Sociodemographic characteristics

Our study population, drawn from an addictology center in Fez city, was divided into two samples. A total of 400 questionnaires were distributed, with 160 distributed for phase I and 240 for phase II. After filtering the data and removing invalid or unreliable observations, 150 valid responses were retained for phase I and 220 for phase II.

The two samples in our study exhibited similar sociodemographic characteristics. The mean age was 27.66 ± 7.96 (range 15–45) and 26.99 ± 7.94 (range 15–46), respectively, for the first and second samples. The dominant gender is male, representing 89% in the first sample and 85.90% in the second. Regarding marital status, 74% of participants in the first sample were unmarried, compared to 74.54% in the second sample. In both samples, approximately two-thirds of the participants had completed secondary education, accounting for 60.67%, and 57.27%, respectively. The majority of respondents were from urban areas: 84.67% in the first sample, and 79.10% in the second (Table 1).

Table 1 Characteristics of participants (N = 370)Exploratory factor analysis results

Before conducting the EFA, sampling adequacy and factorability were estimated by the Kaiser-Meyer-Olkin test and Bartlett’s test for the first sample. The total KMO value was 0.79, and all KMO values for each individual element ranged from 0.72 to 0.85, well above the permissible limit of 0.60 [29, 30]. Bartlett’s test of sphericity (χ2 = 282.313, df = 15, p < 0.001) demonstrated that the interitem correlations within the data were substantial enough to proceed with the EFA.

The scale’s factor number was established through examination of the scree plot, Horn’s parallel analysis, and adherence to the eigenvalue greater than one criterion [31]. EFA was performed using PAF extraction method with “promax” rotation to determine the factor structure of the scale. The PAF, one of the most effective estimation methods in EFA, was selected for several advantages. First, PAF does not rely on distributional assumptions [32]. Second, PAF demonstrates greater robustness in situations with unequal factor loadings, limited indicators per factor, and small sample sizes [33]. Finally, PAF excels in recovering weak factors, a quality shared a few other methods [33, 34].

A saturation threshold of at least 0.40 was applied. Items whose loadings did not surpass this threshold or loaded significantly on multiple factors were excluded from factors. After each iteration, the rotated factor matrix analysis displayed significant loadings and alterations in communality values.

As a result, a model consisting of three factors of the Moroccan version of the CAST subscales was adopted rather than sticking to a single-factor model, which was analyzed in depth. The unsatisfactory results of the single-factor model, as shown by fit indicators (e.g., CFI = 0.78; 216 TLI = 0.64; NFI = 0.77; GFI = 0.94; RMSEA = 0.26; SRMR = 0.07) reinforced the choice of the three-factor model.

The first factor, explaining 22% of the variance, comprises two items (smoking alone and smoking before noon) named “use patterns [UP]”, which pertain to the initiation phase of cannabis consumption, whether solitary or in a social context. The second factor, termed “use reduction [UR]”, explains 22% of the variance and is represented by two items (friends or family and attempted to reduce or stop) that address attempts to reduce cannabis consumption. Last, the third factor, “use disorders [UD]”, accounts for 26% of the variance and is loaded by two items (memory disorders and problems) related to potential disorders that may emerge because of cannabis use.

These three factors were then evaluated using CFA. With eigenvalues of 0.202, 0.130 and 3.502 for “UP”, “UR” and “UD” respectively, the factors each comprised two elements and represented a total variance of 70% (Table 2).

Table 2 Factor structure of the Moroccan version of CAST (6 items)Test of reliability

The construct’s internal consistency and reliability were assessed using Cronbach’s alpha coefficient. Cronbach’s α and item-total correlations were calculated for each construct and individual item, as shown in Table 2. The reliability statistics provide the true value of the overall Cronbach’s coefficient (α = 0.86). Additionally, the alpha values for the items within each subscale ranged from 0.83 to 0.84, indicating a high level of internal consistency. These results confirm that the Moroccan version of the CAST in our sample demonstrated strong internal consistency. It is noteworthy that alpha values of at least 0.70 and ideally above 0.80 are considered indicative of good consistency [35, 36]. Therefore, the obtained alpha values suggest that all the concepts assessed in the study were reliable.

Confirmatory factor analysisInterscale correlations

The three factors exhibited strong and statistically significant correlations (p < 0.001). The highest correlation was observed between “UP” factor and the “UR” factor (r = 0.77). Furthermore, the “UD” factor showed a positive correlation with the “UP” factor (r = 0.66) and was also positively correlated with the “UR” factor (r = 0.56) (Table 3).

