1 of 45

STAT 131A: Statistical Methods for Data Science

Instructor: Josh G

1

Come to the front to grab a worksheet + hi-chew + say hi!

Starting with Lecture 2, every lecture will start and end with an ungraded conceptual question + attendance check.

You may want to practice finding your seat

and filling out the form! → → →

shorturl.at/rt5m6

Q? pollev.com/jdgg

2 of 45

💻❌ Attendance + tech policy

Lecture attendance is required for 131A.�[ Starting with Lecture 2 ]

No laptops or tablets w/ attached keyboards are allowed during lecture, unless we are coding. Phones are allowed.�[ If you need to use a laptop for accessibility, that's OK! ]

See stat131a.berkeley.edu/fall-2024 for more details.

2

Q? pollev.com/jdgg

3 of 45

What will you be able to do after taking 131A? 🤷🏽

3

Q? pollev.com/jdgg

4 of 45

🗺️ Naviance

Naviance is an online, proprietary tool designed to guide college search and application decisions.

More than 40% of U.S. high school students have access.

Q? pollev.com/jdgg

5 of 45

📈 The scattergram

5

Q? pollev.com/jdgg

6 of 45

📈 The scattergram

6

You are here. Should you apply? Why or why not?�[ 🗣️Discuss with neighbor ]

shorturl.at/Hycut

Submit answer here!

Q? pollev.com/jdgg

7 of 45

📈 The scattergram

Suppose your ACT score is below the average score of past students who were admitted.

You may feel dissuaded from applying, even if you are academically qualified to attend.

7

Q? pollev.com/jdgg

8 of 45

🤔 Undermatching

Undermatching occurs when a student applies solely to colleges for which they are overqualified.

Extreme example: Perfect GPA + SAT, only community colleges

We find that showing past admissions outcomes [ as scattergrams ] may increase undermatching for strong students.

8

Q? pollev.com/jdgg

9 of 45

🔬 Methodology

We filed public records requests on Naviance adoption for 220 public high schools.

We also obtained college application data for 70,000 students from these high schools, spanning 2014–2020.

9

Q? pollev.com/jdgg

10 of 45

🏇 Pronounced effect on strong students

10

Margin of error (?)

ACT ≥ 29

Q? pollev.com/jdgg

11 of 45

📈 Naviance adoption in Florida

11

Q? pollev.com/jdgg

12 of 45

🏫 Aggregated results�In other words, not just Florida!

Access to Naviance appears to approximately double the odds of undermatching among high-achieving students.

�Result is robust to adjustment for potential confounders, such as test scores, GPA, gender, and first gen status.

12

?

Q? pollev.com/jdgg

13 of 45

📉 Key takeaways

1. Data visualization choices may have unintended behavioral consequences.

2. After taking 131A, you will have the tools to replicate everything presented so far, and a lot more!

13

Q? pollev.com/jdgg

14 of 45

14

Q? pollev.com/jdgg

15 of 45

🍎 STAT 131A Teaching Team

15

GSI: Van Hovenga

Instructor: Josh Grossman

Q? pollev.com/jdgg

16 of 45

🥅 Course goals

Learn core statistical concepts and, more generally, learn to reason with data.

Additional goals:�- Build practical statistical intuition�- Become scrappier + more independent.�- Learn R/tidyverse�- Prep for interviews, internships, and full-time jobs�

16

Q? pollev.com/jdgg

17 of 45

🥗 Tentative course outline

Week 1: Visualization�Week 2: Data distributions�Week 3: Probability�Week 4: Quantifying uncertainty�Week 5: Confidence intervals�Week 6: Hypothesis testing�Week 7: Midterm 1 (Weeks 1-5) + Linear regression�

17

Q? pollev.com/jdgg

18 of 45

🥗 Tentative course outline

Week 8: More linear regression�Week 9: Logistic regression�Week 10: Buffer�Week 11: Non-parametric methods�Week 12: Midterm 2 (Weeks 1-10) + non-parametric methods�Week 13: Decision trees and random forests�Week 14: Buffer�Week 15: More buffer�RRR: Extra help sessions.�Finals period: Final exam

18

Q? pollev.com/jdgg

19 of 45

🏢 Course logistics

~6 homework assignments [ 20% of grade ]�Due every other week, roughly.�5 slip days. Can use at most 2 slip days per assignment.

