Is There Evidence of P-Hacking in Imaging Research?

BACKGROUND: P-hacking, the tendency to run selective analyses until they become significant, is prevalent in many scientific disciplines. PURPOSE: This study aims to assess if p-hacking exists in imaging research. METHODS: Protocol, data, and code available here https://osf.io/xz9ku/?view_only=a9f7c2d841684cb7a3616f567db273fa. We searched imaging journals Ovid MEDLINE from 1972 to 2021. Text mining using Python script was used to collect metadata: journal, publication year, title, abstract, and P-values from abstracts. One P-value was randomly sampled per abstract. We assessed for evidence of p-hacking using a p-curve, by evaluating for a concentration of P-values just below .05. We conducted a one-tailed binomial test (α = .05 level of significance) to assess whether there were more P-values falling in the upper range (e.g., .045 < P < .05) than in the lower range (e.g., .04 < P < .045). To assess variation in results introduced by our random sampling of a single P-value per abstract, we repeated the random sampling process 1000 times and pooled results across the samples. Analysis was done (divided into 10-year periods) to determine if p-hacking practices evolved over time. RESULTS: Our search of 136 journals identified 967,981 abstracts. Text mining identified 293,687 P-values, and a total of 4105 randomly sampled P-values were included in the p-hacking analysis. The number of journals and abstracts that were included in the analysis as a fraction and percentage of the total number was, respectively, 108/136 (80%) and 4105/967,981 (.4%). P-values did not concentrate just under .05; in fact, there were more P-values falling in the lower range (e.g., .04 < P < .045) than falling just below .05 (e.g., .045 < P < .05), indicating lack of evidence for p-hacking. Time trend analysis did not identify p-hacking in any of the five 10-year periods. CONCLUSION: We did not identify evidence of p-hacking in abstracts published in over 100 imaging journals since 1972. These analyses cannot detect all forms of p-hacking, and other forms of bias may exist in imaging research such as publication bias and selective outcome reporting.

Thoracic imaging tests for the diagnosis of COVID-19.

