Identifying the Big Shots—A Quantile-Matching Way in the Big Data Context

Abstract

The prevalence of big data has raised significant epistemological concerns in information systems research. This study addresses two of them—the deflated p -value problem and the role of explanation and prediction. To address the deflated p -value problem, we propose a multivariate effect size method that uses the log-likelihood ratio test. This method measures the joint effect of all variables used to operationalize one factor, thus overcoming the drawback of the traditional effect size method (θ), which can only be applied at the single variable level. However, because factors can be operationalized as different numbers of variables, direct comparison of multivariate effect size is not possible. A quantile-matching method is proposed to address this issue. This method provides consistent comparison results with the classic quantile method. But it is more flexible and can be applied to scenarios where the quantile method fails. Furthermore, an absolute multivariate effect size statistic is developed to facilitate concluding without comparison. We have tested our method using three different datasets and have found that it can effectively differentiate factors with various effect sizes. We have also compared it with prediction analysis and found consistent results: explanatorily influential factors are usually also predictively influential in a large sample scenario.