Pre-post surveys work best when you need to estimate change that is plausibly attributable to a defined intervention. They are most useful when you can survey the same people twice and when outcomes are expected to shift over a realistic timeframe. To make them decision-ready, plan design choices up front:

• Tie items to objectives: Translate each program objective into measurable outcomes. For example, if the goal is skill adoption, include behavior-frequency items, not just attitudes.
• Keep instruments comparable: Use identical wording, scales, and ordering across pre and post. If you must revise an item, document the change and analyze separately.
• Right timing: Field the pre close enough to the intervention to capture baseline, and the post after a realistic window for change. For fast-learning modules, 1–2 weeks may be sufficient; for behavior change, plan longer follow-up.
• Sampling and tracking: Aim for the same respondents at both waves. Use unique, privacy-safe identifiers or recontact links to support pairing without revealing identity unnecessarily.
• Scale choices: Prefer 5- or 7-point labeled Likert scales for attitudes and confidence; use knowledge checks with keyed correct answers for learning; use frequencies for behavior. Keep direction consistent to reduce respondent error.
• Power and sample size: Estimate needed N using expected effect size and desired power. Paired designs usually require smaller samples than independent post-only comparisons.
• Missing data plan: Decide in advance how to handle partial completes and attrition. Define inclusion criteria for paired analysis and thresholds for attention checks.

Pre-post data should be analyzed at the paired level when possible to isolate within-person change. Pairing increases statistical power and reduces noise from individual differences. Use methods that match the data type:

• Continuous or Likert outcomes: Report mean pre, mean post, mean change (Δ), standard deviation, and confidence intervals. Use paired t-tests or linear mixed models when assumptions are met. For non-normal or ordinal data, consider Wilcoxon signed-rank tests.
• Binary outcomes: Use McNemar's test for paired proportions. Report the discordant pair counts to show direction of change.
• Knowledge items: Summarize percent correct before and after; compute change scores and, if multiple items, create a scaled index with reliability checked via Cronbach's alpha or omega.
• Effect sizes: Include standardized effect sizes such as Cohen's dz for paired designs. These help stakeholders compare impacts across programs.
• Bias controls: Address common threats to validity: regression to the mean, testing effects, social desirability, history/maturity, and nonresponse bias. Use comparable instruments, attention checks, anchoring vignettes when appropriate, and sensitivity analyses that test robustness to attrition.
• Multiple comparisons: If you track many outcomes, pre-specify primary endpoints and control false discovery (e.g., Holm-Bonferroni).
• Interpreting impact: Combine statistical and practical significance. Convert changes into clear decision metrics, such as percentage moving from "unlikely" to "likely," or expected ROI drivers if linked to downstream KPIs.

Move from plan to field with pragmatic tools that save time and prevent avoidable errors.

• Question templates:
  • Attitude: "How confident are you in doing X?" [1 Not at all to 5 Extremely]
  • Knowledge (keyed): "Which of the following is correct about Y?" [one correct option; shuffle order]
  • Behavior frequency: "In the past 2 weeks, how many times did you do Z?" [0, 1–2, 3–5, 6+]
• Paired tracking methods: Use respondent-created codes (e.g., first letters + birth month) or secure recontact links to match responses without storing PII in the dataset.
• Minimize testing effects: Randomize item order within topic blocks, hide correct answers in feedback until after post, and separate pre from intervention start.
• Field quality: Include attention checks, trap items with low base rate, and device/type monitoring. Set minimum exposure times to discourage random responses.
• Reporting structure: Provide an executive summary, a change table (pre, post, Δ, CI), effect sizes, and a short narrative interpreting what the changes mean for decisions. Add a technical appendix with methods, power, and assumptions.
• Ethics and privacy: Keep data minimization in mind. Collect only what you need, store pair keys separately when possible, and be transparent in consent language.