New Insights About the Psychometric Properties of the Montgomery‐Asberg Depression Rating Scale

ABSTRACT

Objectives

This study aimed to comprehensively investigate the psychometric properties of the Montgomery‐Åsberg Depression Rating Scale (MADRS) in a sample from Northeast Brazil.

Methods

Data were obtained from a cross‐sectional study of 303 participants with depressive disorders assessed using the MINI 7.02, MADRS, GAD‐7, and ASSIST 3.0. We evaluated reliability, construct validity, convergent and discriminant validity, and diagnostic accuracy.

Results

Regarding reliability, the instrument demonstrated robust internal consistency, supported by strong Cronbach's alpha and McDonald's omega coefficients. In terms of construct validity, Exploratory Factor Analysis (EFA) indicated a single‐factor solution, while the H‐index supported a replicable latent structure. However, formal assessments of closeness to unidimensionality suggested that the scale is not strictly unidimensional. For convergent and discriminant validity, correlational analyses showed a moderate association with anxiety symptoms and weak associations with substance use measures, respectively. Regarding accuracy, ROC curve analysis indicated that a cutoff score of 9 is useful for screening, whereas 18 may better identify a depressive episode.

Conclusion

While the MADRS remains a valuable tool, structural refinements could further enhance its psychometric performance across distinct clinical settings.

1. Introduction

Mood disorders, including Major Depressive Disorder (MDD) and Bipolar Disorder (BD), are characterized by significant mood disturbances that manifest cognitive, emotional, and neurovegetative symptoms (American Psychiatric Association 2013). The pooled lifetime prevalence of a mood disorder stands at an estimated 9.6% (ranging from 8.5% to 10.7%) based on 83 surveys (Steel et al. 2014). Furthermore, depressive symptoms contribute significantly to the overall burden of disease within both the unipolar as well as the bipolar spectrum (Bogie et al. 2024). Consequently, depressive symptom measures, such as the Montgomery‐Asberg Depression Rating Scale (MADRS), have been used to assess the severity of depression and track changes along the treatment (Sforzini et al. 2022).

The MADRS is a ten‐item tool designed to assess mood, cognitive, and somatic symptoms related to a depressive state (Montgomery and Asberg 1979). While the Hamilton Depression Rating Scale (HAM‐D) is the most used clinician‐rated measure of depression severity, MADRS has gained considerable popularity due to its robust inter‐rater reliability and remarkable sensitivity in discerning treatment‐related changes (Khan et al. 2004; Montgomery and Asberg 1979; Quilty et al. 2013). It can be administered relatively quickly and evaluates core depressive symptoms (Sajatovic et al. 2015). Despite its widespread use, there is no consensus on the most appropriate cutpoints for defining remission or clinical thresholds, with proposed values ranging from 6 to 20 (Hawley et al. 2002; Ketharanathan et al. 2016; Kjaergaard et al. 2014; Leontjevas et al. 2009; Müller‐Thomsen et al. 2005; Thase et al. 2021). However, evidence supports considering a score of 10 points or lower aligns with the remission benchmark on the HAM‐D (Zimmerman et al. 2004).

Satisfactory internal consistency of the MADRS and its strong correlation with depression assessment scales have been confirmed consistently, such as the HAM‐D and Beck Depression Inventory‐BDI (Hawley et al. 2013; Lahlou‐Laforêt et al. 2015; Santi et al. 2023). However, some degree of convergence has also been reported between MADRS scores with scales that focus on other constructs (e.g., anxiety) (Batmaz et al. 2014; Legendre et al. 2005; Seemüller et al. 2012). Additionally, while a single‐factor structure has been replicated as the best goodness‐of‐fit model across different samples, alternative solutions with two, three, or four factors have also been deemed plausible (Quilty et al. 2013).

In Brazil, the MADRS was adapted by Dratcu et al. (1987), and other studies have examined its psychometric properties within Brazilian samples (Carneiro et al. 2015; Duarte‐Guerra et al. 2016; Portugal M et al. 2012). These investigations have confirmed key aspects of the scale's reliability and validity, expanding psychometric assessments on the Brazilian version of MADRS. However, existing research has concentrated on specific aspects, such as reliability and criterion validity (Carneiro et al. 2015; Duarte‐Guerra et al. 2016; Knapskog et al. 2013; Portugal M et al. 2012). In addition, the psychometric evaluations of the MADRS have primarily been conducted in Brazil's south and southeast regions (Carneiro et al. 2015; Dratcu et al. 1987; Fernandes et al. 2019; Fleck et al. 2004; Portugal M et al. 2012). As a result, a more comprehensive understanding of the scale's properties within the Brazilian population is necessary.

