Multivariable Model - Building: A Pragmatic Approach to Regression Anaylsis based on Fractional Polynomials for Modelling Continuous VariablesPatrick Royston, Willi Sauerbrei ISBN: 978-0-470-02842-1 322 pages John Wiley & Sons Ltd, Chichester, England May 2008
For datasets see the end of the page. For stata programs see our original website of the book. |
Everything on this page was reproduced with permission from John Wiley & Sons Ltd.
From the preface:
"[...] Our general objective is to provide a readable text giving the rationale of, and practical advice on, a unified approach to multivariable modelling which aims to make such models simpler and more effective. [...] No multivariable model-building strategy has rigorous theoretical underpinnings. Even those approaches most used in practice have not had their properties studied adequately by simulation. In particular, handling continuous variables in a multivariable context has largely been ignored. Since there is no consensus among researchers on the ‘best’ strategy, a pragmatic approach is required. Our book reflects our views derived from wide experience. The text assumes a basic understanding of multiple regression modelling, but it can be read without detailed mathematical knowledge. [...] As expressed in a very readable paper by Chatfield (2002), we aim to ‘encourage and guide practitioners, and also to counterbalance a literature that can be overly concerned with theoretical matters far removed from the day-to-day concerns of many working statisticians’. [...]"
Table of Contents
1 | Introduction | 1 | ||
1.1 | Real-Life Problems as Motivation for Model Building | 1 | ||
1.1.1 | Many Candidate Models | 1 | ||
1.1.2 | Functional Form for Continuous Predictors | 2 | ||
1.1.3 | Example 1: Continuous Response | 2 | ||
1.1.4 | Example 2: Multivariable Model for Survival Data | 5 | ||
1.2 | Issues in Modelling Continuous Predictors | 8 | ||
1.2.1 | Effects of Assumptions | 8 | ||
1.2.2 | Global versus Local Influence Models | 9 | ||
1.2.3 | Disadvantages of Fractional Polynomial Modelling | 9 | ||
1.2.4 | Controlling Model Complexity | 10 | ||
1.3 | Types of Regression Model Considered | 10 | ||
1.3.1 | Normal-Errors Regression | 10 | ||
1.3.2 | Logistic Regression | 12 | ||
1.3.3 | Cox Regression | 12 | ||
1.3.4 | Generalized Linear Models | 14 | ||
1.3.5 | Linear and Additive Predictors | 14 | ||
1.4 | Role of Residuals | 15 | ||
1.4.1 | Uses of Residuals | 15 | ||
1.4.2 | Graphical Analysis of Residuals | 15 | ||
1.5 | Role of Subject-Matter Knowledge in Model Development | 16 | ||
1.6 | Scope of Model Building in our Book | 17 | ||
1.7 | Modelling Preferences | 18 | ||
1.7.1 | General Issues | 18 | ||
1.7.2 | Criteria for a Good Model | 18 | ||
1.7.3 | Personal Preferences | 19 | ||
1.8 | General Notation | 20 |
2 | Variable Selection | 23 | ||
2.1 | Introduction | 23 | ||
2.2 | Background | 24 | ||
2.3 | Preliminaries for a Multivariable Analysis | 25 | ||
2.4 | Aims of Multivariable Models | 26 | ||
2.5 | Prediction: Summary Statistics and Comparisons | 29 | ||
2.6 | Procedures for Selecting Variables | 29 | ||
2.6.1 | Strength of Predictors | 30 | ||
2.6.2 | Stepwise Procedures | 31 | ||
2.6.3 | All-Subsets Model Selection Using Information Criteria | 32 | ||
2.6.4 | Further Considerations | 33 | ||
2.7 | Comparison of Selection Strategies in Examples | 35 | ||
2.7.1 | Myeloma Study | 35 | ||
2.7.2 | Educational Body-Fat Data | 36 | ||
