Four dimensions have been included in the HLCI – namely diet, physical activity, alcohol consumption, and tobacco use. In defining the dimensions, we undertook a process that aimed to optimally match the available data from the AVQ survey with guidelines and existing literature. These dimensions were structured as ordinal variables, following the ordinal nature of several elementary indicators.

A proper dietary regime is closely linked to a health condition and the pursuit of healthy longevity. An inappropriate diet, coupled with a sedentary lifestyle, can lead to overweight, a condition characterised by excessive body fat accumulation which can significantly impact the health of a population. Moreover, the composition and quality of diet, not merely its impact on body weight, can play a protective role against the onset of cardiovascular, neurodegenerative, and neoplastic conditions [20].

To define the diet dimension, we considered the available information from the AVQ survey regarding various categories of food and beverages (Table 1). The consumption frequency for various food groups was recorded through the following possible answers: more than once a day, once a day, several times a week, less than once a week, never. The consumption frequency for non-alcoholic beverages, particularly water and carbonated drinks, is instead assessed through the question: ‘In what quantity do you habitually consume the following beverages?’ with categorical responses ranging from no consumption to more than a litre a day. Starting from these responses, the daily consumption frequency of foods is calculated through an approximation (see Table S1 in Supplementary Appendix 1) and, if necessary, the overall consumption frequency for a food category is calculated by summing the consumption frequency of its sub-food items. Then, for each food/beverage category a threshold value above/below which the amount of food is considered at risk is identified from the dietary guidelines provided by the Council for Research in Agricultural and Analysis of Agricultural Economy (CREA), last revised in 2018 [21, 22]. Finally, the diet dimension is created from the sum of individual risky behaviours among the 10 food categories and non-alcoholic beverages considered, defining the following categories: 0, 1, 2, 3, 4 and 5+.

The consumption of alcoholic products is associated with the development of numerous non-communicable chronic diseases (NCDs) and it can create dependency. Established effects include increased high-density lipoprotein cholesterol and antithrombotic activity [25]. Overall, the evidence does not indicate any safe level of alcohol consumption [25].

In the context of alcohol consumption assessment, the focus lies on quantifying the daily intake of alcoholic units. An alcoholic unit is defined as 12 grams of ethanol, equivalent to the alcohol content typically found in:

In order to distinguish individuals having daily alcohol intake from those who consume it sporadically, the AVQ survey poses the question: ‘In what quantity do you typically consume the following alcoholic beverages?’ If an individual reports daily consumption within any of the specified alcoholic beverage categories, s/he is further asked: ‘How many glasses of wine, beer, alcoholic aperitifs, liqueurs, or spirits do you usually consume per day?’ For those who reported drinking alcohol more often than sporadically but less than daily, a value of 2 drinks per week was assigned. To derive the total daily alcoholic unit count, we aggregated the quantities of wine, beer, aperitifs, and spirits consumed daily by the respondents. The dimension related to alcohol consumption refers to the total number of alcoholic units consumed per day, grouped in the following categories: 0, (0-1], (1-2], 3+. In this notation, parentheses indicate exclusion of the boundary value, while square brackets indicate inclusion. For example, (0-1] represents more than 0 and up to and including 1 unit.

Tobacco smoking represents a major global public health concern, with the WHO estimating it is responsible for 8 million deaths annually [26]. Cigarette smoking is the primary risk factor for cancers (strongly associated with lung, oral cavity, throat, oesophagus, pancreas, colon, bladder, prostate, kidney, breast, ovaries, and certain types of leukaemia) and non-neoplastic respiratory diseases, such as chronic obstructive pulmonary disease (COPD) and it is also a significant cardiovascular risk factor (hypertension, stroke, and heart attack). Moreover, smoking negatively impacts the reproductive systems of both men and women, reducing fertility [27].

In the context of assessing smoking behaviour, we focus on the number of cigarettes smoked per day. In the ISTAT questionnaire, each individual is asked the following question: ‘Do you currently smoke?’ For those who report current smoking, additional questions are posed. These include: ‘How frequently do you smoke?’ with a focus on daily or occasional smoking, and ‘What is your primary smoking choice?’ with options such as cigarettes, cigars, or pipes for respondents to choose from. Only individuals who predominantly smoke cigarettes are further asked: ‘On average, how many cigarettes do you smoke per day?’. The classification of individuals within this dimension is as follows:

The total number of cigarettes was grouped into 6 categories: 0, 1-4, 5-9, 10-12, 13-15 and 16+.

