An Empirical Investigation of The Effect of Proxy Response and The Merits of Its Remedial Measures

Sammanfattning: In the event of missing data, substitution of data from proxy sources are usually considered a very useful alternative when available to avoid the problem of missingness. Nonetheless, research has also shown that this approach often induces “response bias”. This bias has been known to vary significantly from study to study depending on what is being evaluated. As an extension to the study of Lapin et al. (2021), this study aims to evaluate the effect of binary proxy response under varying degrees of biasness and the merits of its usage in comparison with a few commonly used methods for handling missing data. Specific questions around the comparison of proxy information to self-responses, proxy bias issues and decision making in the absence of self-responses under the missing at random (MAR) mechanism were evaluated. In this study, three levels of bias (i.e., 10%,30%, and 50%) obtainable in a binary proxy response were investigated (proxy substitution) alongside a few commonly used remedial measures (i.e., complete case analysis, multiple imputation, and inverse probability treatment weighting). A Monte Carlo simulation experiment was conducted with a logistic regression model of three explanatory variables (consisting of a binary, discrete and continuous data types). The experiment was conducted under different MAR mechanism with varying sample sizes (100, 500, 1000, 5000, 10000). The various methods were compared using Mean Square Error (MSE) criterion and the relative MSE. The findings of this study show that the performance of each method is highly dependent on the sample size, proportion of missing data under MAR mechanism, data type and error-in-variable. However, in the absence of proxy response, this study recommends the use of inverse probability treatment weighting (IPTW) provided the sample size is large. The findings of the simulation study were used invalidating the results of an existing study conducted with data obtained from the Swedish National Board of Health and Welfare Survey (2017) which consisted about 43% proxy response.