Choose a narrow question and define the unit of observation
Use the fictional comparison-group exercise as a starting point: how did a measured outcome change in two groups before and after an event? This is deliberately small enough that the calculation can be checked by hand. Your goal is to demonstrate careful research practice, not to manufacture a publishable causal result.
Write one sentence defining an observation, such as one group in one period. State the outcome's unit, the event date and how group membership is assigned. If the definition changes between periods, stop and reconcile it. A beautiful chart of incomparable records is not a useful analysis.
Ship the dictionary with the data
Create a raw input file that you do not edit silently. Put transformations in a separate script or workbook sheet. Record where the input came from and label simulated data as simulated. You do not need confidential client data or an unauthorized scraped dataset to show that you understand reproducibility.
| Field | Meaning | Validation |
|---|---|---|
| group | Treatment or comparison group | Exactly one recognized group label |
| period | Before or after the event | One observation per group-period |
| outcome | Fictional measured outcome in arbitrary units | Numeric, finite; missing is not zero |
| synthetic | Teaching data rather than an observed sample | True for every row |
A minimal dictionary for this exercise is below. Additional real-world fields would require their own definitions. Missing values should have an explicit representation, not a convenient zero that changes the average.
Test the file before calculating an effect
Check the row count, duplicate keys, missing values and numeric ranges. For the four-row exercise, there must be exactly one observation for each group-period combination. If an extra treatment-after row appears, determine whether it is a duplicate or a different observational unit before averaging it.
Keep a validation log. A check that fails should stop the analysis or be accompanied by a documented resolution. Do not delete an inconvenient observation merely because it changes the result. If you exclude records for a principled reason, report the rule and show how many observations it removes.
For a larger extension, aggregate individual records only after deciding whether each row represents a person, transaction, organization or repeated measurement. Mixing these units can give large customers or frequently observed people unintended weight. These are analysis-design decisions, not cosmetic cleaning steps.
Calculate the descriptive contrast and challenge it
The fictional treatment outcome changes from 100 to 120 and comparison from 80 to 88. The difference in changes is (120 − 100) − (88 − 80) = 12 units. Include this formula in the memo and make the code or cell references visible.
Then write a limitations paragraph before a conclusion. The file has no individual observations, no earlier trend history and no sampling model. It cannot support a credible standard error or an empirical test of parallel trends. Do not attach a confidence interval invented from four aggregates.
A worthwhile extension would add a clearly specified simulated panel with a known data-generating process, then show how an omitted time-varying factor changes the estimate. Keep the simulation's assumptions separate from any claim about actual markets. Another extension is a public descriptive dataset, but the move from fictional to observed data requires new provenance and permission checks.
Present a small, inspectable package
Deliver the raw file, dictionary, reproducible calculation, validation log and a one-page memo. The memo should state the question, result, limitations and next evidence needed. Add a short README telling a reviewer how to regenerate the output and which software version you used. A reviewer should not have to guess which of six final files is authoritative.
Analysis Group's published description includes quantitative and statistical work; this project is our independent preparation suggestion, not an employer-approved assessment. In an application, describe what you personally built and checked. A truthful line might say that you produced a reproducible fictional comparison-group analysis and documented identification limits. It should not claim that you proved a real business effect or advised a client.
Before sharing, remove personal identifiers and verify any license or course restrictions. An anonymized label does not necessarily make a private dataset safe to publish. When permission is unclear, use the fictional inputs supplied here instead.
Data and worked inputs
Sources[1]