Data analysis graphs and charts: choose by question

Data analysis graphs and charts work best when each display answers one question from a checked table. Use lines for ordered periods, bars for categories, histograms for distributions and scatter plots for paired measurements. The chart should make the comparison easier to see while preserving the table's meaning.

Start with the business question

Write the decision before selecting a chart. A sales leader asking whether revenue is rising needs an ordered time comparison. An operations manager asking why some orders take longer needs the distribution of individual delivery times, rather than another monthly revenue line.

This is chart selection, rather than software selection. For spreadsheet setup, the guide to charts for analysis and reporting covers that neighboring task. Keep the initial question narrow enough that the reader can tell what action the chart supports.

Work the trend and category views from the same rows

Consider an illustrative US distributor's closed monthly sales export. Each row below is the sum of net posted sales for a month and product group. Amounts include credits, use USD throughout and exclude sales tax.

The monthly total line rises from $120,000 to $165,000. Endpoint growth is (165,000 - 120,000) / 120,000 = 37.5%. That statement is correct but incomplete: August falls even while core-product revenue continues to rise. Two product-group lines reveal that the interruption comes from specialty products.

An ordered plot is useful because sequence carries information. NIST's run-sequence guidance describes plotting observations in order to inspect shifts and unusual observations. Preserve chronological order and show any missing month explicitly; a blank month is not evidence of zero sales.

For a category question, use bars for the period totals: core products at $380,000 and specialty products at $170,000. The shares are 380,000 / 550,000 = 69.1% and 170,000 / 550,000 = 30.9%, rounded to one decimal place. Bars compare dollar size; a share display answers the separate question of composition.

The reconciliation works both ways. Monthly totals sum to $550,000, and category totals also sum to $550,000. Tie that amount to the closed sales control total before discussing the chart. For a longer ranked list with cumulative shares, use Pareto analysis instead of squeezing every category into a crowded composition chart.

Show the distribution when averages hide the work

Suppose the same distributor reviews completed delivery durations. The individual observations, in days, are 1, 2, 2, 3, 3, 3, 4, 6. Their sum is 24 days across 8 deliveries, giving a mean of 24 / 8 = 3.0 days.

Each share divides the bin count by 8; for example, 3 / 8 = 37.5%. These equal-width ranges support a histogram. NIST defines a histogram as counting observations within numerical bins to show a distribution's center, spread and shape.

The average does not mean every delivery takes three days. The underlying observations show that 6 of 8 deliveries, or 75.0%, finish within three days, while one takes six days. The operations owner should inspect that delayed delivery before changing the standard promise. With so few observations, describe what happened rather than treating the shape as a reliable picture of all future deliveries.

The bin counts must add back to the number of valid delivery records. Investigate missing completion dates separately; silently excluding unfinished orders can make service look faster.

Use paired measurements for relationship questions

A scatter plot requires two measurements for each same observation. This illustrative branch table pairs weekly support hours with resolved requests.

Plot hours horizontally and requests vertically. East and South show why a simple upward story needs checking: South records more hours but fewer resolved requests. Case complexity, staff experience or different recording rules could explain the difference.

NIST's scatter-plot guidance distinguishes association from proof of cause. Do not infer that adding hours will necessarily produce the same output at every branch. If the question moves from these observed branches to a larger population, inferential data analysis explains the sampling assumptions that must also hold.

Where graphs and charts go wrong

• Comparing incomplete periods. A partial month beside a closed month creates a misleading decline. Mark the cutoff and compare equivalent periods.
• Changing the denominator. A revenue-share chart and a customer-count share chart answer different questions. Name the denominator in the subtitle or table.
• Using truncated bars. Cropping the value axis exaggerates category differences. Start bars at zero and label any unusual scale explicitly.
• Combining incompatible rows. Joining customer revenue to invoice-level service costs can repeat costs. Establish the common reporting grain before aggregating.
• Smoothing away exceptions. A trend average can hide the very interruption management needs to investigate. Show the underlying periods alongside any smoother.

The analyst owns the reconciliation and labels. The sales or operations owner supplies the business explanation and records the action. A useful chart caption states the comparison, the period, the checked result and what still needs investigation.

Do this on your own file in Covirage

Bring a workbook or Google Sheet with period, category and value columns; Covirage maps the columns and its tools compute and check the underlying comparisons. The external AI model explains the result with every figure traced to its rows, so you can review the evidence before deciding how to present it. Create a free Covirage account and bring a sample of your own data.

See self-service analytics with Covirage for the supporting workflow.