What needs to be finished first
A hypothesis about a suspected cause and effect with two measurable variables exists, ideally from a Fishbone analysis or a 5 Whys chain.
What needs to be ready before start
Spreadsheet or statistics tool (Excel, Google Sheets, R, Python with pandas/matplotlib, JMP); dataset with paired values; timestamp per measurement; optional Pareto tool for comparison.
One analyst who prepares and visualizes data; one domain expert who interprets context and anomalies; one data-source owner for data-quality questions.
Hypothesis as a sentence ("we suspect that X influences Y"); definition of variables with units; measurement interval and collection period; known confounders; minimum data volume (30+ pairs).
30-90 min for creation, plus data preparation
Data in two columns: X (cause), Y (effect). Timestamp as third column. Tool open. Hypothesis visible at the top of the sheet.
The one question this method answers
Do the paired data show a pattern that supports, contradicts or leaves open the suspected cause-effect hypothesis?
Marker: Phase
| Step | Duration | Action | Hint |
|---|---|---|---|
1Phase 1: Sharpen hypothesis and variables | 10 min | Formulate hypothesis as an if-then sentence with concrete variables and units. Record expected direction (positive, negative). Define at least 30 data pairs as the threshold. | If the hypothesis mixes several variables ("bad weather makes employees tired"), specify one variable per pair. Otherwise the diagram cannot be interpreted. |
2Phase 2: Collect or prepare data | Variable, often 30-60 min | Extract or collect the dataset. Ensure clean pairing (same time, same context). Check data quality: outliers, missing values, collection conditions. | Data context must be documented: who measured when, under which conditions. Without context, outliers will later be misinterpreted. |
3Phase 3: Create plot | 15-20 min | Scatter plot with X = cause, Y = effect. Axis labels with units. Optionally color points by time or context. Trendline optional (use cautiously). | Trendlines often suggest more certainty than the data supports. First show the plot without a line, add the line only as a second layer if correlation is strong. |
4Phase 4: Describe pattern | 20-30 min | Systematically review patterns: positive trend, negative trend, clusters, outliers, no pattern, U-shape. Calculate correlation coefficient (Pearson, Spearman), but do not overinterpret. | Correlation 0.3 is not "light evidence", it is often noise. Speak of a robust relationship only from around 0.7, and even then only in context of sample size. |
5Phase 5: Interpretation and next step | 15-20 min | Classify result as hypothesis supported, contradicted or open. Discuss confounders. Derive action or further analysis (for example controlled experiment, Pareto, regression). | Correlation is not causation. Before deriving measures, plan at least a controlled test or check stratification, otherwise symptoms get treated as causes. |
What comes out at the end
PNG or PDF export of the plot with axis labels, correlation coefficient, sample size and collection period; plus Markdown note with hypothesis, observation, confounders and next step.
Date, data source and collection period in the header. Save follow-up analyses with newer data as a new plot in the same repo. Archive raw data as CSV with the plot, not only the image.
Compact working template for Scatter Diagram with context, input, output artifacts, and next step.
What is this method used for?
Which question should be answered at the end?
Which data, observations, or materials are available?
Owner, date, success signal.
Concrete filled scenario, fictional example
Concrete filled scenario, fictional example
Hypothesis: If room temperature rises above 25°C, then the typo rate in incoming tickets increases.
Data: 124 shifts between 2026-03-01 and 2026-04-30. Source: office weather station + ticket system.
Variables: X = average temperature per shift (°C). Y = typos per 100 tickets.
Plot: Point cloud shows weak positive trend up to 24°C, strongly positive above 26°C. Pearson r = 0.62 (n=124). Spearman ρ = 0.71.
Clusters: Clearly higher variance above 26°C. Three outliers at 22°C (sickness wave).
Interpretation: Hypothesis supported for range above 26°C. Confounders: shift staffing, weekday, sick leave. Correlation is not causation.
Next step: Pilot with air conditioning in two rooms over 4 weeks, A/B comparison of typo rate. Owner @lisa, start 2026-06-01.
Recognize symptoms and steer against them
Plot with 12 points shows apparent trend, every sample adjustment changes the picture.
Use at least 30 pairs as threshold. With fewer, make no correlation claim. Instead extend collection period or broaden source.
Stakeholders derive measures directly ("lowering temperature lowers typos") without checking confounders.
Before recommending measures, at least discuss confounders and plan a controlled experiment. Explicitly mark reports with "correlation, not proven causation".
Three extreme points are included, correlation flips, nobody knows why these points differ.
Research context for each outlier (weather, special event, defect). If cause can be clarified, report separately or exclude. Document rationale.
Data contain typos, wrong units or time offset, plot is distorted.
Data check before plot: min/max plausible, duplicates, missing values. Review anomalies with the source owner.
Correlation coefficient r is low, although plot shows a clear U-shape.
Interpret plot visually before Pearson r. For non-linear patterns, use Spearman ρ and polynomial fit, or stratify into ranges.
Done signals checkable in under a minute
Go to the profile for purpose, similar methods, and sources or continue to the next method in the catalog.