A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference

Area of Science:

• Biostatistics
• Data Science
• Clinical Research Methodology

Background:

• Data integration enhances study efficiency and generalizability.
• A common challenge is differing outcome measures across datasets.
• Opioid use disorder (OUD) studies often use varied withdrawal symptom scales.

Purpose of the Study:

• To investigate the conditions under which integrating studies with disparate outcome measures yields efficiency gains.
• To analyze the trade-offs between efficiency gains and potential bias when fusing datasets with non-identical outcome measures.
• To provide guidance on selecting appropriate assumptions for data fusion in OUD research.

Main Methods:

• Developed three sets of assumptions with varying strengths to link disparate outcome measures.
• Employed theoretical analysis and empirical evaluation.
• Conducted a case study integrating the XBOT and POAT datasets using different linking assumptions.

Main Results:

• Asymptotic efficiency gains from data integration are achievable only under the strongest linking assumption, which risks bias if misspecified.
• Milder assumptions may offer finite-sample efficiency gains, but these diminish with increasing sample size.
• Varying assumptions linking the Subjective Opiate Withdrawal Scale (SOWS) and Clinical Opiate Withdrawal Scale (COWS) demonstrated potential efficiency gains and risks of bias.

Conclusions:

• Careful selection of assumptions is crucial when integrating datasets with differing outcome measures.
• Data integration for OUD studies requires balancing potential efficiency improvements against the risk of introducing bias.
• Findings offer practical guidance for researchers addressing heterogeneity in outcome measures during data fusion.