1. LandOffer
  2. Blog
  3. All articles
  4. Choose a Chart That Makes the Comparison Clear Without Distorting the Data

Data & Decisions

Choose a Chart That Makes the Comparison Clear Without Distorting the Data

Compare three visual forms for the same queue rates, then repair a misleading response-rate slide. Completed examples show how baselines, date spacing, denominators and missing observations change the claim a chart supports.

10 min readLandOffer team

Choose a chart by the comparison your reader needs to make, then check what its position, length and connecting lines imply. Category rates usually need a common comparison scale; a time series needs real dates; a distribution needs individual observations or an honest summary of their spread. Treat the title, units, denominator, date spacing and axis limits as part of the evidence, because each affects what a reader can reasonably conclude.

The practical test is whether a reader can recover the claim without being pushed toward a larger difference, a smoother trend or greater certainty than the data support. This LandOffer.ai guide works through one dataset in three chart forms, then repairs a separate misleading presentation. All counts and workplace decisions below are original synthetic examples.

An exhibition designer checks wooden plinths on a studio table beside a brass straightedge.

Write the comparison before choosing a visual form

Before opening a chart editor, decide what the reader needs to compare; a request to “make a dashboard” does not settle that question. Write a sentence such as, “Compare the share of eligible requests answered within one day across four queues during the same month.” That sentence specifies the unit, denominator, period and comparison. It also exposes claims the chart cannot establish, such as which team worked hardest or why a queue performed differently.

The ONS chart-selection guidance starts with the relationship the reader should see and favors familiar forms. Apply that principle by separating four common questions. For category comparisons, consider horizontal bars or dots. For change over ordered time, use a dated series. For a distribution, show how observations spread. For the relationship between two quantitative variables, consider a scatterplot with each point's unit defined.

These are starting choices, not a rule that every dataset needs a graph. A small labeled table can be better when exact numbers matter. A histogram requires actual observations or valid bins; four queue averages cannot reveal the distribution of individual response times. A scatterplot requires paired measurements, and a visible association would still not establish causation.

Check definitions before drawing. If one queue excludes weekends while another counts elapsed hours, styling cannot repair the comparison. Resolve that discrepancy or disclose it and narrow the question. This article assumes consistent eligibility and elapsed-time definitions; it does not claim to audit a real reporting system.

Completed comparison: four queue rates, three chart choices

Fictional company Beacon Loom reviews October 2026 request records captured through October 27. The packet includes all 350 eligible requests in that specified extract, assigned to four mutually exclusive queues. Its fictional review is dated October 29, after the last included request has had a full 24-hour follow-up; no request is classified as late before that window closes. A successful response means the first substantive answer arrived within 24 elapsed hours of receipt. Each request counts once; reassigned requests retain their initial queue for this example.

Queue Eligible requests Answered within 24 hours Rate Not answered within 24 hours
Cedar 100 80 80% 20
Elm 50 45 90% 5
Pine 20 18 90% 2
Willow 180 171 95% 9

The three candidates below use exactly the same four rates. They do not switch between percentages and volumes to make a preferred form look convincing. The compact text sketches preserve exact labels; each block's scale is stated explicitly.

Candidate 1: horizontal bars, zero to 100%. Each full block represents five percentage points. The endpoints correspond to 80%, 90%, 90% and 95%, not to numbers of requests.

Within 24 hours; October 1–27 synthetic extract
Scale: 0% to 100%; one block = 5 percentage points
Cedar   ████████████████     80% (80/100)
Elm     ██████████████████   90% (45/50)
Pine    ██████████████████   90% (18/20)
Willow  ███████████████████  95% (171/180)

Nia chooses this briefing view because a reviewer can compare rates and recover the underlying request counts without changing scales or estimating slices. The shared baseline supports comparison of magnitudes, and the counts prevent the identical Elm and Pine rates from hiding their different volumes. Cedar has the lowest observed rate and 20 late requests, the largest late count. That supports asking about Cedar's queue conditions; it does not demonstrate poor staff performance or an effect of staffing levels.

Candidate 2: unconnected dots on the same zero-to-100% scale. The left and right pipes mark 0% and 100%, with 19 interior positions at five-point increments; a dot marks a rate. This uses position rather than filled length and keeps the same ordering and labels.

         0%                 100%
Cedar    |...............●...| 80%
Elm      |.................●.| 90%
Pine     |.................●.| 90%
Willow   |..................●| 95%

The dot view is a reasonable alternative when the audience can read it comfortably. A carefully labeled narrower numeric scale could make small differences easier to inspect, because dot positions do not represent filled bar lengths. This packet keeps the full range to avoid introducing a second scale into the comparison meeting. A zoomed view would need visibly labeled limits and context.

Candidate 3: connected points across queue names. The proposed path is Cedar 80% → Elm 90% → Pine 90% → Willow 95%. It preserves the numbers but adds an implied progression between categories. Reordering the same names would change the shape. The team rejects it because queues are neither successive dates nor positions on a continuous measurement axis.

A pie chart is also unsuitable for these four rates: 80% + 90% + 90% + 95% equals 355%, not one whole. A pie made from the 314 timely-response counts would answer a different question about their distribution across queues. It would not compare the probability of a timely answer within each queue.

