Building your network using ORCiD and ROpenSci

Our neotoma package is part of the ROpenSci network of packages.  Wrangling data structures and learning some of the tricks we’ve implemented wouldn’t have been possible without help from them throughout the coding process.  Recently Scott Chamberlain posted some code for an R package to interface with ORCiD, the rORCiD package.

To digress for a second, the neotoma package started out as rNeotoma, but I ditched the ‘r’ because, well, just because.  I’ve been second guessing myself ever since, especially as it became more and more apparent that, in writing proposals and talking about the package and the database I’ve basically created a muddle.  Who knows, maybe we’ll go back to rNeotoma when we push up to CRAN.  Point being, stick an R in it so that you don’t have to keep clarifying the differences.

So, back on point.  A little while ago I posted a network diagram culled from my cv using a bibtex parser in R (the bibtex package by Roman François).  That’s kind of fun – obviously worth blogging about – and I stuck a newer version into a job application, but I’ve really been curious about what it would look like if I went out to the second order, what does it look like when we combine my publication network with the networks of my collaborators.

Figure 1. A second order co-author network generated using R and ORCiD’s public API.  Because we’re using the API we can keep re-running this code over and over again and it will fill in as more people sign up to get ORCiDs.

Enter ORCiD.  For those of you not familiar, ORCiD provides a unique identity code to an individual researcher.  The researcher can then identify all the research products they may have published and link these to their ID.  It’s effectively a DOI for the individual.  Sign up and you are part of the Internet of Things.  In a lot of ways this is very exciting.  The extent to which the ORCiDs can be linked to other objects will be the real test for their staying power.  And even there, it’s not so much whether the IDs can be linked, they’re unique identifiers so they’re easy to use, it’s whether other projects, institutions and data repositories will create a space for ORCiDs so that the can be linked across a web of research products.

Given the number of times I’ve been asked to add an ORCiD to an online profile or account it seems like people are prepared to invest in ORCiD for the long haul, which is exciting, and provides new opportunities for data analysis and for building research networks.

So, lets see what we can do with ORCiD and Scott’s rorcid package. This code is all available in a GitHub repository so you can modify it, fork, push or pull as you like:

The idea is to start with a single ORCiD, mine in this case (0000-0002-2700-4605).  With the ORCiD we then discover all of the research products associated with the ID.  Each research product with a DOI can be linked back to each of the ORCiDs registered for coauthors using the ORCiD API.  It is possible to find all co-authors by parsing some of the bibtex files associated with the structured data, but for this exercise I’m just going to stick with co-authors with ORCiDs.

So, for each published article we get the DOI, find all co-authors on each work who has an ORCiD, and then track down each of their publications and co-authors.  If you’re interested you can go further down the wormhole by coding this as a recursive function.  I thought about it but since this was basically a lark I figured I’d think about it later, or leave it up to someone to add to the existing repository (feel free to fork & modify).

In the end I coded this all up and plotted using the igraph package (I used network for my last graph, but wanted to try out igraph because it’s got some fun interactive tools:

library(devtools)
install_github('ropensci/rorcid')

You need devtools to be able to install the rOrcid package from the rOpenSci GitHub repository

library(rorcid)
library(igraph)

# The idea is to go into a user and get all their papers, 
# and all the papers of people they've published with:

simon.record <- orcid_id(orcid = '0000-0002-2700-4605', 
                         profile="works")

This gives us an ‘orcid’ object, returned using the ORCiD Public API. Once we ave the object we can go in and pull out all the DOIs for each of my research products that are registered with ORCID.

get_doi <- function(x){
  #  This pulls the DOIs out of the ORCiD record:
  list.x <- x$'work-external-identifiers.work-external-identifier'
  
  #  We have to catch a few objects with NULL DOI information:
  do.call(rbind.data.frame,lapply(list.x, function(x){
      if(length(x) == 0 | (!'DOI' %in% x[,1])){
        data.frame(value=NA)
      } else{
        data.frame(value = x[which(x[,1] %in% 'DOI'),2])
      }
    }))
}

get_papers <- function(x){
  all.papers <- x[[1]]$works # this is where the papers are.
  papers <- data.frame(title = all.papers$'work-title.title.value',
                       doi   = get_doi(all.papers))
  
  paper.doi <- lapply(1:nrow(papers), function(x){
    if(!is.na(papers[x,2]))return(orcid_doi(dois = papers[x,2], fuzzy = FALSE))
    # sometimes there's no DOI
    # if that's the case then just return NA:
    return(NA)
  })

  your.papers <- lapply(1:length(paper.doi), function(x){
      if(is.na(paper.doi[[x]])){
        data.frame(doi=NA, orcid=NA, name=NA)
      } else {
        data.frame(doi = papers[x,2],
                   orcid = paper.doi[[x]][[1]]$data$'orcid-identifier.path',
                   name = paste(paper.doi[[x]][[1]]$data$'personal-details.given-names.value',
                                paper.doi[[x]][[1]]$data$'personal-details.family-name.value', 
                                sep = ' '),
                   stringsAsFactors = FALSE)
      }})
  do.call(rbind.data.frame, your.papers)
  