BACKGROUND: Our March 2021 edition of this review showed thoracic imaging computed tomography (CT) to be sensitive and moderately specific in diagnosing COVID-19 pneumonia. This new edition is an update of the review. OBJECTIVES: Our objectives were to evaluate the diagnostic accuracy of thoracic imaging in people with suspected COVID-19; assess the rate of positive imaging in people who had an initial reverse transcriptase polymerase chain reaction (RT-PCR) negative result and a positive RT-PCR result on follow-up; and evaluate the accuracy of thoracic imaging for screening COVID-19 in asymptomatic individuals. The secondary objective was to assess threshold effects of index test positivity on accuracy. SEARCH METHODS: We searched the COVID-19 Living Evidence Database from the University of Bern, the Cochrane COVID-19 Study Register, The Stephen B. Thacker CDC Library, and repositories of COVID-19 publications through to 17 February 2021. We did not apply any language restrictions. SELECTION CRITERIA: We included diagnostic accuracy studies of all designs, except for case-control, that recruited participants of any age group suspected to have COVID-19. Studies had to assess chest CT, chest X-ray, or ultrasound of the lungs for the diagnosis of COVID-19, use a reference standard that included RT-PCR, and report estimates of test accuracy or provide data from which we could compute estimates. We excluded studies that used imaging as part of the reference standard and studies that excluded participants with normal index test results. DATA COLLECTION AND ANALYSIS: The review authors independently and in duplicate screened articles, extracted data and assessed risk of bias and applicability concerns using QUADAS-2. We presented sensitivity and specificity per study on paired forest plots, and summarized pooled estimates in tables. We used a bivariate meta-analysis model where appropriate. MAIN RESULTS: We included 98 studies in this review. Of these, 94 were included for evaluating the diagnostic accuracy of thoracic imaging in the evaluation of people with suspected COVID-19. Eight studies were included for assessing the rate of positive imaging in individuals with initial RT-PCR negative results and positive RT-PCR results on follow-up, and 10 studies were included for evaluating the accuracy of thoracic imaging for imagining asymptomatic individuals. For all 98 included studies, risk of bias was high or unclear in 52 (53%) studies with respect to participant selection, in 64 (65%) studies with respect to reference standard, in 46 (47%) studies with respect to index test, and in 48 (49%) studies with respect to flow and timing. Concerns about the applicability of the evidence to: participants were high or unclear in eight (8%) studies; index test were high or unclear in seven (7%) studies; and reference standard were high or unclear in seven (7%) studies. Imaging in people with suspected COVID-19 We included 94 studies. Eighty-seven studies evaluated one imaging modality, and seven studies evaluated two imaging modalities. All studies used RT-PCR alone or in combination with other criteria (for example, clinical signs and symptoms, positive contacts) as the reference standard for the diagnosis of COVID-19. For chest CT (69 studies, 28285 participants, 14,342 (51%) cases), sensitivities ranged from 45% to 100%, and specificities from 10% to 99%. The pooled sensitivity of chest CT was 86.9% (95% confidence interval (CI) 83.6 to 89.6), and pooled specificity was 78.3% (95% CI 73.7 to 82.3). Definition for index test positivity was a source of heterogeneity for sensitivity, but not specificity. Reference standard was not a source of heterogeneity. For chest X-ray (17 studies, 8529 participants, 5303 (62%) cases), the sensitivity ranged from 44% to 94% and specificity from 24 to 93%. The pooled sensitivity of chest X-ray was 73.1% (95% CI 64. to -80.5), and pooled specificity was 73.3% (95% CI 61.9 to 82.2). Definition for index test positivity was not found to be a source of heterogeneity. Definition for index test positivity and reference standard were not found to be sources of heterogeneity. For ultrasound of the lungs (15 studies, 2410 participants, 1158 (48%) cases), the sensitivity ranged from 73% to 94% and the specificity ranged from 21% to 98%. The pooled sensitivity of ultrasound was 88.9% (95% CI 84.9 to 92.0), and the pooled specificity was 72.2% (95% CI 58.8 to 82.5). Definition for index test positivity and reference standard were not found to be sources of heterogeneity. Indirect comparisons of modalities evaluated across all 94 studies indicated that chest CT and ultrasound gave higher sensitivity estimates than X-ray (P = 0.0003 and P = 0.001, respectively). Chest CT and ultrasound gave similar sensitivities (P=0.42). All modalities had similar specificities (CT versus X-ray P = 0.36; CT versus ultrasound P = 0.32; X-ray versus ultrasound P = 0.89). Imaging in PCR-negative people who subsequently became positive For rate of positive imaging in individuals with initial RT-PCR negative results, we included 8 studies (7 CT, 1 ultrasound) with a total of 198 participants suspected of having COVID-19, all of whom had a final diagnosis of COVID-19. Most studies (7/8) evaluated CT. Of 177 participants with initially negative RT-PCR who had positive RT-PCR results on follow-up testing, 75.8% (95% CI 45.3 to 92.2) had positive CT findings. Imaging in asymptomatic PCR-positive people For imaging asymptomatic individuals, we included 10 studies (7 CT, 1 X-ray, 2 ultrasound) with a total of 3548 asymptomatic participants, of whom 364 (10%) had a final diagnosis of COVID-19. For chest CT (7 studies, 3134 participants, 315 (10%) cases), the pooled sensitivity was 55.7% (95% CI 35.4 to 74.3) and the pooled specificity was 91.1% (95% CI 82.6 to 95.7). AUTHORS' CONCLUSIONS: Chest CT and ultrasound of the lungs are sensitive and moderately specific in diagnosing COVID-19. Chest X-ray is moderately sensitive and moderately specific in diagnosing COVID-19. Thus, chest CT and ultrasound may have more utility for ruling out COVID-19 than for differentiating SARS-CoV-2 infection from other causes of respiratory illness. The uncertainty resulting from high or unclear risk of bias and the heterogeneity of included studies limit our ability to confidently draw conclusions based on our results.

