Reanalysis shows there is not an extreme decline effect in fish ocean acidification studies

Abstract

This Formal Comment uses re-analysis after appropriate corrections to claim that the extreme decline effect reported by Clements et al. is a statistical artefact caused by the way they corrected for zeros in percentage data, exacerbated by errors in data compilation, selective data inclusions and missing studies with strong effects.

Fig 1. Reanalysis of effect sizes in studies on the impacts of ocean acidification on fish behaviour.

(a, b) Original analysis by Clements and colleagues using 0.0001 to replace zero values in percentage and proportional data and (c, d) reanalysis with the corrected, updated, and screened data set using 0.1 to replace zero values in percentage data and 0.001 to replace zero values in proportional data. Top row (a, c) shows all calculated effect sizes (lnRR) fitted with a Loess curve and 95% confidence bounds. Bottom row (b, d) shows the modelled variance–weighted average effect sizes by year. Experiments with smaller variance are given greater weight in calculating the model means in the bottom row. Panel (e) shows how effect sizes are exaggerated when values below 1 are used to replace zeros in percentage data. The calculated effect size magnitude (lnRR = ln(treatment mean/control mean)) is shown for treatment means of 100% and 50% and control mean values between 5% and 0%, illustrating the inflation of lnRR caused by values below 1% in the denominator. The data underlying this figure (a–d) may be found in https://doi.org/10.25903/jw8m–9007. The data underlying panel (e) are found in S1 Data.

Fig 2. Comparison of decline effect and variance–weighted mean effect sizes with different methods of correcting for zero values in percentage and proportional data.

(a, b) Original data from Clements and colleagues using 0.0001 and reanalysis with (c, d) corrected, updated, and screened data using 0.0001; (e, f) corrected, updated, and screened data using 0.1 for percentage data and 0.001 for proportional data; and (g, h) corrected, updated, and screened data using 1 for percentage data and 0.01 for proportional data. Data underlying this figure may be found at https://doi.org/10.25903/jw8m–9007.