data linkage strategies

27
1 Data Linkage Strategies Shihfen Tu, Ph.D. University of Maine [email protected]

Upload: randy

Post on 19-Jan-2016

25 views

Category:

Documents


0 download

DESCRIPTION

Data Linkage Strategies. Shihfen Tu, Ph.D. University of Maine [email protected]. Faculty Disclosure Information. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Data Linkage Strategies

1

Data Linkage Strategies

Shihfen Tu, Ph.D.

University of Maine

[email protected]

Page 2: Data Linkage Strategies

2

Faculty Disclosure Information

In the past 12 months, I have not had a significant financial interest or other relationship with the manufacturer(s) of the product(s) or provider(s) of the service(s) that will be discussed in my presentation.

This presentation will not include discussion of pharmaceuticals or devices that have not been approved by the FDA.

Page 3: Data Linkage Strategies

3

Acknowledgements

• University of Maine– Quansheng Song– Cecilia Cobo-Lewis

• Maine Bureau of Health– Kim Church– Pat Day– Ellie Mulcahy– Toni Wall

Page 4: Data Linkage Strategies

4

Page 5: Data Linkage Strategies

5

Data Linkage

Page 6: Data Linkage Strategies

6

Data Linkage

Page 7: Data Linkage Strategies

7

Data Linkage

Page 8: Data Linkage Strategies

8

Data Linkage - Probabilistic

Page 9: Data Linkage Strategies

9

Data Linkage - Probabilistic

Page 10: Data Linkage Strategies

10

Data Linkage - Probabilistic

Page 11: Data Linkage Strategies

11

Data Linkage - Probabilistic

Page 12: Data Linkage Strategies

12

Data Linkage - Inconsistency

Page 13: Data Linkage Strategies

13

Data Linkage - Inconsistency

Inconsistency DetectedInconsistency DetectedCorrecting….Correcting….

Message

Page 14: Data Linkage Strategies

14

Inconsistencies

• Record in EHDI links to two records in other database

• The other source indicates the records belong to different people

• How to address depends on processing of other database

EHDI_ID=394Brad A. Graham

ID=4484Brad A. Graham

ID=7354Brad Graham

Page 15: Data Linkage Strategies

15

Inconsistencies

• Other source not de-duplicated ?• Other source de-duplicated, but insufficient evidence to conclude

ID=4484 and ID=7354 are the same person ?– BD may provide additional information so that these probabilities have changed

ID=4484Brad A. Graham

ID=7354Brad Graham

EHDI_ID=394Brad A. Graham

Page 16: Data Linkage Strategies

16

Inconsistencies

EHDI_ID=394John A. Graham

ID=4048John A. Graham

ID=4048Jon A. Graham

EHDI_ID=948Jon A. Graham

ID=9324Jon Graham

EHDI_ID=948 Jon Graham

Page 17: Data Linkage Strategies

17

• How this "cross-over" is resolved depends on whether one or neither file is given precedence

• Influenced by probabilistic de-duplication process performed after a linkage

EHDI_ID=394John A. GrahamEHDI_ID=394

John A. GrahamID=4048

John A. GrahamID=4048

John A. Graham

ID=4048Jon A. Graham

ID=4048Jon A. Graham

EHDI_ID=948Jon A. GrahamEHDI_ID=948Jon A. Graham

ID=9324Jon Graham

ID=9324Jon Graham

EHDI_ID=948Jon Graham

EHDI_ID=948Jon Graham

Inconsistencies

Page 18: Data Linkage Strategies

18

Linkage Creep

• EHDI Database contributes an individual,Catherine A. Sampson

ID Source FirstName MiddleInitial LastName PMatch

113 EHDI Catherine A Sampson

Page 19: Data Linkage Strategies

19

Linkage Creep

• Link the Electronic Birth Certificate– Name is Catherine A. Simpson– Are these the same person?– Perform probabilistic match

• Require .85 probability of a match to conclude two similar records are the same (Critical p = .85)

• Probability is .90, we conclude they’re the same person

ID Source FirstName MiddleInitial LastName PMatch

113 EHDI Catherine A Sampson113 EBC Catherine A Simpson 0.90

Page 20: Data Linkage Strategies

20

Linkage Creep

• Link Birth Defects Registry Data– Name is Kathy A. Simpson– Are these the same person?– Perform probabilistic match (require .85)

• PMatch is .90, we conclude they’re the same person

ID Source FirstName MiddleInitial LastName PMatch

113 EHDI Catherine A Sampson113 EBC Catherine A Simpson113 BDR Kathy A Simpson 0.90

Page 21: Data Linkage Strategies

21

Linkage Creep

• If we compare to Catherine A. Sampson– PMatch=.81

– Conclude they are NOT the same individual– Would not assign same ID

• Which is correct?

ID Source FirstName MiddleInitial LastName PMatch

113 EHDI Catherine A Sampson113 EBC Catherine A Simpson113 BDR Kathy A Simpson 0.81

Page 22: Data Linkage Strategies

22

Linkage Creep• When is this a problem?

– Over time, two distinct individuals may project “tendrils” composed of combinations of identifiers that statistically overlap in probabilistic space

Page 23: Data Linkage Strategies

23

Linkage Creep• When is this a problem?

– Linkage creep will result in the two distinct individuals being erroneously combined under a single ID

Page 24: Data Linkage Strategies

24

Linkage Creep• When is this not problem?

– Over time, certain key identifiers for an individual are expected to change

– This phenomenon will increase as a historical database grows, and as additional sources are input into a centralized system

Page 25: Data Linkage Strategies

25

Linkage Creep• Complexity of “creep” in longitudinal datasets

– Black records are related to all records– Yellow and Blue records are NOT related to White

record– Yellow record is also not related to Red record at

Page 26: Data Linkage Strategies

26

Linkage Creep• Forbidding “creep” will result in a single

individual being divided into two IDs over time

• Further challenge—where to divide records into additional IDs?

Page 27: Data Linkage Strategies

27

Tools for Evaluating Linkage• Inconsistencies can occur in deterministic linkage, but

are more common in probabilistic linkages• Probabilities that create potential for problems provide

a valuable tool for evaluating linkages– Instead of a “are two records the same person ?” Yes/No– Estimates or indices of how likely it is that two records are

the same person

• Should be able to estimate the number of erroneous linkages

• Possible to conduct a detailed examination of quality by ignoring very strong and very weak pairings, and only focusing on pairings that are ambiguous