Recently, network-based approaches have provided an important contribution for the understanding of mental disorders. A growing number of statistical models, developed in the context of continuous variables in high-dimensional settings, are currently being used to infer dependencies between network elements (e.g., symptoms or behavioral elements) in psychometrics. However, psychometric datasets typically correspond to low-dimensional statistical settings, namely with a low number of variables collected from a large enough sample size and the variables collected are ordinal rather than Gaussian. In this large-scale simulation study, we tested and compared the performance of 14 methodological approaches including several that, to our knowledge, have never been tested in the context of psychometrics network inference. We assessed the impact of various factors such as the sample size, the number of variables (i.e., network elements), the density of the true underlying graph and the number of ordinal levels. We conclude that the simple and classic statistical methods are undervalued in the current practice, while polychoric correlations appear to have limited additional benefits. We recommend researchers to systematically rely on more than one method in their analyses.