NutriNet-Santé is an ongoing general population prospective cohort initiated in France in 2009 (https://etude-nutrinet-sante.fr/). Its design and specific aims are detailed elsewhere [24]. In short, recurrent multimedia calls target adults who can follow an online research protocol in French. NutriNet-Santé was approved by the Institutional Review Board of the French Institute for Health and Medical Research and by the National Commission on Informatics and Liberty. Eligible volunteers provide informed consent prior to enrollment.

At inclusion and yearly thereafter, participants complete questionnaires regarding socio-demographics and lifestyle, anthropometrics, physical activity and sedentary behavior, dietary intake (every 6 months), and health status. Additional nutrition- or health-related questionnaires are sent on a monthly basis.

Given the large number of retired participants (49.5%) to whom four WHODAS 2.0 items regarding work and school activities were not relevant, we excluded these items from the total score, as in prior studies [42]. We calculated scores for the remaining 32 items based on the WHODAS 2.0 algorithm: 100 × (total score for each domain)/92 [23]. Participants with invalid or incomplete mental health or disability data were excluded. To avoid potential acquiescence bias, we also excluded n = 56 for CES-D and n = 39 for STAI-T for marking the same response number for all questions even for inversely-scored items. A total of 27,529 participants were available for the analyses. Using SAS Proc Surveyselect and guided by prior multimorbidity research [43], participants were randomly assigned in a 3 to 1 ratio into one of two groups: a development subsample N = 20,000 to develop the indexes and a comparison subsample N = 7,529 to test and compare the performance of the developed indexes (Supplementary Figure S1). Participant characteristics are presented as number (%) from chi² tests (categorical variables) and median (1st and 3rd quartile) from Wilcoxon’s rank sum tests (continuous variables).

For the index development, we first checked the outcome distribution, expected to be right-skewed [23]. Next, we applied Tobit regression and Tweedie regression (logarithm link function and variance power function = 1.4). Both regression models included the 6 mental health conditions as independent variables and the WHODAS 2.0 score as the dependent variable, with adjustment for age and sex. The former is a regression model appropriate for a dependent variable having either left- or right-censoring [44]. The latter is a regression model appropriate for a dependent variable with positive values and exact zeros, and has been previously used to analyze the WHODAS 2.0 score as an outcome [45]. We compared the performance of the two models given the characteristics of the WHODAS 2.0 score distribution (i.e., dependent variable with positive values and exact zeros, right-censoring), to select the one with a better model fit using Akaike’s Information Criterion (AIC; smaller values reflect a better fit) [46]. Next, interactions of each mental health condition with sex and age were tested with a significance level p < 0.10, allowing to balance the reduced statistical power of interaction tests [47] and the risk of committing type I error. Afterwards, two weighting schemes used in the multimorbidity literature were compared by fitting Tweedie regressions with the developed index as the exposure and WHODAS 2.0 total score as the outcome. The first weighting scheme was the beta coefficient (β) multiplied by 10 for each mental health condition [48] and the second was the β divided by the β for age modelled in 5-year increments [49]. Model fit was compared using AIC and the coefficient of determination (R²) [50]. Finally, mental multimorbidity indexes calculated with two different rounding systems for the weights (integer and one decimal place) were compared using AIC and R². All model fit comparisons were conducted in the development subsample. The score of the developed index represents the sum of the weights attributed to each mental health condition for each individual.

Next, Tweedie regression models were fitted in the comparison subsample to test whether the developed indexes were associated with disability. Model 1 was unadjusted; Model 2 was adjusted for age and sex; Model 3 was additionally adjusted for confounders selected using a directed acyclic graph (BMI, education, employment, residential area, children in household, lack of affection during childhood, current household financial situation, self-perceived health status, smoking, physical activity, sedentariness, and lifetime prevalence of major chronic disease). Missing covariate values were imputed using the Multiple Imputation by Chained Equations method (20 imputed data sets) [51]. Model fit parameters and predictive performance of the indexes were compared against a simple count of mental health conditions using AIC, R², and calibration plots, respectively [52]. Finally, we assessed discrimination of the indexes (ability to discriminate among individuals with and without disability) by dichotomizing the WHODAS 2.0 score (disability defined as ≥90th percentile in the full sample [53, 54] = 23.9 points) and using Harrell’s concordance index (C-index) corresponding to the area under the receiver operating curve for binary outcomes [52].

All tests were two-sided and p < 0.05 reflected statistical significance. The variance power function of the Tweedie regression was determined using the Tweedie package of R version 2.3.5. SAS^® version 9.4 (SAS Institute, Inc., Cary NC, United States) was used for all other statistical analyses.

NutriNet-Santé participants included in this study were more likely to be male, older, married or cohabiting, to have lower education, higher physical activity, depressive symptoms, anxiety, subjective cognitive difficulties, and lifetime prevalence of chronic conditions compared to their excluded counterparts. Likewise, they were less likely to be retired, to live in urban areas, to have financial difficulties, children in the household, eating disorders, chronic insomnia, or AUD than those excluded from the analyses (data not tabulated; all p < 0.01).

