Part 1: Where We’ve Been

• Recap of Lectures 1–5
• The empirical researcher’s toolkit

Part 2: The Paper

• Huck (2024) in RFS
• Every section mapped to your toolkit

Part 3: Your Turn

• Assignment 6: Dutch replication
• Walkthrough

Part 4: Wrap-Up

• Thesis connection
• The role of AI in your workflow

Key takeaways

• Research starts with data wrangling, not regressions
• Raw data is messy: missing values, inconsistent formats, multiple sources
• The pipeline: import ⤏ clean ⤏ transform ⤏ merge ⤏ analyze
• Tools: Stata, merge, reshape, collapse, gen/replace

From this course

• Compustat Global via WRDS
• The “tidy data” principle: one row per observation unit
• Optional: SQL DB for efficient data management

Key takeaways

• Visualization is diagnosis, not decoration
• Summary statistics reveal distributions, outliers, patterns
• Histograms, scatter plots, time series — each serves a purpose
• Always check your data before running regressions

Application

Key takeaways

• Identification = isolating causal variation
• Endogeneity: X correlated with the error term
  • Selection bias: who self-selects into treatment
  • Omitted variable bias: what you forgot to control for
• Experiments as the gold standard
• OLS: the workhorse, but only as good as your identification

The DiLLMa lesson

Event Studies

• Short-window test of market reactions
• Estimation window ⤏ expected return model ⤏ abnormal returns ⤏ CARs
• The “counterfactual” is the factor model prediction
• Same logic as DiD, but at daily frequency
• Leads and lags reveal timing of information incorporation

Key picture

Event Studies

• Short-window test of market reactions
• Estimation window ⤏ expected return model ⤏ abnormal returns ⤏ CARs
• The “counterfactual” is the factor model prediction
• Same logic as DiD, but at daily frequency
• Leads and lags reveal timing of information incorporation

Fama–MacBeth

• Two-step procedure for cross-sectional pricing tests
• Step 1: Cross-sectional regression each period
• Step 2: Time-series average of coefficients
• Handles cross-sectional correlation by construction
• The standard approach for testing factor premia

Citation

Huck (2024) — “The Psychological Externalities of Investing: Evidence from Stock Returns and Crime”

The Review of Financial Studies, 37(4), 1172–1213.

The question

Do stock market returns affect the psychological well-being of both investors and noninvestors?

Why this paper?

• Published in a top-5 finance journal
• Uses methods you’ve learned in Lectures 1–5
• Clean identification strategy
• Surprising result: the stock market affects people who don’t even own stocks!

Investors

• Stock market rises ⤏ absolute wealth increases
• Also better off relative to average person
• Psychological well-being improves
• Violent crime decreases

Noninvestors

• Stock market rises ⤏ no change in own wealth
• But now relatively worse off compared to investors
• Psychological well-being declines
• Violent crime increases

Why income terciles?

• Malloy, Moskowitz, and Vissing-Jørgensen (2009) show probability of holding stocks increases significantly with income
• Top tercile ≈ likely investors
• Bottom tercile ≈ likely noninvestors
• Available for all agencies

Why violent crime specifically?

• Aggregate crime ⤏ measurement problem: theft in rich areas likely committed by low-income individuals
• Violent crime (assaults, homicides): typically occurs at home, offender known to victim
• 54% of assault victims in middle/high-income households
• 44% assaulted at or near their home
• Minimizes the misattribution problem between offender income and location income

Leads and lags (Table 4)

Only the contemporaneous (day \(t\)) coefficients are significant!

What this tells us

• No pre-trends: market returns on t−1, t−2, t−3 don’t predict crime
• No delayed effects: market returns don’t affect crime in coming days
• Within the day: effect is strongest after markets close
• This is consistent with: see news ⤏ emotional response ⤏ same-day behavior

What stays the same

• The research question and hypothesis
• The empirical specification structure
• Income terciles as investor proxy
• Interaction terms for differential effects
• Fixed effects and robustness checks

What changes

• Country: Netherlands instead of US
• Market: AEX index instead of S&P 500
• Crime data: CBS municipality data instead of FBI NIBRS
• Frequency: Monthly instead of daily
• Geography: Municipalities instead of cities
• Income: CBS municipality median income instead of ACS

Three raw datasets

1. crime_data.dta: Municipality × month panel of crime incidents by type
2. income_data.dta: Cross-sectional average median income per municipality (2011–2018)
3. aex_daily.dta: Daily AEX index closing prices

Your pipeline

• Section 2: Load and examine
• Section 3: Calculate standardized returns (trailing 252-day SD, then aggregate to monthly)
• Section 4: Reshape crime data (long ⤏ wide)
• Section 5: Merge all three datasets
• Section 6: Create analysis variables (terciles, interactions)

Before regressions, you create:

• Summary statistics by income tercile
• Time series of crime rates (trending down? seasonal patterns?)
• Crime rates overlaid with AEX returns
• Crime trends split by income group

Structure to follow

1. Main result: Do you find differential effects?
2. Signs: Consistent with Huck’s theory?
3. Statistical significance: Are key coefficients significant?
4. Economic magnitude: How large are effects?
5. Robustness: Do results hold across specifications?
6. Comparison to Huck: Same or different from US results?

Common mistakes to avoid

🤦‍♀️ Just listing coefficients without interpretation

🤦‍♀️ Over-claiming causality (“AEX returns cause crime”)

🤦‍♀️ Ignoring null results (“we didn’t find it, so we won’t discuss it”)

💰 Null results are informative! If you don’t find the same pattern, that’s a finding worth discussing.

Data-related reasons

• Monthly ≠ daily: effects may wash out at lower frequency
• AEX is a smaller market with fewer household investors
• CBS crime data has different reporting standards than FBI NIBRS

Context-related reasons

• Dutch stock participation rate may differ from US
• Welfare state ⤏ less relative deprivation?
• Cultural differences in violence as stress response
• Netherlands is geographically smaller ⤏ less local variation

Good limitations

“Monthly data cannot test within-day timing effects (Huck’s Table 4–5), limiting our ability to distinguish same-day from lagged effects.”

“Municipality-level income is an imperfect proxy for investor participation; Huck uses census micro-data allowing for more precise classification.”

“A single market return means no cross-sectional variation in ‘treatment’ within a given month — all municipalities face the same AEX return.”

Bad limitations

“We don’t have enough data” ⤏ Too vague. What data? Why does it matter?

“Results might be wrong” ⤏ Undermines everything without explaining what could be wrong.

“The Netherlands is different from the US” ⤏ How is it different, and why does that affect your specific test?

Your MSc thesis modules

1. Module 1: Literature review & research question
2. Module 2: Replication of published results with new/similar data
3. Module 3: Extension or new analysis building on Module 2

What you did in A6

• Read a published paper ⤏ Module 1
• Replicated the methodology with Dutch data ⤏ Module 2
• Discussed extensions and improvements ⤏ Module 3 ideas

Technical skills

• Build a data pipeline from raw sources to analysis-ready panels
• Run and interpret OLS, panel FE, DiD, event studies, and Fama–MacBeth regressions
• Create publication-quality tables and figures
• Conduct and explain robustness checks

Research skills

• Read and critically evaluate an empirical finance paper
• Choose methods appropriate to your research question and data
• Identify threats to identification and address them
• Write honest interpretation and limitation discussions
• Replicate published work and evaluate generalizability