Diagnostic accuracy of thoracic imaging modalities for the detection of COVID-19.

The ongoing coronavirus disease 2019 (COVID-19) pandemic continues to present diagnostic challenges. The use of thoracic radiography has been studied as a method to improve the diagnostic accuracy of COVID-19. The 'Living' Cochrane Systematic Review on the diagnostic accuracy of imaging tests for COVID-19 is continuously updated as new information becomes available for study. In the most recent version, published in March 2021, a meta-analysis was done to determine the pooled sensitivity and specificity of chest X-ray (CXR) and lung ultrasound (LUS) for the diagnosis of COVID-19. CXR gave a sensitivity of 80.6% (95%CI: 69.1-88.6) and a specificity of 71.5% (95%CI: 59.8-80.8). LUS gave a sensitivity rate of 86.4% (95%CI: 72.7-93.9) and specificity of 54.6% (95%CI: 35.3-72.6). These results differed from the findings reported in the recent article in this journal where they cited the previous versions of the study in which a meta-analysis for CXR and LUS could not be performed. Additionally, the article states that COVID-19 could not be distinguished, using chest computed tomography (CT), from other respiratory diseases. However, the latest review version identifies chest CT as having a specificity of 80.0% (95%CI: 74.9-84.3), which is much higher than the previous version which indicated a specificity of 61.1% (95%CI: 42.3-77.1). Therefore, CXR, chest CT and LUS have the potential to be used in conjunction with other methods in the diagnosis of COVID-19.

Evaluating the Impact of Peer Review on the Completeness of Reporting in Imaging Diagnostic Test Accuracy Research.

BACKGROUND: Despite the nearly ubiquitous reported use of peer review among reputable medical journals, there is limited evidence to support the use of peer review to improve the quality of biomedical research and in particular, imaging diagnostic test accuracy (DTA) research. PURPOSE: To evaluate whether peer review of DTA studies published by imaging journals is associated with changes in completeness of reporting, transparency for risk of bias assessment, and spin. STUDY TYPE: Retrospective cross-sectional study. STUDY SAMPLE: Cross-sectional study of articles published in Journal of Magnetic Resonance Imaging (JMRI), Canadian Association of Radiologists Journal (CARJ), and European Radiology (EuRad) before March 31, 2020. ASSESSMENT: Initial submitted and final versions of manuscripts were evaluated for completeness of reporting using the Standards for Reporting Diagnostic Accuracy Studies (STARD) 2015 and STARD for Abstracts guidelines, transparency of reporting for risk of bias assessment based on Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2), and actual and potential spin using modified published criteria. STATISTICAL TESTS: Two-tailed paired t-tests and paired Wilcoxon signed-rank tests were used for comparisons. A P value <0.05 was considered to be statistically significant. RESULTS: We included 84 diagnostic accuracy studies accepted by three journals between 2014 and 2020 (JMRI = 30, CARJ = 23, and EuRad = 31) of the 692 which were screened. Completeness of reporting according to STARD 2015 increased significantly between initial submissions and final accepted versions (average reported items: 16.67 vs. 17.47, change of 0.80 [95% confidence interval 0.25-1.17]). No significant difference was found for the reporting of STARD for Abstracts (5.28 vs. 5.25, change of -0.03 [-0.15 to 0.11], P = 0.74), QUADAS-2 (6.08 vs. 6.11, change of 0.03 [-1.00 to 0.50], P = 0.92), actual "spin" (2.36 vs. 2.40, change of 0.04 [0.00 to 1.00], P = 0.39) or potential "spin" (2.93 vs. 2.81, change of -0.12 [-1.00 to 0.00], P = 0.23) practices. CONCLUSION: Peer review is associated with a marginal improvement in completeness of reporting in published imaging DTA studies, but not with improvement in transparency for risk of bias assessment or reduction in spin. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY STAGE: 1.

Association of Accuracy, Conclusions, and Reporting Completeness With Acceptance by Radiology Conferences and Journals.

