Contents
Executive Summary 5
Preface 7
Part 1: Probabilistic Record Linkage 10
Chapter 1. Linkage Methodology 11
Introduction 11
Types of Linkages 11
Current CODES Methodology 18
Survey of Linkage Software 19
References 24
Chapter 2. A Comparison of High-Probability, Multiply Imputed, and Maximum a Posteriori Matched Sets 25
Introduction 25
Methods 25
Results 28
Conclusions 37
References 38
Chapter 3. Analysis of Match Probability in Probabilistic Linkage 40
Introduction 40
Definitions 40
Hypothesis 40
Methods 41
Results and Discussion 41
Conclusion 48
References 49
Chapter 4. A Comparison and Demonstration of Multiple Imputation of Missing Data 50
Introduction 50
Missing Data 50
Patterns of Missing Data 51
Methods to Handle Missing Data 52
Producing Multiply Imputed Datasets 52
Multiple Imputation Using a Sequence of Regression Models 53
Analyzing Multiply Imputed Datasets 53
Multiple Imputation Demonstration 55
Comparison of Multiple Imputation Methods 64
References 67
Part 2: Applications From Multiple State CODES Data 68
CODES General Use Model Overview 69
Analysis 1. Comparison of Medical Consequences of Motor Vehicle Crashes among Older Occupants 72
Abstract 72
Introduction 72
Methods 73
Results 73
Conclusions 79
References 79
Analysis 2. Comparison of Medical Outcomes by Reported Safety Restraint Use among Children Ages 1 to 7 Years 80
Abstract 80
Introduction 80
Methods 81
Results 81
Conclusions 84
References 85
Analysis 3. Comparing Medical Outcomes by Helmet Use Laws in 11 States Using CODES Data 86
Abstract 86
Introduction 86
Methods 87
Results 88
Conclusions 91
References 92
Analysis 4. Graduated Driver Licensing and Teenage Driver Involvement in Injury Crashes 93
Abstract 93
Introduction 93
Methods 94
Results 95
Conclusion 97
References 98
Part 2 Summary 99
Table 1.1.1. Summary of Linkage Software Capabilities 23
Table 1.2.1. Summary of linkage variable combinations and corresponding minimum potential match probability for each linkage performed 27
Table 1.2.2. Total matches, sensitivity, and specificity for high-probability and imputed match sets 29
Table 1.3.1. List of variables used in the linkage model by eight CODES States 42
Table 1.3.2. Median percentage of match probabilities across all CODES States 44
Table 1.4.1. Descriptions and Examples of Missingness Types 50
Table 1.4.2. Description of datasets 55
Table 1.4.3. Missingness Rates of Variables 57
Table 1.4.4. Comparison of IVEware and PROC MI 65
Table 2.1.1. Occupant characteristics by age group 74
Table 2.1.2. MVC characteristics by driver age group 74
Table 2.1.3. MVC configuration by driver age group 75
Table 2.1.4. Maximum Abbreviated Injury Scale (MAIS) by age group 75
Table 2.1.5. Top five injured body regions by age group 77
Table 2.1.6. Top five natures of injury by age group 78
Table 2.1.7. Top five injured body regions by age group among unrestrained occupants 78
Table 2.1.8. Median hospital charges and 95% confidence intervals in 2008 dollars by age group 79
Table 2.2.1. Driver safety restraint use by reported child safety restraint use 82
Table 2.2.2. Highest level of care by reported child safety restraint use 83
Table 2.2.3. MAIS by reported child safety restraint use 83
Table 2.2.4. Neck, back, or abdomen injuries and TBI of children that went to the hospital by reported child safety restraint use 84
Table 2.2.5. Median hospital charges in 2008 dollars by reported child safety restraint use 84
Table 2.3.1. Description of the study population 88
Table 2.3.2. Rates and Relative Risks of Medical Outcomes with 95% Confidence Intervals 89
Table 2.3.3. Adjusted Relative Risks of Medical Outcomes for No Helmet Versus Helmet with 95% Confidence Intervals 90
Table 2.3.4. Body regions injured among motorcycle operators seen in the emergency department or admitted to the hospital.1 P-values compare partial to universal law States 90
Table 2.3.5. Body regions injured among motorcycle operators covered by a helmet law according to age and State (partial law State), or under 21 years-old (universal law State) seen in the emergency... 91
Table 2.4.1. Description of the teenage driver study population. Teen driver characteristics by IIHS GDL rating in State 95
Table 2.4.2. Per capita rates and rate ratios of teen driver involvement in injury MVCs by driver characteristic and IIHS GDL rating 96
Table 2.4.3. Age-specific per capita rate ratios of teenage driver involvement in injury MVCs comparing ideal GDL components to non-ideal 97
Figure 1.1.1. Distribution of match weights from an EMS and MVC probabilistic linkage 14
Figure 1.1.2. Distribution of match weights from an EMS and MVC probabilistic linkage by determination of final match status 15
Figure 1.2.1. Distribution of high-probability, imputed, and MAP matched sets for occupant age 30
Figure 1.2.2. Distribution of high-probability, imputed, and MAP matched sets for MVC county 31
Figure 1.2.3. Distribution of high-probability, imputed, and MAP matched sets for MVC hour 32
Figure 1.2.4. Coefficient estimates and and corresponding 95-percent confidence intervals used to model simulated log hospital charges for high-probability, imputed, and MAP matched... 34
Figure 1.2.5. Coefficient estimates and and corresponding 95-percent confidence intervals used to model hospitalization status for high-probability, imputed, and MAP matched... 36
Figure 1.3.1. Percentage of match probabilistic for each of eight OECDS States 43
Figure 1.3.2. Boxplot of match weights for each linkage identifier 46
Figure 1.3.3. Match probability Percentile for each MVC File Size Group 47
Figure 1.3.4. Match probability and number of imputations each record appears 48
Figure 1.4.1. Patterns of missingness 51
Figure 1.4.2. Overview of Analyzing data uaing multiple imputation methods 54
Figure 2.1.1. Highest level of care by age group 76
Figure 2.1.2. Safety restraint usage by highest level of care by age group 76
Figure 2.1.3. Discharge status by age group of those not discharged home 77
Figure 2.2.1. Child age by reported child safety restraint use 82