MUSA 5080 - Public Policy Analytics
Due Date: October 27, 2025
In-Class Presentations: October 27, 2025 (5 minutes per team)
Weight: 15% of final grade
Team: You’ll work with your table-mates as a team. Feel free to delegate. Everyone should upload their final products onto their own portfolio websites. Be sure to acknowledge your team-mates.
Submission Format:
You are consultancy (please name your consultancy) competing to win the bid to work for a project for the Philadelphia Office of Property Assessment. The city wants to improve its Automated Valuation Model (AVM) for property tax assessments. Your task is to build a predictive model for residential sale prices and present your findings to city officials in a 5-minute briefing.
Deliverables:
Your goal: Predict 2023-2024 home sale prices accurately while communicating findings clearly to a policy audience.
Format: Quarto revealjs presentation
Example YAML:
---
title: "Philadelphia Housing Price Prediction"
subtitle: "Improving Property Tax Assessments"
author: "Your Name"
format:
revealjs:
theme: simple
slide-number: true
smaller: true
---Content: ~10-15 (could be less!!) slides covering: 1. Research question & motivation (1-2 slides) 2. Data overview (1 slide) 3. Key visualizations (2-3 slides) 4. Model comparison results (1-2 slides) 5. Main findings (2-3 slides) 6. Policy recommendations (1-2 slides)
Audience: City officials who don’t know R or care about the nitty gritty of stats. They just want the best estimates possible.
No code in these slides - just polished visualizations and key takeaways
Format: Quarto HTML document
Example YAML:
---
title: "Philadelphia Housing Model - Technical Appendix"
author: "Your Name"
format:
html:
code-fold: show
toc: true
toc-location: left
theme: cosmo
---Content: All the technical details: - Complete data cleaning code - All EDA visualizations - Feature engineering code - Full model outputs - Diagnostic plots - Detailed interpretations
Audience: Data scientists and technical reviewers
All code visible - this is where you show your work
Source: Philadelphia Property Sales
This dataset contains actual property sales with:
You will need to:
Required: Browse the OpenPhily Data portal and use Census Data to incorporate spatial features into your model.
Your task: Think like an urban planner. What location factors matter for housing prices in Philadelphia?
Your work should follow this workflow, with results split between presentation and appendix:
Load and clean Philadelphia sales data:
Load secondary data:
Deliverable (Appendix only):
Create at least 5 professional visualizations:
For presentation slides: Select your best 2-3 visualizations that tell a compelling story
For appendix: Include all visualizations with detailed interpretations
Example presentation slide:
## Where Are Expensive Homes in Philadelphia?
[Beautiful map showing price patterns]
**Key Findings:**
- Center City and University City command premium prices
- River wards show emerging appreciation
- Northeast Philadelphia remains most affordableCreate spatial features: (these are examples below, but how you construct your model is up to your team)
Buffer-based features:
k-Nearest Neighbor features:
Census variables:
Interaction terms:
Deliverable (Appendix only):
Build models progressively: (for example)
For presentation slides: Show one comparison table (RMSE, R² for 4 different models you constructed in your process)
For appendix:
Example presentation slide:
## Model Performance Improves with Each Layer
| Model | CV RMSE (log) | R² |
|-------|---------------|-----|
| Structural Only | 0.42 | 0.61 |
| + Census | 0.38 | 0.69 |
| + Spatial | 0.31 | 0.78 |
| + Interactions/FE | 0.26 | 0.84 |
**Bottom line:** Neighborhood effects matter most!Use 10-fold cross-validation: - Compare all 4 models - Report RMSE, MAE, R² for each - Create predicted vs. actual plot
For presentation slides: Final CV results table (shown above) + one compelling visual
For appendix:
Check assumptions for best model:
Deliverable (Appendix only):
Note: Don’t include diagnostic plots in presentation - too technical!