In our cross‐sectional survey about disordered eating behaviors among individuals with mood disorders from a city in the Northeast of Brazil, we identified an opportunity to perform a psychometric evaluation of the MADRS using the collected data. Thus, this study adds evidence regarding the clinical and research utility of the MADRS and aims to assess its reliability, factor structure, convergent and discriminant validity, as well as accuracy in a sample of subjects with depressive disorders.

2. Methods

2.1. Participants and Study Design

This study consisted of a psychometric analysis of the MADRS using data from a cross‐sectional survey on mood disorders. A convenience sample was recruited through social media announcements and from public psychiatric outpatient clinics (Walter Cantídio University Hospital and Professor Frota Pinto Mental Health Hospital) between November 2021 and July 2023. Individuals attending outpatient care or who responded to an online interest form were invited to participate in an eligibility assessment. The inclusion and exclusion criteria are described in detail elsewhere (Lima et al. 2024).

Trained medical students conducted structured interviews based on DSM‐5 criteria and applied a standardized checklist to assess eligibility. A total of 448 volunteers underwent this in‐person evaluation. Of these, 145 did not meet the inclusion criteria or fulfilled one or more exclusion criteria. Therefore, 303 participants were considered eligible and completed the full assessment protocol under the supervision of an experienced psychiatrist. Data collection was carried out using survey instruments hosted on the REDCap (Research Electronic Data Capture) platform (Harris et al. 2019).

Psychometric analyses were conducted in the total sample, including participants with unipolar and bipolar mood disorders. This approach was chosen because the study aimed to evaluate the MADRS as a measure of depressive symptom severity across mood disorders rather than to compare its psychometric properties between diagnostic groups. Separate analyses would substantially reduce the sample size of the bipolar subgroup and could decrease the stability and precision of the exploratory factor, dimensionality, construct replicability, and accuracy estimates. Accordingly, formal comparisons of the psychometric structure of the MADRS between unipolar and bipolar disorders were beyond the scope of the present paper.

2.2. Measures and Procedures

Sociodemographic data, including sex, age, and educational level, were gathered from all participants using a structured questionnaire.

The Mini International Neuropsychiatric Interview version 7.02 (MINI 7.02) is a gold‐standard structured diagnostic interview commonly used for diagnostic confirmation in research settings (Sheehan et al. 1998). The depression and (hypo)mania modules of the MINI were utilized to determine the presence or absence of a current or past mood episode in accordance with DSM‐5 criteria (Amorim 2000). Operationally, this study considered having a Unipolar Mood Disorder when participants met the criteria for a current or past depressive episode only. Conversely, those who experienced both a current or past depressive episode and also a past of (hypo) mania were classified within the Bipolar Spectrum.

The MADRS is a 10‐item questionnaire that assesses the depressive symptoms experienced over the preceding week, with each item rated on a scale from 0 to 6 (Montgomery and Asberg 1979). The total score is determined by adding the scores of each patient's responses, with higher scores indicating greater severity of depressive symptoms. (Dratcu et al. 1987; Montgomery and Asberg 1979). In this study, we employed the version of the MADRS validated by Dratcu et al. (1987).

The GAD‐7 is a validated and widely used self‐report instrument designed to identify and assess the intensity of symptoms related to Generalized Anxiety Disorder (GAD) (Hinz et al. 2017). This instrument comprises seven items that evaluate the frequency of anxiety symptoms over the past 2 weeks, with each item rated on a scale from 0 (never) to 3 (almost every day) (Hinz et al. 2017). The total score is obtained by summing the scores of each item. GAD‐7.The Brazilian adaptation of the GAD‐7 has demonstrated strong reliability (Bártolo et al. 2017).

The Alcohol, Smoking, and Substance Involvement Screening Test (ASSIST 3.0), developed by the World Health Organization (WHO), is used to assess the risk associated with substance use. The Brazilian version has demonstrated good validity (Ali et al. 2013; Henrique et al. 2004). In this study, we analyzed the nicotine and alcohol subscales to examine the discriminant validity of the MADRS.

The choice of external measures was necessarily limited to the instruments available in the cross‐sectional original study. Therefore, the GAD‐7 was selected to assess convergent validity because anxiety and depression are distinct yet clinically related constructs. The alcohol and nicotine subscales of the ASSIST 3.0 were selected to assess discriminant validity because substance use is theoretically more distinct from depressive psychopathology, as measured by the MADRS.

All participants included in the final sample completed the proposed instruments in full, with no missing data for these measures.

2.3. Statistical Analysis

Descriptive statistics were used to characterize the sample. Categorical variables were presented as percentages and compared using the Chi‐square test. Continuous variables were reported as means ± standard deviations and compared using Student's t‐test or the Mann‐Whitney test. Data normality was assessed using the Shapiro‐Wilk test, while Levene's test verified the variance homogeneity. A two‐tailed significance level of 0.05 was adopted for all statistical analyses.