Many methodological approaches are proposed in the literature for calculating weights. Decancq and Lugo [28] distinguish three classes of approaches to set the weights in the case of multidimensional indices: data-driven, normative and hybrid. Data-driven weights (based on statistical analyses) are a function of the distribution of the dimensions analysed and are not based, at least explicitly, on any value judgment. Normative weights, on the other hand, are set from value judgments (e.g. expert opinion and equal or arbitrary). Finally, hybrid approaches’ weights are both data-driven and based on some form of evaluation. In this context of analysis and given the need for medical expertise, we believe that the choice of weights should be entrusted to experts in the field.

The methodology employed in this study for deriving weights is the Analytic Hierarchy Process (AHP). Its originator, Thomas Saaty, an American mathematician, authored the first book on AHP in 1987 [29]. Saaty’s process allows incorporating experts’ judgments and evaluations in a structured way through a pairwise comparison process facilitating the evaluation of the relative importance between the dimensions and simplifying the task for experts. Moreover, it includes a consistency index to assess the logical consistency of the judgments expressed by experts during the comparisons. For each pair of dimensions, each expert is asked to determine which one is more important and by what degree on a scale from 1 to 9 (Table 2). The questions posed to the experts are listed in Section S2 in Supplementary Appendix 1.

Four experts have been involved in the evaluation process. The experts are PhD medical doctors, three of whom are specialised in Epidemiology and Biostatistics, and one has an extensive clinical experience. All of them are experts in Public Health, healthcare service organisation and actively engaged in research related to lifestyle, health promotion, and the assessment of community preventive interventions.

For each expert who performed the pairwise comparison of the dimensions, an individual AHP preference matrix Ai,i=1,…,4  was constructed. The corresponding inconsistency index was computed as a transformation of the Perron-Frobenius eigenvalue, here denoted as λmax:

where k is the matrix order, in this case equal to 4. To make this quantification of the inconsistency more intelligible, the CI was divided by the so-called Random Index RIk, which is the average CI calculated on a large number of randomly generated matrices of order k. This division yields the Consistency Ratio CR=CI/RIk [29]. By assessing consistency for each individual matrix, it is possible to detect and address inconsistent judgments, by excluding them from further analysis. As proposed by Saaty [29], a threshold of 10% for the inconsistency level was established. Consequently, one expert whose responses exceeded this limit (18%) was excluded from the subsequent analyses to maintain the overall consistency of the judgments. The resulting sample size was of three experts, which can be considered an appropriate number, since the problem is rather simple and involves only one level of comparisons [30]

Assuming that all the experts have the same expertise level, individual evaluations have been aggregated in a new matrix representing the overall preferences of the expert panel through the geometric mean, which satisfies the Arrow’s four conditions [31]. The weights attributed to the 4 dimensions correspond to the components of the principal eigenvector associated with the largest eigenvalue of the matrix constructed from these aggregated individual evaluations.

To aggregate the four dimensions composing the Index, a weighted Borda Rule was employed. The Borda rule is appropriate for working with ordinal variables. Originally, Borda scores were used to determine rankings in voting situations (see, e.g., [32–35]). Nowadays, they are applied in a variety of fields such as discrete multi-criteria analysis, composite indices, artificial intelligence, database queries, and Internet multiple search engines [36]. In the context of the Borda rule, and all scoring methods in general, the intensity of preference is measured by the scores given according to the rank positions.

Each dimension has a varying number of levels ordered from the healthiest behaviour to the riskiest (lower is better), as follows:

For each dimension, individuals can be ranked into different levels. The number of levels varies by dimension. The scores are obtained by assigning a value of 0 to the last rank, 1 to the second-to-last rank, and so on, with the top rank receiving a value equal to the number of levels minus one. Therefore, for the diet, physical activity, and tobacco consumption dimensions, the possible scores are 0, 1, 2, 3, 4, and 5, with a score of 0 representing the least healthy behaviour and a score of 5 corresponding to the healthiest behaviour. To handle the alcohol dimension, which has 4 categories, we adopted a proportional scoring approach, in order to preserve the overall scoring scale. With the minimum value always being 0 and the maximum value always being 5, for the intermediate positions, we distributed the remaining points equidistantly: 3.34 points for the second position and 1.67 points for the third.

