The compression index is an important soil property that is essential to many geotechnical designs. Over the decades, a number of empirical correlations have been proposed to relate the compressibility to other soil index properties, such as the liquid limit, plasticity index, in situ water content, void ratio, specific gravity, etc. The reliability and thus predictability of these correlations are always being questioned. Moreover, selection between simple and complicated models is a difficult task and often depends on subjective judgments. A more complicated model obviously provides "better fit" to the data but not necessarily offers an acceptable degree of robustness to measurement noise and modeling error. In the present study, the Bayesian probabilistic approach for model class selection is used to revisit the empirical multivariate linear regression formula of the compression index. The criterion in the formula structure selection is based on the plausibility of a class of formulas conditional on the measurement, instead of considering the likelihood only. The plausibility balances between the data fitting capability and sensitivity to measurement and modeling error, which is quantified by the Ockham factor. The Bayesian method is applied to analyze a data set of 795 records, including the compression index and other well-known geotechnical index properties of marine clay samples collected from various sites in South Korea. It turns out that the correlation formula linking the compression index to the initial void ratio and liquid limit possesses the highest plausibility among a total of 18 candidate classes of formulas. The physical significance of this most plausible correlation is addressed. It turns out to be consistent with previous studies and the Bayesian method provides the confirmation from another angle.