2. Evaluating the data to determine feasibility before commencing the study. Purchasing data often comes at a cost. Before embarking on a study, researchers typically want to find overlapping patient cohorts to understand the size of each dataset, the overlap, and the depth. For the size of patient records, researchers often want to know the denominator and specifically, how many patients and records are in each dataset. For the overlap, understanding how many patients are in common across the datasets is helpful information for researchers to know before moving forward with a study. It also accelerates a path toward what the Yale researchers were set to do: analyze the linked data. Lastly, knowing the depth of the dataset, that is, knowing how many records exist per patient, provides a more complete picture of the patient’s journey. In the case of the mortality data and voter profile data, the idea is to link one patient to each dataset.

The Yale team wanted to understand whether the mortality data had significant coverage across Florida and Ohio voter profile data, and whether it was significant enough to pursue a linkage. It was important for the researchers to examine voter data before and during the COVID-19 pandemic. Between 2018 and 2021, over 577,000 Florida and Ohio resident deaths were matched to the state voter profile data from 2017. These deaths represent 80% of deaths in the US, a limitation of the study. The 2019 data was used as a benchmark to determine expected death rates based on age, time of year, location, and party affiliation, that is, determining how many Democrats and Republicans in a given age bracket would die in a given season. Below normal was considered “excess death.” In the COVID-19 post-vaccine period when all adults were eligible for the vaccine, the excess death rate was 76% higher for Republicans in Florida and Ohio counties with lower vaccination rates.

3. Matching the data. In working with large datasets, a third-party PPRL software tool to match data on behalf of researchers reduces bias from the alternative: using internal mechanisms. Expanding research innovation with the use of de-identified, connected data must be fit-for-purpose and provide high accuracy. Collision scores may be used to ensure high precision for common names. For instance, if there are two John Smiths born on the same day, the records are assigned a high frequency score and the matching algorithm can identify additional information in the data set to declare a positive match.

Health IT Homepage