Test a relationship without declaring a cause

Test a relationship without declaring a cause

A scatter plot preserves the pairing between two variables. Label both scales and make clear what one point represents. Pair observations from the same relevant unit or time window; arbitrary pairing can manufacture a pattern. An apparent slope is an invitation to investigate, not proof that changing one variable will change the other. Product mix, time, measurement error or an omitted factor may explain the relationship. Preserve source labels and inspect unusual points before fitting a model. A useful next step is a safe, discriminating test that would produce different predictions under the competing explanations.

Follow the method

1. Paired observations
2. Response score (index)
3. Synthetic setting (index)

Move from an observed association to a defensible test

Fictional case: a training simulation pairs setting indices 1–6 with response scores 4,5,5,7,8,8. These exact pairs remain in the core scatter plot. Learner Ben is asked whether increasing the setting causes a better response. The numbers are synthetic indices, with no machine operating meaning or calibrated prediction rule.

Follow the method

1. Fact about supplied data
2. Hypothesis
3. Rival hypothesis
4. Next comparison

An upward pattern without a causal verdict

Preserve observation pairs, describe an association accurately, and design a next comparison that can distinguish rival explanations.

Fictional case: a training simulation pairs setting indices 1–6 with response scores 4,5,5,7,8,8. These exact pairs remain in the core scatter plot. Learner Ben is asked whether increasing the setting causes a better response. The numbers are synthetic indices, with no machine operating meaning or calibrated prediction rule.

Role: Improvement analyst with the process owner

Normal condition

Each plotted point represents one intact paired observation. Units, context and possible shared influences are recorded; conclusions match the evidence.

The gap

The team sorts setting and response separately, then proposes a machine adjustment from the positive-looking graph.

• The fictional setting is not an equipment setpoint.
• Six observational pairs cannot establish a safe intervention, causal effect or reliable prediction.

1. Define one paired observation

   Ben identifies the record that links each setting index to its response score. He checks that no measurement has been matched to another observation merely because the columns were sorted.

   Why: A scatter plot analyzes joint observations. Independently sorting columns can manufacture a positive relationship even when none existed in the actual records.

   Evidence: Every plotted point retains an observation identity.

2. Plot the supplied pairs honestly

   He labels the horizontal axis synthetic setting index and the vertical axis response score. He plots all six pairs with no causal fitted line or real-machine units.

   Why: The graphic should reveal the observations rather than create authority through an impressive trend line. A synthetic scale has no direct operating interpretation.

   Evidence: The repeated scores at settings 2/3 and 5/6 remain visible.

3. State the supported description

   Ben describes an upward association in this small supplied set, with equal responses at some adjacent setting values. He does not say every increase produces an improvement.

   Why: Association is a property of these observations. It is weaker than a causal or predictive claim and does not resolve the meaning or reliability of the response score.

   Evidence: A written interpretation distinguishes the plotted pattern from an intervention claim.

4. Name competing explanations

   He asks whether observation order, a material change or a different operator moved with the setting. The record lacks those contextual fields, so the competing explanations remain open.

   Why: A common influence can move both variables. Missing context should become an evidence request, not a convenient assumption that all other conditions were identical.

   Evidence: The investigation record lists setting effect, order effect and source mix as hypotheses.

5. Propose an authorized discriminating comparison

   For the training exercise he sketches repeated comparisons with randomized order or an appropriate blocked design and consistent measurement. A qualified owner must approve any real experiment and operating range.

   Why: An experiment needs a design that separates plausible influences and protects the process. This proposal is a reasoning exercise, not permission to alter equipment or choose a PID setting.

   Evidence: The next step states what would be held comparable, what would vary and what outcome would challenge the hypothesis.

One point was paired to the wrong record

The response 8 originally paired with setting 5 is found to belong to another observation.

Correct the source pairing with a traceable amendment and replot before interpreting. Retain the original mistaken version as an error record if required locally.

The strength or direction of association depends on correct pairs. A visually convenient point must not remain because it supports the preferred story.

The amended observation ID explains which point changed and why.

Run a practice session

Materials

• Paired observation cards
• Blank axes
• Two colors for observed and hypothesized statements

1. Define a point · 4 minutes

   What makes two values a pair?

2. Interpret the supplied graph · 8 minutes

   Which sentence is supported and which is causal?

3. Expose the sorted-column error · 10 minutes

   Compare actual and invented pairs.

4. Debrief the next test · 5 minutes

   What design would separate setting from order?

Debrief

• Reject safe-setting recommendations from synthetic indices.
• Ask what result would weaken the preferred hypothesis.

Plot the changed pairs by hand, then label each conclusion as observation, hypothesis or proposed test.

Transfer into the work

Owner: Process analyst with an authorized experiment owner

Record: Paired source record, inference note and approved comparison plan

Review: Before any trial and after each reviewed result

Evidence: Pair integrity, controlled context and evidence that discriminates between explanations

Collect missing context or redesign the study rather than treating correlation as a command.