Article · AI & ML
AI-assisted programming needs an evidence trail
The valuable question is not how quickly a model writes code. It is how confidently a team can review the result.
Read insightTHE WORKBENCH
Ideas worth putting to work.
Practical perspectives on SAS, R, data science, clinical programming and the responsible use of AI.
Article · R & Shiny
From a Shiny prototype to a trusted review tool
A dashboard is only useful when users understand the data, the calculations and the decisions it supports.
Read insightArticle · Real-world evidence
Real-world data: define the cohort before the model
Index dates, eligibility and follow-up can change the answer before a statistical model is fitted.
Read insightArticle · AI & ML
What I would examine before an AI model goes live
A practical technical review looks beyond a headline accuracy score to the evidence supporting a particular use.
Read insightArticle · Data science
The quiet ways data leakage inflates model performance
Split by the unit that matters, and fit preprocessing only where it belongs.
Read insightArticle · R
Reproducible R is more than set.seed()
Capture the environment, inputs and execution order—not just the random-number seed.
Read insightMini blog · SAS
SAS tip: make BY-group assumptions explicit
FIRST. and LAST. are useful only when your grouping and ordering reflect the intended derivation.
Read insightMini blog · SAS
SAS tip: check row counts before trusting a join
Unexpected duplication is often a key-definition problem rather than a syntax problem.
Read insightMini blog · SAS
SAS tip: keep partial dates and imputation separate
A technically valid date can still encode an undocumented assumption.
Read insightMini blog · SAS
SAS tip: a clean log is a starting point
Automated log scanning helps, but silent analytical errors need separate checks.
Read insightMini blog · R
R tip: declare the relationship in a dplyr join
Let a join check one of your assumptions before it changes your dataset.
Read insightMini blog · R
R tip: distinguish all-missing from zero
Removing missing values can produce a summary that looks more informative than the data.
Read insightMini blog · R & Shiny
Shiny tip: export the context with the table
An exported result should explain which population and filters produced it.
Read insightMini blog · CRAN watch
CRAN watch: ellmer 0.5.0 and controlled LLM experiments
A September 2026 release worth evaluating for structured language-model workflows in R.
Read insightMini blog · CRAN watch
CRAN watch: admiral 1.5.0 for ADaM workflows
Reusable derivations are valuable when the team can explain how they meet its study specification.
Read insightMini blog · CRAN watch
CRAN watch: pointblank 0.12.4 for explicit data checks
Turn quality expectations into readable checks with a defined response to failure.
Read insightMini blog · CDISC
CDISC watch: distinguish the roadmap from a released standard
A projected publication date should prompt planning, not an unreviewed standards migration.
Read insightMini blog · CDISC
CDISC tip: terminology updates deserve a change assessment
Changes to controlled terminology can affect more than a lookup table.
Read insightMini blog · Data science
Data science tip: assess calibration as well as ranking
A model can rank patients well while systematically overstating their predicted risk.
Read insightMini blog · Data science
Data science tip: a drift alert needs an owner
Monitoring is useful only when an alert leads to a defined investigation.
Read insightNo matching insights. Try another topic or a shorter search.
Published 20 September 2026. Technical perspectives and illustrative examples; assess suitability within your organisation’s own standards and review processes. Release-watch entries reflect the source on the date checked.