ASSESSING SELECTION BIAS IN REGRESSION COEFFICIENTS ESTIMATED FROM NONPROBABILITY SAMPLES WITH APPLICATIONS TO...

Area of Science:

• Statistics
• Biostatistics
• Social Sciences

Background:

• Selection bias poses a significant threat to the validity of inferences drawn from nonprobability samples.
• This bias is particularly relevant in genetic studies using volunteer samples for polygenic score (PGS) analysis and in surveys of specific user groups, like smartphone users.

Purpose of the Study:

• To derive novel measures for quantifying selection bias in regression models (linear and probit) fitted to nonprobability samples.
• To enable sensitivity analyses of inferences to assumptions about nonignorable selection when aggregate-level auxiliary data are available.

Main Methods:

• Development of selection bias measures derived from normal pattern-mixture models.
• Application of these measures to assess bias in polygenic score-phenotype relationships in a Facebook volunteer sample.
• Quantification of bias in subgroup mean differences (employment duration) within a smartphone user survey.

Main Results:

• The proposed measures effectively quantify selection bias in both simulated and real-world nonprobability samples.
• Bias was identified and quantified in estimated polygenic score-phenotype relationships and subgroup employment duration differences.
• Performance evaluation using benchmark estimates from large probability samples confirmed the measures' utility.

Conclusions:

• The novel measures provide a robust framework for assessing and understanding selection bias in nonprobability samples.
• These tools are essential for improving the reliability of findings in genetic epidemiology and survey research relying on convenience samples.