Meta-analysis challenges a textbook example of status signalling: evidence for publication bias

The status signalling hypothesis aims to explain conspecific variation in ornamentation by suggesting that some ornaments signal dominance status. Here, we use multilevel meta-analytic models to challenge the textbook example of this hypothesis, the black bib of house sparrows (Passer domesticus). We conducted a systematic review, and obtained raw data from published and unpublished studies to test whether dominance rank is positively associated with bib size across studies. Contrary to previous studies, our meta-analysis did not support this prediction. Furthermore, we found several biases in the literature that further question the support available for the status signalling hypothesis. First, the overall effect size of unpublished studies was zero, compared to the medium effect size detected in published studies. Second, the effect sizes of published studies decreased over time, and recently published effects were, on average, no longer distinguishable from zero. We discuss several explanations including pleiotropic, population- and context-dependent effects. Our findings call for reconsidering this established textbook example in evolutionary and behavioural ecology, raise important concerns about the validity of the current scientific publishing culture, and should stimulate renewed interest in understanding within-species variation in ornamental traits.


Introduction
Here we meta-analytically assessed the textbook example of the status signalling 114 hypothesis in the house sparrow. Specifically, we combined summary and primary 115 data from published and unpublished studies to test the prediction that dominance 116 rank is positively associated with bib size across studies. We found that the meta- 117 analytic mean was small, uncertain and overlapped zero. Hence, our results 118 challenge the status signalling function of the male house sparrow's bib. Also, we 119 identified several biases in the published literature that call for substantial changes in 120 scientific publication culture. Finally, we discuss potential biological explanations for 121 our results, and provide advice for future studies testing the status signalling 122 hypothesis.

124
Overall, we obtained the primary data for seven of 13 (54%) published studies, and 125 we provided data for six additional unpublished studies (Table 1, S1 and S2 126 Appendix).

161
Moderators of the relationship between dominance rank and bib size 162 None of the three biological moderators studied (season, group composition and 163 type of interactions) explained differences among studies (Table 3). Sampling effort 164 (i.e. the ratio of interactions to individuals recorded) was not an important moderator 165 either (Table 3). 166

175
There was no clear asymmetry in the funnel plots (Fig 2). Also, Egger's regression 176 tests did not show evidence of funnel plot asymmetry in any of the meta-analyses 177 ( Table 2 and S4 Table). However, published effect sizes were larger than 178 unpublished ones, and the latter were not different from zero (Table 4; Fig 3).

179
Additionally, we found evidence for a time-lag bias in the published literature as 180 effect sizes decreased over time (Table 4; Fig 4).

208
Estimates are presented as standardized effect sizes using Fisher's transformation (Zr  examining the full texts. We followed the preferred reporting items for systematic 385 reviews and meta-analyses (PRISMA: [89]; see S1 Appendix). We only included 386 articles in which dominance was directly inferred from agonistic dyadic interactions 387 over resources such as food, water, sand baths or roosting sites (S1 Table).

Summary data extraction 389
Some studies had more than one effect size estimate per group of birds studied.  Most studies recorded data from more than one group of birds ( When not reported directly, the number of individuals (n) was estimated from the 430 degrees of freedom. The variance in Zr was calculated as: VZr = 1/(n-3). Estimates

431
(k) based on less than four individuals were discarded (k = 33 estimates discarded). 433 We ran two multilevel meta-analyses to test whether dominance rank and bib size 434 were positively correlated across studies. The first meta-analysis, "meta 1", included  Second, three studies reported "statistically non-significant" results without showing 440 either the magnitude or the direction of the estimates (Table 1). Receipt of primary 441 23 data allowed us to recover some but not all the originally non-reported estimates.

442
Two "non-significant" estimates were still missing. To test the robustness of the 443 results to those two cases of selective reporting, we ran an additional meta-analysis 444 (see [92] for a similar approach). This second meta-analysis, "meta 2", was like meta 445 1 but included the two non-significant non-reported estimates, which were assumed 446 to be zero. Note that non-significant estimates can be either negative or positive, and 447 thus, assuming that they were zero may have either underestimated or  Meta-regressions 454 We tested if season, group composition and/or the type of interactions recorded had 455 an effect on the meta-analytic mean. For that, we ran two multilevel meta-456 regressions that included the following moderators (hereafter "biological 457 moderators"): (1) "season", referring to whether the study was conducted during the  Two studies tested the prediction twice for the same groups of birds (Table 1) and, 487 within each population, some individuals may have been sampled more than once.

25
However, we could not include "group ID" and/or "individual ID" as additional random 489 effects due to either limited sample size or because the relevant data were not 490 available.

492
For the meta-analyses, we assessed publication bias using two methods that are