Statistical Analysis or Machine Learning Model for the Urban Question

Choosing the Right Analytical Method

Before writing a single line of code, you must match your method to your urban research question. The wrong method produces misleading results, no matter how clean your data is. Use the decision-tree flowchart your instructor provides as a checklist, then write a short justification in a Markdown cell explaining why the chosen method fits your variable types, sample size, and research goal.

A few common mappings:

• Predicting a continuous outcome (e.g., permit wait time in days, median rent): use linear regression. Check that your residuals are roughly normally distributed and that predictors are not perfectly correlated with each other (multicollinearity).
• Classifying neighborhoods into groups with no pre-existing labels: use k-means clustering or another unsupervised method. You supply the number of clusters k and evaluate cohesion with the silhouette score.
• Testing whether two categorical variables are independent (e.g., transit access level vs. income bracket): use a chi-square test of independence. Confirm that expected cell counts are at least 5 before trusting the result.
• Comparing means across two or more groups: consider a t-test (two groups) or ANOVA (three or more groups), checking normality and equal-variance assumptions first.

Your justification should name the method, state the outcome variable, state the predictor(s) or grouping variable(s), and explain why the assumptions are likely satisfied by your dataset.

Implementing the Model or Test

Once your method is chosen, implement it in Python (scikit-learn, statsmodels, or scipy) or R (caret or tidymodels). Keep all code in clearly labeled cells inside your Jupyter Notebook so the instructor can re-run it from top to bottom without errors.

For supervised learning (e.g., linear regression):

1. Split your data into training and test sets using `train_test_split` or an equivalent.

2. Fit the model on training data only.

3. Compute evaluation metrics on the test set: R-squared tells you the proportion of variance explained; RMSE (root mean square error) tells you average prediction error in the original units.

For unsupervised learning (e.g., k-means):

1. Scale your features first — k-means is sensitive to units.

2. Run the algorithm across a range of k values and plot the elbow curve or silhouette scores to select k.

3. Report the silhouette score for your chosen k; values above 0.50 generally indicate reasonable cluster separation, though this is data-dependent.

For hypothesis tests (e.g., chi-square):

1. State your null hypothesis (H₀: the variables are independent) and your significance level (commonly α = 0.05) before looking at results.

2. Compute the test statistic and p-value.

3. Report effect size alongside the p-value. For chi-square, Cramér's V is a standard effect-size measure. A statistically significant result with a tiny effect size may not matter practically for city policy.

Interpreting Results in Plain Language

Every metric must be translated back into the context of your urban question. A p-value of 0.03 means nothing to a city planner without a sentence like: "Neighborhoods in the lowest transit-access tier were significantly more likely to be in lower-income brackets (χ² = 14.2, df = 4, p = 0.03, Cramér's V = 0.18), suggesting a small but consistent association that warrants further investigation."

Avoid claims of causation unless your study design supports it. Observational city data almost never supports causal language — use words like associated with, predicts, or differs across.

The Peer-Review Stand-Up

In the two-minute stand-up you will state: (1) your urban question, (2) the method you chose and why, and (3) one key metric with a plain-language sentence. Peers will ask at least one question about assumptions or limitations. This is deliberate practice for the workplace, where data professionals must defend their choices clearly and concisely.

After the session, add a Limitations section to your notebook. Common limitations include: small or non-random sample, data collected for administrative rather than research purposes, unobserved confounders, or the temporal mismatch between datasets joined from different years.

Data and Computational Safety

Always work on a copy of your cleaned dataset, never overwrite the analysis-ready file you spent hours preparing. Use version control (Git commits or dated file names) so you can roll back if a transformation produces unexpected results. If your dataset contains any personally identifiable information (PII), confirm with your instructor that it has been appropriately de-identified before sharing your notebook during peer review. Do not upload raw city data containing PII to public repositories.