Airport Data Role Play: Missing Values, Model Choice, and Explaining Regression Output

In a role-play round, you walk the interviewer through a modeling problem built from airport data. The discussion covers how you handle missing values and the data's format, whether you would use linear regression or a random forest, and, having chosen linear regression, what the coefficients, p-values, t-values and F-value tell you and how you would improve the model. The report calls this a well-known airport case and does not describe the data. Assume a table of historical flights at an airport, with a numeric outcome the business wants to predict, for example departure delay in minutes, and a mix of numeric and categorical predictors such as airline, scheduled departure time, day of week, aircraft type and weather measurements, some of them with missing values. ### Clarifying Questions - What is the outcome, and how will the results be used: to explain what drives it, to forecast it for individual flights, or both? - What is one row: a flight, a day, or a gate? - Which predictors are known at the moment a prediction would have to be made? ### Part 1 — Missing values and data format How would you handle missing values, and what would you check and change about the data's format before modeling? ```hint Why is it missing? Before choosing a fill method, ask what a missing value means in each column, for example a sensor outage versus a flight that never departed. ``` #### What This Part Should Cover - Diagnosing how much is missing and why, column by column - Handling choices matched to the cause, without leaking information into evaluation - Types, encodings, time fields and the row grain needed for modeling ### Part 2 — Linear regression or random forest Choose between linear regression and a random forest, and justify the choice. The rest of the discussion assumes linear regression is chosen. ```hint Who uses the model? Think about what the business side needs from the model: a forecast for each flight, an explanation of what drives the outcome, or both. ``` #### What This Part Should Cover - A comparison on interpretability, assumptions, nonlinearity and data handling - A choice tied to the purpose of the analysis - How the other model can still be useful ### Part 3 — Explain the regression output After fitting the linear regression, explain what the coefficients, their t-values and p-values, and the model's F-value tell you, in terms a business audience can follow and precisely enough for a technical reviewer. ```hint Each number answers a different question For each statistic, ask which hypothesis it is about: one predictor, or all of them together. ``` #### What This Part Should Cover - Correct interpretation of coefficients, including categorical predictors and the meaning of holding other predictors fixed - What t-values, p-values and the F-test test, and what they do not say - The assumptions behind the p-values and where this data may violate them ### Part 4 — Improve the model How would you improve the model? ```hint Let the residuals talk Look at where the model's errors are large or follow a pattern before adding complexity. ``` #### What This Part Should Cover - Diagnostics that show what to fix - Feature, transformation and regularization changes - How improvements are validated ### What a Strong Answer Covers - Treats missing data according to its cause, and never imputes the outcome - A model choice justified by the purpose of the analysis, with the other model as a benchmark - Precise and plain-language interpretation of coefficients, t, p and F, including what p-values do not mean - Awareness of correlated errors, heavy-tailed delays and correlated predictors in this data - Improvements driven by diagnostics and validated on held-out data ### Follow-up Questions - A predictor has a large coefficient but a p-value of 0.4. What do you tell the business side? - The F-test is significant, but no individual coefficient is. How can that happen, and what do you do? - How would you check whether the random forest predicts better, and what would you lose by switching?