*2.3. Data Analysis*

First, we checked for the presence of a common method bias in the responses using two tests: the Harman's one-factor test and the correlation matrix test [27]. In the first test we computed the variance explained by a single-factor exploratory model including all the items administered, considering the bias to be present if the proportion of variance explained by this single factor was higher than 50%. In the second test we considered the correlation matrix between all assessed variables, considering the bias to be present if correlations were higher than 0.90. In both tests, for each variable we considered all the observations (i.e., each participant assessed both before and during the lockdown).

Secondly, we assessed the effect of lockdown on the measured variables through a repeated-measures MANOVA followed by a series of one-way repeated-measures AN-COVAs to assess the effect of time on each dependent variable, while controlling for the effects of sex, age, and education level. As a measure of effect size, we used the partial eta squared which is recommended in order to improve the comparability of effect sizes between studies [28].

Then, we tested our hypothesis that lockdown onset impacted mindfulness, which affected psychological distress, in turn influencing sleep problems. To this aim, we tested the indirect effect of mindfulness on the effect of time on distress/sleep by using path analysis with the Huber-White robust standard errors estimator and bias-corrected confidence intervals that test indirect or mediated effects [29]. In particular, we tested two models. In the first one we included the total score for each scale, in order to assess the relationships between mindfulness, general distress, and sleep problems. In the second model, we used the subscale scores of each questionnaire to investigate the differential contribution of each facet or aspect to the considered effects. Regarding mindfulness, following the literature on MAT theory [19,20], we took into account only the variables considered to be related to either monitoring (i.e., observing) or acceptance (i.e., non-judging and non-reacting). In both models, we controlled for the effects of age, sex, education level, and chronotype (rMEQ score).

As both models were fully saturated (i.e., they perfectly fitted the data because they had as many parameters as there were values to be fitted) no goodness of fit scores could be calculated. In order to both obtain interpretable goodness of fit statistics and reduce the number of free parameters so to counterbalance the small numerosity of the sample, we also analyzed simplified versions of the models where all non-significant path (and covariates) were removed. For each model, we calculated the following fitting indexes: χ2

statistics, comparative fit index (CFI), Tucker–Lewis index (TLI), root mean square error of approximation (RMSEA), and standard root mean square residual (SRMR). Model fit was considered as adequate with the following values: non-significant χ2, CFI and TLI above 0.95, RMSEA of 0.06 or less, SRMR of 0.08 or less [30]. Raw data are available as Supplementary Materials.