BACKGROUND: Preferential publication of studies with positive findings can lead to overestimation of diagnostic test accuracy (i.e. publication bias). Understanding the contribution of the editorial process to publication bias could inform interventions to optimize the evidence guiding clinical decisions. PURPOSE/HYPOTHESIS: To evaluate whether accuracy estimates, abstract conclusion positivity, and completeness of abstract reporting are associated with acceptance to radiology conferences and journals. STUDY TYPE: Meta-research. POPULATION: Abstracts submitted to radiology conferences (European Society of Gastrointestinal and Abdominal Radiology (ESGAR) and International Society for Magnetic Resonance in Medicine (ISMRM)) from 2008 to 2018 and manuscripts submitted to radiology journals (Radiology, Journal of Magnetic Resonance Imaging [JMRI]) from 2017 to 2018. Primary clinical studies evaluating sensitivity and specificity of a diagnostic imaging test in humans with available editorial decisions were included. ASSESSMENT: Primary variables (Youden's index [YI > 0.8 vs. <0.8], abstract conclusion positivity [positive vs. neutral/negative], number of reported items on the Standards for Reporting of Diagnostic Accuracy Studies [STARD] for Abstract guideline) and confounding variables (prospective vs. retrospective/unreported, sample size, study duration, interobserver agreement assessment, subspecialty, modality) were extracted. STATISTICAL TESTS: Multivariable logistic regression to obtain adjusted odds ratio (OR) as a measure of the association between the primary variables and acceptance by radiology conferences and journals; 95% confidence intervals (CIs) and P-values were obtained; the threshold for statistical significance was P < 0.05. RESULTS: A total of 1000 conference abstracts (500 ESGAR and 500 ISMRM) and 1000 journal manuscripts (505 Radiology and 495 JMRI) were included. Conference abstract acceptance was not significantly associated with YI (adjusted OR = 0.97 for YI > 0.8; CI = 0.70-1.35), conclusion positivity (OR = 1.21 for positive conclusions; CI = 0.75-1.90) or STARD for Abstracts adherence (OR = 0.96 per unit increase in reported items; CI = 0.82-1.18). Manuscripts with positive abstract conclusions were less likely to be accepted by radiology journals (OR = 0.45; CI = 0.24-0.86), while YI (OR = 0.85; CI = 0.56-1.29) and STARD for Abstracts adherence (OR = 1.06; CI = 0.87-1.30) showed no significant association. Positive conclusions were present in 86.7% of submitted conference abstracts and 90.2% of journal manuscripts. DATA CONCLUSION: Diagnostic test accuracy studies with positive findings were not preferentially accepted by the evaluated radiology conferences or journals. EVIDENCE LEVEL: 3 TECHNICAL EFFICACY: Stage 2.

Thoracic imaging tests for the diagnosis of COVID-19.