2 midterms + final exam [ 10%+15%+20%=45% of grade ]

Final project [ 15% of grade ]Tentatively, groups of 3. More details to come.

Labs [ 10% of grade ]

Lecture attendance + participation [ 10% of grade ]

19

Q? pollev.com/jdgg

20 of 45

🏢 Course logistics

In general, do not email us. Make a private post on Ed.

Course materials�stat131a.berkeley.edu/fall-2024 + Ed + bcourses

Office hours (OH) + experimental 15-min coffee chats�See website. No office hours during Week 1.��Lab sections�See website. No lab in Week 1. But, try Lab 0 on your own! Post to Ed with questions.

20

Q? pollev.com/jdgg

21 of 45

🤖 Large language model (LLM) policy

In 131A, you can only use the PingPong LLM. �[ Unless otherwise indicated. ]

Using any other LLM is considered cheating.

Invites to PingPong coming soon.��See syllabus for full LLM policy.

�Note: PingPong is experimental. Provide feedback! We may adjust the parameters + settings based on feedback.

21

Q? pollev.com/jdgg

22 of 45

📝 To do

1. Read the syllabus!!! Please!!! 😭🙏🏽�2. Student survey�3. Complete Lab 0��If you’d like help assembling a study group, please complete the form on the website by Monday at midnight.

If you have a Letter of Accommodation (LoA), please make a private Ed post ASAP.

This is all on the website: stat131a.berkeley.edu

22

Q? pollev.com/jdgg

23 of 45

23

Q? pollev.com/jdgg

24 of 45

Closing concept check

24

Starting with Lecture 2, every lecture will start and end with an ungraded conceptual question + attendance check.

For example, I may have asked "What is undermatching?"

You may want to practice finding your seat

and filling out the form! → → →

shorturl.at/rt5m6

Q? pollev.com/jdgg

25 of 45

⚠️ Assessing police discrimination

Assessing discrimination in policing is a critically important but also challenging topic.

It demonstrates both the power and limits of statistical reasoning.

Feel free to participate in the discussion — or take a break from it — to the extent that you are comfortable.

25

Q? pollev.com/jdgg

26 of 45

🛑 Stop and Frisk

Officers stop and question pedestrians when there is “reasonable suspicion” of criminal activity.

Until not long ago, 500,000 stops conducted annually in NYC�[ Substantially curtailed at the end of 2013 ]

26

Q? pollev.com/jdgg

27 of 45

🛑 Stop and Frisk

If officers suspect that a stopped pedestrian is armed or dangerous, they can conduct a frisk.�[ frisk = a brief pat down of outer clothing ]

27

Q? pollev.com/jdgg

28 of 45

🛑 Stop and Frisk

Fact: 80% of stops involved Black or Hispanic individuals.

Fact: 50% of NYC population is Black or Hispanic.

28

Q? pollev.com/jdgg

29 of 45

🛑 Stop and Frisk

Fact: 80% of stops involved Black or Hispanic individuals.

Fact: 50% of NYC population is Black or Hispanic.

Is this persuasive evidence of discrimination?

If yes, explain why. �If not, what would be persuasive?�[ Feel free to chat with neighbor ]

� ��

29

shorturl.at/6B9nA

Q? pollev.com/jdgg

30 of 45

⚖️ Prima facie evidence

The large raw difference in stop+population proportions is sufficient to initiate a legal claim of discrimination.

On its own, this finding does not prove discrimination.

