Linear mixed-effects models (LMEMs) have become increasingly popular in psycholinguistics. However, many researchers do not fully understand how random effects structures affect generalizability. The authors argue that researchers using LMEMs for confirmatory hypothesis testing should use maximal random effects structures, which best capture the variability in the data. They show through theoretical arguments and Monte Carlo simulations that maximal models yield better generalization, especially when sample sizes are small. Random-intercepts-only models perform poorly in within-subjects or within-items designs, often worse than separate F1 tests. Maximal LMEMs are recommended as the "gold standard" for confirmatory hypothesis testing in psycholinguistics. The paper also discusses the importance of specifying random effects structures based on the experimental design, rather than data-driven methods, which can lead to incorrect conclusions. It highlights the need for principled model specification to ensure accurate generalization of findings. The authors emphasize that maximal models are more powerful and less prone to Type I errors, even in worst-case scenarios. They also note that data-driven approaches to random effects specification are often insufficient and can lead to overfitting or underfitting. The paper concludes that maximal LMEMs are the best approach for confirmatory hypothesis testing in psycholinguistics.Linear mixed-effects models (LMEMs) have become increasingly popular in psycholinguistics. However, many researchers do not fully understand how random effects structures affect generalizability. The authors argue that researchers using LMEMs for confirmatory hypothesis testing should use maximal random effects structures, which best capture the variability in the data. They show through theoretical arguments and Monte Carlo simulations that maximal models yield better generalization, especially when sample sizes are small. Random-intercepts-only models perform poorly in within-subjects or within-items designs, often worse than separate F1 tests. Maximal LMEMs are recommended as the "gold standard" for confirmatory hypothesis testing in psycholinguistics. The paper also discusses the importance of specifying random effects structures based on the experimental design, rather than data-driven methods, which can lead to incorrect conclusions. It highlights the need for principled model specification to ensure accurate generalization of findings. The authors emphasize that maximal models are more powerful and less prone to Type I errors, even in worst-case scenarios. They also note that data-driven approaches to random effects specification are often insufficient and can lead to overfitting or underfitting. The paper concludes that maximal LMEMs are the best approach for confirmatory hypothesis testing in psycholinguistics.