Investigate missingness and alternative explanations
Learn the reasoning. Try the task. Compare your answer.
Prepared for DataCareerHub · Curriculum version 2026-10-15.2
By the end of this lesson
- Profile missing ratings by group rather than only overall.
- Keep arrival cohorts separate from completion-day activity.
- Distinguish observations, hypotheses and unsupported explanations.
Harbor Helpdesk · UTC · Snapshot 22 June 2026 00:00 · Prepared fictional extract · No Python or SQL required.
Understand the method
Exploration is not just a search for a high or low number. Ask which observations are absent, how records entered the extract and which alternative explanations remain compatible with the evidence. A result among respondents can be accurate while saying little about nonrespondents. A date comparison can be correct while using the wrong event for the business question.
The missingness pattern matters
Each week has eight ratings among twelve eligible tickets: 66.67% coverage. Both respondent means are 4.0 out of 5, based on a sum of 32 divided by eight. Those identical headline summaries hide different group coverage. In Week 1 all eight email tickets have ratings and none of the four chat tickets do. In Week 2 all four email tickets and four of eight chat tickets have ratings. Across both weeks, email coverage is 12/12, chat coverage 4/12.
A Week 1 chat satisfaction mean is unavailable because its observed denominator is zero. It is not zero out of five, and it cannot be recovered by copying the overall mean. You do not know whether missing ratings reflect the survey workflow, willingness to respond or another mechanism. The group pattern gives a useful question to ask the data owner; it does not identify that mechanism by itself.
The mean of 4.0 applies to observed respondents. Do not claim that every ticket or customer was equally satisfied. Replacing blanks with zero would produce a different, invented measure. Dropping unrated tickets from response-timeliness analysis would also change its population for no justified reason. If exploring a relationship between response time and rating, use only valid pairs, report the paired count and respondent selection, and do not interpret an association as the effect of faster service.
Use the correct clock
The cohorts in this exercise are defined by creation time. Counting first resolutions by resolution date instead describes completion activity, which can include earlier arrivals and is a different question. A ticket resolved next week does not migrate out of its creation cohort. All supplied tickets have adequate follow-up, but a live rolling report would need an explicit maturity rule for recent arrivals.
The file has no arrivals on the final day of either seven-day window. That is a known property of this small constructed extract, not proof of zero business demand on those dates. Daily rates can have very small denominators or no eligible tickets. Display an empty day's rate as unavailable with count zero. Do not join two weekly summary points with a smooth curve and call it a stable trend or seasonal pattern.
Keep an explanation ledger
Write three columns: observation, possible explanation and evidence needed. “Chat share rose” is observed. “The changed entry page shifted channel choice” is a hypothesis compatible with the case notes. “Routing caused faster service” is not established. Request exposure logs, consistent survey instrumentation, staffing/arrival coverage and a longer comparable history as appropriate. Do not invent a survey-system defect just because coverage is uneven.
The most useful conclusion may be a narrower one: observed timeliness and duration summaries changed, channel composition also changed, and ratings do not represent all tickets. That statement protects the decision-maker from confusing a valid descriptive pattern with a confirmed operational explanation.
Practical exercise
Calculate rating coverage and respondent means by cohort and channel. Identify the group with no ratings and state its correct display. Draft three rows for an explanation ledger: an observed pattern, a plausible hypothesis and the evidence request that would test it.
- Explain why the same 4.0 mean does not establish equal overall satisfaction.
- Describe the harm from deleting unrated tickets before response analysis.
- Distinguish no arrivals in the training file from complete business-day demand evidence.
- State why two weekly values cannot establish seasonality.
Open the worked solution
W1 email: 8/8 ratings, mean 4.0; chat: 0/4, mean unavailable. W2 email: 4/4, mean 4.0; chat: 4/8, mean 4.0. Overall per week: 8/12 observed, mean 4.0 among respondents. Across both cohorts, email has 12/12 coverage and chat 4/12. The unobserved opinions remain unknown, so equal respondent means do not establish equal satisfaction for all tickets or customers.
Deleting unrated tickets would discard valid response observations and alter channel composition. Explanation ledger examples: changed chat share → entry-page visibility may affect channel choice → request entry/exposure logs; uneven rating coverage → survey delivery may differ → request survey-offer/delivery events; lower mean duration → mix or staffing may matter → request roster/exposure and longer event history. These are questions, not discovered causes. Missing training arrivals on the last day of each window do not prove zero operational demand; two cohort summaries cannot establish a seasonal pattern.
One more question.
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 →Harbor Helpdesk is a fictional teaching case. Descriptive comparisons do not establish routing causality. 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.