Compare the two observations that actually belong together

State why the two observations form a pair

Write the underlying unit and the relationship that makes its observations belong together. For a permitted participant study, this may be the same respondent at two specified occasions. For an operational comparison, it may be the same branch measured under two defined conditions. Record the time window and measurement definition on both sides. Similar names or the same spreadsheet position do not establish that relationship.

NIST defines paired data through the natural correspondence between the first and second measurements and describes analysis of within-pair differences. The key review question is therefore whether each difference joins the intended observations. Equal sample sizes alone do not answer it, and separately sorting the two columns can destroy the correspondence.

Reconcile the pairing before computing a change

Create a controlled ledger with the approved unit key, first observation, second observation and pairing status. Check one-to-one uniqueness within the intended occasions. Resolve duplicate submissions and repeated visits through the study’s documented rule rather than choosing whichever record produces the largest improvement. Keep the original records traceable without exposing participant names in a widely shared worksheet.

Separate complete pairs, first-only records, second-only records and ambiguous matches. Report their counts and reasons where known. Do not fill an absent second observation with zero, carry forward the first value without justification or invent a match to preserve the headline sample size. Ask the analyst how incomplete records affect the target population and analysis; a complete-pair result may describe a different set of people from the original recruited group.

Make the difference mean the same thing for every unit

Choose and label the subtraction direction, units and scale. An increase in task success and an increase in time required have different interpretations. Check whether either measure changed wording, coding or collection method between occasions. A numerical difference across incompatible definitions cannot be repaired by using a paired statistical test.

Inspect the distribution of the paired differences with the analyst, including unusual values and patterns linked to baseline level or missing follow-up. Preserve legitimate observations while investigating data errors. Avoid comparing only the two aggregate averages and discarding the paired ledger, because the relationship among observations matters to the uncertainty and interpretation. Let the specialist select an appropriate method and assumptions rather than applying one test to every kind of scale.

Report observed change without inventing attribution

State the complete-pair population, observation windows and excluded or unresolved records beside the result. Keep the individual-change question distinct from a repeated cross-sectional question about two different samples. Both can be useful, but they answer different questions and should not inherit each other’s labels merely because the survey wording stayed similar.

Explain what the design can and cannot establish about cause. A before-and-after change can coincide with outside events, practice, selection or other experiences. Pairing identifies which observations belong together; it does not by itself supply an untreated counterfactual. Preserve the ledger and analysis version so later corrections can be traced to affected pairs. The acceptance outcome is a defensible comparison of matched observations, not a guarantee of causal impact or a claim about every original participant.

Make the reasoning concrete.

A fictional professional association measures members’ ability to find an online resource before and after a navigation change. Some members answer only once. An export is sorted by completion time separately for each wave, then its two columns are subtracted row by row.

The association repairs the pairing before analysis and reports who is represented by the complete records. It does not call the observed change a causal effect of navigation without a suitable design.

Take the questions into your next brief.

Use these prompts in your own notes. Nothing is submitted from this page.

1. Relationship

   What makes these two observations a legitimate pair?

2. Key

   Which permitted identifier and occasion rules establish one-to-one matching?

3. Gaps

   How many records are incomplete, duplicated or ambiguous?

4. Difference

   What subtraction direction and compatible measurement units apply?

5. Claim

   Which population and causal limits belong beside the result?

Watch for the wrong conclusion.

• Pairing independently sorted columns by row position.
• Treating missing follow-up as zero improvement.
• Calling a paired before-and-after difference a causal effect by itself.

Sources & scope

Original editorial analysis and fictional study examples. Primary references support only the attributed statistical distinctions. Qualified analysts choose assumptions and methods; no actual participant records, code execution or causal business result is supplied.

• NIST/SEMATECH e-Handbook: Analysis of paired observations →Primary statistical reference, checked 7 November 2026 · Supports: Supports naturally corresponding measurements and within-pair differences. Ledger, missing-record review and fictional example are original analysis.