Operations & Systems

Scatter Plots for Business Reports: Inspect Paired Observations Before Explaining a Relationship

The dots rise across the page. Someone explains why complex requests take longer.

Then the same dots receive request-type labels. Inside each group, the durations fall as the complexity measure rises.

Start with the paired records. Keep the explanation open.

NIST's scatter-plot guidance defines the plot as corresponding values of two variables. It also states that association does not establish causation.

The example below uses fictional requests to make that boundary visible. The records, worksheet, grouping rule, and investigation steps are editorial proposals. They establish no business result or employee-performance finding.

Make One Point Mean One Request

For this example, one point represents one completed request with two measurements. The horizontal coordinate records required checklist checks at intake. The vertical coordinate records elapsed review duration in minutes.

The checklist count is a deliberately narrow proxy for request complexity. It counts distinct required checks under one fictional checklist version. It does not measure difficulty, quality, effort, or the requester's competence.

Duration starts when review begins and ends at the first completed response. Waiting after review starts remains inside the interval. Waiting before review starts stays outside it. Neither variable measures a person's working time.

The population contains twelve eligible requests completed during one fictional seven-day reporting window. Every row has both measurements. Reopened requests are outside this example. The two subgroup labels are standard and external-evidence, recorded at intake.

Keep the pairing key with the extract. Join each request's checklist count to that same request's duration. Do not sort the two measurement columns independently. That operation would create coordinates for combinations that never occurred.

The measurement moments differ by design. Complexity is recorded at intake; duration becomes available at completion. Document that timing before interpreting a pattern. If the intake checklist changes, preserve its version instead of quietly mixing definitions.

Before plotting, count unique request identifiers on both sides of any join. Check whether one request can carry several intake events or completion events. Choose the intended event through a documented rule.

For these twelve synthetic rows, each identifier has one intake checklist count and one first-completion duration. The join therefore retains twelve paired records. It neither multiplies a request into several points nor drops an unmatched request.

If a source produces two intake versions for S-04, stop and resolve which version represents the defined measurement moment. Taking the first visible row would let export ordering decide the variable. Keep the rejected version in the working evidence.

Also check the pair after formatting. A spreadsheet sort that moves only the duration column can leave twelve plausible numbers and twelve valid identifiers. The row count still reconciles. Comparing the joined pairs against the source is what catches the changed relationships.

Read the Complete Paired Dataset

Every row below is synthetic. The coordinate column states exactly what enters the chart. It includes the unit for each axis in the heading.

Request Intake subgroup Required checks, count Review duration, minutes Coordinate: checks, minutes
S-01 Standard 1 8 (1, 8)
S-02 Standard 2 7 (2, 7)
S-03 Standard 2 7 (2, 7)
S-04 Standard 3 6 (3, 6)
S-05 Standard 4 5 (4, 5)
S-06 Standard 5 4 (5, 4)
S-07 External-evidence 6 22 (6, 22)
S-08 External-evidence 7 21 (7, 21)
S-09 External-evidence 8 20 (8, 20)
S-10 External-evidence 8 20 (8, 20)
S-11 External-evidence 9 19 (9, 19)
S-12 External-evidence 10 18 (10, 18)

Twelve rows produce ten distinct coordinate locations. S-02 overlaps S-03. S-09 overlaps S-10. A chart with ten visible positions still represents twelve requests.

Keep the duplicates when the requests have different valid identities. Matching measurements do not make separate requests duplicate records. A source check must establish an actual duplicate before exclusion.

The standard group spans one through five required checks and four through eight minutes. The external-evidence group spans six through ten checks and eighteen through twenty-two minutes.

These constructed ranges make the example easy to inspect. They provide no expected range for a real process. The neat arrangement is part of the demonstration.

Compare Pooled and Labeled Views

The left panel below treats every request as the same kind of point. The right retains all coordinates and adds subgroup symbols. Both panels use the same axes.

Pooled and labeled views of twelve fictional requests Both plots use required checklist checks from zero to ten horizontally and review minutes from zero to twenty-five vertically. Six standard requests have lower check counts and durations that decrease within their group. Six external-evidence requests have higher counts and durations that also decrease within their group. Two coordinate positions each contain two requests. The table supplies all twelve paired rows. Pooled requestsLabeled subgroups MinutesMinutes 2525 00 0510 0510 Required checks, countRequired checks, count 2 requests2 requests 2 requests2 requests Circle: standardSquare: external-evidence Identical paired rows and scales. Symbols add subgroup identity.
Every mark uses fictional paired measurements. Repeated coordinates are labeled. Symbols distinguish subgroups without relying on color. The table preserves every individual request.

Read across the pooled view. Requests with higher checklist counts occupy the higher-duration cluster. A broad positive association is visible across the two groups.

Now read inside each labeled subgroup. The standard sequence runs from one check and eight minutes to five checks and four minutes. External-evidence runs from six checks and twenty-two minutes to ten checks and eighteen minutes.

