class: center, middle, inverse, title-slide # Non-traditional data sources in demography ### Monica Alexander <br> University of Toronto ### DemoPop Workshop, 29 April 2022 --- # Overview -- ### Motivation -- ### How do we get data? -- ### A whirlwind intro to text analysis -- ### Code tutorials --- class: center, middle, inverse # Motivation --- # Motivation ### What do we mean by 'non-traditional'? -- - In short, not censuses and surveys! -- - The nature in which data is collected and stored has changed dramatically. An increasing share of human interaction, communication and culture is being recorded in a digital way + Digital records = data that is (potentially) accessible and analyzable - 'Non-traditional' generally means not primarily collected for research purposes (but still has potential to be useful in this context) -- - Examples + **Social Media data** + Digital trace data + Administrative data + **Corpus data** --- # Social media data -- - With the rise of social media comes the rise of data -- - Data about people + characteristics, movements, connections, interests, views... -- - Potentially rich data sources for population research -- - Strengths that traditional data sources often don't have + timely + large sample sizes + granular information -- - But potential drawbacks as well --- # Types of social media data -- - Population data from advertising platforms + aggregate level population estimates by subgroup + useful in demographic estimation + e.g. [Estimating out-migration from Puerto Rico using Facebook ads data](https://onlinelibrary.wiley.com/doi/abs/10.1111/padr.12289) -- - Individual digital trace data + Individual profiles, friends and behavior + Text / media data, networks + Use to study networks, information spread, opinions, etc + e.g. [Assessing the Russian Internet Research Agency’s impact on the political attitudes and behaviors of American Twitter users in late 2017](https://www.pnas.org/content/pnas/117/1/243.full.pdf) + e.g. [Geography of Twitter networks](https://www.sciencedirect.com/science/article/pii/S0378873311000359?casa_token=_ihyPFLiN3wAAAAA:wlhmucP0-sdp_vwynl23QWEPWjNdj0w1lVAgzA8UjOQzcArTguH2BFt48MVXQc3Vkc-K3cui7Ws) -- - Survey data + Use social media platforms to reach and survey people + Timely, hard-to-reach populations, cheap? + e.g. [Behaviours and attitudes in response to the COVID-19 pandemic: insights from a cross-national Facebook survey](https://epjdatascience.springeropen.com/articles/10.1140/epjds/s13688-021-00270-1) --- # Corpus data -- - Large collections or spoken or written language - Treating text as data; nonreactive content analysis - Information encoded in text is a complement to more traditional forms of data -- - Can be text from social media data, other online sources (blogs, Wikipedia), newspapers, Hansard, bills, scholarly articles... - Examples: + ['Migration Discourse in Sweden: Frames and Sentiments in Mainstream and Social Media'](https://journals.sagepub.com/doi/full/10.1177/2056305120981059) + ['Neutral or Framed? A Sentiment Analysis of 2019 Abortion Laws'](https://link.springer.com/article/10.1007/s13178-022-00690-2) + ['A mixed-methods framework for analyzing text data: Integrating computational techniques with qualitative methods in demography'](https://www.jstor.org/stable/26332229?seq=3) --- # Example: *Demography* articles  Example research questions? --- class: center, middle, inverse # How do we get data? --- # How do we get data off the web -- ### Manually? - I guess you could copy paste info you view on a webpage - Slow, hard, not reproducible -- ### Webscraping - Write code that visit a webpage and extracts data - More automated and reproducible - Legality and ethical considerations depend on the specifics of what you're doing - Some good tips and principles are [here](https://www.tellingstorieswithdata.com/gather-data.html#scraping). - `rvest` is a super useful R package here --- # How do we get data -- ### Application Programming Interfaces (APIs) -- - A tool that makes it easier for a computer to query a website's data - Most social media websites have APIs, that let you programmatically query and extract data - e.g. [Facebook