Investigators are faced with many challenges in designing efficient, ethical randomized trials due to competing demands: a trial must collect enough information to identify meaningful benefits or harms with a desired probability, while also minimizing potential harm and suboptimal treatment of participants. Satisfying these competing demands is further complicated by the limited and imprecise information available during the design of a study.
Studies designed around a fixed sample size are inflexible, requiring investigators to wait until the end of data collection to perform statistical analyses. Group sequential designs are more flexible, allowing studies to be stopped for efficacy or futility according to a pre-planned analyses. Covariate adjustment allows investigators to potentially gain additional precision and power from variables collected on individuals prior to randomization. Unfortunately, the amount of precision gained from covariate adjustment is not known precisely at the outset of a study, complicating the ability to use covariate adjustment to reduce the required sample size. Additionally, some methods of covariate adjustment do not provide the independent increments property required by group sequential design methods.
Rather than planning analyses based on a specific number of participants, investigators can pre-specify when analyses reach pre-specified levels of precision: this is known as information monitoring (Mehta and Tsiatis 2001). This allows investigators to adapt their study to the precision in the accruing data, reducing the risk of under- or overpowered trials. This also allows investigators to use covariate adjustment to shorten the trial duration, rather than just providing additional power and precision.
The independent increments property can also be obtained by orthogonalizing estimates and their covariance, allowing covariate adjustment to be included in both group sequential and information monitoring designs.
Installation
You can install the development version of impart from GitHub using:
# install.packages("devtools")
devtools::install_github("jbetz-jhu/impart")
#>
#> ── R CMD build ─────────────────────────────────────────────────────────────────
#> ✔ checking for file 'C:\Users\jbetz\AppData\Local\Temp\RtmpMxHB2F\remotes532460593196\jbetz-jhu-impart-f8934f5/DESCRIPTION'
#> ─ preparing 'impart':
#> checking DESCRIPTION meta-information ... checking DESCRIPTION meta-information ... ✔ checking DESCRIPTION meta-information
#> ─ checking for LF line-endings in source and make files and shell scripts
#> ─ checking for empty or unneeded directories
#> ─ building 'impart_0.1.0.tar.gz'
#>
#>
Vignettes
There are three vignettes on how to use impart
, each focusing on a different phase of a study. Each vignette is sequential, building upon the previous ones.
vignette("impart_study_design", package = "impart") # Design
vignette("impart_monitoring", package = "impart") # Information Monitoring
vignette("impart_analyses", package = "impart") # Analyses
Background for Information Monitoring
We can estimate the precision required to achieve power to identify a treatment effect with a -sided test with type I error rate using:
This uses the square of the empirical standard error (or the empirical variance estimate) to measure the precision to which the treatment effect can be measured with the data in hand. A precision-adaptive design can reduce the risk of under- or overpowered trials by collecting data until the precision is sufficient to conduct analyses.
Approximate Precision vs. Sample Size
Let denote treatment and denote control, and denote the outcome of interest under treatment assigment , where indicates assignment to the treatment arm and denotes assignment to the control arm.
For a continuous outcome, the required information to estimate the difference in means depends on the sample size and variances of outcomes in each treatment arm:
For a binary outcome, the required information to estimate the risk difference depends on the response rate in the control arm :
For an ordinal outcome with categories, let denote the probability of an outcome in category under treatment . The Mann-Whitney estimand is the probability of having an outcome as good or better under the treatment arm relative to control with an adjustment for ties:
This is also known as the competing probability. The precision/information depends on (Fay and Malinovsky 2018):
Alternatively, the precision/information can be obtained from the distribution of outcomes under each treatment arm (Zhao, Rahardja, and Qu 2008). Let :
Expressions for the information for other estimands can be obtained elsewhere (Jennison and Turnbull 1999). In practice, the parameters in these expressions are not precisely known a priori. The advantage of an information monitoring design is that the sample size is not fixed a priori based on estimates of these parameters, but adapts automatically to the precision of the accruing data.
Covariate Adjustment in Randomized Trials
Covariate adjusted analyses can also give greater precision than an unadjusted analyses without introducing more stringent assumptions, however the amount of precision gained in adjusted analyses are also not precisely known a priori (Benkeser et al. 2020). Instead of predicating the design on assumptions about the potential gain in precision from covariate adjustment, a precision-adaptive design automatically adjusts the sample size accordingly.
The relative efficiency of a covariate adjusted estimator to an unadjusted estimator is . The relative change in variance of a covariate-adjusted analysis to an unadjusted analysis is:
Alternatively, : If a covariate adjusted analysis has a relative efficiency of 1.25, the relative change in variance would be -0.2, or a -20% change in variance. Since precision is the inverse of variance, the relative change in precision of a covariate-adjusted analysis to an unadjusted analysis is:
Alternatively, : If a covariate adjusted analysis has a relative efficiency of 1.25, the relative change in precision would be 0.25, or a 25% change in precision.
Sequential Analyses
Pre-planned interim analyses allow investigators to stop a randomized trial early for efficacy or futility (Jennison and Turnbull 1999). Precision-adaptive trials can integrate both interim analyses and covariate adjustment, using a broad class of methods (Van Lancker, Betz, and Rosenblum 2022). Mehta and Tsiatis (2001) illustrate information-adaptive designs in practice. For a tutorial on implementing interim analyses, see Lakens, Pahlke, and Wassmer (2021).