Matching Traffic Crash Victims to Hospital Patients: Probabilistic v. Deterministic Record Linkage Methods

Transcription

1 Matching Traffic Crash Victims to Hospital Patients: Probabilistic v. Deterministic Record Linkage Methods Richard Miller Research Scientist Wisconsin Office of Health Informatics October 28, 2014

2 Crash Outcome Data and Evaluation System CODES Crash outcomes are the medical consequences of traffic crashes. Crash outcomes are not well captured in traffic crash reports. Hospital patient records contain detailed diagnosis and treatment data. Most states have patient records available as abstracted and coded extracts from the standard Uniform Billing format. 2

3 CODES The Crash Outcome Data and Evaluation Systems o Match records for crash victims to hospital inpatients and emergency department patients. o Combine traffic crash report data and hospital patient information. o Enable detailed tracing and analysis of the medical and financial consequences of crashes. 3

4 A Little History Some states initially developed CODES as part of public health injury surveillance. Key public policy issues: encouraging seatbelt use, child restraints, motorcycle helmet use. NHTSA nurtured this data and analytical effort among dozens of states. 4

5 The Official Matching Method NHTSA supported and encouraged a rigorous methodological approach to matching crash and patient records. Probabilistic record linkage methods use sophisticated developments in Bayesian statistics. Complex models identify the pairs of victim and patient records with a high probability of being the same person and crash. 5

6 Why Probabilistic Linkage? No common unique identifier across data systems. Recorded data on names, dates, and so on are subject to many sources of error. Probabilistic matching techniques can deal with missing and inexact data items. 6

7 Probabilistic Record Matching Wisconsin s probabilistic models use a long list of crash and victim or patient characteristics. The probability of observing each pair of values in every pair of records is calculated and weighted. The results are the probability that the pair is a true match. 7

8 16 Characteristics Compared Dates of crash and admission Role of victim or patient: vehicle occupant, pedestrian, bicycle or motorcycle rider Injured? Died? Ambulance transport? County of crash and of residence Month, day and year of birth, sex, initials, ZIP3 and ZIP2 of residence 8

9 Probabilistic Linkage in Action 267,042 people in Wisconsin s 2012 traffic crash reports o 41,287 reported injured 23,925 inpatient stays for injuries in Wisconsin hospitals o 4,198 coded for traffic crash cause of injury Matching result: 3,248 pairs of records had at least a 90 percent probability of being true matches. 9

10 Probabilistic Linkage Has Done Well Wisconsin has constructed CODES datasets annually for 20 years using probabilistic linkage methods. The results have been extremely useful for many policy studies and for informing state and local traffic safety programs and initiatives. Hospital medical diagnoses can be assigned dollar amounts for costs of care and rehabilitation and lost wages 10

11 Probabilistic Linkage Also Presents Problems The process requires a high level of specialized technical skill with advanced proprietary software. The knowledge set is difficult to transfer to other staff. The process is difficult to describe to general audiences, making research less accessible. Recent examination suggests some limits to the validity of probabilistic matches. 11

12 Match Validity Questions 58 of 3248 inpatient matches (2%) have admission before crash date. o 90% are more than three weeks before. o 79% are not coded as crash-related injuries. 160 matches (5%) have admit date more than six days after crash date. o 90% are more than 10 days after, up to 335 days. o 57% are not coded as crash-related injuries. 12

13 Why do these matches have a high probability? Inspection shows strong agreement on personal characteristics: right person but wrong crash. o A person may be involved in more than one crash report. o A crash victim may be hospitalized for other reasons over the course of a year. The matched hospitalization may be for another crash or for an unrelated injury. 13

14 Is Deterministic Record Matching a Reasonable Alternative? Can records be determined to be a match if key characteristics are an exact match? Can the quality of the data support confidence in valid deterministic matching? If so, then deterministic matching offers a more efficient, transparent and accessible alternative. 14

15 Iterative Deterministic Matching Limit potential matches to admissions within six days following crash. Define strongest combination of identifiers. Identify exact matches. Sort out multiple crash reports matched to same hospitalization and vice versa. Repeat on the residual records with next best identifier combination. 15

16 Four Deterministic Combinations 1. First and last initial + DOB + Sex + ZIP 40,853 crash victims and 23,891 inpatients 2,603 initial matches 2 had multiple crash reports for same inpatient: selected the one with admit date = crash date 160 had multiple inpatients matched to same crash: selected the one with earliest admit date 2,524 final matches 16

17 Four Deterministic Combinations 2. First and last initial + DOB + Sex + ZIP3 252 initial matches, 252 final matches 3. First and last initial + DOB + Sex 147 initial matches, 140 final matches 4. DOB + Sex + ZIP3 ** Included only the inpatients with a traffic crash E-code (7%) 109 initial matches, 106 final matches 17

18 Deterministic Match Results Deterministic matching process identified 3,022 matches. 95% had the cause of injury coded by the hospitals as a motor vehicle traffic crash. 18

19 Comparative Results 3,030 probabilistic matches had the inpatient admission within six days of the crash date. 3,022 deterministic matches fit the same plausible window. 92 of 3,030 probabilistic matches (3%) were not found with the deterministic matching process. 84 of 3,022 deterministic matches (3%) were not found with the probabilistic matching process. 19

20 Comparative Results 97% of the matches for each method were also found by the other method. 94% of all matches combined were found by both methods. 20

21 Conclusions Deterministic record linkage methods offer a valid alternative for CODES-related record matching. Deterministic matching may be efficiently coded and performed with straightforward SAS or similar software. Deterministic matching methods are transparent and transferrable. 21

22 Caveats to the Conclusions Staff should be familiar with and confident in the quality control for each data system. Missing data should be minimal for the most likely records. Some (minimal) manual judgment may be required for deterministic linking. Results of prior probabilistic matching can inform development of a deterministic process. 22

23 Matching Traffic Crash Victims to Hospital Patients Richard Miller Research Scientist Wisconsin Division of Public Health Office of Health Informatics