Evaluation of statistical methods to study flexural strength of dental CAD-CAM composites

Determining the optimal statistical method is key for reliable interpretation of the flexural strength test. To date, statistical recommendations regarding the best approach and sample size are based on theoretical assumptions that may not hold in practice. Therefore, this study identified the optimal statistical approach for analyzing the flexural strength of computer-aided design/computer-aided manufacturing (CAD-CAM) composites using a real dataset. In this perspective, the flexural strength dataset of commercial CAD-CAM composites, Cerasmart 270 (CER), Katana Avencia (KAT), Grandio (GRN), and polymer-infiltrated ceramic network Vita Enamic (ENA) were used (10 blocks/material; 15 samples/block). Leave-K-out cross-validation was performed on this dataset with K = 3 to 9 blocks (a total of 967 scenarios per material). The goal was to compare seven common statistical methods: Normal distribution (1), Lognormal distribution (2), the 2-parameter Weibull distribution calculated by maximum likelihood estimation (MLE) (3) and least squares (LSQ) estimation (LSQ; with three different estimators 4-5-6); (4) 3-parameters Weibull distribution. These methods were assessed based on comparative performance, standalone performance, and lower tail prediction. The results highlight that all statistical approaches have limited reliability with small sample sizes (e.g., n = 30) employed in dental materials research. Factors such as inter-block heterogeneity and physical characteristics of CAD-CAM composites could affect the efficacy of the statistical methods. The 2P-Weibull distribution was less prone to overestimating strength at low failure probability. LSQ with the mean estimator (i/(n+1)) comparatively outperformed others, particularly with small sample sizes. Weibull analysis would be more reliable with over 60 samples, but ideally more than 100 is recommended.