Explaining Statistics and ML Decisions

Move From Definition to Decision

Definition-only answers often sound memorized. If asked when to use median instead of mean, "The median is the middle value" is true but incomplete. A stronger answer is: "For a strongly skewed distribution or one with extreme outliers, median may better represent a typical observation because it is less sensitive to those values. I would still inspect the distribution and use mean when total magnitude is the decision concern."

Use this pattern:

1. Define the concept briefly.
2. State when it matters.
3. Explain the trade-off.
4. Connect it to the data or business decision.
5. Say what you would check.

Statistics Answers Need Context

For an A/B-test question, do not stop at statistical significance. Explain the effect size, uncertainty, data quality, sample design, and whether the change is large enough to justify cost or risk. For confidence intervals, distinguish sampling uncertainty from sampling bias. For any statistic, describe the condition under which its interpretation becomes useful.

This does not require a lecture on hypothesis testing. It requires showing that a number is evidence within an analysis, not an automatic decision.

Model Choices Are Comparisons

Asked, "Why start with logistic regression?" explain that it can be a fast, interpretable baseline with probability outputs. Then add the boundary: it may miss nonlinear patterns, so compare it against sensible alternatives using the same validation process.

Asked, "Why Random Forest?" mention nonlinear relationships, interactions, and reduced single-tree instability, then avoid claiming it always wins. Model choice depends on data type and volume, the metric, interpretability, compute, latency, maintenance, and the consequence of errors.

"It depends" is useful only after naming what it depends on. For example: "Threshold choice depends mainly on class balance, false-positive and false-negative cost, and whether the output triggers a fixed action or ranks cases for review."

Preprocessing and Validation

A strong answer to "Why scale features?" says that scale-sensitive methods can be dominated by larger numeric ranges, while tree-based models are usually less sensitive. A strong answer to "Why use a pipeline?" connects consistent preprocessing, train/inference consistency, and leakage-safe cross-validation.

For leakage questions, use a risk-consequence-workflow structure: identify a post-outcome or future feature, explain why it creates unrealistically strong validation, then describe a safer prediction-time feature contract and split. Likewise, tuning on the final test set turns it into development evidence and weakens its role as an unbiased final check.

Translate Into Consequences

Technical choices become credible when connected to outcomes: false positives can create unnecessary manual review; false negatives can miss fraud; a larger model can increase cost and latency; extra features can add maintenance burden. Do not give causal meaning to a predictive feature simply because it appears important.

Use Practice as Rehearsal

Use existing Statistics, Machine Learning, and Model Evaluation Practice problems verbally: solve the scenario, explain why, state assumptions, name one alternative, and describe one failure mode. This turns a correct answer into a decision-oriented explanation without creating a duplicate problem bank.

Rehearsal Prompts

1. When is median preferable to mean?
2. Why establish a baseline first?
3. When should a model be scaled?
4. Random split or chronological split?
5. Logistic regression or a tree model for a tabular decision?
6. What would make a validation result untrustworthy?

Interview Perspective

What this demonstrates: Many technical questions test whether you can choose and defend a decision under assumptions, not whether you can recite a glossary definition.

Next Lesson

Next, apply this decision-oriented reasoning to the common metric questions covered in How to Answer Metric Questions in Interviews.