Unit of analysis errors in meta-analyses of exposure to plastic chemicals

Article type
Authors
Stone J1, Maurits J2, Symeonides C3, Swinton P4, Dunlop S5, Aromataris E1
1JBI, University of Adelaide, North Adelaide, South Australia, Australia
2Radboud University Medical Center, Nijmegen, The Netherlands
3Minderoo Foundation, Perth, Western Australia, Australia
4Robert Gordon University, Aberdeen, Scotland, United Kingdom
5Minderoo Foundation, Perth, Western Australia, Australia; The University of Western Australia, Perth, Western Australia, Australia
Abstract
"Background: Unit of analysis errors are a common pitfall in meta-analytic research. We examined the impact of plastic chemicals exposure on health outcomes through an Umbrella review. We found 16% of included meta-analyses were difficult to interpret due to unit of analysis errors.

Objectives: To present alternative approaches to statistical analysis to avoid unit of analysis errors.

Methods: Reanalysis of the association between phthalate metabolites, polychlorinated biphenyls (PCB) and risk of several health outcomes, such as uterine leiomyoma, shortened anogenital distance (AGD), prostate cancer, and non-Hodgkin lymphoma (NHL) using meta-analytic methods of increasing complexity including: 1) selecting one effect estimate; 2) creating a composite effect estimate; and 3) robust variance estimation (RVE).

Results: Differences were observed in some pooled estimates, increasing uncertainty and expected differences between studies with greater heterogeneity when accounting for dependent data. When selecting one effect estimate, the odds of uterine leiomyoma moved from 1.07 (95%CI: 0.96 to 1.19) to 1.77 (95%CI: 0.96 to 3.27). The odds of prostate cancer per 1 ug/g lipid of sum of PCBs level was 1.49 (95%CI: 1.07 to 2.06) and moved to 1.62 (95%CI: 1.00 to 2.63). Using the beta coefficient, AP distance -0.91 (95%CI: -1.63 to -0.20) and AS distance -0.86 (95%CI: -1.46 to 0.26) in male infants moved to -0.27 (95%CI: -0.44 to -0.11) using the best measure of AGD from each study.

Using composite effect estimates, odds of leiomyoma moved from 1.16 (95%CI: 0.89 to 1.50) to 1.29 (95%CI: 0.88 to 1.89). The association between DEHP and AGD using the beta coefficient moved to -0.29 (-0.46 to -0.13).

Using RVE, there was no significant association between DEHP and uterine leiomyoma, nor between PCBs and NHL.

Conclusions: Different approaches for avoiding unit of analysis errors in meta-analysis require various levels of statistical expertise. In the absence of statistical expertise, we recommend meta-analysts select only one exposure-outcome effect estimate when multiple estimates are available for analysis.

Relevance and importance to patients: This study enhances methodological guidance for conducting meta-analyses, aiming to promote more robust evidence production that informs public health practices and for meta-analysts and journal editors to consider."