Internal consistency of the MADRS was assessed via Cronbach's alpha (α) and McDonald's omega (ω), calculated for the full scale and with item deletion. Reliability of GAD‐7 and ASSIST 3.0 subscales was also reported. Interpretation followed George and Mallery's thresholds (≥ 0.9 excellent; ≥ 0.8 good; ≥ 0.7 acceptable; ≥ 0.5 questionable; < 0.5 unacceptable) (George and Mallery 2003).

An Exploratory Factor Analysis (EFA) was performed to determine the number of factors for the MADRS. Factors could be extracted if Kaiser‐Meyer‐Olkin (KMO) > 0.8 (Kaiser 1974) and Bartlett's Test yielded a significant result (p < 0.05) (Bartlett 1950). Furthermore, the Measure of Sampling Adequacy (MSA) assessed the adequacy of the polychoric correlation matrix, with an MSA < 0.50 suggesting a potential need for item removal (Kaiser 1970; Timmerman and Lorenzo‐Seva 2011). In the case of an MSA value less than 0.5, the model will be tested without items with inadequate MSA.

To extract the factors, an Oblimin Rotation with Diagonally Weighted Least Squares (DWLS) was applied to the polychoric matrix (Bartlett 1950; Kaiser 1974). An optimized parallel analysis utilizing bootstrapping supported the decision about the number of factors to be retained (Timmerman and Lorenzo‐Seva 2011). We considered acceptable factor loads ≥ 0.3 and communalities ≥ 0.5 for each factor.

The goodness‐of‐fit of the model was evaluated using the Root Mean Square Error of Approximation (RMSEA), the Root Mean Square Residual (RMSR), and the Comparative Fit Index (CFI). Thresholds were set at RMSEA ≤ 0.08, an RMSR ≤ 0.08, and CFI > 0.90, which are in line with the literature (Brown 2015).

The Generalized H index (G‐H Index) quantifies how well a set of items represents a common factor and its potential for replication across different studies (“construct replicability”) (Ferrando and Lorenzo‐Seva 2018; Hancock and Mueller 2001; Rodriguez et al. 2016). Additionally, H‐Latent points out whether the theoretical model (the latent factor) is adequately supported by the continuous latent variables underlying the observed items. At the same time, H‐Observed indicates how well the latent factor can be identified based on collected data (Ferrando and Lorenzo‐Seva 2018). The H value ranges from 0 to 1, and values exceeding 0.8 suggest a stable factor model and replicable in other samples (Ferrando and Lorenzo‐Seva 2018; Rodriguez et al. 2016).

We assessed the closeness to unidimensionality using the Unidimensional Congruence (UniCo), Explained Common Variance (ECV), and the Item Residual Absolute Loadings (IREAL) for the full scale and each item. Data can be considered essentially unidimensional when UniCo > 0.95, ECV > 0.85, and MIREAL < 0.30 (Ferrando and Lorenzo‐Seva 2018).

Spearman's correlation coefficient (ρ) was calculated between the MADRS scores and measures expected to be convergent (GAD‐7) and discriminant (nicotine and alcohol subscales of the ASSIST 3.0). The strength of the correlations was interpreted as follows: negligible: 0.00–0.10; weak: 0.10–0.39; moderate: 0.40–0.69; strong: 0.70–0.89; very strong: 0.90–1.00. Convergent validity was indicated if |ρ| ≥ 0.4, while discriminant validity was determined by |ρ| ≤ 0.10 (Schober et al. 2018).

Receiver Operating Characteristic (ROC) analysis was conducted to evaluate the accuracy of MADRS in identifying depressive episodes, using the MINI as the gold standard. For each cutoff, we calculated sensitivity, specificity, positive and negative predictive values (PPV, NPV), area under the curve (AUC), Youden's index, the PPV × NPV product, and the F1 score. AUC values were interpreted as follows: ≥ 0.9 (excellent), 0.8–0.9 (good), 0.7–0.8 (fair), 0.6–0.7 (poor), and < 0.6 (fail) (Nahm 2022). Optimal cutpoints were determined using three methods: Youden's index (to balance sensitivity and specificity), the PPV × NPV product (López‐Ratón et al. 2014), and the F1 score, a harmonic mean of PPV and sensitivity that evaluates the model efficacy by effectively balancing precision and recall, particularly in the presence of class imbalance (Yacouby and Axman 2020).

We used Jamovi to conduct descriptive, reliability, correlation, and ROC analyses (The Jamovi Project 2023), and Factor for EFA, MSA, unidimensionality, and H index computations (Lorenzo‐Seva and Ferrando 2006).

3. Results

Descriptive statistics of the final sample (n = 303), including age and sex distribution, are presented in Table S1. The mean age was 37.03 years (± 11.72), and most participants were female (84.16%; n = 255). We observed that 67.99% (n = 206) of the participants were classified as Unipolar Disorder and 32.01% (n = 97) as Bipolar Disorder. There was only a significant difference regarding ASSIST 3.0 Alcohol subscale (p = 0.007) scores between unipolar and bipolar groups, while MADRS scores were very close to the threshold of the significance level between these groups (p = 0.049).

