How many participants do I need for a clinical study?
Sample size follows from the research question, the primary outcome, the design and the planned analysis. What drives the number, how designs differ, a worked example, and allowing for dropout.
Blog and insights
Short, practical notes on statistical methods and research practice, written for researchers rather than statisticians.
About this blog
The JH Stat blog covers applied statistics, health research methods, responsible use of AI in health research and reproducible analysis.
The first article is published below. The remaining cards are planned topics, shown to indicate the kind of writing this section will contain — they are clearly marked and cannot yet be read.
If you would like to be told when new articles are published, or if there is a topic you would find useful, get in touch.
Planned categories
Writing will be organised under the following headings.
- Applied statistics
- Health research methods
- Responsible AI in health research
- Statistical reporting
- Reproducible research
Articles
Articles and topics in preparation
Published articles can be read in full. Cards marked in preparation have not been written yet.
Why complete-case analysis is rarely the right default
Dropping incomplete records is the most common way missing data are handled, and usually the least defensible. This piece will set out the mechanisms behind missingness and what changes when multiple imputation is used instead.
When machine learning beats regression — and when it does not
Prediction models in health research are increasingly built with machine learning methods. This piece will look at where that genuinely improves performance, where a well-specified regression does just as well, and why calibration matters more than discrimination for clinical use.
Have a question that deserves an article?
If there is a statistical problem your team keeps running into, it is probably worth writing about. Suggestions are welcome — and if you need an answer sooner than the blog can provide one, ask directly.