The result should survive a second pair of hands.

A medicines dataset can pass through downloads, exclusions, classifications, joins and calculations before it becomes a headline. Small choices along that journey can change its interpretation. Our proposed work examines those choices explicitly. The aim is a useful analytical record linking the question, the source and the transformation, so a reviewer can distinguish an observed value from a derived measure and an assumption from an established fact.

We propose an independently reproduced analysis using the same frozen source files and a documented specification. One analyst could prepare the initial result while another follows the specification without relying on the first person's working notes. Differences would be reconciled at the level of filters, units and calculations. The comparator is the current analytical workflow, including the time needed to resolve disagreements and make the result usable by someone else.

A robust analysis should cope with missing categories, renamed fields, changed coverage and rounding without silently producing a plausible mistake. We would create checks around totals, denominators and permitted transformations. Sensitivity analysis would show what happens when reasonable classification choices differ. The goal is not a perfect-looking line: it is an explanation that survives scrutiny and makes clear which conclusions are stable and which depend on a particular choice.

UKRI's data-sharing guidance allows for legitimate restrictions, while its integrity guidance sets expectations for responsible research conduct (UKRI, n.d.; UKRI, n.d.). Our project-specific plan would establish licences, attribution and what can be shared before processing begins. Public source data does not automatically permit unrestricted use of every accompanying asset. Any organisational data introduced later would require a separate review of access, purpose and disclosure rather than inheriting the public-data arrangement.

The first collaboration should end with an independent reproduction report and a list of unresolved differences. We would expand only if the analysis can be recreated, its limitations remain understandable and the update effort is proportionate to the decision it supports. We would stop or narrow the question if source changes cannot be reconciled, definitions make comparison misleading or the apparent insight disappears under reasonable alternative assumptions. A corrected conclusion is a valuable result.

One useful output would distinguish changes in item counts from changes in total ingredient cost, then show where the direction differs across categories. A user could inspect the underlying totals and definitions before choosing a more detailed question. That exploration should explain the observed pattern without pretending to identify its cause from national aggregates alone.