In our first meeting, we dive into data straightaway, conducting a class-wide survey experiment to measure how race and gender affect our reactions to #MeToo allegations. We analyze our responses in a spreadsheet, using randomization inference to determine the statistical significance of our results. Along the way, we learn about p-values, vignette experiments, and our own implicit biases, while discovering why writing code can save a lot of time compared to doing statistics in Google Sheets or Excel. In the second half of class, we go over course logistics and install Python 3 and Atom code editor on our laptops. Before we leave, we write our first piece of Python code, run it, and submit it as our first in-class assignment.

In our second meeting, we discuss how cell phone metadata can be used to predict outbreaks of violence in the aftermath of civil war, but why that's also double edged sword. We take a deep dive into vignette experiments, examining ongoing research into ethnic/religious bias and looking for clever ways to avoid respondents attempting to please the survey-taker. We also practice using variables & lists, conditional evaluation, and for loops to create a currency converter and a primitive caller id app.

With all the hype about big data in the past five years, have surveys and other traditional forms of “small data” become obsolete? We examine three papers that demonstrate how when we match up these two types of data sources we can learn new things that neither could teach us on its own. Applications include estimating regional patterns in wealth in developing countries, identifying friendships among cell phone users, and forecasting the spread of the flu (or so we thought…). We also learn about the pitfalls of “big data hubris” and the difficulties in basing scientific discoveries on search engines and apps whose algorithms are constantly changing. All of these studies raise questions about privacy, but they also open the door to new opportunities to help people live happier, healthier lives. How do we get the balance right? We get a taste of how this research is performed by searching for outliers in our own cell phone data, inferring a friendship network from call records, and identifying search terms that best predict pregnancy. To do so, we practice using for/while loops, if/else statements, and iterating through parallel and nested lists. New statistics concepts covered include outliers, percentiles, regression, variable selection, confidence intervals, false positives, ROC curves, and out-of-sample prediction.

Which racial groups are most likely to intermarry? Simply counting the number of intermarriages may not give us the best answer, even if that count comes from as reliable a source as the U.S. Census. Is intermarriage between African-Americans and Asian-American so rare because of racial preference/prejudice or because they tend to live in different communities? To answer these questions, we learn about weighted estimates and multiple regression. We also discuss the historical misuses of census data to prosecute draft-dodgers in WWI and intern Japanese citizens in WWII, as well as the risk of misuse today against immigrants and suspected terrorists. We consider who is most likely to be undercounted in a census and the political implications of including a citizenship question. Finally, we look at censuses in other countries and how breakdowns in data collection can themselves be a useful form of data for measuring state capacity. To reinforce our understanding of weighting and practice writing functions, we write a script that adjusts intermarriage rates to reflect each group’s share of the population.

Can data science help us understand history? Do archives and historical records contain data we might be able to use Python programming to analyze? In our first article, we consider how using geolocated data from Nazi records of Jewish residences in the occupied Netherlands can lend us insight into who is likely to protect potential victims during persecution and genocide. In our second article, we take a deep dive into the social lives of voters in two 19th century American towns to see how social networks influence voting. In our discussion, we consider the ethical tradeoffs of studying the dead without their consent. The Nazis collected this data in order to round up Jews for execution. Does that make it problematic for researches to use it? Is there a difference between using data about where their victims lived and using data collected from sadistic Nazi medical experiments? Meanwhile, the authors of the voting article used public records and archival data to find out not merely where voters lived but who they socialized with. Is it ethical to try to probe the personal lives of the deceased for research purposes? Does it matter if these are ordinary citizens and not public figures? Does it matter how much time has passed? Instead of a new in-class coding exercises, we practice commenting, unit-testing, and debugging three previous exercises.

Building on last week’s readings, we take a foray into the fast-growing field of network analysis and try our hand at some network exercises. We also discuss regression coefficients, standard errors, and the differences between description, prediction, and causal inference. In preparation for our semester-long project, we examine the differences between a topic, research question, and hypothesis and how to proceed from one to the next. Groups will have time to meet and refine their research questions with instructor feedback. Lastly, to reinforce our Python skills and better understand geospatial data analysis, we replicate Robert Braun’s approach to analyzing the role of minority churches in saving Jews during the Holocaust using a simplified, simulated data set. In doing so, we practice writing list comprehensions, indexing and slicing, and importing Python libraries.

In the past decade, Twitter has revolutionized the public sphere from political mass mobilization to middle school bullying to corporate marketing strategies. Yet the pace at which behavior on Twitter is changing and the evolution of the platform itself present challenges to social scientists. Will a dataset collected today still be relevant by the time its results are published? Can those findings be verified or will changes to the underlying software render replication impossible? Are we learning anything useful about humanity if the phenomena we study are barely recognizable five years hence? In addition to the issues of temporal validity described above, we also discuss the difference between experiments and large-scale observational studies, as well as what can be learned in both cases from a non-representative sample of the population. Are Twitter users a population worth studying in their own right? Do they have something to teach us about people in general? To what extent do behaviors on Twitter reflect human behavior offline? For our in-class Python exercise, we analyze a simulated Twitter experiment aimed at reducing middle school bullying to practice working with text as data, retrieving data stored in dictionaries, and reading data from files.

We begin by discussing the oft-overlooked skill of how to deal with feedback when presenting—namely, how to respond to critiques while simultaneously taking them to heart. Groups then present their project proposals, including research question, hypotheses, data sources, and analysis plan. Providing useful feedback to other groups is included in the grade. Time permitting, we do an object-oriented programming exercise related to last-week’s Twitter studies.

Just as we’ve seen historical and archival data repurposed, so too can we make use of data collected by governments and physical/natural scientists to answer social science questions. We first look the impact of rainfall on American elections and learn about natural and quasi-experiments. We contrast this framework with an observational study that uses images of the earth at night to monitor the Syrian refugee crisis. Returning to our theme of prediction versus causal inference, we discuss how each type of research design is best suited to different questions and where their weaknesses lie. We also discuss the problems with forecasting the effects of climate change on social problems. As we saw earlier in the semester when we looked at Twitter, it is difficult to use past or even present-day data when the phenomena we are studying are rapidly changing in ways we have never experienced. And yet to not study these pressing issues is unconscionable. How should social scientists go about studying an ever-changing world? Are there fundamental laws to be derived or only trends to observe and extrapolate? In preparation for our research projects, we briefly examine how a scholarly report aimed at practitioners (second reading) differs stylistically from articles aimed at other academics. Lastly, we see why hungry sharks did not, in all likelihood, get Woodrow Wilson elected, and why the world would be hopelessly unpredictable place if this finding, and a hundred others like it, were all true simultaneously. We visualize this finding using Mathplotlib in a Jupyter Notebook.

Surveys are ubiquitous in our society: university offices, stores, and apps are constantly asking us to evaluate them, while the media frequently report on public opinion and election polls. When it comes to interpreting the answers, we’re often so focused on what’s that we often neglect three equally fundamental questions: Who’s being asked? How are they being asked? Who is doing the asking? After enumerating the many biases that can arise in the survey process (selection bias, social desirability bias, recall & salience bias, etc.) we learn about at how those biases can best be averted, minimized, or corrected. Multiple regression, though far from a panacea, allows us to control for observed confounders, which is an important step toward improving our estimates. We practice working with survey data in Pandas data frames (imported with the csv module), examine how biases affect inference, and try out the multiple regression tools available in the scikit-learn and NumPy libraries.

Companies like Facebook have an unprecedented capacity to conduct social science research. For one thing, they are conducting market research on us every time we use their services. Is it ethical for them to also manipulate their users to address questions of scientific interest, given that they’re already doing so to maximize profits? Facebook faced a massive PR debacle when it reportedly made millions of its users ever-so-slightly less happy to answer questions about how emotions are contagious (first article). Yet the media hardly commented when, in the second article, it ever-so-slightly nudged millions of its users to go vote. Is there a difference? What should our ethical guidelines be here? There’s a deeper scientific problem when it comes to working with Facebook data, however: you have to work with Facebook to get it. Our third article, published a full year before the 2016 election, treats Facebook merely as a microscope through which we watch polarization and media bubbles. Yet if the authors had found that Facebook caused these negative trends, would the company have allowed them to publish their findings? How can other researchers hope to replicate these findings if the data and code are not publicly available? Ultimately, Facebook and other proprietary platforms are too important to ignore, both as tools and objects of study. We discuss new partnership models, such as the Social Science One partnership, that seek to create greater transparency and replicability, while still protecting the anonymity of users (and, presumably, Facebook’s trade secrets).

With enough sources of data, can we finally succeed in reliably predicting rare and dramatic events? Psychologist Philipp Tetlock and his collaborators believe the answer is yes, but by using people rather than machines. In our final regular class meeting, we discuss the successes of Tetlock’s so-called superforecasters, the reliability of betting markets, and the ethics of Amazon Mechanical Turk. In doing so, we contrast the wisdom of crowds experts, the wisdom of experts, and the wisdom of “crowds of experts.” We then turn to critiques by Kai Jäger (and many other scholars) of why fancy computer algorithms, even with human supervision, face fundamental limitations when it comes to predicting human activity. We go over various categories of machine learning—supervised, unsupervised, deep neural networks—and try to get a sense of why they may be highly reliable for some tasks but very bad at others. As we implement our own machine learning algorithm in Python, we consider the common critique that such algorithms present a “blackbox” which fails to shed light on causal mechanisms. In the end, the machines have learned a lot about us, but what have we managed to learn from them? We conclude by reviewing the differences between description, prediction, and causal inference.

Groups will present their Jupyter notebooks to the class as if soliciting feedback at a conference prior to submitting a journal article for publication. Students are expected to give and incorporate feedback. Final version of Jupyter notebook due 5 days after final class.