Estimands, Estimators and Sensitivity Analysis in Clinical Trials

Craig Mallinckrodt, Geert Molenberghs, Ilya Lipkovich, Bohdana Ratitch

• Publisher: Chapman and Hall/CRC
• Year: 2019
• ISBN:
  9780429950056
• Edition: 1
• Topic: Medicine
• Subtopic: Biostatistics

The concepts of estimands, analyses (estimators), and sensitivity are interrelated. Therefore, great need exists for an integrated approach to these topics. This book acts as a practical guide to developing and implementing statistical analysis plans by explaining fundamental concepts using accessible language, providing technical details, real-world examples, and SAS and R code to implement analyses. The updated ICH guideline raises new analytic and cross-functional challenges for statisticians. Gaps between different communities have come to surface, such as between causal inference and clinical trialists, as well as among clinicians, statisticians, and regulators when it comes to communicating decision-making objectives, assumptions, and interpretations of evidence.

This book lays out a path toward bridging some of these gaps. It offers

A common language and unifying framework along with the technical details and practical guidance to help statisticians meet the challenges

A thorough treatment of intercurrent events (ICEs), i.e., postrandomization events that confound interpretation of outcomes and five strategies for ICEs in ICH E9 (R1)

Details on how estimands, integrated into a principled study development process, lay a foundation for coherent specification of trial design, conduct, and analysis needed to overcome the issues caused by ICEs:

A perspective on the role of the intention-to-treat principle

Examples and case studies from various areas

Example code in SAS and R

A connection with causal inference

Implications and methods for analysis of longitudinal trials with missing data

Together, the authors have offered the readers their ample expertise in clinical trial design and analysis, from an industrial and academic perspective.

Section IV

Technical Details on Selected Analyses

Each chapter in Section IV provides technical details on an analytic approach that can be used as the main estimator or a sensitivity analysis consistent with one or more of the five strategies for dealing with intercurrent events as outlined in the ICH E9(R1) Addendum. An example data set is used, and the example code is provided to implement the analyses.

17

Example Data

17.1 Introduction

Throughout this section, technical details of selected analyses are illustrated using an example data set that includes 50 patients (25 per arm) with three post-baseline assessments. This data set is based on data from real clinical trial subjects but represents a subset of the actual trial-enrolled population. The following section provides more details on the data set and how it was created. Aspects of these data are admittedly arbitrary, and the intent is not to mimic any specific clinical trial setting but rather to provide a data set for illustration and convenience. A listing of these data is provided for those readers who wish to replicate and/or expand on the analyses presented here.

17.2 Details of Example Data Set

Two versions of the data set were created. The first version (complete data) had complete data where all patients adhered to the originally assigned study medication. The second version (missing data) was identical to the first except some data were missing such as would arise from patient dropout. Each data set had 50 subjects, 25 per arm, and 3 post-baseline assessments. A complete listing of these data sets is provided at the end of this chapter. The outcome data comprised of the HAMD17 (Hamilton 17-item rating scale for depression; Hamilton 1960) and Patient Global Impression of Improvement scale (PGIIMP; Guy 1976). The HAMD is a continuous variable. The PGI has seven ordered categories from “very much improved” (1) to “very much worse” (7), with “not improved” corresponding to the midpoint score =4.

17.2.1 Complete Data Set

The complete data set was created by extracting patients from a clinical trial in major depressive disorder (Detke et al., 2004). All subjects that completed the trial and were from the investigational site with the largest enrollment were selected. Additional subjects, who were also completers, were selected...