The MADRS showed good internal consistency (α = 0.862; ω = 0.869), and reliability remained stable across all items (Table 1). Other instruments also demonstrated adequate reliability (GAD 7: α = 0.779 and ω = 0.902; ASSIST 3.0 Nicotine Subscale: α = 0.918 and ω = 0.920; ASSIST 3.0 Alcohol Subscale: α = 0.829 and ω = 0.835).

TABLE 1.

Cronbach's α and McDonald's ω for the complete MADRS and each item dropped.

Criteria	Item	α	ω
Apparent sadness	1	0.839	0.847
Reported sadness	2	0.834	0.841
Inner tension	3	0.852	0.860
Reduced sleep	4	0.858	0.864
Reduced appetite	5	0.872	0.878
Poor concentration	6	0.847	0.855
Lassitude	7	0.842	0.850
Inability to feel	8	0.842	0.850
Pessimism	9	0.845	0.852
Suicidal thoughts	10	0.851	0.860
	Full scale	0.862	0.869

Factorability was confirmed with a KMO value of 0.86 (95% CI 0.78–0.88), and a significant Bartlett's Sphericity Test (p < 0.001). Overall, MSA, factor loadings, and communalities are satisfactory (Table 2). Although item 5 (“reduced appetite”) showed suboptimal properties regarding factor loading and MSA, its removal resulted in an increase in RMSEA from 0.103 to 0.113, a virtually unchanged CFI (0.975 vs. 0.976), and a slight decrease in RMSR from 0.074 to 0.069 (Supporting Information S1: Digital Content 2).

TABLE 2.

Measure of sampling adequacy (MSA), factor loadings, and explained variance by factors.

Item	Measure of sampling adequacy		Factor loadings	Communality	Explained variance by factors
	MSA	(95% CI)			Factor	Real data	Random data
1	0.82	0.72–0.86	0.81	0.85	1	55.03	24.15
2	0.83	0.74–0.86	0.88	1.00	2	11.76	20.36
3	0.88	0.75–0.91	0.62	0.63	3	7.94	17.46
4	0.91	0.76–0.94	0.54	0.51	4	6.91	14.75
5	0.67	0.35–0.81	0.29	0.54	5	6.07	12.59
6	0.90	0.81–0.92	0.70	0.64	6	4.46	10.59
7	0.89	0.81–0.92	0.75	0.75	7	3.45	8.61
8	0.92	0.84–0.94	0.74	0.72	8	2.57	6.59
9	0.85	0.75–0.89	0.72	0.85	9	1.75	4.37
10	0.87	0.74–0.91	0.64	0.70	—	—	—

Note: One factor should be retained because only factor 1 explained the variance of the real data more than the random data.

The optimized parallel analysis showed that only one factor explained more of the variance of the real data (53.03%) than in the simulated data (24.15%) (Table 2). For the full MADRS, the RMSEA was 0.103 (95% CI: 0.077–0.115), above the acceptable threshold of 0.08. However, the Root Mean Square Residuals (RMSR) stood at a value of 0.074 (95% CI: 0.059–0.080), and CFI was 0.975 (95% CI: 0.966–0.987), both within acceptable ranges (CFI > 0.90 and RMSR < 0.08). Additionally, a construct replicability analysis provided an H‐latent Index of 0.924 (95% CI: 0.905–0.936) and an H‐observed Index of 0.934 (95% CI: 0.912–1.074), both of which surpass the threshold of 0.8. Therefore, H indices suggest a well‐defined latent variable is likely found in other samples.

Table 3 reports the data on the closeness to unidimensionality assessment. The analysis has provided a UniCo of 0.91 (95% CI: 0.89–0.97) and IREAL of 0.23 (95% CI: 0.18–0.28), both according to the expected values for unidimensionality. However, we observed an ECV exactly of 0.85, and its lower limit of the CI (95% CI: 0.82–0.90) is slightly below 0.85. In addition, only item 5 displayed unsatisfactory values concerning the cut‐offs adopted for all indices (UniCo: 0.26 [95% CI: 0.04–0.87]; ECV: 0.21 [95% CI: 0.03–0.64]; IREAL: 0.63 [95% CI: 0.24–0.95]). Overall, these results suggest that the scale does not have robust evidence of closeness to unidimensionality.

TABLE 3.

Closeness to unidimensionality assessment for MADRS and their items.

