Domain

Analytic Methods

Type

Empirical Study

Theme

effectiveness; population

Start Date

7-6-2014 2:55 PM

End Date

7-6-2014 4:15 PM

Structured Abstract

Introduction: For privacy and practical reasons, it is sometimes necessary to minimize sharing of individual-level information in multi-site comparative effectiveness and patient-centered outcomes research studies. However, individual-level information is often needed to adjust for biases inherent in studies that analyze observational data.

Objectives: To compare three analytic methods that only require sharing of summary-level information in multi-center studies to perform statistical analysis that have traditionally required access to detailed individual-level data from each site.

Methods: We analyzed data from 7 sites participating in the Scalable PArtnering Network for Comparative Effectiveness Research (SPAN) study between 2005 and 2009. We compared the long-term risk of re-hospitalization between the laparoscopic adjustable gastric banding and Roux-en-y gastric bypass procedures. We generated propensity scores at each site to summarize the baseline confounders, including age, sex, race and ethnicity, year of procedure, body mass index, Charlson comorbidity score, individual comorbidities (e.g., asthma, diabetes), smoking status, blood pressure, and type of insurance. We used three methods that require only summary-level information from the 7 participating sites to obtain the propensity score-adjusted estimates: (1) stratified analysis of propensity score-defined strata, (2) case-centered analysis of risk set data, and (3) meta-analysis of site-specific effect estimates. Results from these analyses were compared with the result from an individual-level data analysis (i.e., the “gold-standard”).

Findings: The study included 7,342 eligible patients. In the individual-level data analysis, the adjusted hazard ratio was 0.71 (95% confidence interval: 0.59, 0.84) comparing adjustable gastric banding with Roux-en-y gastric bypass procedures. The corresponding effect estimate was 0.70 (0.59, 0.83) in the propensity score-stratified analysis, 0.71 (0.59, 0.84) in the case-centered analysis, and 0.71 (0.60, 0.84) in both the fixed-effect and random-effects meta-analysis.

Discussion: This empirical study shows that methods requiring only summary-level information produce results that are identical or very similar to results from individual-level data analysis when adjustment for multiple confounders is required.

Conclusion: Propensity score-stratified analysis, case-centered analysis, and meta-analysis produced results that are highly comparable to individual-level data analysis. These very practical and efficient analytic approaches can be considered when sharing of individual-level information is not feasible or not preferred in multi-site studies.

Acknowledgements

This project was supported by grant number 1R01HS019912 from the Agency for Healthcare Research and Quality. The content is solely the responsibility of the authors and does not necessarily represent the official views of the Agency for Healthcare Research and Quality.

The authors would like to thank the programmer/analysts from the following SPAN sites for providing data for this project: Essentia Institute for Rural Health (EIRH), Group Health Research Institute (GHRI), Harvard Pilgrim Health Care (HPHC), HealthPartners Institute for Education and Research (HPIER), Kaiser Permanente Northern California (KPNC), Kaiser Permanente Colorado (KPCO), and Kaiser Permanente Hawaii (KPHI).

We are indebted to the following SPAN Investigators for the collaborations that made this study possible: KPCO (Matthew F. Daley, MD [SPAN PI]; Elizabeth A. Bayliss, MD, MSPH; Ella Lyons, MS), EIRH (Thomas E. Elliott, MD), HPIER (Pamala A. Pawloski, PharmD), KPHI (Cynthia Nakasato, MD; Rebecca Williams, DrPH, MPH; Vinutha Vijayadeva, PhD), and KPNC (Lisa Herrinton, PhD).

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.

Share

COinS
 
Jun 7th, 2:55 PM Jun 7th, 4:15 PM

Empirical assessment of privacy-preserving confounding adjustment methods in multi-center comparative effectiveness studies

Introduction: For privacy and practical reasons, it is sometimes necessary to minimize sharing of individual-level information in multi-site comparative effectiveness and patient-centered outcomes research studies. However, individual-level information is often needed to adjust for biases inherent in studies that analyze observational data.

Objectives: To compare three analytic methods that only require sharing of summary-level information in multi-center studies to perform statistical analysis that have traditionally required access to detailed individual-level data from each site.

Methods: We analyzed data from 7 sites participating in the Scalable PArtnering Network for Comparative Effectiveness Research (SPAN) study between 2005 and 2009. We compared the long-term risk of re-hospitalization between the laparoscopic adjustable gastric banding and Roux-en-y gastric bypass procedures. We generated propensity scores at each site to summarize the baseline confounders, including age, sex, race and ethnicity, year of procedure, body mass index, Charlson comorbidity score, individual comorbidities (e.g., asthma, diabetes), smoking status, blood pressure, and type of insurance. We used three methods that require only summary-level information from the 7 participating sites to obtain the propensity score-adjusted estimates: (1) stratified analysis of propensity score-defined strata, (2) case-centered analysis of risk set data, and (3) meta-analysis of site-specific effect estimates. Results from these analyses were compared with the result from an individual-level data analysis (i.e., the “gold-standard”).

Findings: The study included 7,342 eligible patients. In the individual-level data analysis, the adjusted hazard ratio was 0.71 (95% confidence interval: 0.59, 0.84) comparing adjustable gastric banding with Roux-en-y gastric bypass procedures. The corresponding effect estimate was 0.70 (0.59, 0.83) in the propensity score-stratified analysis, 0.71 (0.59, 0.84) in the case-centered analysis, and 0.71 (0.60, 0.84) in both the fixed-effect and random-effects meta-analysis.

Discussion: This empirical study shows that methods requiring only summary-level information produce results that are identical or very similar to results from individual-level data analysis when adjustment for multiple confounders is required.

Conclusion: Propensity score-stratified analysis, case-centered analysis, and meta-analysis produced results that are highly comparable to individual-level data analysis. These very practical and efficient analytic approaches can be considered when sharing of individual-level information is not feasible or not preferred in multi-site studies.