2.7.3 | Glioma Study | 38 | ||
2.8 | Selection and Shrinkage | 40 | ||
2.8.1 | Selection Bias | 40 | ||
2.8.2 | Simulation Study | 40 | ||
2.8.3 | Shrinkage to Correct for Selection Bias | 42 | ||
2.8.4 | Post-estimation Shrinkage | 44 | ||
2.8.5 | Reducing Selection Bias | 45 | ||
2.8.6 | Example | 46 | ||
2.9 | Discussion | 47 | ||
2.9.1 | Model Building in Small Datasets | 47 | ||
2.9.2 | Full, Pre-specified or Selected Model? | 47 | ||
2.9.3 | Comparison of Selection Procedures | 49 | ||
2.9.4 | Complexity, Stability and Interpretability | 49 | ||
2.9.5 | Conclusions and Outlook | 50 |
3 | Handling Categorical and Continuous Predictors | 53 | ||
3.1 | Introduction | 53 | ||
3.2 | Types of Predictor | 54 | ||
3.2.1 | Binary | 54 | ||
3.2.2 | Nominal | 54 | ||
3.2.3 | Ordinal, Counting, Continuous | 55 | ||
3.2.4 | Derived | 55 | ||
3.3 | Handling Ordinal Predictors | 55 | ||
3.3.1 | Coding Schemes | 55 | ||
3.3.2 | Effect of Coding Schemes on Variable Selection | 56 | ||
3.4 | Handling Counting and Continuous Predictors: Categorization | 58 | ||
3.4.1 | ‘Optimal’ Cutpoints: A Dangerous Analysis | 58 | ||
3.4.2 | Other Ways of Choosing a Cutpoint | 59 | ||
3.5 | Example: Issues in Model Building with Categorized Variables | 60 | ||
3.5.1 | One Ordinal Variable | 61 | ||
3.5.2 | Several Ordinal Variables | 62 | ||
3.6 | Handling Counting and Continuous Predictors: Functional Form | 64 | ||
3.6.1 | Beyond Linearity | 64 | ||
3.6.2 | Does Nonlinearity Matter? | 65 | ||
3.6.3 | Simple versus Complex Functions | 66 | ||
3.6.4 | Interpretability and Transportability | 66 | ||
3.7 | Empirical Curve Fitting | 67 | ||
3.7.1 | General Approaches to Smoothing | 68 | ||
3.7.2 | Critique of Local and Global Influence Models | 68 | ||
3.8 | Discussion | 69 | ||
3.8.1 | Sparse Categories | 69 | ||
3.8.2 | Choice of Coding Scheme | 69 | ||
3.8.3 | Categorizing Continuous Variables | 70 | ||
3.8.4 | Handling Continuous Variables | 70 |
4 | Fractional Polynomials for One Variable | 71 | ||
4.1 | Introduction | 72 | ||
4.2 | Background | 72 | ||
4.2.1 | Genesis | 72 | ||
4.2.2 | Types of Model | 73 | ||
4.2.3 | Relation to Box–Tidwell and Exponential Functions | 73 | ||
4.3 | Definition and Notation | 74 | ||
4.3.1 | Fractional Polynomials | 74 | ||
4.3.2 | First Derivative | 74 | ||
4.4 | Characteristics | 75 | ||
4.4.1 | FP1 and FP2 Functions | 75 | ||
4.4.2 | Maximum or Minimum of a FP2 Function | 75 | ||
4.5 | Examples of Curve Shapes with FP1 and FP2 Functions | 76 | ||
4.6 | Choice of Powers | 78 | ||
4.7 | Choice of Origin | 79 | ||
4.8 | Model Fitting and Estimation | 79 | ||
4.9 | Inference | 79 | ||
4.9.1 | Hypothesis Testing | 79 | ||
4.9.2 | Interval Estimation | 80 | ||
4.10 | Function Selection Procedure | 82 | ||
4.10.1 | Choice of Default Function | 82 | ||
4.10.2 | Closed Test Procedure for Function Selection | 82 | ||
4.10.3 | Example | 83 | ||
4.10.4 | Sequential Procedure | 83 | ||
4.10.5 | Type I Error and Power of the Function Selection Procedure | 84 | ||
4.11 | Scaling and Centering | 84 | ||
4.11.1 | Computational Aspects | 84 | ||
4.11.2 | Examples | 85 | ||
4.12 | FP Powers as Approximations to Continuous Powers | 85 | ||
4.12.1 | Box–Tidwell and Fractional Polynomial Models | 85 | ||
4.12.2 | Example | 85 | ||
4.13 | Presentation of Fractional Polynomial Functions | 86 | ||
4.13.1 | Graphical | 86 | ||
4.13.2 | Tabular | 87 | ||
4.14 | Worked Example | 89 | ||
