Category Archives: open science

Preprints Exist!

(This post originally appeared on the Coast and Ocean Collective Blog)

Preprints are non-peer-reviewed scholarly documents that precede publication in a peer-reviewed journal. Several disciplines like Physics, Astronomy and computer Science have been using preprints through arXiv for decades. Other disciplines are catching on, notably the biological sciences (See bioRxiv), and a variety of other discipline specific preprint services (e.g., here). There are many great articles and blog posts discussing preprints recently— common questions, critiques, misconceptions, concerns, etc. — here are three especially useful introductions:

1) Bourne PE, Polka JK, Vale RD, Kiley R (2017) Ten simple rules to consider regarding preprint submission. PLoS Comput Biol 13(5): e1005473.

2) Sarabipour S, Debat HJ, Emmott E, Burgess SJ, Schwessinger B, Hensel Z (2019) On the value of preprints: An early career researcher perspective. PLoS Biol 17(2): e3000151.

3) A recent comprehensive look by Sheila Saia for the Young Hydrologic Society website is particularly useful for the Earth and Environmental scientists.

Full disclosure: I am a big advocate for preprints, interested in preprint adoption as a topic of study, and I am a current member of the EarthArXiv community — EarthArXiv is a community run preprint server for the Earth sciences (Narock et al. 2019). We have a very active community (especially on twitter!) so please bring us your questions/comments/concerns/clarifications.

To me, there are too many interesting facets about preprints to discuss in a single post. Here, I focus on some ways in which preprints compliment existing, more traditional ways of publishing — so we need to start by looking at scholarly communication and scholarly publishing, specifically journals.

A recent comment by Wouters et al (2019) outlined 5 roles for journals:

[Graphic from Wouters et al (2019) ]

I want to focus on discussing preprints in relation to task 1 (Registering) and 5 (Archiving). These are tasks that a scientific journal currently does by giving submission dates and assigning a persistent identifier to a journal article (i.e., the digital object identifier; DOI). In my opinion, these are tasks that we do not necessarily need a scientific journal to do. Instead, a preprint can accomplish these tasks — establishing precedent for an idea, and providing a means of citing the idea via a DOI.

If we rely on journals to do these tasks, the process can be attenuated. Peer-review can take months (or even years) before an article is published and visible to a community of peers. This is not a complaint against peer-review or the peer-review process — I am arguing here that several steps can occur before peer-review. My opinion is that bundling the registration of an idea (Task 1) and the archiving of idea (Task 5) with the peer-review process is suboptimal for one key reason:

No one can read, cite, or respond to an idea when the paper is hidden in review — only the editor, AE, reviewers, and coauthors can read, engage with, explore, and think about the work. These ideas may be presented at conferences, but in the written record, they do not exist (e.g., many journals have policies discouraging citations to conference abstracts). Ideas that are preprinted have persistent identifiers (DOIs) and can (and should!) be cited and discussed by others — preprints exist.

As an early career scientist, this is especially important. Scholarly work in review with no preprint remains invisible to the broader community. Early career scientists often mention ‘in prep’ or in review articles on CVs — I’d argue that this is far less meaningful than linking to a preprint version (where people could actually read and cite you work). Again, preprints exist.

Being unable to read and cite articles that are in review in a transparent way hampers our ability to do science. Hiding articles through the review process is a form of information asymmetry — and a bizarre, imperfect hiding. I know about lots of work that remains hidden — I read about it as a grant or paper reviewer, I hear about it in passing from colleagues, and conference presentations give a glimpse of what will be published in the next few years — but I cannot cite these ideas or these works unless there are preprints. Put another way — there is a subset of ideas that I know about, but can’t share with colleagues. This is strange.

This is where preprints come into the picture. Preprint services like EarthArXiv can 1) store papers (i.e., registering intellectual claims associated with author names and submission timestamps), and 2) assign DOIs and archive scholarly artifacts. Therefore preprint services accomplish Task 1 (registering) and Task 5 (archiving) in the Wouters et al (2019) taxonomy. Preprints leave the other tasks (curating, evaluating, and disseminating) for other services such as scientific journals.

My argument here is that we should unbundle the services that journals provide to increase the flow of information. Preprints can accomplish some of these tasks faster, cheaper, and better than traditional journals.

My Time at CSDMS 2019


(This post originally appeared on the Coast and Ocean Collective Blog)

In May I went to my first annual meeting of CSDMS— the Community Surface Dynamics Modeling System. It was great to see old friends and meet new ones.

