Keep missingness, coverage and denominators visible
Learn the reasoning. Try the task. Compare your answer.
Prepared for DataCareerHub · Curriculum version 2026-10-15.4
By the end of this lesson
- Report observed delivery measures with their eligible counts.
- Distinguish missing values from zero and not applicable.
- Choose status-rate denominators that match the stated question.
April–May teaching extract · Order values in EUR · Delivery blanks remain missing · Use a local spreadsheet; Python is optional.
Understand the method
Missing data changes what an average describes. A delivery average calculated over observed records is not necessarily the average for every completed order. First state which orders could contribute a delivery measurement, then count how many actually do. For this exercise, delivery_days refers to a completed-order duration. Returned and cancelled orders have no supplied delivery measurement and are outside that comparison. This is a defined analytical eligibility rule, not a claim about all real returns.
Count the eligible and observed populations
There are 20 completed orders in April, of which 14 have recorded delivery durations and six do not. May also has 20 completed orders, with 19 recorded durations and one missing. Coverage is 14/20 = 70% and 19/20 = 95%. The observed means are about 3.64 and 3.53 days. Attach both observed and eligible counts to these numbers. A slight difference between means does not establish an operational improvement when the observed populations and missingness differ.
The missing completed durations are A001, A004, A007, A010, A013, A016 and Y001. All are web orders in the supplied data. This concentration is a reason to investigate the collection process. You cannot infer the unobserved durations from their channels, neighbouring rows or the mean of observed records. Missingness mechanisms require evidence; this exercise shows a pattern, not proof of why the measurements are absent.
Why zero filling distorts the result
April’s recorded durations sum to 51 days across 14 measured orders. Dividing that sum by all 20 completed orders would give 2.55 days, implicitly treating the six unknown durations as zero. May’s observed sum is 67 across 19 measured orders; dividing by 20 would give 3.35 days. This faulty operation can create an apparent deterioration when the observed-only average moves slightly in the other direction. Neither method recovers the unknown true all-order mean.
Maintain three concepts: a known numeric zero, a missing measurement and a field outside the analytical population. A blank in a CSV represents missing input under this dictionary; it must not silently become zero during numeric conversion. If a different dataset legitimately allows same-day zero duration, retain that numeric zero in the observed population. Test for absence, not for truthiness.
Name the status denominator
Each month has one returned order among 22 all-status orders. The returned share of all supplied orders is 1/22 = 4.55%. If the question explicitly compares returned orders with completed plus returned orders, the denominator is 21 and the share is 4.76%. These are different metrics. Neither is a customer return probability, a refund amount or a cohort-based rate after a fixed follow-up window. The file lacks those definitions and timing details.
Use coverage to guide the next action
Report the measured delivery average as descriptive of observed completed orders and flag that April’s coverage is lower. Ask how durations were captured, whether missingness is channel-specific in a wider extract, and whether later completion changes the available values. Do not claim “delivery improved” from a small difference alone. A useful operational recommendation can be to investigate data capture before making a service-performance claim.
Practical exercise
For each month, calculate eligible completed orders, observed delivery counts, missing counts, coverage, observed duration sum and observed mean. Demonstrate the effect of incorrectly filling missing durations with zero, keeping that example clearly labelled as wrong. Then calculate returned share using both all-status and completed-plus-returned denominators.
Deliverable: a coverage table, two accurately named status shares and a sentence explaining why no all-completed-order delivery improvement is established. List the IDs with missing completed delivery and inspect their channel distribution.
Open the worked solution
| Measure | April | May |
|---|---|---|
| Eligible completed orders | 20 | 20 |
| Observed / missing durations | 14 / 6 | 19 / 1 |
| Observed coverage | 70% | 95% |
| Observed duration sum | 51 days | 67 days |
| Observed mean | 3.6429 days | 3.5263 days |
| Incorrect zero-filled mean | 2.55 days | 3.35 days |
| Returned / all statuses | 1/22 = 4.55% | 1/22 = 4.55% |
| Returned / (completed + returned) | 1/21 = 4.76% | 1/21 = 4.76% |
Keep missing durations excluded from the observed-duration denominator and visible in the coverage denominator. All seven missing completed measurements belong to web orders. This is evidence of a missingness pattern, not permission to impute channel means or a demonstration of its cause. The four returned/cancelled blank durations belong outside the stated completed-delivery comparison.
An appropriate conclusion is: “Observed completed-order delivery durations average 3.64 days for 14/20 orders in April and 3.53 for 19/20 in May. Different coverage, concentrated missingness and a small teaching extract prevent an all-order improvement claim.” Status proportions also require careful labels: a share of supplied order statuses does not measure customer behaviour over a defined follow-up window.
One more question.
Method references: NIST exploratory data analysis and box plots and IQR fences. The case and exercises are authored for DataCareerHub; the quartile convention is stated in the dictionary.
How these learning materials are prepared and maintained · Report a correction: support@datacareerhub.io
Read and practice without an account. Sign in with an existing account to save lesson completion.
Next lesson →Fictional training data. Portfolio practice is not a claim of client work or an accredited qualification. Explore practical lessons, worked examples and projects. Follow a published path, practise the method and build work you can explain.