The scores, calculated for each individual, are subsequently multiplied by the relative weight assigned to the corresponding dimension and computed using Saaty’s approach [29]. After performing this procedure for each of the dimensions, the four weighted scores are summed to obtain the overall score for each individual. So, the score associated to the i-th individual, si,  can be expressed by the following formula:

where:

Finally, to normalise the obtained score, a Min-Max normalisation approach was applied, using the theoretical minimum and maximum values. The normalised score, si′, was scaled to the interval [0, 100], according the following:

where 0 corresponds to the absolute worst lifestyle and 100 to the best.

The Borda Rule offers a simple and intuitive approach, whose computational efficiency becomes particularly advantageous when dealing with large datasets, as in our case. The Borda rule is often criticised for its violation of the independence of irrelevant alternatives (IIA), as the addition or removal of a non-winning candidate can affect the outcome, even if it shouldn’t logically influence the ranking of the remaining candidates [32]. However, the impact of IIA violations can appear less significant in larger datasets. As the number of units increases, the likelihood of ties (where multiple alternatives receive the same number of points) also increases, and the relative ranking between alternatives becomes less sensitive to small changes in individual rankings [32].

A descriptive analysis of the level of HLCI and its four dimensions was conducted for several demographic and socio-economic characteristics. The descriptive analysis of the four dimensions is important, because the same level of HLCI may be the result of different combinations of the four dimensions. Since the distributions do not conform to the assumptions of normality and homoscedasticity, the non-parametric Wilcoxon test is employed for comparing two groups (e.g. males vs females), while the non-parametric Kruskal-Wallis test is selected in the case of three or more groups (e.g. educational level). Conducting a large number of tests places our analysis within the context of multiple testing, in which the probability that at least one rejected hypothesis is a false positive is much higher than the nominal significance level. This probability is called the family wise error rate (FWER). To control it, we apply the Wilcoxon rank-sum test with continuity correction using the Bonferroni adjustment method.

After that, we proceeded with a multivariate analysis to determine which demographic and socio-economic characteristics significantly affect HLCI. Given that we are interested in capturing the effect of the demographic and socio-economic variables on the individuals who have a particularly low index, we adopted a quantile regression [37], in which we model the first decile of the index distribution.

The quantile regression model is:

where:

A a consequence, the p-th quantile is given by:

and the p-th quantile of the conditional distribution of yi given xi is:

The quantile regression can be applied to the data regardless of the distribution of the variable of interest, thus making unnecessary any hypothesis about it.

The initial sample size of the AVQ 2022 survey is 42,022, from which 6,106 individuals under the age of 18 are subtracted. The sample consisting only of adults has a sample size of 35,916 individuals, representative of the Italian adult population, which amounts to approximately 47.5 million individuals. In the data quality check, about 3,316 units are identified to have at least one missing value for one of the key variables for defining the dimensions described above. These units are removed, resulting in a final sample size of 32,600 units. Table 3 shows the main descriptive characteristics of the sample. All the analyses are performed using R 4.3.0 software [38].

As it can be observed in Table 3, the sample is fairly evenly split between males and females, with a slight majority of females (52.25%). The largest age group is 45-54 years (18.99%), followed by those 75 and older (15.16%) and 65-74 (14.63%). Married individuals form the largest group (48.72%), followed by singles (31.59%). As regards educational level, high school diploma holders form the largest group (39.54%), followed by those with middle school diplomas (27.41%). The sample shows a diverse geographical representation across Italy, with the South having the largest share (26.88%).

The estimated population size after applying sampling weights can be found in Table S3 in the Appendix. As can be observed by comparing the distribution of characteristics in the sample and in the population (last two columns of Table S3), after removing observations with missing data, the sample’s characteristics remain largely consistent with those of the population, with all differences falling within a ±5% range.