Within both constructed groups, duration decreases as the check count rises. The pooled pattern and the within-group patterns answer different questions about the same records.

Do not select whichever view supports a preferred explanation. Show the pooled view and the labeled view together. State that group membership is associated with different ranges of both variables in this sample.

The labels do not explain why those ranges differ. They identify a condition worth inspecting. The source record still needs evidence about timing definitions, request paths, and how required checks were assigned.

Neither subgroup covers the other's complexity range. This dataset therefore supplies no same-count comparison between standard and external-evidence requests. A sentence claiming the extra duration at a fixed complexity would exceed the available records.

Use Shape as an Investigation Prompt

NIST identifies scatter-plot questions about nonlinear relationships, changing variation, and unusual observations. It also describes views conditioned on another variable. Treat those features as questions to investigate.

For this proposed review, first ask whether a single straight description misses visible structure. A curve or bend would call for checking the source conditions around that region. This fictional example shows separated groups instead of a smooth curve.

Next ask whether the points form clusters. Here the intake subgroup labels correspond to the separated locations. Keep those labels because they were defined independently of chart appearance. Avoid inventing a new operational category solely to explain a gap.

Inspect overlapping points before calling a region sparse. Our chart labels two coincident requests at each repeated coordinate. In a denser real report, provide a count or a separate table of repeated pairs. If you offset points for readability, disclose the offset and preserve exact coordinates in the table.

Ask whether vertical spread changes across comparable horizontal ranges. The constructed table has too little repeated coverage to establish a durable change in spread. Two identical pairs do not make that question settled.

Finally, inspect observations that carry much of a visual conclusion. As a separate hypothetical, an additional request at ten checks and two minutes would complicate the upper group's pattern. Verify its timing events and subgroup before explaining it.

Do not insert that hypothetical point into the twelve-row chart. Keep proposed investigations separate from observed records. Use an outlier review log if a real record's validity remains unresolved.

Preserve the Pair When Values Are Missing

Suppose a later extract contains a thirteenth eligible request with a known check count and an unavailable duration. That request cannot occupy a complete coordinate in this view. Keep it in the eligibility reconciliation with the missing measurement identified.

Report thirteen eligible requests and twelve plotted pairs. Do not replace the missing duration with zero. Do not borrow another request's duration to complete the row.

The earlier synthetic dataset has no missing measurements. This scenario describes how the handoff would change if an additional incomplete record appeared. Retain the earlier version so readers can distinguish a new source discovery from a chart edit.

Check exclusions as paired decisions too. If a request falls outside the defined population, remove the entire paired observation from the scoped view. Record the rule and identity in the working notes. Preserve the original extract.

Use consistent chart axes across any subgroup panels. Changing the vertical range between panels would add another visual decision to an already conditional comparison.

Hand Off the Finding and the Next Question

Copy this proposed handoff into the report. A reader should be able to locate the paired rows without reconstructing your chart settings.

Handoff field Entry for this fictional report
Pair identity S-01 through S-12; one completed request per row
Dataset version synthetic-pairs-v1; twelve eligible and twelve plotted
Horizontal variable Required checklist checks at intake, count; same checklist version
Vertical variable Elapsed review start to first completed response, minutes
Population Eligible completions in one seven-day window; reopened requests outside scope
Exclusions and missing values No missing measurements; no additional eligible rows excluded
Subgroup Standard or external-evidence, recorded at intake
Overlapping coordinates Two requests at (2, 7); two at (8, 20)
Observed pattern Higher-count subgroup has higher durations; both within-group sequences decrease
Limits Synthetic sample; separated ranges; no causal test or employee comparison
Next question Do both subgroups use the same start and completion event definitions?
Evidence needed Checklist version, event dictionary, and paired source records

Assign the next question to the person who can retrieve its evidence. Keep the answer unresolved until that check returns. A chart annotation should not silently turn a proposed explanation into a verified condition.

If the underlying question concerns change through time, retain timestamps and use a separate run-sequence view. Horizontal position in this scatter plot means check count. It does not mean observation order or elapsed calendar time.

Stop the Explanation at the Evidence Boundary

A positive pooled pattern does not establish that adding checklist checks lengthens a request. A negative within-group pattern does not establish that adding checks shortens one. This example contains no intervention that tests either claim.

It also contains no employee identity, workload measure, quality assessment, or working-time record. Turning the duration axis into an employee score would add a claim the data never measured.

Keep regression fitting and forecasting outside this review. The immediate deliverable is a correctly paired dataset, two honest views, and an answerable source question.

Preserve the pairs. Label the groups. Carry the unresolved question forward.

Sources

Topic map: Business reporting and statistics

Articles in this cluster, then the 3 clusters that share the most topics with it.

This cluster: Business reporting and statistics

Neighbor: Spreadsheets, CSV files, and data handoffs

Neighbor: Records, access, and continuity

Neighbor: APIs, webhooks, and automation