CSDMS is involved in a range of different projects and provides a suite of different services to the earth surface processes modeling community. You might know about CSDMS from its model repository (with metadata and links to source code) and the handy tools developed by CSDMS to link models together. For more background on CSDMS, check out their webpage.

One nice aspect of CSDMS is that the keynotes and panels are recorded and put on YouTube, and many poster presenters upload PDFs of their poster. I have spent a few hours skimming through these videos and PDFs from past meetings — lots of interesting ideas.

The annual meeting theme this year was ‘Bridging Boundaries’, and there was a range of interesting talks, posters, clinics, breakout sessions, and panels. I want to just mention a few highlight during those 3 packed days.

  • I really enjoyed the wide range of keynotes. Two particularly interesting ones were:
  • I really enjoyed the 2 panel discussions:
  • A real highlight for me was Dan Buscombe’s deep learning clinic. Dan walked us through a comprehensive Jupyter notebook based on his work on pixel-scale image classification. It was great to hear Dan explain his workflow, and it was great to meet him in person. I urge you to check out his work!
  • There were too many amazing posters to cover in one post. I recommend scrolling through the abstracts and poster pdfs online.
  • I live-tweeted the 3rd day through the CSDMS and AGU EPSP twitter accounts. This was really fun and I’m grateful for the opportunity from the AGU EPSP social media team.
  • I am very grateful to CSDMS for inviting me to give a keynote this year — it was exciting to share my ideas with such a talented group of people. My talk — video, slides — focused on ML work that I have done with the Coast and Ocean Collective (and others), specifically work on swash, runup, ‘hybrid’ models, and the ML review paper that was just published.
  • Lastly, I ate a lot of (good) pizza.



Data Collection: getting data for GRL articles

In previous posts I have looked at several aspects of Earth and Space Science citations in Wikipedia. As part of a project I am working on, I’m interested in expanding this work to look at some mechanics of citations in Wikipedia to articles in Geophysical Research Letters (e.g., when do they appear, who writes the edit, on what Wikipedia pages, etc.). In this post, I want to walk through my method for getting the data that I will analyze. All the code is available (note that I am not a good R programmer).

Data on Wikipedia mentions are aggregated by Altmetric. rOpenSci built a tool to get altmetric data (rAltmetric) using the Altmetric API. rAltmetric works by retrieving data for each paper using the paper’s DOIs — so I need the DOIs for any/all papers before I do anything. Fortunately, rOpenSci has a tool for this too — rcrossref — which queries the Crossref database for relevant DOIs given some condition.

Since my project is focused on Geophysical Research Letters, I only need the DOIs for papers published in GRL. Using the ISSN for GRL, I downloaded 36,960 DOIs associated with GRL and then the associated Altmetric data (using rAltmetric).

The data from rAltmetric returns the number of times a given article is cited in Wikipedia. But I want some extra detail:

  • The name of the Wikipedia article where the GRL citation appears
  • When the edit was made
  • and Who made the edit

This information is returned through the Altmetric commercial API — you can email Stacy Konkiel at Altmetric to get a commerical API key through Altmetric’s ‘researcher data access program’ (free access for those doing research). I got the data another way, via webscraping. To keep everything in R, I used rvest to scrape the Altmetric page (for each GRL article) to get Wikipedia information — the Wikipedia page that was edited, the author, and the edit date. Here is an example of an page for a GRL article:


The Wikipedia page (‘Overwash’), the user (‘Ebgoldstein’ — hey that’s me!), and the edit date (’10 Feb 2017′) are all mentioned… this is the data that I scraped for.

Additionally I scraped the GRL article page to get the date that the GRL article first appeared online (not when it was formally typeset and published). Here is an exampLE of a GRL article landing page:


Notice that the article was first published on 15 Dec 2016. However, if you click the ‘Full Publication History’ link, you find out that the article first appeared online 24 Nov 2016 — so potential Wikipedia editors could add a citation prior to the formal ‘publication date’ of the GRL article.

So now that I have that data, what does it look like? Out of 36,960 GRL articles, 921 appear in Wikipedia, some are even cited multiple times. Below is a plot with the number of GRL articles (y-axis) that appear in Wikipedia, tallied by the number of times they are cited in Wikipedia — note the log y-axis.


GRL articles are spread over a range of Wikipedia pages, but some Wikipedia pages have many references to GRL articles (note the log scale of the y-axis):


553 Wikipedia Articles have a reference to only a single GRL article, while some articles contain many GRL references. Take for instance the ‘Cluster II (spacecraft)‘ page, with 25 GRL citations, or ‘El Niño‘ with 11 GRL references).

