Extreme original data yield extreme decline effects

Abstract

Clements et al. respond to Munday's claim that his "reanalysis shows there is not an extreme decline effect in fish ocean acidification studies". They contend that extreme data reported in early studies authored by Dixson and Munday indeed result in an "extreme" decline effect in this field, and conclude that the decline effect is primarily driven by papers by particular authors.

Fig 1

(A–C) Distributions (vertical bars) of the probability of obtaining a particular variance around a mean value of “percent time in chemical alarm cue” when examining groups of juvenile spiny chromis (Acanthochromis polyacanthus) with sample sizes of 15 (A), 20 (B), and 30 (C). Distributions were produced from 10,000 bootstrapping simulations per panel when using the control (white bars) and high CO₂ (grey bars) data obtained from behavioral videos from Munday’s lab (see Table A1 in Supp. Appendix) and defining the stated sample sizes. Sample sizes were selected to match those reported in previous papers by Munday and Dixson (colored circles from a subset of references noted in each panel), where a range of species have been tested using various chemical cues (not all associated with ocean acidification). Circle symbols are sometimes hidden behind each other. Note that all papers by Munday and Dixson contain data with variances that are lower than plausible based on the available video evidence. Full references for the papers identified in the figure panels can be found in Supp. Appendix. (D, E) Mean effect size magnitude (absolute lnRR ± upper and lower confidence bounds) for each year of publication online estimated from our revised dataset (i.e., after including or excluding Munday’s data “corrections”; see Supp. Data 1) including all studies (D) and the dataset excluding studies authored or coauthored by Dixson and/or Munday (all papers before 2012 were authored by Dixson and/or Munday) (E). Mean effect size magnitudes and confidence bounds were estimated using Bayesian simulations and a folded normal distribution. Colors in panels D and E are aesthetic in nature and follow a gradient according to year of publication. Effect size magnitudes and their uncertainty were estimated from our revised dataset using the same analytical approach as detailed in [1]. Annotated R code and raw data files for the meta-analysis in panels D and E are in Supp. Code, Supp. Data 2, and Supp. Data 3. Panels A–C were created using SigmaPlot Version 11.0 (Systat Software, San Jose, CA), while panels D and E were created using R (ggplot2 package; [13]). Underlying data for each figure panel can be found in Supp. Data 2. All supplementary files can be found at https://osf.io/7spzx/.