Jérémie F Cohen1, Martin Chalumeau2, Robert Cohen3, Daniël A Korevaar4, Babak Khoshnood5, Patrick M M Bossuyt4. 1. INSERM U1153, Obstetrical, Perinatal and Pediatric Epidemiology Research Team, Center for Epidemiology and Biostatistics Sorbonne Paris Cité (CRESS), Paris Descartes University, 53, avenue de l'Observatoire, 75014 Paris, France; Department of Pediatrics, Necker-Enfants Malades Hospital, Assistance Publique-Hôpitaux de Paris, Paris Descartes University, 149, rue de Sèvres, 75015 Paris, France; Department of Clinical Epidemiology, Biostatistics and Bioinformatics, Academic Medical Center, University of Amsterdam, PO Box 22700, 1100 DE Amsterdam, The Netherlands. Electronic address: jeremie.cohen@inserm.fr. 2. INSERM U1153, Obstetrical, Perinatal and Pediatric Epidemiology Research Team, Center for Epidemiology and Biostatistics Sorbonne Paris Cité (CRESS), Paris Descartes University, 53, avenue de l'Observatoire, 75014 Paris, France; Department of Pediatrics, Necker-Enfants Malades Hospital, Assistance Publique-Hôpitaux de Paris, Paris Descartes University, 149, rue de Sèvres, 75015 Paris, France. 3. Association Clinique et Thérapeutique Infantile du Val-de-Marne (ACTIV), 27, rue d'Inkermann, 94100 Saint-Maur-des-Fossés, France; Unité Court Séjour Nourrissons, Centre Hospitalier Intercommunal de Créteil, 40, avenue de Verdun, 94000 Créteil, France. 4. Department of Clinical Epidemiology, Biostatistics and Bioinformatics, Academic Medical Center, University of Amsterdam, PO Box 22700, 1100 DE Amsterdam, The Netherlands. 5. INSERM U1153, Obstetrical, Perinatal and Pediatric Epidemiology Research Team, Center for Epidemiology and Biostatistics Sorbonne Paris Cité (CRESS), Paris Descartes University, 53, avenue de l'Observatoire, 75014 Paris, France.
Abstract
OBJECTIVES: Empirical evaluations have demonstrated that diagnostic accuracy frequently shows significant heterogeneity between subgroups of patients within a study. We propose to use Cochran's Q test to assess heterogeneity in diagnostic likelihood ratios (LRs). STUDY DESIGN AND SETTING: We reanalyzed published data of six articles that showed within-study heterogeneity in diagnostic accuracy. We used the Q test to assess heterogeneity in LRs and compared the results of the Q test with those obtained using another method for stratified analysis of LRs, based on subgroup confidence intervals. We also studied the behavior of the Q test using hypothetical data. RESULTS: The Q test detected significant heterogeneity in LRs in all six example data sets. The Q test detected significant heterogeneity in LRs more frequently than the confidence interval approach (38% vs. 20%). When applied to hypothetical data, the Q test would be able to detect relatively small variations in LRs, of about a twofold increase, in a study including 300 participants. CONCLUSION: Reanalysis of published data using the Q test can be easily performed to assess heterogeneity in diagnostic LRs between subgroups of patients, potentially providing important information to clinicians who base their decisions on published LRs.
OBJECTIVES: Empirical evaluations have demonstrated that diagnostic accuracy frequently shows significant heterogeneity between subgroups of patients within a study. We propose to use Cochran's Q test to assess heterogeneity in diagnostic likelihood ratios (LRs). STUDY DESIGN AND SETTING: We reanalyzed published data of six articles that showed within-study heterogeneity in diagnostic accuracy. We used the Q test to assess heterogeneity in LRs and compared the results of the Q test with those obtained using another method for stratified analysis of LRs, based on subgroup confidence intervals. We also studied the behavior of the Q test using hypothetical data. RESULTS: The Q test detected significant heterogeneity in LRs in all six example data sets. The Q test detected significant heterogeneity in LRs more frequently than the confidence interval approach (38% vs. 20%). When applied to hypothetical data, the Q test would be able to detect relatively small variations in LRs, of about a twofold increase, in a study including 300 participants. CONCLUSION: Reanalysis of published data using the Q test can be easily performed to assess heterogeneity in diagnostic LRs between subgroups of patients, potentially providing important information to clinicians who base their decisions on published LRs.
Keywords:
Bayes theorem; Data interpretation, statistical; Diagnostic techniques and procedures/statistics and numerical data; Likelihood functions; Predictive value of tests; Sensitivity and specificity