Item	UniCo (95% CI)	ECV (95% CI)	IREAL (95% CI)
1	0.99 (0.94–1.00)	0.91 (0.73–1.00)	0.24 (0.00–0.45)
2	0.99 (0.98–1.00)	0.95 (0.83–1.00)	0.19 (0.01–0.38)
3	0.96 (0.73–1.00)	0.78 (0.51–0.98)	0.32 (0.09–0.59)
4	0.99 (0.95–1.00)	0.96 (0.77–1.00)	0.11 (0.00–0.30)
5	0.26 (0.04–0.87)	0.21 (0.03–0.64)	0.63 (0.24–0.95)
6	0.99 (0.94–1.00)	0.91 (0.74–0.99)	0.22 (0.03–0.40)
7	1.00 (0.99–1.00)	0.97 (0.87–1.00)	0.11 (0.00–0.27)
8	1.00 (0.99–1.00)	0.99 (0.92–1.00)	0.06 (0.00–0.21)
9	0.96 (0.75–0.99)	0.79 (0.53–0.91)	0.38 (0.22–0.62)
10	1.00 (0.96–1.00)	0.98 (0.79–1.00)	0.08 (0.00–0.32)
Full scale	0.91 (0.89–0.97)	0.85 (0.82–0.90)	0.23 (0.18–0.28)

Table 4 presents the level of linear correlation between MADRS scores and GAD‐7, as well as ASSIST 3.0 Nicotine and Alcohol subscales scores. A moderate correlation was observed between MADRS and GAD‐7 (ρ = 0.55; p < 0.001). Meanwhile, MADRS showed weak but significant correlation with ASSIST 3.0 Nicotine (ρ = 0.13, p = 0.017) and Alcohol (ρ = 0.12, p = 0.034) subscales.

TABLE 4.

Spearman's Rho (ρ) of the MADRS scoring regarding GAD‐7 and ASSIST 3.0 subscales.

Scales	MADRS	p‐value
GAD‐7	0.55	<  0.001
ASSIST 3.0
Nicotine	0.13	0.017
Alcohol	0.12	0.034

The ROC analysis using MINI 7.02 as the reference standard showed an excellent AUC of 0.902, indicating high accuracy (Figure 1; Table 5). Based on the optimal value of Youden's index (0.688), the score of 18 of MADRS emerges as the point of best balance between sensitivity and specificity (Se: 85.71%; Spe: 83.08%; PPV: 94.88%; NPV: 61.36%%). In another way, the cutpoint method selection based on predictive values (PPV × NPV:1.843) indicates that a nine threshold on MADRS has high Se (99.58%), PPV (87.45%), and NPV (96.88%), although it shows a loss in Spe (47.69%). The largest value of F1 (0.931) indicates that nine points on MADRS would have the best balance between Se (99.58%) and PPV (87.45%).

FIGURE 1.

ROC curve for MADRS using the MINI as a diagnostic tool. The ROC curve shows the sensitivity versus 1‐specificity for MADRS scores in relation to a depression episode diagnosed by MINI.

TABLE 5.

Results of the ROC test for MADRS regarding MINI 7.02.

