Confidence assessment (CA) involves students stating alongside each of their answers a confidence rating (e.g. 0 low to 10 high) to express how certain they are that their answer is correct. Each student's score is calculated as the sum of the confidence ratings on the items that they answered correctly, minus the sum of the confidence ratings on the items that they answered incorrectly; this scoring system is designed to incentivize students to give truthful confidence ratings. Previous research found that secondary-school mathematics students readily understood the negative-marking feature of a CA instrument used during one lesson, and that they were generally positive about the CA approach. This paper reports on a quasi-experimental trial of CA in four secondary-school mathematics lessons (N = 475 students) across time periods ranging from 3 weeks up to one academic year, compared to business-as-usual controls. A meta-analysis of the effect sizes across the four schools gave an aggregated Cohen's d of -0.02 [95% CI -0.22, 0.19] and an overall Bayes Factor B[subscript 01] of 8.48. This indicated substantial evidence for the null hypothesis that there was no difference between the attainment gains of the intervention group and the control group, relative to the alternative hypothesis that the gains were different. I conclude that incorporating confidence assessment into low-stakes classroom mathematics formative assessments does not appear to be detrimental to students' attainment, and I suggest reasons why a clear positive outcome was not obtained.