I’ll dive into data I collected over the next few weeks in a series of blog posts, but I want to leave you with some caveats about the code and the data so far. (Edited after the initial posting) only shows the data for up to 5 Wikipedia mentions for a given journal articles unless you have paid (instituitonal) access. Several GRL articles were cited in >5 Wikipedia articles, so I manually added the missing data. Hopefully i will make a programmatic work-around sometime. After I wrote this post, I was informed that the commerical Altmetric API gives all of the Wikipedia data (edit, editor, date). To get a commerical API key through Altmetric’s ‘researcher data access program’ (free access for those doing research), email Stacy Konkiel at Altmetric (thanks Stacy!).

Furthermore, many of the edit times that you see here could be re-edits, therefore ‘overprinting’ the date and editor for the first appearance of the wikipedia citation. This will be the subject of a future post, though I haven’t yet found an easy way to get the original edit…

Retaliation in the Peer Review Model

(The full motivation, rule set, and previous results for this model are collected here)

Today I am adding a new rule into my toy peer review model. Suppose some fraction of the population is set to ‘retaliate’ when called upon to be a reviewer. Specifically, these agents assign review scores based on their feelings (postive or negative) toward the author. This is an example of biases that might influence a reviewers decision (e.g., Lee et al., 2012).

So the new rule is:

  • If a ‘retaliator’ is assigned a review, and they feel positively or negatively toward the author, the review is postive or negative, respectively (overriding a random review).

(n.b.: A more gentle statement of this review could instead focus solely on ‘cliques’ — if a reviewer feels positively toward the author, the review is positive. if the review feels negatively, the review is random. )

The issue is now there are 4 types of peple in the model:

  • Those who sign reviews, and do not retaliate
  • Those who sign reviews, and retaliate
  • Those who do not sign reviews, and do not retaliate
  • Those who do not sign reviews, and retaliate

Again I will use the range of incoming and outgoing node weights to visualize model results. As a reminder:

R_{i}^{in} is the maximum incoming weight minus the minimum incoming weight. This represents the range of feelings all other scientists have about scientist i.

R_{i}^{out} is the maximum outgoing weight minus the minimum outgoing weight. This represents the range of feelings scientist i has about all other scientists in the discpline.

So here are the results with 30% of all scientists being ‘retaliators’.


  • Compared to the previous results, the same trends hold: Rin is larger for signed reviewers (blue), and Rout is roughly the same for signed vs unsigned. (ranges are different for the previous results because of a change in the number of model timesteps).
  • Unsigned retaliators (empty orange markers) are similar to Unsigned non-retaliators. If you never sign reviews, no author will end up knowing that you are a retaliator (the editor is a different story).
  • Signed retaliators (empty blue markers) have a large Rin — they are polarizing figures. Authors are either on the good side of these people (they are friends) or on the bad side (they are enemies).

Signed and Unsigned Reviews in Earth Surface Dynamics

All of the reviews for Earth Surface Dynamics are open, published, and citable. Today I do a bit of webscraping to determine the % of mix of signed and blind reviews for the 198 paper reviewed in EsurfD. Also, since reviews occur in sequence (i.e.,  R1 submits their review before R2), we can exame how R1’s decision to sign a review influences the decision of R2.

The code to do the webscraping is here. Note that R is not my best language, but I am using it because of all the cool packages written for R to interface with Crossref (rcrossref, for obtaining publication DOIs), and the easy webscraping (rvest).

The code works by:

  1. Pulling (from Crossref) details for all ESurf Discussion publications using the ISSN number.
  2. Going to every EsurfD page (following the DOI link)
  3. Scraping the webpage for author, editor, and reviewer comment (see this helpful tutorial on using rvest).
  4. Checking for descriptive words, for instance “Anonymous Comment #1”, to determine if Reviewer 1 and/or Reviewer 2 were anonymous.
  5. Check to see if a Reviewer 3 exists (to exclude the data… I only want to deal with papers with 2 reviewers for this initial study).

I imagine some specific pathological cases in review comments may have slipped through this code, but a cursory check shows it captures relevant information. After the code runs, I am left with 135 papers with 2 reviewers, for a total of 270 reviews. In total, 41% reviews are signed — this matches previous reports such as 40% reported by Okal (2003) and the 40% reported by PeerJ

  • Reviewer 1 totals are 74 unsigned, 61 signed —55% unsigned, 45% signed
  • For the 74 papers where Review 1 is unsigned,
    • Reviewer 2 data is 59 unsigned, 15 signed — 80% unsigned, 20% signed
  • For the 61 papers where Review 1 is signed,
    • Reviewer 2 data is 27 unsigned,  34 signed — 44% unsigned, 54% signed.

