knitr::opts_chunk$set( collapse = TRUE, comment = "#>" )
Data from JSTOR/DfR is unlike most other data you encounter when doing text
analysis. First and foremost, the data about articles and books come from a
wide variety of journals and publishers. The level of detail and certain formats
vary because of this. jstor
tries to deal with this situation with two
strategies:
An example for the first case are references. Four different ways how
references can be specified are known at this time, and all are imported in
specific ways to deal this variation. There might however be other formats,
which should lead to an informative error when trying to import them via
jst_get_references()
.
An example for the latter case are page numbers. Most of the time, the entries
for page numbers are simply 42
, or 61
. This is as expected, and could be
parsed as integers. Sometimes, there are characters present, like M77
. This
would pose no problem either, we could simply extract all digits via regex and
parse as character. Unfortunately, sometimes the page is specified like this:
v75i2p84
. Extracting all digits would result in 75284
, which is wrong by
a long shot. Since there might be other ways of specifying pages, jstor
does
not attempt to parse the pages to integers when importing. However, it
offers a set of convenience functions which deal with a few common cases
(see jst_augment()
and below).
There are many other problems or peculiarities like this. This vignette tries to list as many as possible, and offer solutions for dealing with them. Unfortunately I have neither the time nor the interest to wade through all the data which you could get from DfR in order to find all possible quirks. The following list is thus inevitably incomplete. If you encounter new quirks/peculiarities, it would be greatly appreciated if you sent me an email, or opened an issue at GitHub. I will then include your findings in future version of this vignette, so this vignette can be a starting point for everybody who conducts text analysis with data from JSTOR/DfR.
After importing data via jst_get_article()
, there are at least two tasks you
might typically want to undertake:
There are four functions which help you to streamline this process:
jst_clean_page()
attempts to turn a character vector with pages into an integer
vector.jst_add_total_pages()
adds a column with the total number of pages per
article.jst_unify_journal_id()
merges different identifiers for journals into one.jst_augment()
wraps the above functions for convenience.In the following sections, known issues with data from DfR are described in greater detail.
Page numbers are a mess. Besides the issues mentioned above, page numbers might
sometimes be specified as "pp. 1234-83" as in
this article from the American Journal of
Sociology.
Of course, this results in first_page = 1234
and last_page = 83
, and the
computed number of total pages from jst_get_total_pages()
will be
negative. There is currently no general solution for this issue.
As outlined above, page numbers come in very different forms. Besides this
problem, there is actually another issue. Imagine you would like to quantify
the lengths of articles. Obviously you will need information on the first
and the last page of the articles. Furthermore, the pages need to be
parsed properly: you will run into troubles if you calculate page numbers like
75284 - 42 + 1
, in case the number was parsed badly. jst_clean_page()
tries
to do this properly, based on a few known possibilities:
Parsing correctly is unfortunately not enough. Things like "Errata" might come
to haunt you. For example there might be an article with first_page = 42
and
last_page = 362
, which
would leave you puzzled as to if this can be true^[Although it sounds absurd,
this can actually be true. There are some articles which are 200 pages long.
Obviously, they are not standard research articles. You will need to decide
if they fall into your sample or not.]. There could be a simple explanation:
the article might start on page 42, and end on page 65, and there is furthermore
an erratum on page 362. Technically, last_page = 362
is true then, but it
will cause problems for calculating the total number of pages. Quite often,
there is information in another column which could resolve this: page_range
,
which in this case would look like 42 - 65, 362
.
A small helper to deal with those situations is jst_get_total_pages()
. It
works for page ranges, but also for first and last pages:
library(jstor) library(dplyr) input <- tibble::tribble( ~first_page, ~last_page, ~page_range, NA_real_, NA_real_, NA_character_, 1, 10, "1 - 10", 1, 10, NA_character_, 1, NA_real_, NA_character_, 1, NA_real_, "1-10", NA_real_, NA_real_, "1, 5-10", NA_real_, NA_real_, "1-4, 5-10", NA_real_, NA_real_, "1-4, C5-C10" ) input %>% mutate(n_pages = jst_get_total_pages(first_page, last_page, page_range))
This is actually identical to using jst_add_total_pages()
:
input %>% jst_add_total_pages()
Identifiers for the journal usually appear in three columns:
journal_doi
journal_jcode
journal_pub_id
sample_article <- jst_get_article(jst_example("article_with_references.xml")) knitr::kable(sample_article)
From my samples, it seems that the information in journal_pub_id
is
often missing, as is journal_doi. The most important identifier is thus
journal_jcode
. In cases where both journal_jcode
and journal_pub_id
are
present, at least in my samples, the format of journal_jcode
was different.
For consistency, jst_unify_journal_id()
thus takes content of
journal_pub_id
if it is present, and that of journal_jcode
otherwise.
With this algorithm, it should be possible to reliably match them to
general information about the respective journals, which are available from
jst_get_journal_overview()
:
sample_article %>% jst_unify_journal_id() %>% left_join(jst_get_journal_overview()) %>% tidyr::gather(variable, value) %>% knitr::kable()
|Source |time span |Part | |:---------------------------------|-----------------:|-----------------:| |American Journal of Sociology |Unknown |Book Reviews |
For the AJS, ngrams for book reviews are calculated per issue. There are numerous reviews per issue, and each of them has an identical file of ngrams, containing ngrams for all book reviews of this issue.
A possible strategy for dealing with this is either not to use those ngrams, since they are calculated on all reviews in the issue, irrespective of whether actually all reviews of the given issue are in the sample or not. Alternatively, one could group by issues, and only take one set of ngrams per issue.
Information on langues is not consistent. For the sample article, language
is
eng
.
sample_article %>% pull(language)
In other cases it might be en
. It is thus advisable to take a quick look at
different variants via distinct(meta_data, language)
or
count(meta_data, language)
.
When analysing data about references and footnotes, you will encounter many inconsistencies and errors. Most of them are not due to errors from DfR, but stem simply from the fact, that humans make mistakes when creating manuscripts, and not all errors with references are caught before printing.
A common problem are names with non-english characters like german umlauts (Ferdinand Tönnies) or nordic names (Gøsta Esping-Andersen). These will appear in many different variations: Tonnies, Tönnies, Gosta, Gösta, etc.
For older articles, you might encounter issues that stem from digitising text
with OCR-software. A common problem is distinguishing I
from l
, like in
the phrase "In love". Depending on which names appear in your data, this might
lead to inconsistencies.
There are many examples where authors make mistakes and your summary statistics end up being skewed. This article about "Ethics Education in the Workplace" cites the same items multiple times, which is possibly an artifact. The advantage of using JSTOR/DfR data is, that you can inspect all sources and check, if a specific pattern you see is an artifact or genuine.
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.