Cutpoint	Se (%)	Spe (%)	PPV (%)	NPV (%)	AUC	YI	PPV × NPV	F1
0	100.00%	0.00%	78.55%	NaN a	0.902	0.000	NaN a	0.880
1	100.00%	6.15%	79.60%	100.00%	0.902	0.062	1.796	0.886
2	99.58%	6.15%	79.53%	80.00%	0.902	0.057	1.595	0.884
3	99.58%	13.85%	80.89%	90.00%	0.902	0.134	1.709	0.893
4	99.58%	18.46%	81.72%	92.31%	0.902	0.180	1.740	0.898
5	99.58%	21.54%	82.29%	93.33%	0.902	0.211	1.756	0.901
6	99.58%	24.62%	82.87%	94.12%	0.902	0.242	1.770	0.905
7	99.58%	29.23%	83.75%	95.00%	0.902	0.288	1.787	0.910
8	99.58%	35.38%	84.95%	95.83%	0.902	0.350	1.808	0.917
9	99.58%	47.69%	87.45%	96.88%	0.902	0.473	1.843	0.931
10	98.32%	50.77%	87.97%	89.19%	0.902	0.491	1.772	0.929
11	96.64%	52.31%	88.12%	80.95%	0.902	0.489	1.691	0.922
12	95.38%	56.92%	89.02%	77.08%	0.902	0.523	1.661	0.921
13	94.96%	58.46%	89.33%	76.00%	0.902	0.534	1.653	0.921
14	92.86%	58.46%	89.11%	69.09%	0.902	0.513	1.582	0.909
15	92.44%	63.08%	90.16%	69.49%	0.902	0.555	1.597	0.913
16	90.34%	70.77%	91.88%	66.67%	0.902	0.611	1.585	0.911
17	88.66%	80.00%	94.20%	65.82%	0.902	0.687	1.600	0.913
18	85.71%	83.08%	94.88%	61.36%	0.902	0.688	1.562	0.901
19	83.19%	84.62%	95.19%	57.89%	0.902	0.678	1.531	0.888
20	77.73%	86.15%	95.36%	51.38%	0.902	0.639	1.467	0.856
21	73.95%	87.69%	95.65%	47.9%	0.902	0.616	1.436	0.834
22	68.07%	89.23%	95.86%	43.28%	0.902	0.573	1.391	0.796
23	62.61%	89.23%	95.51%	39.46%	0.902	0.518	1.350	0.756
24	56.3%	89.23%	95.04%	35.8%	0.902	0.455	1.308	0.707
25	51.68%	92.31%	96.09%	34.29%	0.902	0.440	1.304	0.672
26	48.32%	95.38%	97.46%	33.51%	0.902	0.437	1.310	0.646
27	44.96%	96.92%	98.17%	32.47%	0.902	0.419	1.306	0.617
28	40.76%	98.46%	98.98%	31.22%	0.902	0.392	1.302	0.577
29	36.55%	100.00%	100.00%	30.09%	0.902	0.366	1.301	0.535
30	31.09%	100.00%	100.00%	28.38%	0.902	0.311	1.284	0.474
31	28.99%	100.00%	100.00%	27.78%	0.902	0.290	1.278	0.450
32	26.05%	100.00%	100.00%	26.97%	0.902	0.261	1.270	0.413
33	25.21%	100.00%	100.00%	26.75%	0.902	0.252	1.267	0.403
34	20.17%	100.00%	100.00%	25.49%	0.902	0.202	1.255	0.336
35	16.81%	100.00%	100.00%	24.71%	0.902	0.168	1.247	0.288
36	13.03%	100.00%	100.00%	23.9%	0.902	0.130	1.239	0.230
37	10.92%	100.00%	100.00%	23.47%	0.902	0.109	1.235	0.197
38	9.24%	100.00%	100.00%	23.13%	0.902	0.092	1.231	0.169
39	6.30%	100.00%	100.00%	22.57%	0.902	0.063	1.226	0.119
40	4.20%	100.00%	100.00%	22.18%	0.902	0.042	1.222	0.081
41	3.78%	100.00%	100.00%	22.11%	0.902	0.038	1.221	0.073
43	3.36%	100.00%	100.00%	22.03%	0.902	0.034	1.220	0.065
44	1.68%	100.00%	100.00%	21.74%	0.902	0.017	1.217	0.033
47	1.26%	100.00%	100.00%	21.67%	0.902	0.013	1.217	0.025
50	0.42%	100.00%	100.00%	21.52%	0.902	0.004	1.215	0.008

Note: Bold values indicate the optimal cutpoints based on the maximum Youden’s index (sensitivity and specificity), PPV × NPV product (positive and negative predictive values), and F1 score (harmonic mean of sensitivity and positive predictive value).

^a NaN: Not a number (There is no true negative case for the denominator of NPV).

4. Discussion

This is the first psychometric analysis of MADRS involving a Northeast Brazilian sample, and it presents a significantly larger sample size than most previous research on this topic in Brazil. To our knowledge, no Brazilian study has simultaneously evaluated the internal consistency, factor structure, and validity of MADRS in a large sample. We have also introduced innovative methods to assess its dimensionality, replicability, and cutoff point, leading to a deeper understanding of the MADRS.

Reliability analysis demonstrated good Cronbach's alpha and McDonald's Omega values (≥ 0.8) (George and Mallery 2003), which underscores the MADRS's utility as a reliable measure of depressive symptoms. Insofar as Cronbach's alpha has faced criticisms as a reliability measure, we also included the Omega, further confirming the good reliability for all items and full MADRS (Bonniga and Saraswathi 2020). This high internal consistency is aligned with previous research, including studies conducted with Brazilian samples (Carneiro et al. 2015; Dratcu et al. 1987; Kjaergaard et al. 2014; Portugal M et al. 2012). Thus, these findings highlight the significance of the MADRS in reliably assessing depressive symptoms across diverse populations.

Although sampling adequacy was satisfactory overall, item 5 (“reduced appetite”) showed the poorest psychometric performance across several indices, including factor loading, MSA, and item‐level indices of closeness to unidimensionality. Nevertheless, its communality exceeded 0.50, indicating that a substantial proportion of its variance was accounted for by the common factor solution. Previous studies have also reported comparatively weak psychometric performance for the appetite item of the MADRS (Johansson et al. 2025; Quilty et al. 2013). Taken together, these findings raise questions about the performance of item 5 and support further investigation of its formulation. However, our supplementary analysis showed that removing this item did not improve the psychometric solution; instead, RMSEA worsened, while the other psychometric indices (CFI and RMSR) improved slightly. Therefore, the present findings do not provide sufficient evidence to recommend item removal. Future revisions of the MADRS might examine whether item 5 could be reformulated to better capture appetite disturbances or, alternatively, whether its removal would improve scale performance across independent samples. Importantly, the present study evaluated the version validated by Dratcu et al. (1987); therefore, our findings identify a potential limitation of item 5 in this sample rather than support a modification of the validated instrument.

