Show Notes
How do you decide whether a clinical trial “worked”? In Part 2 of our Galleri series, we examine the landmark randomized trial of a blood test designed to detect more than 50 cancers. We explore why different outcome measures led to dramatically different headlines, discuss primary versus secondary outcomes, pre-registration, hierarchical testing, and post hoc analyses, and explain why mortality remains the outcome everyone is waiting for. Along the way, we uncover a statistical mystery involving dozens of missing cancers and discover how a little arithmetic can sometimes reveal more than a press release. Statistical topics cancer screening exploratory analyses hierarchical testing missing data multiple testing outcome measures post hoc analyses pre-registration primary and secondary outcomes randomized clinical trials screening tests
Methodologic Morals “When the simple numbers don't add up, pay attention. The arithmetic may be trying to tell you something.” “The first question should not be, did it work? It should be, what counts as success?”
References Giridhar KV, et al. Safety and performance results from PATHFINDER 2, a registrational study of a multi-cancer early detection test in an intended-use population. Presented at the 2026 American Society of Clinical Oncology (ASCO) Annual Meeting. May 2026. Hubbell E, Clarke CA, Aravanis AM, Berg CD. Modeled Reductions in Late-stage Cancer with a Multi-Cancer Early Detection Test. Cancer Epidemiol Biomarkers Prev. 2021;30(3):460-468. doi:10.1158/1055-9965.EPI-20-1134 Neal RD, Johnson P, Clarke CA, et al. Cell-Free DNA-Based Multi-Cancer Early Detection Test in an Asymptomatic Screening Population (NHS-Galleri): Design of a Pragmatic, Prospective Randomised Controlled Trial. Cancers (Basel). 2022;14(19):4818. Published 2022 Oct 1. doi:10.3390/cancers14194818 ASCO slides: https://grail.com/wp-content/uploads/2026/05/Swanton_ASCO-2026_NHS-Galleri_FINAL-Slides-05.26.2026.pdf UK registry protocol: https://www.isrctn.com/ISRCTN91431511 Clinicaltrials.gov protocol: https://clinicaltrials.gov/study/NCT05611632
Common biases in cancer screening studies Cancer screening studies are subject to several well-known biases that can make a screening test appear more effective than it actually is. Three of the most important are: Lead-time bias: Screening advances the time of diagnosis, making survival from diagnosis appear longer even if the patient's lifespan is unchanged. For example, if a screening test detects a Stage II cancer at age 60 that otherwise would have been diagnosed because of symptoms at age 62, but the patient dies at age 68 regardless, survival from diagnosis appears to increase from 6 years to 8 years even though the patient did not live any longer. Length bias: Screening preferentially detects slower-growing, less aggressive cancers because they remain detectable for longer than fast-growing cancers. For example, a slow-growing cancer that remains in Stage I for 5 years is much more likely to be found by screening than an aggressive cancer that progresses to symptoms within months. This can make screened patients appear to have better survival simply because screening preferentially found the less aggressive cancers. Overdiagnosis: Screening detects cancers that would never have caused symptoms or death during a person's lifetime, leading to unnecessary diagnosis and treatment. For example, a screening test may detect a very slow-growing prostate or thyroid cancer in an older adult that would never have become clinically important if it had remained undiscovered.
Kristin and Regina’s online courses:
Demystifying Data: A Modern Approach to Statistical Understanding Clinical Trials: Design, Strategy, and Analysis Medical Statistics Certificate Program Writing in the Sciences Epidemiology and Clinical Research Graduate Certificate Program Programs that we teach in: Epidemiology and Clinical Research Graduate Certificate Program
Find us on: Kristin - LinkedIn
More from Normal Curves: Sexy Science, Serious Statistics
A famous Harvard researcher. A blockbuster finding about honesty. Suspicious data, failed replications, and a $25 million lawsuit. Welcome to one of the strangest research misconduct scandals in recent memory. In the first of two episodes about the Francesca Gino case, we investigate a famous experiment claiming that signing an honesty pledge at the top of a form makes people less likely to cheat. We follow the trail from a failed replication to some extraordinary detective work involving an Excel spreadsheet—and conduct a little statistical sleuthing of our own. Along the way, we discover how suspicious changes to a handful of observations can alter the story a scientific study tells. Statistical topics Statistical sleuthing Data integrity Data management Replication Sensitivity analysis Effect sizes Statistical significance Reproducibility Spreadsheets Conceptual replication Direct replication CalcChain Methodologic Morals “There's cleaning your data, and then there's laundering your data.” “Good data record keeping is still sexy. Version control, baby.”
Can you spot problems in a scientific study just by reading the published paper? We put our statistical sleuthing skills to the test with a famous study on procrastination and deadlines that was retracted more than 20 years after it was published. But were there clues hidden in plain sight in the published results—and would we have noticed anything suspicious at the time? We put on our data-detective hats and find a few statistical clues of our own before turning to Data Colada’s investigation of the original data, where the story gets even stranger. Along the way, we hunt for missing participants, question suspiciously enormous effects, meet some unlikely data twins, and discover why sometimes the best statistical tool is simply asking: Do these data behave like real people?
What happens when you discover that your published paper is wrong? Does retracting it ruin your reputation—or could owning the mistake actually make things better? In this episode, we meet the inaugural winners of the Ctrl-Z Awards, which recognize scientists for correcting or retracting their own work. Behavioral ecologist Kate Laskowski takes us inside a scientific detective story involving suspicious spider data, three retracted papers, and the fear that going public might tank her career. Then computational social scientist Alyssa Smith tells us how one small error in a figure led her to retract the paper at the center of her PhD and job talk. Along the way, we talk about data forensics, the stigma around retractions, and why self-correcting scientists are the heart of science.
Statistical power is one of the most important ideas in statistics—and one of the most misunderstood. In this episode, we explain what statistical power really means, why researchers need to think about it before they collect their data, and what goes wrong when they calculate “post hoc power” after the study is over. Along the way, we hunt for UFOs, test Kristin’s psychic abilities, explore the ethics of underpowered studies, and discover why post hoc power is really just a p-value wearing a fake mustache. Then statistician and powerlifter Andrew Althouse joins us for the inside story of a years-long scientific feud over post hoc power, where we discuss what happens when being wrong about statistics turns into doubling down, angry letters to the editor, and debates about how science corrects itself.
Can weed protect your brain as you get older? We unpack a study linking lifetime cannabis use to larger brain volumes and better cognitive test scores in older adults—and investigate how those findings became headlines urging readers to “pass the pot to grandma.” Along the way, we explore collapsing categorical variables, confounding, clinically versus statistically significant results, multiple testing, and how a press release subtly changed a story. We also meet permanently stoned mice, learn about Carl Sagan’s surprising connection to cannabis research, and discover how a single missing word can change the entire meaning of a study. Statistical topics Categorical variables Confounding Effect size False discovery rate Healthy volunteer bias Measurement error Multiple testing Observational studies Recoding variables Science communication Statistical vs practical significance