BACKGROUND: The respiratory illness caused by SARS-CoV-2 infection continues to present diagnostic challenges. Our 2020 edition of this review showed thoracic (chest) imaging to be sensitive and moderately specific in the diagnosis of coronavirus disease 2019 (COVID-19). In this update, we include new relevant studies, and have removed studies with case-control designs, and those not intended to be diagnostic test accuracy studies. OBJECTIVES: To evaluate the diagnostic accuracy of thoracic imaging (computed tomography (CT), X-ray and ultrasound) in people with suspected COVID-19. SEARCH METHODS: We searched the COVID-19 Living Evidence Database from the University of Bern, the Cochrane COVID-19 Study Register, The Stephen B. Thacker CDC Library, and repositories of COVID-19 publications through to 30 September 2020. We did not apply any language restrictions. SELECTION CRITERIA: We included studies of all designs, except for case-control, that recruited participants of any age group suspected to have COVID-19 and that reported estimates of test accuracy or provided data from which we could compute estimates. DATA COLLECTION AND ANALYSIS: The review authors independently and in duplicate screened articles, extracted data and assessed risk of bias and applicability concerns using the QUADAS-2 domain-list. We presented the results of estimated sensitivity and specificity using paired forest plots, and we summarised pooled estimates in tables. We used a bivariate meta-analysis model where appropriate. We presented the uncertainty of accuracy estimates using 95% confidence intervals (CIs). MAIN RESULTS: We included 51 studies with 19,775 participants suspected of having COVID-19, of whom 10,155 (51%) had a final diagnosis of COVID-19. Forty-seven studies evaluated one imaging modality each, and four studies evaluated two imaging modalities each. All studies used RT-PCR as the reference standard for the diagnosis of COVID-19, with 47 studies using only RT-PCR and four studies using a combination of RT-PCR and other criteria (such as clinical signs, imaging tests, positive contacts, and follow-up phone calls) as the reference standard. Studies were conducted in Europe (33), Asia (13), North America (3) and South America (2); including only adults (26), all ages (21), children only (1), adults over 70 years (1), and unclear (2); in inpatients (2), outpatients (32), and setting unclear (17). Risk of bias was high or unclear in thirty-two (63%) studies with respect to participant selection, 40 (78%) studies with respect to reference standard, 30 (59%) studies with respect to index test, and 24 (47%) studies with respect to participant flow. For chest CT (41 studies, 16,133 participants, 8110 (50%) cases), the sensitivity ranged from 56.3% to 100%, and specificity ranged from 25.4% to 97.4%. The pooled sensitivity of chest CT was 87.9% (95% CI 84.6 to 90.6) and the pooled specificity was 80.0% (95% CI 74.9 to 84.3). There was no statistical evidence indicating that reference standard conduct and definition for index test positivity were sources of heterogeneity for CT studies. Nine chest CT studies (2807 participants, 1139 (41%) cases) used the COVID-19 Reporting and Data System (CO-RADS) scoring system, which has five thresholds to define index test positivity. At a CO-RADS threshold of 5 (7 studies), the sensitivity ranged from 41.5% to 77.9% and the pooled sensitivity was 67.0% (95% CI 56.4 to 76.2); the specificity ranged from 83.5% to 96.2%; and the pooled specificity was 91.3% (95% CI 87.6 to 94.0). At a CO-RADS threshold of 4 (7 studies), the sensitivity ranged from 56.3% to 92.9% and the pooled sensitivity was 83.5% (95% CI 74.4 to 89.7); the specificity ranged from 77.2% to 90.4% and the pooled specificity was 83.6% (95% CI 80.5 to 86.4). For chest X-ray (9 studies, 3694 participants, 2111 (57%) cases) the sensitivity ranged from 51.9% to 94.4% and specificity ranged from 40.4% to 88.9%. The pooled sensitivity of chest X-ray was 80.6% (95% CI 69.1 to 88.6) and the pooled specificity was 71.5% (95% CI 59.8 to 80.8). For ultrasound of the lungs (5 studies, 446 participants, 211 (47%) cases) the sensitivity ranged from 68.2% to 96.8% and specificity ranged from 21.3% to 78.9%. The pooled sensitivity of ultrasound was 86.4% (95% CI 72.7 to 93.9) and the pooled specificity was 54.6% (95% CI 35.3 to 72.6). Based on an indirect comparison using all included studies, chest CT had a higher specificity than ultrasound. For indirect comparisons of chest CT and chest X-ray, or chest X-ray and ultrasound, the data did not show differences in specificity or sensitivity. AUTHORS' CONCLUSIONS: Our findings indicate that chest CT is sensitive and moderately specific for the diagnosis of COVID-19. Chest X-ray is moderately sensitive and moderately specific for the diagnosis of COVID-19. Ultrasound is sensitive but not specific for the diagnosis of COVID-19. Thus, chest CT and ultrasound may have more utility for excluding COVID-19 than for differentiating SARS-CoV-2 infection from other causes of respiratory illness. Future diagnostic accuracy studies should pre-define positive imaging findings, include direct comparisons of the various modalities of interest in the same participant population, and implement improved reporting practices.