Computing Essentials for Social Scientists: Day One

Today

Introductions
Program Requirements
What is data science? What do we hope to get out of this?
Python vs R
What you need to know to succeed with data science

Let’s get to know one another

Jon Atwell

Michigan PhD, 2017 - Sociology
Studies the emergence of culture and norms
Expertise: Python, text methods, agent-based modeling, experiments

Chris Skovron

Michigan PhD, 2017 - Political Science
Studies public opinion and what people believe about it
Expertise: R, statistical modeling, causal inference

We expect from you

Sustained commitment to integrating data science methods in your research
Be a resource for your colleagues, both students and faculty
Be flexible, open to struggling and learning new things about yourself and your research

Program requirements

Three sessions this week
At least four additional sessions. Intro R and Python count as two sessions.
Optional but suggested to attend afternoon practice sessions for workshops.
Participate in future data science programming put on by Data Science Initiative

Program support

Jon and I will be mentors
We aren’t here to debug but will help you strategize and find resources
Research Computing Services is another source of support

What is data science?

- the analysis of data ¯\(ツ)/¯

Etymology

Naur, “Datalogy” (1974)
Was proposed that statistics rename itself to DS (1997)
Today’s use comes via computer science (Classification Society conf. 1996)

What does it mean in practice today?
Everything and Nothing, IOHO

What do LIDAR and Tweets have in common?
- Only data scientists!
Ok, sophisticated processing and analysis algorithms too

Ok, what does it mean in the academy?

New data sources and types
Non-regression-based quantitative empirical analytical approaches.

Why are we here?

Business got more serious about analytics

Related concepts and disciplines

Information Sciences
“Big data”
Machine learning
Computational Social Science

Northwestern Data Science Initiative definition

New analytic methods
New sources
Defined relative to what has historically been typical in disciplines

New data sources

Text, tweets
Experiments conducted online, in the field at scale
More technical data used to analyze social behavior: satellite lights

New methods

A plethora of new text methods
Machine learning methods
Fine-grained measurement (geolocations, etc)

New scales

Surveys are bigger and more frequent
People leave more traces of themselves online
We can observe social behavior using new technologies

Beware the hype

People have incentives to play up how fancy their data and methods are
Important to keep a critical eye to whether people are being honest about why they use the data they do and the method they do
Don’t become a data science supremacist
OLS is often just as good and orders of magnitude simpler than fancy methods!
It’s important to have a good bullshit detector when working in this area

Pitfalls for inference

As social scientists, we need to remember that using new data science tools doesn’t remove or obligations to do good research design
Bigger data or more computationally intensive approaches also present new challenges for transparency
Just because you have the full population doesn’t mean you can speak to causality, etc
Big data doesn’t always mean your findings are generalizable

Why CESS?

CESS students were chosen because

You already show creativity and promise in your research projects and ideas
Many of these projects could benefit from data science approaches
You don’t already know all the things we are going to teach

Much of what we are working on in CESS is not really taught anywhere in the social sciences

We focus on statistics in methods classes, which is great
But doing that at the scale required for data science requires organization and an understanding of how computing works
Most of us in social science don’t have extensive programming experience
This is tacit knowledge Chris and Jon learned the hard way. We want to help you make it through this thicket faster than we did

Principles for being successful in applying data science to social science

Get organized
Know the landscape of available tools and resources
Know what you need to know
Understand how to teach yourself

Working together is a key

CESS is a community of students facing similar challenges and opportunities
Working collaboratively on programming requires some vulnerability, but community is essential
You should work on getting comfortable with asking peers for help
You should share what you know and tricks you pick up with peers and with faculty

Python vs R

Where do you fit in DS?

or, What you need (and don’t need) to know to thrive using computational tools

or or, Relax and focus on what you need

Part 1: The professional angle

Academic research or industry?

You aren’t effectively preparing for the latter here
Transition is getting harder
Both can be great, or terrible

Part 2: The tools angle

Machine Learning

Breiman, 2001: “Two Cultures”

What generates data?

The problem of statistical inference

Statistics

Machine Learning

Supervised vs. Unsupervised

Supervised => You know values of dependent variable(s)
Unsupervised => You don’t

Specified model vs blackbox

Orthogonal to supervised vs. unsupervised

Supervised

High predictive power
more diversity of generating mechanisms the reg.
easily overfitted
can be blackbox

Unsupervised

Lack groundtruth
define probablistic model/generative process
harder to overfit
can be blackbox

Do you need supervised models?

Regression models, of course!
Classification? umm, probably not
- Yet, lots of hype here

Do you need unsupervised models?

text as data!
clustering, yes
- but not that new

Part 3: The good (academic) programmer angle

No one knows it all, so relax

Be comfortable in that
But don’t stop learning!

Wrong emphasis: Efficiency

a computer science measure of the work to be done as a function of the number of input population size. Whole subfield dedicated to study of complexity, but it boils down to whether a process can run in sublinear, linear, or exponential time. Consider a population of size N. Drawing a random element from the pop doesn’t depend on the pop size (sublinear). Taking the average does depend on N, because you have to visit each element. Finding the standard dev takes twice as long (calc mean, revisit each to calc error.), but it’s still linear because you’re just iterating over the list twice. But say one of your variables is the geocoordinates of the elements and you want to know the average distance between all elements in the population. You need to calculate the distance between all pairs of elements. In a population of size N, there are N(N-1) pairs. N(N-1) is very, very close to N^2. So in pop of 100, there are 10,000 operations to do to calc average distance. In a population of 1000, there are 1 mil operations. The number of observations increased 10 fold, but the number of operations increased 100 fold. That’s exponential growth! You want to avoid exponential grow as much as you can if you want to be efficient.

Why? (more on Day 2)
- clocktime vs human scale
- but complexity matters.

Wrong emphasis: Succinctness

Compactness often isn’t good
(See the translation process)

Correct emphasis: Scientific question

lead with questions, not methods

Correct emphasis: Reproducibility

organized so others can reproduce result
- but working code not the same as. ..

Correct emphasis: Replicability

logic of code is clear
- via comments and other documentation
- via properly expressive code