4.14.1 | Details of all Fractional Polynomial Models | 89 | ||
4.14.2 | Function Selection | 90 | ||
4.14.3 | Details of the Fitted Model | 90 | ||
4.14.4 | Standard Error of a Fitted Value | 91 | ||
4.14.5 | Fitted Odds Ratio and its Confidence Interval | 91 | ||
4.15 | Modelling Covariates with a Spike at Zero | 92 | ||
4.16 | Power of Fractional Polynomial Analysis | 94 | ||
4.16.1 | Underlying Function Linear | 95 | ||
4.16.2 | Underlying Function FP1 or FP2 | 95 | ||
4.16.3 | Comment | 96 | ||
4.17 | Discussion | 97 |
5 | Some Issues with Univariate Fractional Polynomial Models | 71 | ||
5.1 | Introduction | 99 | ||
5.2 | Susceptibility to Influential Covariate Observations | 100 | ||
5.3 | A Diagnostic Plot for Influential Points in FP Models | 100 | ||
5.3.1 | Example 1: Educational Body-Fat Data | 101 | ||
5.3.2 | Example 2: Primary Biliary Cirrhosis Data | 101 | ||
5.4 | Dependence on Choice of Origin | 103 | ||
5.5 | Improving Robustness by Preliminary Transformation | 105 | ||
5.5.1 | Example 1: Educational Body-Fat Data | 106 | ||
5.5.2 | Example 2: PBC Data | 107 | ||
5.5.3 | Practical Use of the Pre-transformation gδ(x) | 107 | ||
5.6 | Improving Fit by Preliminary Transformation | 108 | ||
5.6.1 | Lack of Fit of Fractional Polynomial Models | 108 | ||
5.6.2 | Negative Exponential Pre-transformation | 108 | ||
5.7 | Higher Order Fractional Polynomials | 109 | ||
5.7.1 | Example 1: Nerve Conduction Data | 109 | ||
5.7.2 | Example 2: Triceps Skinfold Thickness | 110 | ||
5.8 | When Fractional Polynomial Models are Unsuitable | 111 | ||
5.8.1 | Not all Curves are Fractional Polynomials | 111 | ||
5.8.2 | Example: Kidney Cancer | 112 | ||
5.9 | Discussion | 113 |
6 | MFP: Multivariable Model-Building with Fractional Polynomials | 115 | ||
6.1 | Introduction | 115 | ||
6.2 | Motivation | 116 | ||
6.3 | The MFP Algorithm | 117 | ||
6.3.1 | Remarks | 118 | ||
6.3.2 | Example | 118 | ||
6.4 | Presenting the Model | 120 | ||
6.4.1 | Parameter Estimates | 120 | ||
6.4.2 | Function Plots | 121 | ||
6.4.3 | Effect Estimates | 121 | ||
6.5 | Model Criticism | 123 | ||
6.5.1 | Function Plots | 123 | ||
6.5.2 | Graphical Analysis of Residuals | 124 | ||
6.5.3 | Assessing Fit by Adding More Complex Functions | 125 | ||
6.5.4 | Consistency with Subject-Matter Knowledge | 129 | ||
6.6 | Further Topics | 129 | ||
6.6.1 | Interval Estimation | 129 | ||
6.6.2 | Importance of the Nominal Significance Level | 130 | ||
6.6.3 | The Full MFP Model | 131 | ||
6.6.4 | A Single Predictor of Interest | 132 | ||
6.6.5 | Contribution of Individual Variables to the Model Fit | 134 | ||
6.6.6 | Predictive Value of Additional Variables | 136 | ||
6.7 | Further Examples | 138 | ||
6.7.1 | Example 1: Oral Cancer | 138 | ||
6.7.2 | Example 2: Diabetes | 139 | ||
6.7.3 | Example 3: Whitehall I | 140 | ||
6.8 | Simple Versus Complex Fractional Polynomial Models | 144 | ||
6.8.1 | Complexity and Modelling Aims | 144 | ||
6.8.2 | Example: GBSG Breast Cancer Data | 144 | ||
6.9 | Discussion | 146 | ||
6.9.1 | Philosophy of MFP | 147 | ||
6.9.2 | Function Complexity, Sample Size and Subject-Matter Knowledge | 148 | ||
6.9.3 | Improving Robustness by Preliminary Covariate Transformation | 148 | ||
6.9.4 | Conclusion and Future | 149 |
7 | Interactions | 151 | ||
7.1 | Introduction | 151 | ||
7.2 | Background | 152 | ||
7.3 | General Considerations | 152 | ||
7.3.1 | Effect of Type of Predictor | 152 | ||