Keep denominators and decision limits attached

The overall extract rate is 314/350, approximately 89.7%. Averaging the four queue percentages gives 88.75%, which gives the 20-request queue the same weight as the 180-request queue. Neither figure should appear without its calculation being identified. For this packet, the overall request rate uses total successes divided by total eligible requests.

The meeting record is deliberately modest. Operations owner Nia selects the bar view, keeps a companion count table, and asks queue owner Dev to review Cedar's request mix and handoff delays. Dev accepts that review, but the record contains no completed investigation or promised improvement. There is no assignment of extra staffing based only on the plotted rate.

The four rates describe all listed requests in this authored extract; they were not estimated from a probability sample of future requests, and they do not establish next month’s service quality. No confidence interval was calculated. That does not make next month's rate certain: request mix, reporting completeness and process changes remain limitations. Pine's 18/20 should not be presented as equally stable future evidence merely because it shares Elm's 90% label.

If the displayed values were sampled estimates, uncertainty could materially change the interpretation. The ONS uncertainty guidance recommends showing meaningful uncertainty ranges and explaining them. Use intervals justified by the design and analysis; do not add decorative error bars or decide that overlapping intervals alone supply a complete hypothesis test.

Wooden viewing frames and a sage cord rest beside ceramic discs and a brass divider on a cream tabletop.

Completed repair: a dramatic rise becomes a two-point difference

A separate fictional briefing slide is titled “Response performance doubles.” It shows three bars labeled October 1, October 2 and October 9. Their values are 92%, 93% and 94%, each based on 100 eligible requests for that date. The slide omits the percent unit, the denominators and the axis start. The hidden lower limit is 90%, so the chart fills only the distance from 90% to each displayed rate rather than the distance from zero.

That baseline makes the visible lengths 2, 3 and 4 units. The third bar is twice the first bar's visible height, but the underlying rates are not doubled. The actual difference is two percentage points, and the relative increase is about 2.17%: (94−92)/92. Both calculations describe these authored observations only.

Tableau's axes guidance explains why a truncated bar baseline can exaggerate differences and why missing intervals can distort interpretation. The repair retains all three observations and changes their presentation rather than altering the data to fit the headline.

Slide element Original problem Completed correction
Title “Response performance doubles” “Observed within-24-hour response share: 92%, 93%, 94% on three recorded dates”
Bar scale Hidden start at 90 Zero-to-100% scale if bars are retained; labels show 92/100, 93/100 and 94/100
Unit and period Neither specified Percent of eligible requests received on each named date in October 2026
Dates between observations October 3–8 omitted without explanation Mark those six dates as not observed in this packet; do not treat them as zero
Trend alternative Three equally spaced points connected Actual calendar positions: day offsets 0, 1 and 8; show isolated points and retain the unobserved gap
Conclusion Sustained improvement implied Three observed rates differ by two points; persistence, mechanism and future performance remain unestablished

Analyst Imani chooses the dated dot view for the corrected slide because the immediate question concerns the three recorded observations. On a true calendar axis, the gap from October 2 to October 9 occupies seven days, compared with one day from October 1 to October 2. A line spanning that gap would invite readers to imagine an unobserved path, so the correction uses points without interpolation.

If someone wants to discuss rates of change, the first observed interval changes by one percentage point per day. The second changes by one point over seven days, about 0.143 points per day. Those endpoint calculations do not establish what happened on the six missing dates. A straight interpolated line would be an additional assumption, not a recovered record.

The review accepts the corrected title, scale and missing-date note. Manager Tomas leaves the “sustained improvement” claim unresolved until the missing-date question is investigated. The corrected slide supports the three dated observations. It supplies neither the six missing rates nor evidence that a process intervention caused a sustained improvement.

Check scale, labels and accessibility together

The ONS axes guidance distinguishes zero-based filled bars from line or scatter displays that may use a cropped scale. The distinction is about visual encoding. A general command to start every possible axis at zero can obscure a real pattern, while a cropped filled bar changes the apparent magnitude ratio. Choose deliberately and show the limits plainly.

Keep comparable panels on consistent scales. Two small charts with unrelated vertical ranges can make a small change look equal to a large one. Avoid independently scaled dual axes when they invite a false relationship between different units. Prefer aligned separate panels with units and dates visible, unless a specific dual-axis use is genuinely necessary and understandable.

Labels should survive separation from the surrounding slide. Include the metric, period, unit, denominator where relevant, source or synthetic-data note, and material omissions. Color should help distinguish a meaningful category, rather than serving as the only way to recover it. Direct labels, adequate contrast and an adjacent data table help people who cannot rely on color or inspect a dense image.

Check the exported mobile or presentation view, not just the editing canvas. Can a reviewer read the labels without hovering? Does a narrow layout hide the missing-date note? Can someone using a text alternative recover the comparison and its limits? A clean chart with an unreadable caveat is still an incomplete communication.

Before sending your next analysis, have the decision owner read the proposed comparison sentence, then reconcile every displayed value, scale limit and missing-date note against the retained data table. Keep unresolved uncertainty visible rather than polishing it away. Explore more workplace analysis guides at LandOffer.ai, then retain the chosen view, its rejected alternatives and the claim the data actually support.

Sources and Further Reading