# Preventive trial withdrawal rates

Recently I’ve been doing some thinking about the nuts and bolts of preventive clinical trials in genetic prion disease, and among other things, doing some power calculations. This led me to the following question: what is a reasonable withdrawal rate to assume in a clinical trial? In other words, what percent of people in a clinical trial either formally withdraw, or are lost to follow-up, or simply stop taking their assigned drugs, per year of the trial? This proved surprisingly hard to answer in either a quick Google search or a quick Google Scholar search, so I spent a few hours digging deep, and now I am reporting back.

Withdrawal rate must depend on a lot of different variables. How tolerable is the drug? How onerous are the clinical visits? What is the study population? How aggressive are the study investigators about following up? There’s no doubt that different trials will have very different withdrawal rates, but I wanted at least a ballpark range to start from.

In my search for a reasonable number, I started out by looking at trials in neurodegenerative diseases. For example, the recently reported aducanumab trial in Alzheimer’s saw 40 out of 165 people (24%) withdraw in the course of a 1-year trial [Sevigny 2016]. (Note that in this post I will use the term “withdraw” to include any reason for not completing the trial in compliance with the study regimen). But it occurred to me that symptomatic trials in neurodegeneration may not be the best comparison for what I want to model. I’m thinking about a preventive trial in healthy, completely asymptomatic people. Most Alzheimer’s trials focus on people who already have dementia, which might make trial participation more onerous.

My next thought was to look specifically for preventive trials in asymptomatic people in neurological indications. I could only find two examples that were already completed and had reported results, though. One study, ADAPT, enrolled cognitively normal people age ≥70 with a family history of Alzheimer’s, and randomized them to celecoxib, naproxen, or placebo for one year [ADAPT Research Group 2008]. They had 411 of 2528 people (16%) withdraw. The other example I found was PRECREST, a study of creatine in pre-symptomatic people with a Huntington’s disease mutation [Rosas 2014]. In that trial, 21 out of 64 people withdrew over only 6 months. The withdrawal rate was higher in the creatine group (13/32 withdrew) than placebo (8/32), perhaps due to tolerability issues, but any way you slice it, it’s a pretty high withdrawal rate. I decided to convert everything to annual rates for comparison. If you calculate it as (21/64)/.5, that’s a 65% annual withdrawal rate, but people who do biostatistics for clinical trials all use exponential models (you need to account for compounding), so the right formula is actually w = 1-exp(log(A)/t) where w is the withdrawal rate, A is the proportion of people who complete the trial, and t is the trial length. In this case, 1-exp(log(43/64)/.5) works out to a 55% annual withdrawal rate.

I couldn’t find any more examples in neurology, so in a search for more data points, I cast further afield, and decided to also look at preventive trials in cardiology. Here’s what I found, sorted by withdrawal rate low to high and including the above examples:

category | trial | drug | how data are described | annual withdrawal rate | ref |
---|---|---|---|---|---|

cardiology | WOSCOPS | pravastatin | 6,695 randomized; cumulative withdrawal rate 25.2% (placebo) and 24.7% (drug) at year 4 | 6.9% | Shepherd 1995 |

cardiology | AFCAPS/TexCAPS | lovastatin | 6,605 randomized; 71% compliance in drug group, 63% in placebo group | 7.4% | Downs 1998 |

cardiology | OSLER | evolocumab | 4,465 randomized; 213+101 of 2976 on drug discontinued or withdrew; 59 of 1489 on standard of care discontinued or withdrew; median follow-up was 11.1 months | 9.0% | Sabatine 2015 |

cardiology | JUPITER | rosuvastatin | 17,802 randomized; “at the time the study was terminated, 75% of participants were taking their study pills”; median follow-up was 1.9 years | 14.1% | Ridker 2008 |

neurology | ADAPT | celecoxib / naproxen | 2,528 randomized, 411 did not contribute to analyses for various reasons, 1 year trial | 16.3% | ADAPT Research Group 2008 |

cardiology | ODYSSEY LONG TERM | alirocumab | 2,341 randomized; 1113/1553 (drug) and 595/788 (placebo) completed 78 week trial | 19.0% | Robinson 2015 |

cardiology | NCT00607373 | mipomersen | 51 randomized; 6/34 (drug) and 0/17 (placebo) discontinued in 26-week trial | 22.1% | Raal 2010 |

neurology | PRECREST | creatine | 64 randomized, 43 completed 6 months | 54.9% | Rosas 2014 |

As you can see, most of the cardiology examples are huge trials which used cardiac event endpoints. For comparison, though, I did include one rare disease drug, mipomersen, a subcutaneously delivered antisense oligonucleotide for people with homozygous *LDLR* mutations. That trial was brief (half a year) and had a biomarker endpoint (LDL cholesterol). You could imagine that possibly people with rare diseases might be more motivated to continue a trial and so would have a lower withdrawal rate, but this trial actually had the highest withdrawal rate of any cardiology trials considered here. 4 out of those 6 withdrawals were due to adverse events.

Although there’s a wide variability in withdrawal rates here, this at least gives some sort of range to start from. Statins are just about as tolerable and easy to take as any drug, so it seems like ~7% is probably the bare minimum withdrawal rate one could expect in any trial. On the high end, the PRECREST trial shows you could lose half your patients in a year. When modeling a trial, it’s probably best to consider a range of scenarios.