Table 3 Results of composite reliability, average variance extracted, and correlations between latent constructsConvergent validity

The findings from the CFA also demonstrated that the standardized regression coefficients were above 0.70, and the factor loadings for the “reduced consumption” factor (UR1) were the lowest, with a value of 0.74. Conversely, the loadings for the other six factors were all greater than 0.77. Additionally, the t-ratios, computed by dividing the parameter estimate by the standard error, were greater than 1.64 for each factor-factor and factor-variable pair. These t-ratios indicated significant relationships between the variables, with p values below 0.001, signifying a high level of statistical significance. Therefore, considering the regression coefficients exceeding 0.50 and the significant relationships associated with the high t scores, the first-order CFA offered statistically acceptable evidence of convergent validity [23] (Fig. 1).

Fig. 1

Moreover, to confirm the reliability and convergent validity of the instrument, the composite reliability (CR) and average variance extracted (AVE) indicators were calculated [37]. With CR values between 0.76 and 0.88 and AVE values ranging from 0.62 to 0.78, the results of the entire factor analysis process were confirmed, suggesting a favorable fit of the CAST instrument to the collected data (Table 3).

Similarly, the correlation analysis between the detection outcomes of the validated instrument (CAST) and the Gold Standard instrument (MINI) indicated a statistically significant and relatively high correlation (r = 0.81, p < 0.001), consistent with the CR and AVE (Table 4).

Table 4 Correlation of CAST score and MINI scoreDiscriminant validity

To assess the discriminant validity of the model, which consists of measuring the degree of differentiation between overlapping concepts [38], two criteria were used: the Fornell & Lacker criterion and the HTMT [26, 27, 37].

Table 3 presents the intercorrelations between the dimensions of the latent factors, with the square root of the average variance extracted (AVE) values highlighted in bold. Among the correlations, the highest value (0.75) was observed between the factor’s “UR” and “UD”, while the lowest value among the square roots of the AVE values was 0.82. Notably, the diagonal values of the matrix were greater than the off-diagonal values in the corresponding rows and columns [25].

The HTMT criterion value should be below 0.85 for strict [39, 40] and 0.90 for liberal discriminant validity [41, 42]. Table 5 reveals that all matrix values are below 0.85, providing further support for the potential discriminant validity among all the concepts in the proposed model. Overall, the reliability tests and tests for convergent and discriminant validity consistently support the justification of the concepts in the measurement model based on both types of tests (Fornell and Larcker criterion and HTMT).

Table 5 Discriminant validity (HTMT Criterion)Fitness of the measurement mode

The results of the CFA showed that the fit indices for the three-factor model were good (Table 6). Specifically, the chi-square to degrees of freedom ratio (χ2/df) was 2.23, indicating an acceptable fit [43, 44]. The comparative fit index (CFI) = 0.99 (> 0.90) suggests that the fitted model is in very good agreement with the observed data [45]. The goodness-of-fit index (GFI) was 0.99 (> 0.90), reflecting a high level of model fit [46, 47]. The standardized root mean square residual (SRMR) was 0.02 (< 0.05), suggesting a small discrepancy between the model and the observed data [48, 49]. The standardized root mean square residual (RMSEA) was 0.07 (< 0.08), suggesting a reasonable fit between the model and the data [50,51,52]. The normed fit index (NFI) was 0.98 (> 0,90), and the non-normed fit index (NNFI) or Tucker-Lewis Index (TLI) was 0.97 (> 0,90) [46, 47].

Detection capability

The examination of CAST’s detection capacity using the MINI as the Gold Standard reveals high sensitivity and Positive Predictive Value (PPV) for CAST, both exceeding 0.90. However, the specificity and Negative Predictive Value (NPV) are relatively low. The optimal balance between sensitivity and specificity for the CAST scale is achieved at a cutoff of 3 and 4, identified by the maximum Youden index (Y = 0.75 and Y = 0.73) (Table 7).

Table 7 Screening characteristics of CAST across different cutoff points

The high discriminatory power of the scale is evident in the high AUC of 0.881 (95% IC: 0.83–0.92) revealed in the CURVE ROC (Fig. 2), which affirms its strong ability to distinguish individuals with a clinical diagnosis from those without it.

Fig. 2

ROC curve and AUC for CAST