The overall structure of the indicator, inclusive of weights, is depicted in Figure 1. The dimension that receives the highest weight is related to tobacco consumption (0.387), followed by diet at 0.249, physical activity at 0.189 and alcohol consumption at 0.174.

The average score for the Healthy Lifestyle Composite Index (HLCI) is 61.77 (Table 4), exhibiting a left-skewed distribution (Figure 2). The analysis reveals that 61 individuals (0.19%) exhibit the minimum value of the composite index (0), indicating they adopt the least healthy behaviour across all four dimensions, while 9 individuals (0.03%) achieve the maximum value, demonstrating the healthiest score in all four dimensions (100).

Turning to the descriptive analysis of HLCI dimension scores, it is important to note that these scores do not involve weights, which are only used to aggregate the dimensions into the final index. Among these unweighted dimensions, the one with the highest average score is related to tobacco consumption (mean = 88.37), followed by alcohol consumption (mean = 72.03). These high scores are attributable to the relatively low number of individuals in the sample who smoke and consume alcohol daily (80.5% of individuals do not smoke, 82.0% do not consume alcohol or consume it very slightly). The lower scores in the physical activity (36.20) and diet (32.60) dimensions indicate areas where there is more room for improvement in the population’s health behaviours, and at the same point potentially less awareness of their importance.

In the formative approach, it is important to consider the correlations between dimensions. These should not be too high to avoid considering the same concept in several dimensions and thus overestimating it. To this aim, the Goodman and Kruskal’s Gamma index, which measures the covariation between two ordinal variables, was calculated (Table 5). Due to the large sample size in this study, even weak correlations emerged as statistically significant, highlighting the high statistical power of the analysis rather than necessarily indicating strong or practically meaningful relationships between variables. Covariations between diet and alcohol consumption, diet and physical activity, and tobacco consumption and physical activity can be considered as negligible. Covariations between diet and tobacco consumption, and alcohol consumption and physical activity, although higher, are weak. A moderate Gamma covariation is between tobacco consumption and alcohol consumption (0.232), indicating a positive monotonic relationship between the two variables. This is largely due to the high proportion of people who neither smoke nor drink. On the contrary, an inverse relationship is observed between alcohol consumption and physical activity. This means that high levels of physical activity are accompanied by low alcohol consumption, and vice versa. Interestingly, the same inverse relationship is not observed between physical activity and smoking habits.

Table 6 illustrates how the HLCI and its dimensions are distributed among individuals according to certain demographic and socioeconomic characteristics. The results of the Wilcoxon rank-sum test with continuity correction using the Bonferroni adjustment method are provided in the Supplementary Appendix 1 from Table S4 to Table S23.

HLCI is significantly higher for females compared to males (65.54 vs. 59.95, p < 0.001). Except physical activity, all the dimensions and in particular alcohol consumption exhibit higher scores for women, meaning lower consumption. The HLCI mean scores across age groups show a significant yet subtle U-shaped trend, with highest values in the 18-19 age group, decreasing until the 35-44 age group, and then increasing again. This non-linear trend is also evident from the results of the pairwise comparison tests (Table S4). The diet dimension scores increase with age, demonstrating the adoption of a healthier diet as age advances, except for a slight decline in the oldest age group (75+). All age groups are significantly different from their subsequent groups, with the sole exceptions of the 18-19 age group not differing from the 20-24 group, and no significant differences are observed among the age groups above 60 years old (Table S5). Physical activity decreases with increasing age, with this decline being gradual, as demonstrated by the pairwise comparison tests not being significant for adjacent age groups (Table S6); the only age group characterised by markedly lower physical activity levels, significantly different from all others, is the 75+ age group. Alcohol consumption appears more prevalent among adults and the older adults, with a drop for the 75+ age group. Also in this case the pairwise comparison tests results to be not significant for some adjacent age groups. Smoking is concentrated in the middle-age groups (Table S7).

Single individuals score lower on the diet dimension, while married individuals do not exhibit significant differences compared to those who are separated or divorced (Table S10). Widowed individuals have lower scores in physical activity, possibly due to age-related factors. Married individuals, analogously to separated or divorced (Table S13), demonstrate the lowest scores in alcohol consumption, while separated or divorced individuals and singles show the lowest scores in smoking (Table S12).