7.3.2 | Power | 153 | ||
7.3.3 | Randomized Trials and Observational Studies | 153 | ||
7.3.4 | Predefined Hypothesis or Hypothesis Generation | 153 | ||
7.3.5 | Interactions Caused by Mismodelling Main Effects | 154 | ||
7.3.6 | The ‘Treatment–Effect’ Plot | 154 | ||
7.3.7 | Graphical Checks, Sensitivity and Stability Analyses | 154 | ||
7.3.8 | Cautious Interpretation is Essential | 155 | ||
7.4 | The MFPI Procedure | 155 | ||
7.4.1 | Model Simplification | 156 | ||
7.4.2 | Check of the Results and Sensitivity Analysis | 156 | ||
7.5 | Example 1: Advanced Prostate Cancer | 157 | ||
7.5.1 | The Fitted Model | 158 | ||
7.5.2 | Check of the Interactions | 160 | ||
7.5.3 | Final Model | 161 | ||
7.5.4 | Further Comments and Interpretation | 162 | ||
7.5.5 | FP Model Simplification | 163 | ||
7.6 | Example 2: GBSG Breast Cancer Study | 163 | ||
7.6.1 | Oestrogen Receptor Positivity as a Predictive Factor | 163 | ||
7.6.2 | A Predefined Hypothesis: Tamoxifen–Oestrogen Receptor Interaction | 163 | ||
7.7 | Categorization | 165 | ||
7.7.1 | Interaction with Categorized Variables | 165 | ||
7.7.2 | Example: GBSG Study | 166 | ||
7.8 | STEPP | 167 | ||
7.9 | Example 3: Comparison of STEPP with MFPI | 168 | ||
7.9.1 | Interaction in the Kidney Cancer Data | 168 | ||
7.9.2 | Stability Investigation | 168 | ||
7.10 | Comment on Type I Error of MFPI | 171 | ||
7.11 | Continuous-by-Continuous Interactions | 172 | ||
7.11.1 | Mismodelling May Induce Interaction | 173 | ||
7.11.2 | MFPIgen: An FP Procedure to Investigate Interactions | 174 | ||
7.11.3 | Examples of MFPIgen | 175 | ||
7.11.4 | Graphical Presentation of Continuous-by-Continuous Interactions | 179 | ||
7.11.5 | Summary | 180 | ||
7.12 | Multi-Category Variables | 181 | ||
7.13 | Discussion | 181 |
8 | Model Stability | 183 | ||
8.1 | Introduction | 183 | ||
8.2 | Background | 184 | ||
8.3 | Using the Bootstrap to Explore Model Stability | 185 | ||
8.3.1 | Selection of Variables within a Bootstrap Sample | 185 | ||
8.3.2 | The Bootstrap Inclusion Frequency and the Importance of a Variable | 186 | ||
8.4 | Example 1: Glioma Data | 186 | ||
8.5 | Example 2: Educational Body-Fat Data | 188 | ||
8.5.1 | Effect of Influential Observations on Model Selection | 189 | ||
8.6 | Example 3: Breast Cancer Diagnosis | 190 | ||
8.7 | Model Stability for Functions | 191 | ||
8.7.1 | Summarizing Variation between Curves | 191 | ||
8.7.2 | Measures of Curve Instability | 192 | ||
8.8 | Example 4: GBSG Breast Cancer Data | 193 | ||
8.8.1 | Interdependencies among Selected Variables and Functions in Subsets | 193 | ||
8.8.2 | Plots of Functions | 193 | ||
8.8.3 | Instability Measures | 195 | ||
8.8.4 | Stability of Functions Depending on Other Variables Included | 196 | ||
8.9 | Discussion | 197 | ||
8.9.1 | Relationship between Inclusion Fractions | 198 | ||
8.9.2 | Stability of Functions | 198 |
9 | Some Comparisons of MFP with Splines | 201 | ||
9.1 | Introduction | 201 | ||
9.2 | Background | 202 | ||
9.3 | MVRS: A Procedure for Model Building with Regression Splines | 203 | ||
9.3.1 | Restricted Cubic Spline Functions | 203 | ||
9.3.2 | Function Selection Procedure for Restricted Cubic Splines | 205 | ||
9.3.3 | The MVRS Algorithm | 205 | ||
9.4 | MVSS: A Procedure for Model Building with Cubic Smoothing Splines | 205 | ||
9.4.1 | Cubic Smoothing Splines | 205 | ||
9.4.2 | Function Selection Procedure for Cubic Smoothing Splines | 206 | ||
