Decide what to collect

Use AI to draft a data dictionary and analysis plan, then check both against the protocol and seek statistical advice.

1. Plan the data and analysis

Try the same assistant

Ask ChatGPT, Microsoft Copilot or Google Gemini to work from the fictional case. The prefilled prompt does not require an upload. The answer is a draft for discussion, not a statistical protocol.

Copy: draft an analysis plan

For a before-and-after evaluation of a pharmacist-led antimicrobial review process, the unit is an adult antimicrobial treatment episode. The outcome is whether a review was documented within 72 hours (Yes, No or unknown). Draft a minimal data dictionary and descriptive analysis plan. Define the numerator, denominator, timepoint, treatment of unknown outcomes and comparison in percentage points. Explain why repeated episodes from the same patient, clustering by ward and changes in documentation may matter. State assumptions and questions for a statistician; do not claim a before-after difference proves causation.

Check: use the synthetic-review-data.csv and checked results. There are two missing outcomes in each period; do not count them as No. Consider how the reason for missingness could affect interpretation. Discuss the plan with a statistician before real use.

No account? Use the prepared data dictionary and results on this page, then decide what you would need to check before applying the plan to a real service.

Here's one we prepared earlier

For each eligible episode: record a synthetic episode ID, period (before/after) and whether a review was documented within 72 hours (Yes/No/Unknown). Agree eligibility, time origin and how missing documentation is handled before collecting real data.

Describe counts and the percentage documented in each period, using known Yes/No values as the denominator and reporting missing values separately. Show the difference in percentage points with an appropriate confidence interval. A two-proportion comparison may be considered if observations are independent; repeated episodes, ward clustering and small samples require different methods or advice. Do not interpret an observed difference as proof that the intervention caused it.

2. Analyse the data in ChatGPT

Download synthetic-review-data.csv. If your account supports file uploads, start a new chat, attach that CSV file and paste this prompt. File analysis and upload limits depend on your plan and account. If upload is unavailable, use expected-results.txt to follow the example.

Copy: analyse synthetic data

The attached file contains antimicrobial treatment episodes. For each Period, count Review_72h values Yes, No, blank and unexpected values separately. Calculate the percentage Yes among records with Yes or No outcomes and report the missing outcomes. Show the after-minus-before difference in percentage points, calculations and assumptions. Do not infer causality. Explain how I can check the counts independently.

Compare every count with the independently checked results; investigate any mismatch. Do not upload real patient or confidential information.

Try it: use the synthetic-review-data.csv and compare your answer with the checked results. In this fictional example, before: 14 of 28 known episodes (50%); after: 21 of 28 (75%). There are two missing values in each period. The descriptive difference is 25 percentage points; it does not establish causation.

Check: AI may suggest a test without knowing how observations were sampled, related or missing. Ask why it is appropriate; never accept a test solely because it looks familiar.

PreviousExplore with Gemini Notebook

NextTry code