Abstract
<p>Ceiling and floor effects occur when proportions of observations fall at measurement boundaries, obscuring variation beyond the scale range. In psychological network analysis, these effects may attenuate covariances and distort Gaussian graphical model edges, topology, and centrality. The current study examined these consequences and proposed a two-step method-of-moments correction that recovers marginal moments and pairwise correlations under censoring before network estimation. A simulation study varied sample size (n = 200, 400, 800) and six censoring profiles across a 10-node network. Networks were constructed using EBICglasso, Bonferroni-corrected significance testing, and false discovery rate (FDR)-corrected testing. The naive covariance estimator showed substantial negative relative bias across conditions (-36.4% to -57.7%), which did not diminish with larger samples. The proposed correction reduced covariance bias to between -1.8% and 1.8% and improved true-edge bias and structural recovery across all methods, sample sizes, and censoring profiles. Improvements were largest at moderate sample sizes and when censoring affected hub-related nodes or involved mixed floor and ceiling effects. Centrality recovery improved consistently for EBICglasso and Bonferroni estimation, whereas FDR estimation showed less consistent gains because restored signal increased both sensitivity and inclusion of weak associations. An empirical illustration using a 17-symptom posttraumatic stress disorder network showed stronger corrected edges among floor-affected avoidance and numbing symptoms and changes in the most central symptom across selection methods. Ceiling and floor effects therefore represent network-relevant measurement problems. The proposed correction offers an accessible approach for reducing censoring-induced distortion, although conclusions remain dependent on sample size, network topology, and edge-selection procedures.</p>