**2. Method**

This is a cross-sectional, population-based study in the Northeast of Brazil, which used secondary data from the National Health Survey (PNS, in Portuguese), conducted by the Brazilian Institute of Geography and Statistics (IBGE, in Portuguese) in 2013. PNS is a household survey with national coverage, which was designed to represent the adult population, based on cluster sampling. Details of the design and sample selection process can be found in the PNS reports [18–24].

Altogether, 60.202 households in Brazil were visited and interviews conducted with individuals aged 18 years or older. A total of 205.546 individuals responded to the survey, among which 23.815 were elderly, and of these, 10.541 were male, 3238 from the Northeast [20]. This study included elderly male individuals (60 years of age or older) living in the Northeast region selected by the PNS-IBGE sampling process. Elderly men with missing information in the database were excluded.

The dependent variable of this study corresponded to the health service profile used by elderly men and was composed of ten questions from the PNS referring to the set of the use of health services. These questions were grouped and categorized through Latent Class Analysis (LCA) and were presented, after analysis, in a single health service use variable that represents a variety of phenomena to explain the outcome.

The LCA is a statistical approach that identifies distinct mutually exclusive groups (latent classes) based on the response patterns of categorical variables [25]. LCA works with heterogeneous data in which individuals are classified in the group by similar characteristics, that is, it is considered that individuals come from the same population and that the trajectory can be extrapolated to an entire population as well as the covariables that affect the trajectory will influence individuals in the same way [26]. Latent classes or trajectories aim to estimate the size and number of latent classes, the probability of the response of each individual and to assign latent class association to individuals in the population [27].

The independent variables were the PNS questions related to predisposing factors, capacity, and health needs of elderly men, organized according to the classic theoretical model for the Use of Health Services [7].

Predisposing factors were: North-eastern states; Condition of responsible or not for the domicile; Age; Color or race; Marital status; and Able to read and write. The capacity factors were: Possession of health insurance; and Participation in organized social activities (clubs, community or religious groups, elderly living centers). The necessity factors were: Diagnosis of any chronic, physical, or mental illness; Degree of difficulty in eating, bathing, use the bathroom, dressing, walking at home, lying, or getting out of bed, sitting or getting up from the chair, shopping, managing finances, taking medicine, go the doctor, and go out alone by transport; Occurrence of a fall in the last 12 months; and Perceived health status.

Regarding the variables referring to Basic and Instrumental Activities of Daily Living (ADLs and IADLs, respectively) there are some validated instruments [28–30] that evaluate these activities, however, not all the questions used in these scales were used in the PNS, such as urinary and fecal continence and the difficulty to use the phone. Thus, the set of questions that assessed ADLs and IADLs could not be grouped according to the instruments. As a result, the twelve variables referring to the degrees of difficulty to perform the ADLs and IADLs were also studied using the LCA method to form a single variable of the difficulty level to perform ADLs and IADLs.

To assess the latent class model and identify the number of classes that better define the object of study, some statistical criteria were considered. The first is entropy, the probability that the individual is perfectly classified in a particular latent class, whose measures can vary between 0 and 1, and the closer to 1 the value is, the more appropriate the model will be, indicating a good classification of the individual in the class [31].

Other criteria were considered such as the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC) and adjusted BIC, used for assessing the model's adjustments. In the analysis, the lower the value of AIC, BIC and adjusted BIC, the more suitable the model will be [32]. To assess the evolution of the testing model, the likelihood tests were used considering p < 0.05 as statistically significant values.

In this article, five models were tested, with two, three, four, five and six latent classes, to identify the number of classes that best represents the object of study according to the aforementioned statistical criteria. Weights and strata from the database for the LCA were considered.

In the descriptive statistical analysis, the quantitative variable corresponding to age was presented as a measure of central tendency and dispersion, with a 95% Confidence Interval (CI) being calculated. Qualitative variables were presented in the form of a frequency table and 95% CI.

In analytical statistics, the presence of an association between independent variables and the dependent variable (categorized through Latent Class Analysis) was investigated using the Rao-Scott test used in complex samples [33]. The significance level was 5%. The effect measures of the factors on the dependent variable were expressed by Odds Ratio (OR) and calculated by simple and multiple models of multinomial logistic regression, following the theoretical model of Health Services Utilization [7] from the assumption of the hierarchical approach.

Initially, a simple analysis was carried out on the blocks of predisposing factors, capacity factors and health needs. Within each block, variables with p < 0.25 were tested in multiple models [34]. In the end, the variables with p < 0.05 remained in the model for each block and were considered adjustment factors for the subsequent blocks.

The PNS-IBGE database is in the public domain and is available on the IBGE website (http://www.ibge.gov.br). The statistical programs used were IBM SPSS Statistics version 20 for data analysis, and Mplus 7.31 to establish latent classes in Latent Class Analysis (LCA).

The National Health Research (PNS) project was approved by the National Commission for Ethics in Research for Human Beings, of the National Health Council, under commission's opinion number 328159, on June 26, 2013. The Informed Consent Forms of the research participants were signed on the interviewers' handheld computers. The research project in this article did not require submission to the Ethics and Research Committee since it was subsidized by secondary data in the public domain.