The optimized parallel analysis supported a single‐factor solution for the MADRS, accounting for 53.03% of the data variance (Timmerman and Lorenzo‐Seva 2011). This statistical approach is more accurate than using eigenvalues to determine the number of factors, thereby reducing the likelihood of retaining spurious factors (Timmerman and Lorenzo‐Seva 2011). However, evidence regarding model fit was mixed. Although the CFI (0.975) and RMSR (0.074) met the predefined criteria, the RMSEA (0.103, 95% CI 0.077–0.115) exceeded the acceptable threshold and may indicate residual model misfit (Brown 2015). Similar discrepancies across fit indices, including RMSEA values above 0.08 despite favorable values for other fit indices, have been reported in previous studies of the MADRS (Johansson et al. 2025; Tran et al. 2025). Thus, the present findings support a predominant single‐factor structure, although the fit indices did not provide uniformly favorable evidence of model fit.

The sample's Latent and Observed H‐indices point to a well‐defined latent variable, likely to be replicated in other studies (Hancock and Mueller 2001). These findings align with other study validations that observed a unidimensional structure for the MADRS (Duarte‐Guerra et al. 2016; Hallit et al. 2019; Vestin et al. 2024). Nevertheless, conflicting results of UniCo, ECV, and IREAL indicate that MADRS is not unquestionably unidimensional. Importantly, our data on closeness to unidimensionality are supported by alternative factor structures that have been demonstrated in distinct clinical contexts (Borentain et al. 2022; Farner et al. 2009; Ketharanathan et al. 2016; Quilty et al. 2013; Seemüller et al. 2023).

We have found a positive moderate correlation between MADRS and GAD‐7, suggesting convergent validity. This finding is consistent with Harris et al. (2024), who also reported a significant correlation between MADRS and GAD‐7 scores (r = 0.68, p < 0.001, 95% CI 0.49–0.81). Associations between anxiety and depressive symptom severity have also been reported using other depression measures, with Kertz et al. (2013) and Dhira et al. (2021) demonstrating significant correlations between GAD‐7 and the PHQ‐9 and CES‐D‐10, respectively. Taken together, these findings are in line with the well‐established relationship between depressive and anxiety symptoms and may reflect shared transdiagnostic processes such as affective distress and rumination (Funk et al. 2022; Toro Tobar et al. 2020).

Significant positive correlations were found between MADRS scores and the nicotine and alcohol subscales of the ASSIST 3.0. It is important to note that the correlations with the nicotine (ρ = 0.13) and alcohol (ρ = 0.12) subscales did not meet the formal threshold adopted for discriminant validity (|ρ| ≤ 0.10). However, both associations were weak and only slightly exceeded this predefined threshold, indicating a limited relationship between MADRS and ASSIST 3.0 subscales scores. In addition, associations of similarly small magnitude between depressive symptom severity and alcohol and nicotine use severity have been reported (Cherenack et al. 2026; Xiong et al. 2024). Notably, validation studies of the ASSIST have primarily examined its relationship with other measures of substance use and dependence, rather than its relationship with measures of depressive psychopathology such as the MADRS (Humeniuk et al. 2008; Rubio Valldolid et al. 2014). Nonetheless, these findings are valuable since studies discussing the discriminant validity of the MADRS in relation to addiction scales remain scarce.

A ROC test assessing MADRS scores concerning a Depressive Episode diagnosis related to the MINI 7.02 was performed. Results displayed an excellent AUC for all cutpoints, suggesting a meaningful accuracy in detecting a Depressive Episode. The best balance between sensitivity and specificity was achieved at a score of 18 in this sample, based on the highest YI. While there is no consensus on the cut‐off for MADRS, much of the existing literature has suggested lower values, such as 10 or 11 as comparable with the threshold remission for HAM‐D (Montgomery 1994; Zimmerman et al. 2004) However, a study in a Parkinson's Disease sample has established the best cut‐off to be 17 (Ketharanathan et al. 2016). In turn, a score of at least 20 has been regarded as the best threshold to diagnose Depression in patients with early‐onset Dementia (Leontjevas et al. 2009). Thus, a higher prevalence of severe depressive episodes in a sample may shift the ROC‐derived optimal cutpoint upward, as the test tends to capture the distributional characteristics of the sample's symptom severity (Nahm 2022).