Individuals with a university degree or higher education exhibit the healthiest lifestyle, as reflected in both the composite index and individual dimensions, particularly in physical activity. Conversely, those with a middle school education show the lowest scores on the composite index and in the dimensions of diet, alcohol consumption, and tobacco use. Notably, the physical activity dimension demonstrates a clear and increasing trend corresponding to higher levels of education. Regarding the diet dimension, the differences are not significant between those with no education or elementary education and individuals with a university degree or postgraduate qualification, as well as between those with a middle school diploma and high school graduates (Table S15). The latter comparison is also not significant for alcohol consumption (Table S17).

Finally, regarding the geographical distribution, the South and Islands regions show the lowest scores, primarily due to low values in physical activity, despite having high scores in alcohol consumption. Conversely, the North-East presents the highest healthy lifestyle index score, mainly thanks to its performance in physical activity and tobacco consumption, in which it exhibits significant differences compared to all other areas except the North-west (Tables S21 and S22).

Table 7 shows the results of the quantile regression, in which we model the first decile of the index distribution.

The quantile regression on the first decile shows interesting results that differ from what emerges in the descriptive analysis. We start from an HLCI value at the first decile of 54.95, corresponding to a female subject, aged 18-19, single, with at most a primary school education and residing in the north-east.

Firstly, the difference between males and females, already evident in Table 6, is even more pronounced here, with males having a first decile value 11 points lower than females. This means that, at the lowest levels of the indicator, the difference between females and males is even greater. The difference between young people and adults is also much more marked, with first decile values more than 12 points lower between 25 and 34 years (even 16.21 between 35 and 44 years) compared to young people aged 18–19. The effect of higher education is also noteworthy: a graduate subject has a first decile of HLCI 9.48 points higher than those with at most a primary school education.

The ‘worst’ profile on the first decile of the distribution is given by a male subject, between 35 and 44 years old, separated or divorced, with a middle school education and residing in South Italy. His first decile HLCI is 30.77.

Stability was tested by splitting the sample initially into two and then into three random subsamples of equal size. The procedures were repeated fifty times, and the mean of the HLCI and its dimensions for each sample was calculated. The Wilcoxon signed-rank test (for two subsamples) and the Kruskal-Wallis test (for three subsamples) were employed to determine the frequency of significant differences among the means of the HLCI and its dimensions among the subsamples. Regarding HLCI stability, just three of the 50 trials based on two subsamples shows statistically significant differences between the HLCI means while, in the test with three subsamples, it was not observed that the means of HLCI differed significantly. As for the stability of the HLCI dimensions, the test with two subsamples provided the following results: one instance of significant differences in diet means (error rate = 2%), three cases of significant difference in physical activity and alcohol consumption (error rate = 6%), and two significant differences for tobacco consumption (error rate=4%). When the sample was divided into three subsamples, the means of the diet, alcohol consumption and tobacco consumption dimensions differ significantly in four cases (error rate = 8%) and for the physical activity in three cases (error rate=6%). In conclusion, we retain that such error rates are acceptable and that the validation results demonstrate good stability of the HLCI.

To further illustrate the potential of the Healthy Lifestyle Composite Index (HLCI) for temporal monitoring, we applied the same methodology to the most recent wave of the Aspects of Daily Life (AVQ) survey, collected in 2023². The 2023 AVQ survey involved 35,681 individuals aged 18 years and over (after excluding respondents with missing values on key variables, the final analytical sample consists of 32,785 individuals). The survey design remains unchanged compared to 2022, ensuring comparability of the results across years.

As expected, the results obtained for 2023 are very similar to those of 2022 (Table 8). This is consistent with the fact that the two survey waves are consecutive and no major changes in population behaviours are typically observed within such a short time frame. The main interest in applying the HLCI longitudinally lies in the possibility of comparing results over longer periods, which would allow the identification of relevant lifestyle trends and shifts. Such an in-depth temporal comparison, however, goes beyond the scope of the present paper, whose main focus is on the construction and illustration of the index.