marketing API](https://developers.facebook.com/docs/marketing-apis/) - [Twitter API](https://developer.twitter.com/en/docs) - e.g. [arXiv API](https://arxiv.org/help/api/) -- ### R packages that help to query APIs -- - Many packages exist to help R users query APIs - We will be using `rtweet` - Others e.g. `rfacebookstat`, `Rlinkedin`, `scholar`... --- class: center, middle, inverse # A brief introduction to text analysis --- # Text as data -- - Inherently high dimensional. Sample of documents, `\(w\)` words long, `\(p\)` possible words: unique representation is `\(p^w\)` -- - We need to simplify (just like any other modeling process) -- - In general, represent raw text `\(\mathcal{D}\)` as a numerical array `\(\boldsymbol{C}\)` + The elements of `\(\boldsymbol{C}\)` are usually counts of *tokens*: words, phrases, etc. -- - We then use `\(\boldsymbol{C}\)` to understand characteristics of the underlying text + what words(/phrases) are common/important? + what is the underlying sentiment? + what topics are being talked/written about? --- # Tokens - Single words, i.e. take corpus and split into single words + simplest, most common + problem of assuming 'bag-of-words'; independence - *n*-grams e.g. bi-grams are two-word phrases + e.g. "cool climate wines" contains two bi-grams; "cool climate" and "climate wines" --- # Descriptive measures - Frequencies i.e. just counting the number of times a word/token appears in a document - tf-idf (term-frequency - inverse document frequency) + For a word or other feature `\(j\)` in document `\(i\)`, term frequency `\(tf_{ij}\)` is the count `\(c_{ij}\)` of occurrences of `\(j\)` in `\(i\)`. + Inverse document frequency `\(idf_j\)` is $$ \log(n/d_j) $$ where `\(d_j = \sum_{i} \mathbf{1}_{\left[c_{i j}>0\right]}\)` and `\(n\)` is the total number of documents. + tf-idf if the product of these two quantities --- # Sentiment analysis -- - Outcome of interest is latent sentiment of document -- - No estimation involved; just use a pre-specified dictionary that relates words to particular categories of sentiment -- - Aggregate sentiment values/ratings across documents to get an idea of the relative sentiment within a corpus -- - For example, is the underlying sentiment in new articles about immigrants associated with the political party in power? --- # Topic models Topic models assume that a document is a realization of a mixture of latent topics. The topics themselves are represented by a set of words that are selected from that topic. -- - We are trying to model a **generative process** -- - E.g. a politician first chooses the topics they want to speak about. After choosing the topics, the politician then chooses appropriate words to use for each of those topics. -- - E.g. a demographer decides the topics they want to write about; then chooses the appropriate words -- - Each document as a mixture of topics, and each topic as a mixture of words. - Statistically, topic models consider each document as having been generated by some probability distribution over topics. Similarly, each topic is considered a probability distribution over words/terms -- - There are many different topic models; probably the most common is Latent Dirichlet Allocation (LDA) --- class: center, middle, inverse # This workshop --- # Plan for the rest of the workshop -- - Run through two modules: Twitter data (API) and *Demography* (webscraping and text analysis) -- - I will share my screen and go through the code in the RMarkdown files -- - Questions / comments welcome -- - Probably too much content?? -- - All materials available, hopefully useful as a starting point for your own learning/research: https://mjalexander.github.io/demopop-workshop/ --- # Let's get started! ### Contact info <a href="mailto:monica.alexander@utoronto.ca"><i class="fa fa-paper-plane fa-fw"></i>&nbsp; monica.alexander@utoronto.ca</a><br> <a href="monicaalexander.com"><i class="fa fa-link fa-fw"></i>&nbsp; monicaalexander.com</a><br> <a href="http://twitter.com/monjalexander"><i class="fa fa-twitter fa-fw"></i>&nbsp; @monjalexander</a><br> <a href="http://github.com/MJAlexander"><i class="fa fa-github fa-fw"></i>&nbsp; @MJAlexander</a><br>