Reading a Dataset Honestly
A workflow that catches the errors the earlier lessons describe: look first, group second, summarise third.
- Run a fixed sequence of checks on an unfamiliar dataset
- Compare mean and median and use the gap diagnostically
- Detect suspicious values with fences and inspect them before removing anything
- Report a summary that states its own limitations
What it means
This lesson is a checklist rather than a new quantity. Every measure in this course is easy to compute and easy to misuse, and the misuse is almost always a skipped step rather than a wrong formula.
The sequence that prevents most of it:
- Look at the data before summarising it: counts, missing values, types, and a histogram.
- Compute the centre two ways and compare them; a large gap is information about the shape.
- Measure spread two ways and compare them; the same gap shows up.
- Flag suspicious values and inspect each one rather than deleting it.
- Group before pooling when the data arrived in batches or categories.
- Report the summary with its uncertainty and its limitations.
Why it adds something
An analysis that skips step 1 is a calculation, not an analysis. The checklist turns the individual lessons into a habit, and the habit is what catches Simpson’s paradox, an outlier that was a typo, and a mean that describes nobody.
Worked example
A dataset of 12 support tickets, resolution times in hours:
Step 1 — look. Twelve values, all numeric, no obvious missing entries. The maximum is far from everything else.
Step 2 — centre two ways. Mean: total 162, so 13.5. Median: . A gap of 9 hours: the data has a long right tail.
Step 3 — spread two ways. Standard deviation about 27.6; IQR: , , so 8.25. The standard deviation is more than three times the IQR, which is another way of saying the tail dominates it.
Step 4 — flag and inspect. Upper fence: . The 96 is flagged. Inspect it: if it is a ticket that was reopened after a month, it may be a genuine case but belongs to a different process; if it is a data-entry error (96 minutes recorded as hours), fix it. Do not silently delete it.
Step 5 — group. Are these tickets from one team, or several? A mix of easy and hard queues would produce exactly this shape.
Step 6 — report. Something like: “Median resolution time was 4.5 hours (IQR 2.75–11). The mean, 13.5 hours, is pulled up by one 96-hour ticket; it does not describe a typical ticket. Excluding that case, the median is unchanged and the mean falls to 6.0.”
🧩 Challenge, think first, then reveal
A colleague reports “average response time 13.5 hours” for these tickets. What single additional number would you ask for first, and why?
💡 Answer: The median. It is 4.5 hours here against a mean of 13.5, and that gap tells the reader more about the dataset than the mean does on its own. Whether the 96-hour case is real is the second question.
The same thing in Python
from statistics import mean, median, stdev, quantiles
hours = [1, 2, 2, 3, 3, 4, 5, 6, 8, 12, 20, 96]
q1, _, q3 = quantiles(hours, n=4)
iqr = q3 - q1
print('n', len(hours))
print('mean', round(mean(hours), 2), 'median', median(hours))
print('sd', round(stdev(hours), 2), 'iqr', iqr)
print('fences', round(q1 - 1.5*iqr, 2), round(q3 + 1.5*iqr, 2))
print('flagged', [h for h in hours if h < q1 - 1.5*iqr or h > q3 + 1.5*iqr])pandas gives the same summary in one call and makes the missing-values check trivial:
import pandas as pd
s = pd.Series(hours)
print(s.describe())
print(s.isna().sum())A reporting template
The shape of a summary that survives review:
- what was measured, with units and sample size;
- centre, giving the median first when the data is skewed;
- spread, choosing the IQR or the standard deviation to match the centre;
- what was excluded and why, never a silent deletion;
- what it cannot tell you, in one sentence.
When to use it
Every time. The checklist is most valuable on data you did not collect, which is most data.
Limitations
- The checklist does not replace domain knowledge. Whether a 96-hour ticket is plausible depends on what a ticket is.
- It catches one-variable problems best. Relationships need a scatter plot, and grouping needs enough data per group to be meaningful.
- It is a starting point, not a full analysis. Questions about change over time, or about a specific causal claim, need their own designs.
Common pitfalls
- Deleting outliers before inspecting. Often the most informative cases, and the deletion biases every number that follows.
- Reporting only the mean, which hides the shape entirely.
- Pooling groups before checking whether the groups differ.
- Reporting two decimal places on a measurement recorded to the nearest hour.
🧩 Challenges
🧩 Challenge, think first, then reveal
Design the three-line summary you would send to a manager who has never heard of a median, using the support-ticket data, without using any statistical jargon.
💡 Answer: "Half the tickets closed within 4.5 hours, and half took longer. A typical ticket sits between about 3 and 11 hours. One ticket took 96 hours, which is what drags the average up to 13.5; that case is worth checking before we quote the average to customers." No jargon, and every number carries its meaning.
🤔 Socratic Questions
- Why is comparing the mean and median a cheap diagnostic for the shape of the data?
- When is the standard deviation a better companion to the mean than the IQR?
- What is the difference between an error and an outlier?
- If you must report a single number for a skewed dataset, which would you choose, and what would you add to it?
✅ Quick check
1. The mean is 40 and the median is 25. What does the gap suggest?
2. You find a value above the 1.5 × IQR fence. What should you do first?