Descriptive characteristics and mental health status in the full sample and in each subsample are presented in Tables 1, 2, respectively. Frequency distributions of mental condition pairs are presented in Supplementary Table S1. In the full sample (N = 27,259), mean age was 59.5 ± 13.7 years, 73.5% of the participants were women, and the median WHODAS 2.0 score was 5.4 (Interquartile Range 1.1–13.0). Overall, 58.1% of the participants had ≥1 mental health condition, with the most prevalent pair being anxiety and depressive symptoms; 15.5% and 17.6% of the participants had 2 and ≥3 mental health conditions, respectively. No significant differences regarding participant profiles were observed between the development and comparison subsamples.

As the Tweedie regression model presented a better fit to the observed data than did the Tobit regression model (AIC = 124,344 for Tweedie; 130,859 for Tobit), the former was retained for subsequent analyses. A significant interaction with sex was observed for AUD (p < 0.05) and cognitive difficulties (p < 0.07). Thus, we fit separate Tweedie models to develop a general mental multimorbidity index and a sex-specific index.

All mental health conditions were associated with the disability score except for AUD in women (Table 3). The indexes were calculated with two weighting schemes (β*10 and β/β of age for 5-year increments) and with two rounding systems (integer and one decimal place). Those with β*10 (AIC 124,560.9 for β*10 vs. 124,561.1 for β/β of age for 5-year increments) and one decimal place (AIC 124,560.5 for one decimal place vs. 124,574.6 for integer) seemed to provide the best fit especially for the sex-specific index. The final weights attributed to each mental health condition in each index are presented in Table 3. Absolute values of the beta coefficients were generally higher in men than in women, except for depressive symptoms. In both sexes, the highest absolute weight was obtained for subjective cognitive difficulties (6.2 for men, 6.4 for women). The lowest absolute weights were observed for AUD in men (weight = 1.6) and for eating disorders in women (weight = 2.1). The value distribution of the general index (score range: 0–20.5) and the sex-specific index (score range: 0–22.4) is presented in Supplementary Figure S2.

The general and the sex-specific indexes were positively associated with disability in the crude, age- and sex-adjusted, and the fully-adjusted Tweedie regression models (Table 4).

Both indexes showed somewhat lower AIC and higher R² values than those obtained with a simple count of mental disorders (Supplementary Table S2). Absolute values of the discrimination parameters (C-index) were also slightly higher than those of a simple count (Table 5). Calibration plots showed that the developed indexes had slightly better calibration than a simple count of mental disorders (Supplementary Figure S3).

We obtained mental health data via self-report questionnaires commonly used in epidemiological research, which is consistent with the intended use of these indexes. Yet, despite the fact that the mental health assessment tools have been validated against diagnostic criteria [25, 27, 31–33, 36], they do not reflect clinical diagnoses; this is a limitation of the study. In addition, CES-D, STAI-T and SCOFF do not distinguish among subtypes of depression, anxiety and eating disorders, respectively. For insomnia, we relied on DSM-5 and ICSD-3 criteria. Future confirmatory research could replicate the analysis using different tools for the same mental disorders assessed in representative adult samples. Overall, mental disorders tend to be under-diagnosed in the general population partly due to stigma and perceived discrimination [61]. Mental health assessment in epidemiological studies can help identify individuals with symptoms even without a clinical diagnosis. Next, consistent with the objectives and with prior research, we used the 90th percentile as cutoff of the WHODAS 2.0 score [53, 54, 62] to identify participants with disability. Another limitation stems from the COVID-19 pandemic, which disrupted the way of life and had a deleterious impact of mental health [63, 64], and which coincided with the 2020–2022 questionnaire administration. However, according to a French national survey, average alcohol and tobacco consumption were fairly stable even during the lockdowns [65]. We can speculate that the context might have had a minimal impact on the measures. Finally, our sample was composed predominantly of women and individuals of higher socio-economic status than the respective proportions in the general population [66, 67], which might have led to an underestimation of the impact of mental disorders on disability. Thus, it will be necessary to replicate and confirm the findings using other study populations.

This study has several important strengths. First, we relied on a rich database with available information on a variety of socio-demographic, lifestyle and mental health characteristics. Next, to our knowledge, this is the first mental multimorbidity index development research based on common mental health conditions meeting specific criteria with relevance to public health, and thus applicable in an epidemiological context across age. Moreover, these are the first multimorbidity indexes based on disability, to which mental disorders have a substantial contribution [23]. Disability was measured using WHODAS 2.0 which is commonly used in epidemiology [23]. Assessment of mental disorders relied on validated instruments [25, 27, 31–33, 36] administered in a large and diverse sample of men and women. The prevalence of each mental disorder and that of mental multimorbidity in this study were consistent with those reported in the literature [5, 68–73]. Further, the performance of two statistical approaches (i.e., Tobit and Tweedie regressions) was compared, allowing us to select the one with the better model-fitting performance (smaller AIC value) to develop the indexes. In addition, several weighting schemes used in the multimorbidity literature were tested to identify the one with the best performance [48, 49]. The developed multimorbidity indexes had good discrimination and calibration, and were significantly associated with disability in fully-adjusted models. These indexes could help identify at-risk individuals without a full clinical examination, thus could provide cheaper and earlier screening.

We developed a general and a sex-specific mental multimorbidity indexes that could help with needs assessment and prevention. These indexes could help identify at-risk individuals without a full clinical examination, thus could provide earlier and cost-efficient screening. Following validation, future research with representative samples could test their predictive potential regarding health status, medication use, primary care consultations, and QOL in longitudinal analyses.