}

So now we’ve got the functions, we’re going to get all my papers, make a list of the unique ORCIDs of my colleagues and then get all of their papers using the same ‘get_papers’ function. It’s a bit sloppy I think, but I wanted to try to avoid duplicate calls to the API since my internet connection was kind of crummy.

simons <- get_papers(simon.record)

unique.orcids <- unique(simons$orcid)

all.colleagues <- list()

for(i in 1:length(unique.orcids)){
  all.colleagues[[i]] <- get_papers(orcid_id(orcid = unique.orcids[i], profile="works"))
}

So now we’ve got a list with a data.frame for each author that has three columns, the DOI, the ORCID and their name. We want to reduce this to a single data.frame and then fill a square matrix (each row and column represents an author) where each row x column intersection represents co-authorship.

all.df <- do.call(rbind.data.frame, all.colleagues)
all.df <- na.omit(all.df[!duplicated(all.df),])

all.pairs <- matrix(ncol = length(unique(all.df$name)),
                    nrow = length(unique(all.df$name)),
                    dimnames = list(unique(all.df$name),unique(all.df$name)), 0)

unique.dois <- unique(as.character(all.df$doi))

for(i in 1:length(unique.dois)){
  doi <- unique.dois[i]
  
  all.pairs[all.df$name[all.df$doi %in% doi],all.df$name[all.df$doi %in% doi]] <- 
    all.pairs[all.df$name[all.df$doi %in% doi],all.df$name[all.df$doi %in% doi]] + 1

}

all.pairs <- all.pairs[rowSums(all.pairs)>0, colSums(all.pairs)>0]

diag(all.pairs) <- 0

Again, probably some lazy coding in the ‘for’ loop, but the point is that each row and column has a dimname representing each author, so row 1 is ‘Simon Goring’ and column 1 is also ‘Simon Goring’. All we’re doing is incrementing the value for the cell that intersects co-authors, where names are pulled from all individuals associated with each unique DOI. We end by plotting the whole thing out:

author.adj <- graph.adjacency(all.pairs, mode = 'undirected', weighted = TRUE)
#  Plot so that the width of the lines connecting the nodes reflects the
#  number of papers co-authored by both individuals.
#  This is Figure 1 of this blog post.
plot(author.adj, vertex.label.cex = 0.8, edge.width = E(author.adj)$weight)

Recovering dark data, the other side of uncited papers.

A little while ago, on Dynamic Ecology, a question was posed about how much self-promotion was okay, and what kinds of self promotion were acceptable.  The results were interesting, as was the discussion in the comments.  Two weeks ago I also noticed a post by Jeff Ollerton (at the University of Northampton, HT Terry McGlynn at Small Pond Science) who also weighed in on his blog, presenting a table showing that up to 40% of papers in the Biological Sciences remain uncited within the first four years since publication, with higher rates in Business, the Social Sciences and the Humanities.  The post itself is written more for the post-grad who is keen on getting their papers cited, but it presents the opportunity to introduce an exciting solution to the secondary issue: What happens to data after publication?

In 1998 the Neotoma Paleoecological Database published an ‘Unacquired Sites Inventory‘.  These were paleoecological sites (sedimentary pollen records, representing vegetation change over centuries or millennia) for which published records existed, but that had not been entered into the Neotoma Paleoecological Database or the North American Pollen Database.  Even accounting for the fact that the inventory represents a snapshot that ends in 1998, it still contains sites that are, on average, older than sites contained within the Neotoma Database itself (see this post by yours truly).  It would be interesting to see the citation patterns of sites in the Unacquired Sites versus those in the Neotoma Database, but that’s a job for another time, and, maybe, a data rescue grant (hit me up if you’re interested!).

Figure 1. Dark data. There is likely some excellent data down this dark pathway, but it’s too spooky for me to want to access it, let’s just ignore it for now. Photo Credit: J. Illingworth.

Regardless, citation patterns are tied to data availability (Piwowar and Vision, 2013), but the converse is also likely to be true.  There is little motivation to make data available if a paper is never cited, particularly an older paper, and little motivation for the community to access or request that data if no one knows about the paper.  This is how data goes dark.  No one knows about the paper, no one knows about the data, and large synoptic analyses miss whole swaths of the literature. If the citation patterns cited by Jeff Ollerton hold up, it’s possible that we’re missing 30%+ of the available published data when we’re doing our analyses.  So it’s not only imperative that post-grads work to promote their work, and that funding agencies push PIs to provide sustainable data management plans, but we need to work to unearth that ‘dark data’ in a way that provides sufficient metadata to support secondary analysis.

