# Session 3 Handout: Data Quality and Fairness

<p class="meta">FAST-05 Applied Data Science and ML for Business Users · 7formations · Instructor: Wadood Chaudhary · All examples fictional</p>

## Why this session matters

A model can't be better than the data it learned from. Duplicates, blanks, inconsistent codes, and skewed samples quietly become wrong predictions, and sometimes unfair ones that land on the same group of people every time. You can spot most of these problems with a spreadsheet and a few good questions, before anyone trains anything.

## Key concepts

### 1. Garbage in, garbage out

| Problem | What it looks like | Why it matters |
| --- | --- | --- |
| Duplicates | Same record twice | Overweights some cases; inflates counts |
| Missing values | Blanks, "N/A", 0 used as blank | If missing isn't random, filling or dropping creates bias |
| Inconsistent formats | 03/04/2026 vs 2026-04-03; Y/N vs 1/0 | Misread dates; mislabeled outcomes |
| Outliers and errors | Lead time of 900 days | Distorts averages and patterns |

### 2. Sampling and bias

**Who's in the data determines who the model works for.** If 90% of training rows come from one office, one customer type, or one season, the model may perform poorly, or unfairly, for everyone else.

**Proxies:** even if a protected characteristic isn't a feature, others (ZIP code, insurance type, commute distance) can stand in for it. Ask: if this model is wrong, who bears the cost, and is it always the same group?

### 3. Drift

Models assume the future looks like the past. When something changes (a new reminder system, a big customer's new ordering pattern, a pandemic, a price change), predictions drift. Watch for errors that jump after a known change, and plan for retraining or human override.

### 4. Practical data checks (no coding)

- Count rows by group (PivotTable / COUNTIFS).
- Count blanks per column, and by group.
- List distinct values in key columns (codes, labels).
- Check date formats and ranges.
- Remove exact duplicates (on a copy).
- Ask the data owner about any process changes during the period.

AI tools can profile a table quickly. Confirm one number by hand.

## Worked example: Charles River Family Dental no-show file (fictional, IDs only)

| Finding | Size | Consequence |
| --- | --- | --- |
| Skewed sample | Newton 90%, Waltham 10% (similar real volumes) | Waltham predictions rest on ~120 rows |
| Missing lead time | 112 rows, 104 Waltham | Filling with 0 would make Waltham look like same-week bookings |
| Mixed formats | Waltham dates MM/DD; no_show 1/0 | Misread dates; mixed labels |
| Duplicates | 18 rows | Inflated counts |
| No Saturdays | Waltham opens Saturdays | Model never saw Saturday patterns |
| Drift | Text reminders since Jan 2025 | Likely overpredicts no-shows now |

**Fairness note:** Double-booking based on these scores would mostly affect Waltham patients with public insurance and long commutes, who could face longer waits. Lower-harm alternative: reminder calls for high scores, tested by office first.

## Checklist: before trusting a dataset

- [ ] Rows counted by key groups; sample matches reality?
- [ ] Blanks counted, and by group.
- [ ] Codes and date formats consistent.
- [ ] Duplicates removed (on a copy).
- [ ] Process changes during or after the data period identified.
- [ ] Proxy features considered; affected groups named.

## Do and don't

| Do | Don't |
| --- | --- |
| Ask why values are missing | Fill blanks with zero by default |
| Compare the sample to reality | Assume "lots of data" means representative |
| Name who bears the cost of errors | Call a model "objective" |
| Prefer lower-harm actions on scores | Use scores to deny or downgrade service without review |
| Log process changes as drift risks | Assume the past predicts the future |

## Try this at work

1. Take one non-sensitive dataset your team relies on.
2. Run the six checks above.
3. Write one finding with size, location, and consequence.

## Glossary

| Term | Meaning |
| --- | --- |
| Skewed sample | Data over-representing some groups or conditions |
| Proxy | A feature that stands in for a protected or sensitive attribute |
| Drift | Change in the world that makes past patterns less valid |
| Missing not at random | Blanks concentrated in a group or condition |
| De-identification | Removing identifiers per specific rules (HIPAA has two methods) |

## Key takeaways

- Most model problems start as data problems you can see in a spreadsheet.
- Who's in the data decides who the model works for.
- Watch for drift after any change.