30

Q? pollev.com/jdgg

31 of 45

🛑 Stop and Frisk

“. . . the police are [not] engaged in racial profiling . . . they are stopping people in those communities who fit descriptions of suspects or are engaged in suspicious activity.”��Michael Bloomberg, former New York City Mayor�Washington Post Op-Ed [ 2013 ]

31

Q? pollev.com/jdgg

32 of 45

🎛️ Adjusting for observables

Officers report stop data on a UF-250 form.�[ e.g., demographics, location, reason(s) for stop ]

32

Q? pollev.com/jdgg

33 of 45

🎛️ Adjusting for observables

Using UF-250 data, we can compare frisk rates for pedestrians who differ only in their recorded race/ethnicity.�[ the same sex, age, stop reason, location, ... ]

Is this an appropriate strategy to test for discrimination?[ Feel free to chat with neighbor ]

33

Answer here: shorturl.at/6B9nA

Q? pollev.com/jdgg

34 of 45

🎛️ Adjusting for observables

What about the factors we do not observe?�[ Omitted-variable bias ]

34

Q? pollev.com/jdgg

35 of 45

🎛️ Adjusting for observables

What about the factors we do not observe?�[ Omitted-variable bias ]

Can we fully trust the data?�[ e.g., UF-250 is filled out after the stop, not before ]

35

Q? pollev.com/jdgg

36 of 45

🎛️ Adjusting for observables

What about the factors we do not observe?�[ Omitted-variable bias ]

Can we fully trust the data?�[ e.g., UF-250 is filled out after the stop, not before ]

Does "differ by only race/ethnicity" even make sense?�[ e.g., location strongly correlated with race+ethnicity ]

36

Q? pollev.com/jdgg

37 of 45

🚪 Outcome tests

Rather than action rates, look at action success rates. �[ An “outcome test” (Becker, 1957) ]

Hit rate�Proportion of frisks that "successfully" recovered a weapon.

37

Q? pollev.com/jdgg

38 of 45

🚪 Outcome tests�Hypothetical scenario

Among frisked Black pedestrians, 2% had a weapon.

Among frisked white pedestrians, 4% had a weapon.

38

Q? pollev.com/jdgg

39 of 45

🚪 Outcome tests�Hypothetical scenario

Among frisked Black pedestrians, 2% had a weapon.

Among frisked white pedestrians, 4% had a weapon.

How might you interpret this result?�[ Feel free to chat with neighbor ]

39

Answer here: shorturl.at/6B9nA

Q? pollev.com/jdgg

40 of 45

🚪 Outcome tests�Hypothetical scenario

Among frisked Black pedestrians, 2% had a weapon.

Among frisked white pedestrians, 4% had a weapon.

On average, frisked white pedestrians were riskier.�[ i.e., twice as likely to have a weapon ]

40

Q? pollev.com/jdgg

41 of 45

🚪 Outcome tests�Hypothetical scenario

Among frisked Black pedestrians, 2% had a weapon.

Among frisked white pedestrians, 4% had a weapon.

On average, frisked white pedestrians were riskier.�[ i.e., twice as likely to have a weapon ]

Therefore, they may have been held to a more lenient standard.�[ i.e., only frisked if they appeared extra risky ]

41

Q? pollev.com/jdgg

42 of 45

42

White

More risky

Less risky

Q? pollev.com/jdgg

43 of 45

43

White

4%

weapon recovery rate

from frisks

Frisked

Not frisked

More risky

Less risky

Q? pollev.com/jdgg

44 of 45

44

2%

weapon recovery rate from frisks

White

Black

4%

weapon recovery rate

from frisks

Frisked

Not frisked

Frisked

Not frisked

More risky

Less risky

Q? pollev.com/jdgg

45 of 45

45

0%

0%

0%

0%

0%

0%

0%

1%

1%

1%

3%

4%

2%

weapon recovery rate

from frisks

White

Black

0%

0%

0%

0%

0%

1%

2%

2%

3%

5%

4%

weapon recovery rate

from frisks

Frisked

Not frisked

Frisked

Not frisked

Perceived chance of carrying weapon

Q? pollev.com/jdgg