USF Health logo

Linking Maternal and Child Health Data to Create a Comprehensive Longitudinal Dataset: The Florida Experience

Hamisu Salihu, Ph.D., Principal Investigator
University of South Florida College of Public Health
R01 HS19997-01
9/30/2010 - 9/29/2013 (3 years)

April 6, 2011

Specific Aims

1. To create an expanded clinically enhanced maternal-infant dataset for the State of Florida by augmenting the current statewide hospitalization data files through linkages to other data sources.
2. To validate the created dataset in Specific Aim 1 through a rigorous process that will establish confidence in the use of the dataset.
3. To demonstrate the utility of the newly created, enriched dataset in conducting comparative effectiveness analysis using early term elective delivery as a case study.

Data Sources

FDOH (Office of Vital Statistics) Data

Live Birth Data 1998-2009

Mortality (Infant and Mom) Data 1998-2009

Fetal Death Data 1998-2009

Agency for Health Care Administration Data

Inpatient Hospitalization Data 1998-2009

Outpatient/Ambulatory Hospitalization Data 1998-2009

Emergency Department Data 2005-2009

Hospital Financial Data 1998-2009

"Follow" Infants Over Time Through Linkage

Arrow pointing to the right to show the passage of time with the following events placed along the arrow from left to right: birth, inpatient hospitalization #1, inpatient hospitalization #2, outpatient hospitalization #1, emergency room visit, death.

"Follow" Moms Over Time Through Linkage

Arrow pointing to the right to show the passage of time with the following events placed along the arrow from left to right: delivery, inpatient hospitalization #1, inpatient hospitalization #2, outpatient hospitalization #1.

Special Challenges to Our Data Linkage

• Birth vital records contain a significant amount of identifying information
• Hospital records (inpatient, ambulatory, ED) for the infant contain limited identifying information
  • No infant SSN, name, address
  • Primary identifier is mother's SSN (INFANTLINK), but it is missing >10% and disproportionately among certain subgroups
  • Previous investigation reveal that maternal SSN has a typo or transposition in over 1,000 instances (ASSUME identifiers have errors)
  • Missing mother's date of birth, a key linking and/or confirmatory variable

Our Approach to Linking AHCA to VS

• Stage I
  • Within the inpatient hospital discharge data, we first attempt to link infants to their mothers (so called dyad links) with the primary goal of obtaining maternal DOB, an important linking variable (FIND other identifying information)
• Stage II
  • Link these dyad pairs to birth vital records, now incorporating infant's and mom's DOB, mom's SSN, and facility of birth as the primary linking variables
• Stage III
  • Attempt to link infant and mom hospitalizations that did not link to a maternal record from Stage #1 directly to the birth record

Example of Overarching Linkage Approach

Complex flow chart illustrating three stages of linkage approach. Stage 1: Within the inpatient hospital discharge data, first attempt to link infants to their mothers (so called dyad links) with the primary goal of obtaining maternal D O B, an important linking variable (FIND other identifying information). Stage 2: Link these dyad pairs to birth vital records, now incorporating infant's and mom's D O B, mom's S S N, and facility of birth as the primary linking variables. Stage 3: Attempt to link infant and mom hospitalizations that did not link to a maternal record from Stage 1 directly to the birth record.

Software To Facilitate Data Linkage

• LinkSolv
• AutoMatch
• LinkageWiz
• FRIL
• LinkPlus
• Link King
• SQL Match
• FEBRL
• SQL Server (SSIS)
• SAS, SPSS, Stata, S-Plus, R
• ...many more!

Decorative graphic showing figure putting a puzzle piece into a puzzle.

SQL Match

Clockwise from upper left:
Screen in S Q L Match software used to set up data linkage.
Screen in S Q L Match software used to view linkage results.
Screen in S Q L Match software used for manual review.

Freely Extensible Biomedical Record Linkage

Clockwise from upper left:
Screen in F E B R L software used to set up data linkage.
Screen in F E B R L software used to view linkage results.
Screen in F E B R L software used for manual review and summary.
Screen in F E B R L software used to set up data linkage.

Link Plus

Clockwise from upper left:
Screen in Link Plus software used to set up data linkage.
Screen in Link Plus software used to view linkage results.
Screen in Link Plus software used for manual review and summary.
Screen in Link Plus software used for manual review and summary.

Link King

Clockwise from upper left:
Screen in Link King software used to set up data linkage.
Decorative graphic of a king.
Screen in Link King software used to select variables.
Screen in Link King software used for manual review.
Screen in Link King software used for manual review.

Fine-Grained Record Linkage (FRIL)

Clockwise from upper left:
Screen in F R I L software used to set up data linkage.
Screen in F R I L software used to select variables and weights.
Screen in F R I L software used for manual review.
Screen in F R I L software used for join methods.

But our choice...SAS

Screen showing SAS code.

Linking Mechanics

• Developed a SAS macro
• Hierarchical, stepwise series of linking stages, using various combinations of variables, proceeding from highest to lowest confidence
  • Exact and partial matching, linking with replacement
  • Primarily deterministic, includes probabilistic elements
  • CREATES potential matches
• Coding algorithm to calculate a "linking confidence" score to GRADE matches
  • Also incorporate a "delivery confidence" score
• Records above a certain score are SELECTED as links, borderline scores require manual validation
  • May find false + we need to CORRECT
  • We try to minimize manual review

Linking Mechanics

• We do not use blocking
  • Too concerned about flawed data
  • Linking approximately 230,000 birth hospitalization records to approximately 1.4 million "women" records using the merging macro takes approximately 1 hour
    • Will sacrifice extra time for greater sensitivity
• SAS
  • Not as automated or "point-and-click" as other software
  • Extremely customizable through coding
  • Easy to incorporate a large number of variables (Link King)
  • Easy to allow "crossover" links
    • Mom's SSN in AHCA links to father's SSN in vital stats
  • Can process extremely large datasets quickly given powerful computers

Screen showing summary output for two merged clauses.

Additional Challenges

• Disentangling multiples (twins, triplets, etc)
  • No infant SSN, no names in hospital data
  • Multiples will share all mom characteristics
  • Ordering of variables in AHCA does not match birth order
  • Can use sex to differentiate between opposite-sex dizygotic twins
  • Can use diagnosis codes that reflect 500 gram birth weight categories to disentangle same sex multiples that may differ in birth weight
  • For multiples that have the same sex, similar birth weights, it may be impossible to determine, given the available data, which hospital record goes with which birth record
    • Investigating other options
      • Random assignment
      • Allocation to "family" as unit

Next Steps

• Finalize enhancements to linkage of birth record to birth hospitalizations
• Link in post-birth hospitalizations, ambulatory records, and ED data
  • Challenging for those with missing/incorrect maternal SSN
• Develop an identifier crosswalk to link in cost-to-charge ratios (CCRs) from CMS to convert hospital charges to costs