Recommender

Thank you for submitting your stage 1 registered report: "Investigating sex differences in learning in a range-expanding bird" to PCI RR. Your work has now been considered by three reviewers, whose comments are enclosed. As you will see, while all of them found merit in the proposed research and would ultimately like to see the stage 1 registered report accepted, they have also raised a number of points that preclude it from being acceptable in its present form. I also liked your paper, and I am therefore willing to consider it further for acceptance provided that you revise it appropriately along the lines recommended by the reviewers. In particular, I would like you to carefully consider the suggestions made about the methods (sample size, code, etc.) and the causality. Please make sure that the design table is included in the text of the main document. I look forward to seeing a revised version. When you submit the revision, please include with it a letter in which you describe how you have responded to each of the referees’ comments. Please number the comments and refer to line numbers in the original and revised paper for easy reference. A marked-up revision is also helpful and easy to join. Please submit your revised paper in an editable format.

1. REPLY: We thank the Recommender for the opportunity to revise and resubmit our preregistration. Our revisions include careful consideration of the Reviewers’ concerns (please see our numbered point-by-point responses below). Specifically, we: (1) now include an additional analysis showing that our reinforcement learning model can detect a range of different effect sizes of sex on both target learning parameters, highlighting how null results are not just due to our anticipated sample size; (2) incorporated the ‘rtools’ package information into our GitHub repository (link below), to streamline code workflow; (3) explicitly introduce the causal framework for our study (by highlighting the potential interactive nature of dispersal and sex on learning ability in great-tailed grackles); and (4) more generally clarified our introductory and methods text in cases where confusion arose due to our previously poor phrasing; in doing so, we feel these changes have substantially improved our manuscript. 

With the above improvements, however, the overarching aim of our preregistration remains the same: to test, for the first time, interactive links between dispersal, sex, and learning ability in range-expanding great-tailed grackles, using a vetted (via agent-based simulations) Bayesian reinforcement learning model. 

Our revised preregistration, which now includes the study design table, is available at the following GitHub repository: 

https://github.com/alexisbreen/Sex-differences-in-grackles-learning

Sincerely,

Dr. Alexis Breen & Dr. Dominik Deffner

Reviewer 1

The great-tailed grackle (Quiscalus mexicanus) is a passerine bird from Central America, which is rapidly expanding its distribution area across the USA. Pioneering individuals facing new environments, they are suspected to have good learning abilities. Yet, most of the pioneering great-tailed grackles seem to be males. The authors of the manuscript therefore propose to compare males and females’ learning strategy, performing experiments on temporarily captive, wild birds from three different study sites.

In my opinion, the subject is relevant, even though sex-biased dispersal could have other causes than sex-related differences in learning strategy.

The experimental protocol is clearly described and sufficiently detailed to enable replications. I have no doubt that it would be suitable for birds held captive for a long time. However, I have concerns about the way wild grackles react to captivity. How do they endure such a situation? And do both sexes endure captivity as well? This is an important question because a difference between the two sexes in the ability to endure captivity might skew the results and lead to erroneous conclusions. Another cause of concern is sample size. I understand that the authors trust in the power of their statistical procedure, but sample size is very small for females (n = 14). I noted that the two colours used in the learning experiments differ between sites; however, I don’t think it is a major issue, especially as the site enters the statistical analyses as a (random-effect) factor.

2. REPLY: Thank you—we are pleased to receive such positive feedback on the clarity of the presented experimental protocol, particularly since we did not design/carry out the work ourselves. Related to this point, we cannot comment from first-hand experience about great-tailed grackles’ tolerance to captive conditions, but we are confident that they endure it reasonably well since published studies show that aviary-housed male and female grackles will participate in a battery of cognitive tasks (Logan 2016, doi: 10.7717/peerj.2215; Logan 2016, doi: 10.7717/peerj.1975; Blaisdell 2021, doi: 10.24072/pcjournal.44). Nonetheless, the Reviewer raises an excellent point concerning unobserved factors that might influence our results. As such, we now include a sentence in our analysis plan highlighting such potential interpretative constraints (please see L184 - 187). We will also advocate for future research efforts to test the replicability of our findings, when we write our discussion. 

Concerning our anticipated sample size (at least 14 females and 35 males), ideally we would wish for a more balanced, in terms of sex, sample size; however, as noted above, data collection is outwith our control. But we hope the following additional analyses reassure the Reviewer of the strength of our analytical approach. Specifically, we have now formally tested our model for different effect sizes of sex on learning for both major parameters—please see our newly included Figure 4. Specifically, in L221 - 227 we now state: 

“this [new analysis] confirmed that, for our anticipated minimum sample size, our reinforcement learning model: (i) detects sex differences in [phi] values >= 0.03 and [lambda] values >= 1; and (ii) infers a null effect for [phi] values < 0.03 and [lambda] values < 1 i.e., very weak simulated sex differences (Figure 4). Both of these points together highlight how our reinforcement learning model allows us to say that null results are not just due to small sample size. Additionally, estimates obtained from step three were more precise in the reversal learning phase compared to the initial learning phase (Figure 4), and we can expect to detect even smaller sex differences if we analyse learning across both phases—an approach we will apply if we detect no effect of phase.” 

We note that the latest Woodland, California field trapping reports show an additional 3 females will likely be included in our validated model. But to reduce potential bias we refrain from monitoring their testing progress. 

Finally, we are pleased to be in agreement with the reviewer that tube-colour variation is accounted for by our model. 

The authors envisage fitting their data with a sophisticated reinforcement-learning model. They used simulations to check that this model provides unbiased estimates. It is unclear, however, how trials with no choice between the two options are taken into account in the model. The authors should highlight this point. Furthermore, I think that it might be interesting to analyse the time elapsed before a choice is made. Mixed-effects survival models might probably be used in this case in order to take into account the trials with no choice before 8 minutes.

3. REPLY: As usefully suggested by the Reviewer, in L181 - 184 we now highlight that: “our reinforcement learning model excludes trials where a bird did not make a tube-choice, as this measure cannot clearly speak to individual learning ability—for example, satiation rather than any learning of ‘appropriate’ colour tube-choice could be invoked as an explanation in such cases.” As such, we hope that our modelling rationale is now sufficiently clear. Because grackles were habituated pre-test to tubes (of a different colour to test tubes) by baiting, here then, latency to tube-choice at test is not biologically meaningful; in other words, all grackles are shaped to feed from tubes prior to testing. 

Specific comments:

Lines 22-23 ‘individuals must…’: Do the authors suggest that individuals behave so for the good of the species? Group selection is controversial.

4. REPLY: No, we do not wish to invoke a group selection explanation—we apologise for our lack of clarity. Here (L22 - 23), we simply wish to highlight that range expansion, by definition, involves movement by individuals away from a population’s core; this dynamic, we hope, is now clearly captured in our rephrasing, pasted below:

“Dispersal and range expansion go ‘hand in hand’; movement by individuals away from a population's core is a pivotal precondition of witnessed growth in species' geographic limits”

Lines 71-72: Are (ii) and (iii) well formulated? ‘Spatial proximity’ or ‘geographical distance’? If dispersal is male-biased, I would expect relatedness to decrease more sharply in females than males when distance increases.

5. REPLY: The Reviewer’s expectation is spot on—thank you for highlighting our mis-wording/reporting of the Sevchick et al. data. In L77 - 81, we now state: “average relatedness [...] (ii) decreases with increasing geographic distance among females; but (iii) is unrelated to geographic distance among males; hence, confirming a role for male-biased dispersal in great-tailed grackles.” We hope our text amendment now makes clear the direction of the relationship (or lack thereof) between average relatedness, sex, and geographic proximity of members in the studied grackle population. 

Lines 81-82: ‘Natural selection’ generally refers to an in situ process. In the present case, I think that it would be more accurate to write that range expansion should lead to a ‘spatial sorting’ of learning ability.

6. REPLY: We have amended our text in accordance with the Reviewer’s useful suggestion. Specifically, in L88 - 90 we now state: “[...] range expansion should disfavour slow, error-prone learning strategies, leading to a spatial sorting of learning ability in great-tailed grackles.”

Reviewer 2

The proposed project tests for sex differences in reversal learning in Grackles sampled from three different sites. The authors have clearly justified the project, and provided code and simulated data. I think the RR should be accepted. I just have some minor comments which are aimed at making the RR and any resulting reports easier for readers to access and understand.

7. REPLY: Thank you—we are delighted to read that the Reviewer recommends acceptance of our proposed study. We hope that our responses, detailed below, satisfy the Reviewer’s minor comments. 

1A. The scientific validity of the research question(s).

The research question is perfectly logical. I would like to see the authors set some of their reasoning out a little more explicitly, though.

For example, when the authors write (lines 36-37): “Given the influence of sex-biased dispersal and learning ability on range expansion, it is perhaps surprising, then, that their potential joint impact on this aspect of movement ecology remains unexamined…” What do they mean by ‘joint impact’? I’m under the impression that the hypothesis of the report is that male Grackles will have faster reversal learning because of sex-biased dispersal. I’m not sure what the implications are of this on range expansion. (I think what I’m asking is “what DAG do you draw from this sentence?”)

8. REPLY: We apologise for forcing guesswork by our poor choice of wording. We have replaced the word ‘joint’ with ‘interactive’, to make clear the synergistic implications of the dynamic being introduced—please see L37. Directed acyclic graphs (DAGs) cannot, unfortunately, represent potential interactions; however, illustrative constraints aside, the simplest DAG for our study would be: sex → reversal learning ← dispersal. We stress “simplest DAG” because, admittedly, a number of unobserved variables could also be at play—we now highlight this point in L184 - 187.

1B. The logic, rationale, and plausibility of the proposed hypotheses, as applicable.  

Again, I think this is perfectly valid. A couple of comments on spelling out the reasoning follow.

Lines 44-52. If I understand correctly, the argument here is that sex difference in learning speed can be estimated in a way that is ‘uncontaminated’ by these various other factors. Given that this argument is based on the absence of a correlation, can the authors comment at all on either: a) the strength of evidence for null correlations? b) whether these correlations have been estimated in the same populations that they are now studying?

9. REPLY: The Reviewer’s understanding is correct. Regarding the Reviewer’s questions: a) many apologies but we admittedly struggled to pin down what the Reviewer means by “strength of evidence for null correlations”; that is, the soundness of the Logan 2016a,b study designs (these followed standard animal cognition paradigms)? Sample sizes (depending on the task, these ranged from 5 - 8)? Effect sizes (none reported)? Replicability (not yet tested)? Regardless, such fine-grained data dissection, we feel, respectfully, is outwith the focus of our second introductory paragraph. We hope that our cautious phrasing in L49 - 50 that “learning ability appears [uncorrelated…]” sufficiently captures the current level of certainty concerning the reported null correlations. b): these correlational (or lack thereof) data were estimated in one of the populations included in the current study—grackles living in Santa Barbara, California (please see Logan 2016a,b). 

Lines 81-82: is the argument here that recent selection pressures for fast learning should have been stronger for males than for females? If so, could this be stated explicitly? If not, could the argument be clarified?

10. REPLY: Here, we have removed our reference to natural selection, in line with Reviewer 1’s suggestion—please see our final response to Reviewer 1. Specifically, in L89 - 91 we now state: “[...] range expansion should disfavour slow, error-prone learning strategies, leading to a spatial sorting of learning ability in great-tailed grackles.” We hope that our argument is now sufficiently clear. 

1C. The soundness and feasibility of the methodology and analysis pipeline (including statistical power analysis or alternative sampling plans where applicable).

This RR describes an analysis of data that are already being collected by a third party. I therefore have no real concerns about the methodology. I would like to see explicit reference to the primary data collection being approved by a relevant ethics board.

11. REPLY: We now include an ethics statement—please see L249. 

Power analysis based on NHST is not applicable here. I would like to see in the final report some consideration of how representative captured grackles are likely to be of grackles in general.

12. REPLY: The Reviewer raises an excellent point concerning factors, such as trappability, that might influence our results. As such, we now include a sentence in our analysis plan highlighting such potential interpretative constraints (please see L184 - 187). We will also advocate for future research efforts to test the replicability of our findings, when we write our discussion. 

1D. Whether the clarity and degree of methodological detail is sufficient to closely replicate the proposed study procedures and analysis pipeline and to prevent undisclosed flexibility in the procedures and analyses.

With regard to method: the authors cite a detailed description of the experimental protocols, which is also published in an open-access journal and is thorough.

With regard to the analysis: I have run the analysis code and got the same outputs as given in the RR (minus a couple of formatting issues). This is great to see, and I commend the clarity of the authors’ code and commenting!

13. REPLY: Thank you—we are delighted to read that the Reviewer reproduced our code outputs. 

Because I’m not great at using the rethinking package (or indeed R more generally), it took me a wee while of flailing and googling to get the code running on my machine. I’d like to suggest that the authors include links to the Rtools installation page (if indeed this was actually necessary and not just something I ended up doing because I had no clue what I was doing…) and Richard McElreath’s rethinking installations instructions page (links below) for the inexperienced rethinking user. The (clunky, inefficient) code I ended up running to get everything working is below.

14. REPLY: We wholeheartedly sympathise with the Reviewer’s experience. In addition to the links to R, stan, rethinking, and tidyverse software installation instructions, we now include on our GitHub repro (https://github.com/alexisbreen/Sex-differences-in-grackles-learning) a link to the rtools installation page—thank you for this helpful workflow suggestion. 

1E. Whether the authors have considered sufficient outcome-neutral conditions (e.g. absence of floor or ceiling effects; positive controls; other quality checks) for ensuring that the obtained results are able to test the stated hypotheses or answer the stated research question(s).

I think so. I would like to see a little more detail about how the size of the sex difference in the simulated data was chosen (was it based on a report from another species, for example?) Also it would be useful to see a slightly clearer treatment of how non-choices are to be treated in the analysis, and mention of any plans the authors might have to investigate choice latency as a variable (given that it seems to be being recorded).

15. REPLY: Thank you for these useful suggestions. 

Regarding our simulated sex effect: the simulated sex difference was chosen arbitrarily in a way that generates realistic, yet distinguishable, learning curves. To better illustrate this point, we have now formally tested our model for different effect sizes of sex on learning for both major parameters—please see our newly included Figure 4. And we now reported in L221 - 227:

“this [new analysis] confirmed that, for our anticipated minimum sample size, our reinforcement learning model: (i) detects sex differences in [phi] values >= 0.03 and [lambda] values >= 1; and (ii) infers a null effect for [phi] values < 0.03 and [lambda] values < 1 i.e., very weak simulated sex differences (Figure 4). Both of these points together highlight how our reinforcement learning model allows us to say that null results are not just due to small sample size. Additionally, estimates obtained from step three were more precise in the reversal learning phase compared to the initial learning phase (Figure 4), and we can expect to detect even smaller sex differences if we analyse learning across both phases—an approach we will apply if we detect no effect of phase.” 

Regarding our analytical treatment of non-choice data: in L181 - 184 we now highlight that “our reinforcement learning model excludes trials where a bird did not make a tube-choice, as this measure cannot clearly speak to individual learning ability—for example, satiation rather than any learning of ‘appropriate’ colour tube-choice could be invoked as an explanation in such cases.” As such, we hope that our modelling rationale is now sufficiently clear. 