Answer these questions:
For presentation slides: 1-2 slides with clear, concise answers (bullet points)
For appendix: 2-3 paragraphs with detailed discussion
Example presentation slide:
## Key Findings & Recommendations
**Model Accuracy:** RMSE = 0.26 (log scale) ≈ 26% typical error
**Top Predictors:**
- Neighborhood fixed effects (largest impact)
- Square footage (β = 0.0003, p < 0.001)
- Distance to transit (β = -0.05, p < 0.001)
**Recommendations:**
✓ Current AVM undervalues transit-accessible properties
✓ Model struggles in rapidly gentrifying neighborhoods Upload to Canvas - A link to your portfolio that Contains
Presentation Slides
LastName_FirstName_Presentation.html (rendered slides)LastName_FirstName_Presentation.qmd (source file)format: revealjsTechnical Appendix
LastName_FirstName_Appendix.html (rendered document)LastName_FirstName_Appendix.qmd (source file)format: htmlData files OR clear download instructions in appendix
For Teams: Use LastName1_LastName2_Presentation.html
Format: 5 minutes per team
What to present:
What NOT to do:
| Component | Points | Criteria |
|---|---|---|
| Research Question | 2 | Clear motivation, Set the stage |
| Data Overview | 2 | Concise description of sources, sample size |
| Visualizations | 3 | 2-3 polished, publication-quality visualizations; clear takeaways |
| Model Comparison | 3 | Clean results table; clear winner; interprets improvement |
| Key Findings | 3 | Top predictors identified; coefficients interpreted correctly |
| Presentation Quality | 3 | Professional design, no typos, flows logically, appropriate for audience |
Key: Slides should tell a compelling story without technical jargon. Imagine presenting to the Deputy Mayor.
| Component | Points | Criteria |
|---|---|---|
| Content | 2 | Covers key points efficiently, answers questions thoughtfully |
| Time Management | 2 | Finishes within 5 minutes without rushing |
| Component | Points | Criteria |
|---|---|---|
| Data Cleaning | 3 | Complete code, proper filtering, missing value handling, documentation |
| EDA | 3 | 5+ visualizations, each with interpretation |
| Feature Engineering | 3 | Buffers , kNN , census , interactions , all properly created |
| Model Building | 3 | 4 progressive models, proper specification, handles sparse categories |
| Cross-Validation | 3 | Proper 10-fold CV, results table, code runs without errors |
| Diagnostics | 3 | Residual plots included, interpreted, violations addressed |
| Code Quality | 3 | Clean, commented, reproducible, follows best practices, no errors |
Key: This is where technical reviewers verify your work. All code must run without errors & be reproducible!!.
Slide 1: Title - Your team name and teammates - Project title - Date
Slide 2: The Problem
Slide 3: Data Sources - Property sales (n = X,XXX, 2023-2024) - Census ACS (income, education, poverty) - OpenDataPhilly (parks, transit, crime)
Slide 4: Where Are Expensive Homes? - [Map visualization] - Key pattern observed
Slide 5: What Drives Prices? - [Best scatter plot or faceted visualization] - Key relationship identified
Slide 6: Model Comparison - [Results table] - “Each layer improves prediction”
Slide 7: Top Predictors - Neighborhood (biggest impact) - Square footage (β = X) - Transit access (β = Y)
Slide 8: Model Performance - Final RMSE: 0.26 (log scale) - Translation: ~26% typical error - Beats baseline by 40%
Slide 9: Hardest to Predict - [Visualization of residuals by neighborhood]
Slide 10: Recommendations
Slide 11: Limitations & Next Steps
Slide 12: Questions? - Thank you - [Contact info]
Q: Do I have to create both slides AND an appendix?
A: Yes! Slides are your main deliverable (present findings). Appendix proves you did the work correctly.
Q: Can code appear in my presentation slides?
A: NO! Slides are for city officials. All code goes in the technical appendix.
Q: How many slides should I have?
A: 10-15 maximum. Quality over quantity. Each slide should have a clear purpose.
Q: Can I use a different city?
A: No, everyone uses Philadelphia for comparability.
Q: How do I make my .qmd render as revealjs slides?
A: Use format: revealjs in your YAML (see template above). Test it early!
Q: My presentation is 8 minutes. Is that okay?
A: NO! You must cut it to 5 minutes. Practice and trim ruthlessly.
Q: Should I include all 5 EDA visualizations in my slides?
A: No! Slides should have your best 2-3 visualizations. Put all 5 in the appendix.
Q: My RMSE is 0.35 in log scale. Is that good?
A: Depends on your data, but 0.25-0.45 is typical for hedonic models. Compare to your baseline or put your data back into dollars!
Q: Should I remove all outliers?
A: No! Only remove obvious errors. Use log transformation to handle legitimate outliers.
Q: What if my code works on my computer but not when I knit?
A: Start fresh, restart R, knit in a clean session. Check for hard-coded paths.
Q: Can I use ChatGPT/Claude to write my analysis?
A: You may use AI for debugging code, but NOT for writing your analysis.
Q: How formal should my presentation be?
A: Professional but not stuffy. Like you’re briefing a city council member who’s smart but doesn’t know statistics.
Q: What happens if I go over 5 minutes?
A: I’ll politely cut you off and you’ll lose points. Practice with a timer!
Week 1: - Download data - Initial cleaning - Basic EDA
Week 2: - Feature engineering - Build 4 models - Run cross-validation - Diagnostics - Write conclusions - Proofread and submit