Scholarlink

Background <p>Growing availability of laboratory test result information in large healthcare databases presents opportunities to improve confounding control. Whilst information on the ordering of tests and the resulting continuous test values are available, their optimal integration into data-driven confounding control strategies like the high-dimensional propensity score (HDPS) remains unclear.</p> Methods <p>We propose methods to incorporate test-related data into HDPS, addressing key concerns around data quality and missing data. We illustrate these approaches using UK Clinical Practice Research Datalink GOLD data comparing COPD-specific mortality in new-users of proton pump inhibitors (PPIs) and H2-receptor antagonists (H2RAs). Hazard ratios (HRs) were estimated using Cox models weighted by inverse HDPS. Integration of 35 blood-test values was achieved via biologically informed cut-offs and using a missing-indicator approach for continuous data. Results were benchmarked against a HDPS model derived from clinical, referral, and prescription data dimensions.</p> Results <p>Among 733,885 new PPI and 124,410 H2RA users, the adjusted HR for PPI use and COPD mortality was 1.36 (95% CI: 1.14–1.64) in the primary HDPS analysis. Incorporating test-requested data and continuous blood test values attenuated estimates towards the expected null association (HR 1.24; 95% CI: 1.02–1.56). Of the 500 HDPS covariates selected in the final model, 46% were derived from test-related data.</p> Conclusion <p>We empirically evaluated methods for incorporating laboratory test results into the HDPS framework, demonstrating their potential to reduce residual confounding in UK EHR studies. We propose principled approaches for incorporating laboratory test result data into data-driven confounding control strategies. Results from our case study highlight the potential for these data to reduce residual confounding in UK EHR studies. When available, test information should be considered within HDPS.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Methods for incorporating test result information within the high-dimensional propensity score framework: application in UK electronic health record data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月30日

查看全文
  • John Tazare,
  • Jeremy P. Brown,
  • Daniel R. Morales,
  • Liam Smeeth,
  • Stephen J. W. Evans,
  • Ian J. Douglas,
  • Elizabeth J. Williamson

摘要

Background

Growing availability of laboratory test result information in large healthcare databases presents opportunities to improve confounding control. Whilst information on the ordering of tests and the resulting continuous test values are available, their optimal integration into data-driven confounding control strategies like the high-dimensional propensity score (HDPS) remains unclear.

Methods

We propose methods to incorporate test-related data into HDPS, addressing key concerns around data quality and missing data. We illustrate these approaches using UK Clinical Practice Research Datalink GOLD data comparing COPD-specific mortality in new-users of proton pump inhibitors (PPIs) and H2-receptor antagonists (H2RAs). Hazard ratios (HRs) were estimated using Cox models weighted by inverse HDPS. Integration of 35 blood-test values was achieved via biologically informed cut-offs and using a missing-indicator approach for continuous data. Results were benchmarked against a HDPS model derived from clinical, referral, and prescription data dimensions.

Results

Among 733,885 new PPI and 124,410 H2RA users, the adjusted HR for PPI use and COPD mortality was 1.36 (95% CI: 1.14–1.64) in the primary HDPS analysis. Incorporating test-requested data and continuous blood test values attenuated estimates towards the expected null association (HR 1.24; 95% CI: 1.02–1.56). Of the 500 HDPS covariates selected in the final model, 46% were derived from test-related data.

Conclusion

We empirically evaluated methods for incorporating laboratory test results into the HDPS framework, demonstrating their potential to reduce residual confounding in UK EHR studies. We propose principled approaches for incorporating laboratory test result data into data-driven confounding control strategies. Results from our case study highlight the potential for these data to reduce residual confounding in UK EHR studies. When available, test information should be considered within HDPS.