BACKGROUND: An influential psychological model of persecutory delusions proposed that they are caused by a bias towards holding others responsible for negative events (an externalising attributional bias), preventing the individual from becoming aware of underlying low self-esteem. An early version of the model predicted self-esteem would, therefore, be preserved in people with these delusions, but a later version suggested it would be unstable, and that there would be a discrepancy between explicit and implicit self-esteem, with the latter being lower. We did a comprehensive meta-analytical test of the key predictions of this model and assessed the quality of evidence. METHODS: We searched PubMed from Jan 1, 1994, to July 31, 2018, and collated systematic reviews of the defensive model's predictions in relation to persecutory delusions. We also searched PsycINFO, MEDLINE, Embase, and Web of Science for articles published from Jan 1, 2012, to Sept 10, 2016. Cross-sectional data from case-control, longitudinal, or experimental studies that examined self-esteem or the externalising attributional bias in individuals diagnosed as having schizophrenia-spectrum disorder were eligible for meta-analyses of group differences if at least 50% of participants with psychosis also had current persecutory delusions. Uncontrolled and longitudinal studies were included in meta-analyses of correlations and self-esteem instability, respectively. Study and outcome quality were assessed with the Agency for Healthcare Research and Quality assessment tool, and a modified version of Grading of Recommendations Assessment, Development and Evaluation, respectively. The study protocol is registered with PROSPERO, number CRD42016032782. FINDINGS: We screened 3053 records, examined 104 full-text reports, and included 64 eligible studies. Consistent with the predictions of both versions of the model, paranoia severity in psychosis was positively correlated with the degree of externalising attributional bias (21 studies involving 1128 individuals; r=0·18, 95% CI 0·08 to 0·27, with moderate quality evidence). People with persecutory delusions also had a greater externalising attributional bias than non-clinical individuals (27 studies involving 1442 individuals; g=0·48, 95% CI 0·23 to 0·73) and depressed individuals (ten studies involving 421 individuals; g=1·06, 0·48 to 1·63), and people with psychosis without persecutory delusions (11 studies involving 480 individuals; g=0·40, 0·12 to 0·68), all based on moderate quality evidence. Contrary to the predictions in the early version of the model, paranoia severity in psychosis was negatively correlated with explicit self-esteem (23 studies involving 1866 individuals; r=-0·26, 95% CI -0·34 to -0·17, with high quality evidence). People with persecutory delusions also had lower explicit self-esteem than non-clinical individuals (22 studies involving 1256 individuals; g=-0·88, 95% CI -1·10 to -0·66, with high quality evidence) and explicit self-esteem similarly low to that in people with psychosis without persecutory delusions (11 studies involving 644 individuals; g=-0·26, -0·54 to 0·02, with moderate quality evidence). Consistent with the predictions in the later version of the model, self-esteem instability was positively correlated with paranoia severity in psychosis (four studies involving 508 individuals; r=0·23, 95% CI 0·11-0·34, with high quality evidence), and people with persecutory delusions had a greater discrepancy between their implicit and explicit self-esteem than depressed individuals (seven studies involving 398 individuals; g=0·61, 95% CI 0·37 to 0·85, with moderate quality evidence). They had higher explicit self-esteem than depressed individuals (13 studies involving 647 individuals; g=0·89, 95% CI 0·51 to 1·28, with moderate quality evidence), but similarly low implicit self-esteem (seven studies involving 398 individuals; g=-0·19, -0·45 to 0·07, with low quality evidence). In contrast to the later predictions, people with persecutory delusions did not have a greater self-esteem discrepancy than non-clinical individuals (ten studies involving 592 individuals; g=-0·17, 95% CI -0·45 to 0·12), although the evidence was very low quality. People with psychosis with or without persecutory delusions did not differ for implicit self-esteem (four studies involving 167 individuals; g=-0·24, 95% CI -0·77 to 0·30, with low quality evidence) or self-esteem discrepancies (four studies involving 165 individuals; g=0·17, -0·19 to 0·53, with moderate quality evidence). INTERPRETATION: The predictions that self-esteem would be preserved in people with persecutory delusions in the early version of the paranoia as defence model and that implicit-explicit self-esteem discrepancy would be greater in people with persecutory delusions than in non-clinical individuals and people with psychosis without persecutory delusions in the later version of the model were not supported. By contrast, the later version correctly predicted that people with persecutory delusions have a greater self-esteem discrepancy than people with depression and a greater externalising attributional bias than all control groups, and that both this bias and self-esteem instability are associated with increased paranoia severity. Nevertheless, the reviewed data had limitations. Experimental studies, which might include interventionist-causal trials, are needed. FUNDING: None.