9.4.3 | The MVSS Algorithm | 206 | ||
9.5 | Example 1: Boston Housing Data | 207 | ||
9.5.1 | Effect of Reducing the Sample Size | 208 | ||
9.5.2 | Comparing Predictors | 212 | ||
9.6 | Example 2: GBSG Breast Cancer Study | 214 | ||
9.7 | Example 3: Pima Indians | 215 | ||
9.8 | Example 4: PBC | 217 | ||
9.9 | Discussion | 219 | ||
9.9.1 | Splines in General | 220 | ||
9.9.2 | Complexity of Functions | 221 | ||
9.9.3 | Optimal Fit or Transferability? | 221 | ||
9.9.4 | Reporting of Selected Models | 221 | ||
9.9.5 | Conclusion | 222 |
10 | How ToWork with MFP | 223 | ||
10.1 | Introduction | 223 | ||
10.2 | The Dataset | 223 | ||
10.3 | Univariate Analyses | 226 | ||
10.4 | MFP Analysis | 227 | ||
10.5 | Model Criticism | 228 | ||
10.5.1 | Function Plots | 228 | ||
10.5.2 | Residuals and Lack of Fit | 228 | ||
10.5.3 | Robustness Transformation and Subject-Matter Knowledge | 229 | ||
10.5.4 | Diagnostic Plot for Influential Observations | 230 | ||
10.5.5 | Refined Model | 231 | ||
10.5.6 | Interactions | 231 | ||
10.6 | Stability Analysis | 232 | ||
10.7 | Final Model | 235 | ||
10.8 | Issues to be Aware of | 235 | ||
10.8.1 | Selecting the Main-Effects Model | 235 | ||
10.8.2 | Further Comments on Stability | 236 | ||
10.8.3 | Searching for Interactions | 238 | ||
10.9 | Discussion | 238 |
11 | Special Topics Involving Fractional Polynomials | 241 | ||
11.1 | Time-Varying Hazard Ratios in the Cox Model | 241 | ||
11.1.1 | The Fractional Polynomial Time Procedure | 242 | ||
11.1.2 | The MFP Time Procedure | 243 | ||
11.1.3 | Prognostic Model with Time-Varying Effects for Patients with Breast Cancer | 243 | ||
11.1.4 | Categorization of Survival Time | 245 | ||
11.1.5 | Discussion | 246 | ||
11.2 | Age-specific Reference Intervals | 247 | ||
11.2.1 | Example: Fetal growth | 247 | ||
11.2.2 | Using FP Functions as Smoothers | 248 | ||
11.2.3 | More Sophisticated Distributional Assumptions | 249 | ||
11.2.4 | Discussion | 249 | ||
11.3 | Other Topics | 250 | ||
11.3.1 | Quantitative Risk Assessment in Developmental Toxicity Studies | 250 | ||
11.3.2 | Model Uncertainty for Functions | 251 | ||
11.3.3 | Relative Survival | 252 | ||
11.3.4 | Approximating Smooth Functions | 253 | ||
11.3.5 | Miscellaneous Applications | 254 |
12 | Epilogue | 255 | ||
12.1 | Introduction | 255 | ||
12.2 | Towards Recommendations for Practice | 255 | ||
12.2.1 | Variable Selection Procedure | 255 | ||
12.2.2 | Functional Form for Continuous Covariates | 257 | ||
12.2.3 | Extreme Values or Influential Points | 257 | ||
12.2.4 | Sensitivity Analysis | 257 | ||
12.2.5 | Check for Model Stability | 258 | ||
12.2.6 | Complexity of a Predictor | 258 | ||
12.2.7 | Check for Interactions | 258 | ||
12.3 | Omitted Topics and Future Directions | 258 | ||
12.3.1 | Measurement Error in Covariates | 258 | ||
12.3.2 | Meta-analysis | 258 | ||
12.3.3 | Multi-level (Hierarchical) Models | 259 | ||
12.3.4 | Missing Covariate Data | 259 | ||
12.3.5 | Other Types of Model | 259 | ||
12.4 | Conclusion | 259 |
Appendix A: Data and Software Resources | 261 | |||
A.1 | Summaries of Datasets | 261 | ||
A.2 | Datasets used more than once | 262 | ||
A.3 | Software | 267 |
Appendix B: Glossary of Abbreviations | 269 | |||
References | 271 | |||
Index | 285 |
Datasets and some information are available for download here
Datasets in available formats - Stata - SAS - Excel - ASCII
For more details about the data see the Appendix A of the book.