While 18 may be considered confirmatory for diagnosis using the MADRS in our sample, a lower cut‐off should be used for screening purposes in various clinical settings. The PPV × NPV and F1 score pointed to 9 as the point of greater efficacy of the model to classify the cases correctly in this survey, despite a specificity of 47.69% (López‐Ratón et al. 2014; Yacouby and Axman 2020). Notably, the F1 score provides the best balance of precision and recall across cut‐off points in a model (Yacouby and Axman 2020). This alternative cut‐off approximates the findings from another Brazilian study involving the elderly, which has found a 10 score as the optimal cut‐off for diagnosing Depression in participants both with and without Dementia (Portugal M et al. 2012). Further, our findings converge with the notion that eight points indicate the absence of a depressive state, as proposed by Müller (Müller‐Thomsen et al. 2005). Finally, a MADRS score ≥ 9 strongly suggests the presence of a depressive state, consistent with similar findings from previous studies, and may serve for screening purposes of a Depressive Episode (Ketharanathan et al. 2016; Müller‐Thomsen et al. 2005).

This study has some limitations that should be mentioned. First, the study used a convenience sample from a cross‐sectional research project on food addiction in mood disorders, which prevented us from making causal inferences. Second, using a non‐randomized sample may introduce potential selection bias and limit the external validity. A third limitation is the absence of the HAM‐D in our assessments. A comparison of MADRS and HAM‐D scores could have provided a more comprehensive psychometric evaluation, particularly regarding the correlation of scores and severity ratings. Fourth, we were unable to determine the severity of depressive symptoms for MADRS based on the MINI evaluation, as this one does not provide severity criteria for a Depression Episode. Finally, although the combined analysis was consistent with the aims of this study, the present study was not designed to investigate diagnosis‐specific differences in the psychometric properties of the MADRS. Such comparisons should be addressed in specifically designed studies with adequately powered unipolar and bipolar samples.

The strengths of this work also deserve recognition. First, the analysis was made based on a large sample using optimized statistical approaches to assess the psychometric properties of the MADRS. Notably, some innovative approaches to reliability, construct validity, and metrics to evaluate cut‐offs have not been seen in studies validating the MADRS. Second, we used a gold‐standard interview (MINI 7.02) to conduct an ROC test. To our knowledge, none of the previous works in Brazil examined the criterion validity comparing MADRS with the MINI. Third, the self‐report scales for anxiety and addiction were previously validated and demonstrated satisfactory reliability within the context of this sample. Finally, rigorous eligibility criteria have contributed to excluding some mental disorders and general medical conditions that could potentially bias the diagnostic process.

This work has provided additional evidence supporting the reliability and validity of the Brazilian MADRS. While we replicated previous findings about this one, valuable insights have emerged from our findings. So, MADRS has remained a brief, useful, and valid instrument for clinical and research contexts since its development in 1979. However, as many years passed and social contexts have changed worldwide, further validation studies should consider reviewing this scale to adapt it to the 21st century and provide a more robust reliability and factor structure.

5. Conclusions

This study provides new psychometric evidence on the Brazilian version of the MADRS, based on the first analysis conducted in a sample from Northeastern Brazil, culturally distinct from those previously evaluated. We confirmed strong internal consistency and a replicable factor structure, although the scale is not strictly unidimensional. Our findings have shown moderate correlations between MADRS and GAD‐7 scores, while weak correlation with nicotine and alcohol ASSIST 3.0 scores. Furthermore, we identified two valid cutoff points: scores ≥ 9 for screening and ≥ 18 for confirmatory purposes. Although the MADRS remains a reliable and valid instrument for clinical and research use, updating certain items—such as the one related to appetite—could improve its performance in contemporary psychiatric settings.

Author Contributions

Samuel Vieira Pinho Neto: writing – original draft, formal analysis, writing – review and editing, conceptualization. Erivan de Olivindo Cavalcante: investigation, writing – review and editing, conceptualization. Mariana Alves Leao: investigation, writing – review and editing, conceptualization. Alexandre Menezes Sampaio: conceptualization, writing – review and editing. Matias Carvalho Aguiar Melo: conceptualization, writing – review and editing. Paulo Rodrigues Nunes Neto: conceptualization, funding acquisition, writing – review and editing, data curation, supervision, formal analysis, methodology, project administration.

Funding

PRNN received funding (grant number P20‐0171‐00053.01.00/20) from the Fundação Cearense de Apoio ao Desenvolvimento Científico e Tecnológico (FUNCAP; Ceará Research Support Foundation) under the Program Pesquisa para o SUS (Research Program for the Unified Health System).

Ethics Statement

The research project received approval from the Research Ethics Committee of the Walter Cantídio University Hospital at the Federal University of Ceará, in compliance with ethical and legal principles, and was approved under the protocol number 4,477,006 and Certificate of Presentation for Ethical Assessment number 38642520.0.0000.5045.

Consent

All participants signed a printed free informed consent form.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting Information S1

Table S1: Descriptive statistics of the sample.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.