Figure 2. PaleoDeepDive body size estimates generated from a publication corpus (gray bars) versus estimates directly assimilated and entered by humans. Results are not significantly different.

Enter Paleodeepdive (Peters et al., 2014).  PaleoDeepDive is a project that is part of the larger, EarthCube funded, GeoDeepDive, headed by Shanan Peters at the University of Wisconsin and built on the DeepDive platform. The system is trained to look for text, tables and figures in domain specific publications, extract data, build associations and recognize that there may be errors in the published data (e.g., misspelled scientific names).  The system then then assign scores to the data extracted indicating confidence levels in the associations, which can act as a check on the data validity, and helps in building further relations as new data is accquired.  Paleodeepdive was used to comb paleobiology journals to pull out occurrence data and morphological characteristics for the Paleobiology Database.  In this way PaleoDeepDive brings uncited data back out of the dark and pushes it into searchable databases.

These kinds of systems are potentially transformative for the sciences. “What happens to your work once it is published” is transformed into a two part question: how is the paper cited, and how is the data used. More and more scientists are using public data repositories, although that’s not neccessarily the case as Caetano and Aisenberg (2014) show for animal behaviour studies, and fragmented use of data repositories (supplemental material vs. university archives vs. community lead data repositories) means that data may still lie undiscovered.  At the same time, the barriers to older data sets are being lowered by projects like PaleoDeepDive that are able to search disciplinary archives and collate the data into a single data storage location, in this case the Paleobiology Database. The problem still remains, how is the data cited?

We’ve run into the problem of citations with publications, not just with data but with R packages as well.  Artificial reference limits in some journals preclude full citations, pushing them into web-only appendices, that aren’t tracked by the dominant scholarly search engines.  That, of course, is a discussion for another time.

Publication metrics and interdisciplinary research.

van Dijk and others have just published an interesting paper in Current Biology “Publication metrics and success on the academic job market”. The main point in their paper is that it’s important to publish, it’s important to publish lots, and that having a highly cited paper can overcome the disadvantage of not publishing in high impact journals.

The last sentence really caught my eye:

Our results suggest that currently, journal impact factor and academic pedigree are rewarded over the quality of publications, which may dis-incentivize rapid communication of findings, collaboration and interdisciplinary science.

Figure 1. If you want to be a PI you’d best spend more time writing, and less time watching birds. “A Girl Writing; The Pet Goldfinch” Henriette Browne (1870)

This tone echos what we said in Goring et al. (2014), where we pointed out that early career researchers may be disadvantaged in interdisciplinary research, both by the Matthew effect and because interdisciplinary research often results in lags to publication as disciplinary bridges need to be overcome. With more support for this argument it becomes clear that, either (1) committees need to take the costs of interdisciplinary research into account when evaluating candidates for hiring or tenure, or (2) they need to specify interdisciplinarity as a key criteria in hiring and reward it explicitly. Our metrics help balance the costs of interdisciplinarity against a number of research outcomes, but if these metrics aren’t evaluated then early career researchers are effectively penalized, as van Dijk et al. point out.

van Dijk et al. don’t cite Petersen et al. (2012), but it’s worth pointing out that people have considered what it takes to make it in academia, which makes the statement “This is the first study that quantifies what is predictive of an academic career in terms of becoming a principal investigator” a bit of a dodgy statement in my opinion. Petersen et al. only study Assistant Professors and Professors, but it’s similar enough in intent and results that a reference to the paper should be included (in my opinion).

Finally, I want to point out a couple of peculiarities about the data set and analysis used in this paper.

1. The paper assumes that the last author is a PI, and so “success” is measured once you get three last author publications. Weltzin and others (2006) have taken this issue on, and made some important contributions. Tscharntke and others (2007) make the point that the last author is not always a PI in some disciplines, and so the blanket application of this method may be problematic. Indeed, it is my understanding that all of the papers in the (open access) Macrosystems Ecology special edition of Frontiers in Ecology and the Environment are ordered by contribution. So maybe this is an assumption that is slowly but surely breaking down over time (and with good reason).
2. PubMed is not an exhaustive database. I have 19 publications on Google Scholar and only 2 on PubMed. I suspect that this is an issue tied largely to whether disciplinary journals are archived by PubMed, but even Shultz (2007) found that Google Scholar often returns a greater number of journal search results than equivalent searches on PubMed. If no effort is made to constrain results to a particular discipline (and it’s not clear to me that that is the case) then it is possible that the results van Dijk present might be compromised.

Part of the reason that van Dijk’s results seem to resonate (check out the paper’s AltMetrics) is that it tells us a lot about what we already know intuitively. Getting papers in good journals matters. Good journals help increase visibility, but even if you can’t get into a good journal, you can still score with a highly cited article. Then, publish. Publish or perish. Finally, and disappointingly, it also doesn’t hurt to be a man (although it had a surprisingly low correlation with success if I’m reading the supplemental material correctly).

So what do you think?