There is one clear confounding factor here, which is how positive/negative reviews impact the likelyhood to sign a review (both for R1 and R2). I imagine referee suggestions to the editor (e.g., minor revisions, major revisions, reject) and/or text mining could provide some details. (I can think of a few other confounds beyond this one)…. Furthermore, I would assume that since many (all?) journals from Copernicus/EGU have open review, this analysis could be scaled…

Peering into the Nature Geoscience author dataset

A recent Nature Geoscience editorial looked at the reviewer suggestions of submitting authors.  The editorial examined many different issues, including:

  • The geographic breakdown of submitting authors.
  • The geographic breakdown of author-suggested reviewers.
  • The geographic and gender breakdown for submitting authors whose paper was sent for review.
  • The gender breakdown of suggested reviewers by submitting author gender.

Fortunately, the data behind the editorial was also provided as a supplement. So let’s take a peek and investigate some other aspects of the data. First, let’s check out the gender breakdown of submitting authors by geographic region


For reference, ‘f’ is female, ‘m’ is male, and ‘u’ is unknown. The disproportion is clear accross all regions (note that Australia and NZ seem to be least disproportionate).

Next, let’s check out the geography of suggested reviewers by submitting author geography. Here is the number of authors who suggested reviewers, by geography:F3.jpeg

Now from this set of authors, the proportion of suggested reviewers broken down by geography:


One major trend I see, aside from the lack of balance across all recommendations, is that North American authors recommend North American reviewers most of the time (~65%). No other geographic location recommends itself as much (see even the European + Middle East authors, who recommend European + Middle East reviewers equally with North Americans)

I can think of data that is missing from this dataset —  in particular, the breakdown of assigned reviewers by geography. However the editorial alludes to some answers:

“Nevertheless, the geographical distribution of editor-assigned reviewers resembles the biases of author-suggested reviewers”

The R code for my analysis is here — this post was a good excuse to continue learning R (keep that in mind that I am learning R as you look at the messy, verbose code).


Peer review model — results part 2

Backstory of the model:

This model is based on networks, so I’ll use some of the language and techniques from the study of networks to analyze the data.This peer review model creates a directed and weighted network. In other words, the ‘scientists’ (nodes) are connected (via edges) to other scientists (other nodes). The connections (edges) have a direction (how ‘scientist A’ feels toward ‘B’) and weight (-3, negatively). The book-keeping for this model is an adjacency matrix.

A=\begin{pmatrix}0&5&-1&0\\ 1&0&2&0\\ 0&-3&0&0\\ 0&0&1&0\end{pmatrix}

Where A_{ij} denotes the an edge from i to j with a given weight. In this model, it is the mood that scientist i has toward scientist j . (Some other texts do the reverse convention).

A measurement for this sort of matrix is incoming and outgoing node strength. The outgoing strength of scientists i — how scientist i feels about all other scientists — can be denoted as:

s_{i}^{out}=\sum_{j \neq i} A_{ij}

And can be calculated by summing rows. The incoming strength of scientists i — how all other scientists feel about scientist i — can be denoted as:

s_{i}^{in}=\sum_{j \neq i} A_{ji}

And can be calculated by summing columns. (for reference, my previous post showed time series plots of the mean of incoming weights, similar to the strength metric we are talking about here (s_{i}^{in})).

Signed reviewers can be polarizing — weights can quickly become very negative and/or very positive. So the strengths (s_{i}^{in} and s_{i}^{out}) will be a sum of extreme positives and negatives — this is not very descriptive because it can lead to 0 strength. Instead I want to look at the range of incoming and outgoing weights, or:

R_{i}^{out}= \max\limits_{j \neq i} A_{ij} - \min\limits_{j \neq i} A_{ij} which denotes the maximum outgoing weight minus the minimum outgoing weight.

R_{i}^{in}=\max\limits_{j \neq i} A_{ji} - \min\limits_{j \neq i} A_{ji} which denotes the maximum incoming weight minus the minimum incoming weight.

Now let’s now look at some model results, R_{i}^{out} and R_{i}^{in}, for each scientist.


  • Both types of reviewers have similar R_{i}^{out} — they tend to have a similar range in their opinions about the scientists in the discipline.
  • Signed reviewers tend to have a larger R_{i}^{in} —  the range of feelings that other scientists have toward the signed reviewers — compared to those who do not sign reviews. Scientists tend to either like or dislike signed reviewers more strongly that unsigned reviewers. 

An added feedback is coming….

Some inspiration for this work comes from: