POSC 8410: Introduction

Will Horne

Who am I?

  • I am a Political Scientist

    • Comparativist: polarization, representation, democracy

    • Methodologist: Text-as-Data, Machine Learning, Causal Inference

  • My third year at Clemson

    • PhD Princeton 2022
    • GSU: Post Doc 2022-2024

  • You can call me Will or Dr. Horne, either is fine!

How did I get here?

Enough About me!

Now let’s talk about you for a minute

What does social science research look like?

Three Kinds of Questions

  • Descriptive: What is happening?

    • How polarized is the American electorate? Has it changed since 1980?

  • Causal: What happens to Y if we change X?

    • Does registering voters by mail increase turnout?

  • Predictive: Given what we know, what comes next?

    • Which countries are most at risk of democratic backsliding?

These need different research designs. A lot of bad research is a design built for one question answering a different one.

From Concept to Evidence

  • The things we care about are abstract: discrimination, polarization, democracy, state capacity

  • The things we can observe are concrete: survey answers, votes, callbacks, budget lines

  • Research is the work of connecting the two

    • Concept → theory → testable implication → measure → data

Measurement is a choice, and choices can be wrong. “How democratic is Hungary?” has no reading on a dial.

Compared to What?

  • Almost every interesting claim is a comparison claim

    • “The program worked” = recipients did better than they would have otherwise

  • The problem: we never observe otherwise

    • We see the person who got the job training, not that same person in the world where they didn’t

  • So we substitute a comparison group — everything depends on whether it is a fair one

People who sign up for job training differ from people who don’t in a hundred ways. Which difference produced the result?

What Makes Research Credible?

  1. A clear question — descriptive, causal, or predictive

  2. Valid measurement — your variable captures the concept you claim

  3. A credible comparison — you can defend the counterfactual

  4. Honest uncertainty — you report how wrong you might be

  5. Transparency — someone else could redo it

A Worked Example

Do Employers Racially Discriminate?

  • A genuinely important policy question

    • Hiring discrimination is illegal. Is it happening anyway?

  • But notice how hard it is to study

    • You cannot ask employers — nobody self-reports discriminating

    • You cannot just compare real applicants: people from different races also differ in schools, neighborhoods, networks, experience

This is the “compared to what?” problem!

The Design

  • Bertrand and Mullainathan (2004) mailed ~5,000 fictitious résumés to help-wanted ads in Boston and Chicago

  • The résumés were identical except for the name at the top

    • “Emily Walsh” / “Greg Baker”

    • “Lakisha Washington” / “Jamal Jones”

  • Names were randomly assigned to résumés

  • Outcome: did the résumé get a callback?

Randomization manufactures the counterfactual. The only systematic difference between the two piles is the name.

The Result

A gap of about 3 percentage points, or roughly 50% more callbacks for an identical résumé.

Does a Better Résumé Help?

The Returns to a Better Résumé

  • White-sounding names: a stronger résumé buys +2.3 points of callback

  • Black-sounding names: a stronger résumé buys +0.5 points

  • The gap does widens as you improve your credentials

This kind of “the effect of X depends on Z” claim has a name. We will look at interactions at the end of the semester

Is this credible?

  1. Clear question? Yes - causal, and stated as such

  2. Valid measurement? Callbacks, not hires. A real limitation, but defensible

  3. Credible comparison? Yes - randomization does the work

  4. Honest uncertainty? Yes - we will get to how to measure this!

  5. Transparent? Yes - the data is public.

We will learn to do this kind of work!

  • Is 3 points a big gap? → Descriptive statistics, measurement

  • Could we have gotten this by chance? → Probability

  • What would we expect in a different sample? → Sampling and estimation

  • How confident should we be? → Hypothesis testing, confidence intervals

  • Does the gap differ by industry? By résumé quality? → ANOVA and regression

We will come back to this study all semester. You will analyze this data yourself.

Goals of the course

The goals of this course are:

  1. Introduce the foundations of Quantitative Social Science (QSS)

    • Descriptive Statistics

    • Probability Theory

    • Statistical inference

    • Regression analysis -> our final destination

  2. Link QSS foundations to policy relevant questions

  3. Build enough R fluency to do your own analyses

Course Sequence

  • In practice, this is the first course in a sequence

    • We will cover roughly intro level stats –> OLS Regression. Roughly 2 semesters of (rigorous) UG stats.

    • Spring: advanced regression models + design based techniques for causal inference (DiD, Experiments, etc)

  • Tentatively: Quant 3 (POSC 8430) will cover machine learning, text-as-data and some other advanced topics.

Why Quantitative Social Science?

  • Not Quant vs Qual

    • Quant and qual

    • Mixed methods research often stronger than pure qual or pure quant

      • I’ve done archival work + Other courses available specifically devoted to qualitative research methods
    • Commonalities in thinking hard about good research design
      • King, Keohane and Verba (1994) Designing Social Inquiry is the classic on this point

  • New techniques –> wider range of research questions

Motivation

Motivation

  • It is not that “social scientist” is a booming job title

    • Political scientists: -3%. Survey researchers: -5%

  • It is that the methods are in demand, everywhere

    • Data scientists: +34%, median pay $112,590 (May 2024)

  • The valuable combination is substantive knowledge + the ability to work with data

    • Plenty of people can run a regression. Far fewer know which regression to run, or what it means.

That combination is what this course is for.

Course Outline

  • Weeks 1-2: Getting up and running in R; describing data

  • Weeks 3-6: Probability theory

  • Weeks 7-8: Sampling, estimation, and inference

  • Midterm

  • Weeks 9-10: Asymptotics, hypothesis testing, confidence intervals

  • Weeks 11-14: ANOVA and regression

This Will Require Some Math

  • Calculus (derivatives, integrals, limits) + very basic matrix algebra

  • We will review! Not a bad idea to check out Khan Academy or similar if rusty/new

  • The math is not there for its own sake

    • You cannot interpret a regression coefficient you don’t understand

    • You cannot tell when a method is the wrong tool if you only know how to call it

  • Grad school –> Learning is your responsibility, be proactive.

Expectations

  • This course will be hard but…

    • Don’t stress, grad school is not about grades!

      • Goal of this course is to give you tools for your future work

  • If you are lost….stop me and ask questions

    • Office hours: By appointment (online or in-person). Please utilize!

    • Please read! Tons of online resources for both statistics and coding. If you don’t like the readings, feel free to supplement w/ something else.

Course Books

Assignments and Grading

  • 25% Midterm Exam

  • 25% Final Exam

  • 15% Problem sets

    • De-emphasizing P-set grades because of gen AI.

  • 35%: Final Project (Research Proposal with Analyses)

Getting Started

  • This course has two main aims

    • Teaching the fundamentals of probability + statistics for social scientists

    • Teaching you to actually do data analysis

  • We move fast on R: our next two meetings get you to the point where you can load, clean, and plot real data

    • There are tons of tools out there to learn R.

A Note on How I Teach R

  • I am not going to lecture at you about syntax for three weeks

    • That is the slowest possible way to learn a language
    • We don’t have time for it, anyways!

  • Instead: two intensive sessions, then you learn by using it

    • Every problem set is an R assignment

    • Readings from R4DS are meant to be worked through

  • If you get stuck, that is normal and expected. Come to office hours.

On Generative AI and Coding

  • You may use it.

  • I use it.

  • But: if you cannot read and explain the code, it is worthless to you

    • You will be asked to explain your code

    • The exams are in person/on paper

  • Use it the way you’d use a very fast, very confident, somewhat sloppy colleague

Why R?

  • Open Source and Free (unlike STATA, SPSS, SAS, etc)

  • Widely used in academic and government research

  • Specifically developed for statistical analysis (unlike Python)

  • Has friendly IDEs

How do I get R?

  • R is maintained by CRAN (Comprehensive R Archive Network).

  • Download here: https://cran.r-project.org/

    • A good idea to keep it relatively up-to-date, but you shouldn’t have to update for this course!

  • You will rarely, if ever, actually open the R that you download from here, but it must be on your machine for anything else to work!

What is an IDE?

  • R by itself is just a language. An IDE (“Integrated Development Environment”) is the program you actually sit in front of

    • Write and save scripts, run code, inspect your data, view plots, manage packages

  • For R, the two main options are R Studio and Positron

    • Both are made by Posit, both free, both open source

  • We will use Positron

Why Positron?

  • It is polyglot: R and Python in the same tool, sometimes the same project

    • You will likely want Python eventually. This way you don’t relearn your editor later.

  • It is where Posit’s development effort is going

    • R Studio still works and is still supported, but Positron is the forward direction

If you already use R Studio and love it, you may keep using it. Everything in this course runs identically in both. But my screen will show Positron.

How to get Positron

Download Positron here:

https://positron.posit.co/download.html

  • Free, and installs like any normal application (Mac, Windows, Linux)

  • Install R first, then Positron — Positron detects the R installations on your machine at startup

  • Detailed install instructions, if you want them: https://positron.posit.co/install.html

Before Next Class

  • Required: install R, then install Positron, and confirm Positron opens and finds R

    • We start working with data immediately. If your setup is broken you will fall behind on day one.

    • If you hit a wall, email me before class.

  • Read: Wickham, R for Data Science, Introduction and Chapter 2

  • Read: Ismay et al, Statistical Inference via Data Science, Chapter 1