Table A.1 Datasets used once in our book. N/A = not applicable. Further details accompany the example in the relevant section (page 261).
Name (and Link) | Outcome | Obs. | Events | Variablesa | Section reference |
Myeloma | Survival | 65 | 48 | 16 | 2.7.1 |
Freiburg DNA breast cancer | Survival | 109 | 56 | 1 | 3.4.1 |
Cervix cancer | Binary | 899 | 141 | 21 | 3.5 |
Nerve conduction | Cont. | 406 | N/A | 1 | 5.7.1 |
Triceps skinfold thickness | Cont. | 892 | N/A | 1 | 5.7.2 |
Diabetes | Cont. | 42 | N/A | 2 | 6.7.2 |
Advanced prostate cancer | Survival | 475 | 338 | 13 | 7.5 |
Quit smoking study | Cont. | 250 | N/A | 3 | 7.11.3 |
Breast cancer diagnosis | Binary | 458 | 133 | 6 | 8.6 |
Boston housing | Cont. | 506 | N/A | 13 | 9.5 |
Pima Indians | Binary | 768 | 268 | 8 | 9.7 |
Rotterdam breast cancer | Survival | 2982 | 1518 | 11 | 11.1.3 |
Fetal growth | Cont. | 574 | N/A | 1 | 11.2.1 |
Cholesterol | Cont. | 553 | N/A | 1 | 11.2.3 |
a Maximum number of predictors used in analyses. Categorical variables count as
>1 predictor, if modelled using several dummy variables.
Table A.2 Datasets used more than once in our book. N/A = not applicable. Further details are given in Appendix A.2 (page 262).
Name | Outcome | Obs. | Events | Variablesa | Section reference |
Research body fat | Cont. | 326 | N/A | 1 | 1.1.3, 4.2.1, 4.9.1, 4.9.2, 4.10.3, 4.12 |
GBSG breast cancer | Survival | 686 | 299 | 9 | 1.1.4,3.6.2, 5.6.2,6.5.2, 6.5.3, 6.5.4,6.6.5, 6.6.6, 6.8.2, 7.6, 7.7.2, 8.8, 9.6 |
Educational body fat | Cont. | 252 | N/A | 13 | 2.7.2, 2.8.6, 5.2, 5.3.1, 5.5.1, 8.5 |
Glioma | Survial | 411 | 274 | 15 | 2.7.3, 8.4 |
Prostate cancer | Cont. | 97 | N/A | 7 | 3.6.2, 3.6.3, 4.15, 6.2, 6.3.2, 6.4.2, 6.4.3, 6.5.1, 6.5.3, 6.6.1, 6.6.2, 6.6.3, 6.6.4, 7.11.3 |
Whitehall I | Survival | 17 260 | 2576 | 10 | 6.7.3 |
Binary | 17 260 | 1670 | 10 | 4.13.1, 4.13.2, 4.14, 7.11.1,7.11.3 | |
PBC | Survival | 418 | 161 | 17 | 5.3.2, 5.4, 5.5.2, 9.8 |
Oral cancer | Binary | 397 | 194 | 1 | 6.7.1, 9.3.1 |
Kidney cancer | Survival | 347 | 322 | 10 | 5.8.2,7.9 |
a Maximum number of predictors used in analyses. Categorical variables count as
>1 predictor, if modelled using several dummy variables.