Finally, regarding latency-to-choice-data: because grackles were habituated pre-test to tubes (of a different colour to test tubes) by baiting, here then, latency to tube-choice at test is not biologically meaningful; in other words, all grackles are shaped to feed from tubes prior to testing.  

 R Code:

#NB this ran on R 4.0.3 (“Bunny-Wunnies Freak Out”) and RStudio 1.3.1093.

#installed latest version of rtools from: https://cran.rstudio.com/bin/windows/Rtools/rtools40.html

#make Rtools work

write('PATH="${RTOOLS40_HOME}\\usr\\bin;${PATH}"', file = "~/.Renviron", append = TRUE)

#verify that Rtools works. correct output is:      make     "C:\\rtools40\\usr\\bin\\make.exe"

Sys.which("make")

library(cowplot)    #For combining plots

library(tidyverse)  #For data wrangling

library(rstan)      #Needed for rethinking - includes ggplot2 for graphing

library(Rcpp)    #Needed for rethinking apparently

#install rethinking as per RM instructions at: #https://www.rdocumentation.org/packages/rethinking/versions/2.13

install.packages(c("coda","mvtnorm","devtools","loo","dagitty"))

devtools::install_github("rmcelreath/rethinking")

library(rethinking) #for making the model

Reviewer 3

The authors present a very clearly thought out and well-planned study of sex differences in reversal learning in great-tailed grackles.

The authors have a well-defined experimental question (do male grackles, who are considered to more often be the dispersing sex, show faster initial and reversal learning in an experimental task, with a lower sampling rate?). The authors plan to analyse largely existing data, with only one of three study sites yet to complete data collection. The authors present a clear analysis plan which will allow them to explore the impact of sex upon individual learning rate and sampling rate, and support this approach by analysing simulated data, demonstrating that their approach is appropriate for the data they will analyse. 

I found this proposal compelling and very convincing in terms of analytical approach. I thus have only very minor comments for the authors.

16. REPLY: Thank you—we are thrilled to receive such supportive feedback, and we hope that we have sufficiently addressed the Reviewer’s minor concerns in our responses below. 

1. The authors suggest they may conduct an extended analysis exploring population-mediated sex effects. It would be good to clarify at this stage whether this would be a post-hoc analysis based on the results of the main analysis, or if this is a pre-planned analysis.

17. REPLY: Thank you for highlighting a gap in clarity in our manuscript. We now explicitly introduce both the background regarding, and our intention to explore, potential population-mediated sex effects on our target learning parameters (please see L45 - 48 ; L56 - 59; and L90 - 92); and we restate this intention when detailing our modelling approach (please see L179 - 181). We hope that these changes to our manuscript now make clear the pre-planned nature of this aspect of our analysis. 

2. Have previous studies of reversal learning in grackles looked for / found any sex effects? I believe the studies cited have relatively small sample sizes, perhaps this hasn't been possible previously. This would be good to clarify in the introduction.

18. REPLY: To our knowledge, no previous study has examined potential sex differences in great-tailed grackles’ (or any other species of grackle) reversal learning. We now highlight this point in L69—thank you for helping us to clarify the state of the literature. 

3. The introduction currently focuses largely on range expansion, and while it is well-reasoned and presents a clear argument and justification for the study, I would suggest incorporating existing literature on sex differences (or lack thereof) in reversal learning, perhaps with particular focus on the dispersing sex.

19. REPLY: The potential interactive influence between sex-biased dispersal and reversal learning (or learning ability more broadly) remains, to our knowledge, unexamined (please see L36 - 40); thus, such citeable literature is lacking. At least two studies have, however, examined potential links between sex-biased dispersal and aggression, which we highlight in L39 - 40. We note that we also highlight in L32 - 35 a study showing links between reversal learning and dispersal (urban vs. rural-dwelling individuals). We hope that the Reviewer finds the degree of studies cited from the available literature sufficient.