Statistical Thinking in Epidemiology
Published July 27th 2011 by Chapman and Hall/CRC – 231 pages
While biomedical researchers may be able to follow instructions in the manuals accompanying the statistical software packages, they do not always have sufficient knowledge to choose the appropriate statistical methods and correctly interpret their results. Statistical Thinking in Epidemiology examines common methodological and statistical problems in the use of correlation and regression in medical and epidemiological research: mathematical coupling, regression to the mean, collinearity, the reversal paradox, and statistical interaction.
Statistical Thinking in Epidemiology is about thinking statistically when looking at problems in epidemiology. The authors focus on several methods and look at them in detail: specific examples in epidemiology illustrate how different model specifications can imply different causal relationships amongst variables, and model interpretation is undertaken with appropriate consideration of the context of implicit or explicit causal relationships. This book is intended for applied statisticians and epidemiologists, but can also be very useful for clinical and applied health researchers who want to have a better understanding of statistical thinking.
Throughout the book, statistical software packages R and Stata are used for general statistical modeling, and Amos and Mplus are used for structural equation modeling.
Uses of Statistics in Medicine and Epidemiology
Structure and Objectives of This Book
Nomenclature in This Book
Vector Geometry of Linear Models for Epidemiologists
Basic Concepts of Vector Geometry in Statistics
Correlation and Simple Regression in Vector Geometry
Linear Multiple Regression in Vector Geometry
Significance Testing of Correlation and Simple Regression in Vector Geometry
Significance Testing of Multiple Regression in Vector Geometry
Path Diagrams and Directed Acyclic Graphs
Directed Acyclic Graphs
Direct and Indirect Effects
Mathematical Coupling and Regression to the Mean in the Relation between Change and Initial Value
Why Should Change Not Be Regressed on Initial Value? A Review of the Problem
Proposed Solutions in the Literature
Comparison between Oldham’s Method and Blomqvist’s Formula
Oldham’s Method and Blomqvist’s Formula Answer Two Different Questions
What Is Galton’s Regression to the Mean?
Testing the Correct Null Hypothesis
Evaluation of the Categorisation Approach
Testing the Relation between Changes and Initial Values When There Are More than Two Occasions
Analysis of Change in Pre-/Post-Test Studies
Analysis of Change in Randomised Controlled Trials
Comparison of Six Methods
Analysis of Change in Non-Experimental Studies: Lord’s Paradox
ANCOVA and t-Test for Change Scores Have Different Assumptions
Collinearity and Multicollinearity
Introduction: Problems of Collinearity in Linear Regression
Mathematical Coupling and Collinearity
Vector Geometry of Collinearity
Geometrical Illustration of Principal Components Analysis as a Solution to Multicollinearity
Example: Mineral Loss in Patients Receiving Parenteral Nutrition
Solutions to Collinearity
Is ‘Reversal Paradox’ a Paradox?
A Plethora of Paradoxes: The Reversal Paradox
Background: The Foetal Origins of Adult Disease
Hypothesis (Barker’s Hypothesis)
Vector Geometry of the Foetal Origins Hypothesis
Reversal Paradox and Adjustment for Current Body Size: Empirical Evidence from Meta-Analysis
Testing Statistical Interaction
Introduction: Testing Interactions in Epidemiological Research
Testing Statistical Interaction between Categorical Variables
Testing Statistical Interaction between Continuous Variables
Partial Regression Coefficient for Product Term in Regression Models
Categorization of Continuous Explanatory Variables
The Four-Model Principle in the Foetal Origins Hypothesis
Categorization of Continuous Covariates and Testing Interaction
Finding Growth Trajectories in Lifecourse Research
Current Approaches to Identifying Postnatal Growth Trajectories in Lifecourse Research
Partial Least Squares Regression for Lifecourse Research
Dr Yu-Kang Tu is a Senior Clinical Research Fellow in the Division of Biostatistics, School of Medicine, and in the Leeds Dental Institute, University of Leeds, Leeds, UK. He was a visiting Associate Professor to the National Taiwan University, Taipei, Taiwan. First trained as a dentist and then an epidemiologist, he has published extensively in dental, medical, epidemiological and statistical journals. He is interested in developing statistical methodologies to solve statistical and methodological problems such as mathematical coupling, regression to the mean, collinearity and the reversal paradox. His current research focuses on applying latent variables methods, e.g. structural equation modeling, latent growth curve modelling, and lifecourse epidemiology. More recently, he has been working on applying partial least squares regression to epidemiological data.
Prof Mark S Gilthorpe is professor of Statistical Epidemiology, Division of Biostatistics, School of Medicine, University of Leeds, Leeds, UK. Having completed a single honours degree in mathematical Physics (University of Nottingham), he undertook a PhD in Mathematical Modelling (University of Aston in Birmingham), before initially embarking upon a career as self-employed Systems and Data Analyst and Computer Programmer, and eventually becoming an academic in biomedicine. Academic posts include systems and data analyst of UK regional routine hospital data in the Department of Public Health and Epidemiology, University of Birmingham; Head of Biostatistics at the Eastman Dental Institute, University College London; and founder and Head of the Division of Biostatistics, School of Medicine, University of Leeds. His research focus has persistently been that of the development and promotion of robust and sophisticated modelling methodologies for non-experimental (and sometimes large and complex) observational data within biomedicine, leading to extensive publications in dental, medical, epidemiological and statistical journals.