096093,0.0005831067,0.0001556268,0.0002128144,0.00000510135,0.000008370866,0.00003455752],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1267496,"threshold_uncertainty_score":0.9855103,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null}]}