A hierarchy of correlates for objectively measured physical activity, sedentary time, and physical fitness in older adults: A CHAID analysis

ABSTRACT The aging process reflects, in many cases, not only a decline in physical activity (PA) and physical fitness (PF), but also an increase in overall levels of sedentary time (ST). In order to hierarchically identify the most powerful correlates related to low and high levels of objectively assessed PA, ST, and PF during the late adulthood, a total of 2666 older adults were cross-sectionally evaluated. Multidimensional correlates were obtained through interview. Using chi-squared automatic detection analysis to identify the cluster of correlates with most impact on PA (<21.4 min/day), ST (≥8 h/day), and PF (<33.3th percentile), was found that the most likely subgroup to be physically inactive consisted of widowers not owning a computer and sport facilities in the neighbourhood (94.7%), while not being widowed, reporting to have a family that exercises and a computer at home (54.3%) represented the subgroup less likely to be inactive. Widowers without sidewalks in the neighbourhood were the most sedentary group (91.0%), while being a married woman and reporting to have space to exercise at home (40%) formed the most favourable group of correlates regarding ST. Men reporting a financial income <500€ and physical problems frequently formed the group with the lowest PF level (70.3%). In contrast, the less likely subgroup to have low levels of PF level consisted of having a financial income ≥1000€ and a computer at home (3.4%). Future interventions should target widowers with limited accessibility to computer and urban/sport-related infrastructures, as well as impaired older adults with low financial income. Highlights Chi-squared automatic interaction detection was used to identify and hierarchise correlates of objectively measured physical activity, sedentary time, and fitness. Widowers not having a computer at home and sport facilities in the neighbourhood were the most likely to be physically inactive, while not being widowed, having a family that exercises and a computer at home represented the subgroup less likely to be physically inactive. The most likely to be classified as sedentary were widowers without sidewalks in the neighbourhood, while the most favourable group of correlates regarding ST was formed by married women and reporting to have space to exercise at home. Individuals with a low financial income and physical problems comprised the population subgroup with the lowest PF levels, while having a medium-high financial income and a computer at home represented the less likely subgroup to have low levels of PF.


Introduction
Physical activity (PA), sedentary time (ST), and physical fitness (PF) are known to be important determinants of healthy aging and physical functioning among older adults (Caspersen et al., 1985;Patterson et al., 2018). Previous investigations have shown that, in many cases, the aging process reflects not only a decline in PA and PF levels, but also an increase overall ST (Gomez-Bruton et al., 2020;Patterson et al., 2018;Santos et al., 2018). Since PA, PF, and ST are complex behaviours that are shaped by the interrelationships between individual, social, and environmental factors (Mooney et al., 2017;Peralta et al., 2018), identifying different correlates that influence these behaviours is a necessary step to develop effective healthcare policies.
When looking at the literature on correlates, most investigations have assessed PA and ST through self-report (Cleland et al., 2019;Peralta et al., 2018), an assessment method with known limitations in terms of response bias and misperceptions (Troiano et al., 2020). Thus, there is a need for further investigations on correlates that include objective assessments of both PA and ST. Moreover, little is known regarding correlates of PF, which are important to consider since it compromises a set of health-related components (Caspersen et al., 1985), and is considered a major determinant of healthy aging and physical independence during late adulthood (Cress et al., 1999).
To date, most investigations on sociodemographic and environmental correlates among older adults have used ecological methodologies to identify single correlates as possible targets for interventions addressing specific risk factors (Barnett et al., 2017;Cleland et al., 2019;Peralta et al., 2018). By contrast, few studies have followed data mining approaches based on classification methods (i.e. decision trees) to better understand the inter-relationships between PA and ST correlates at different levels. This information would be useful for public health workers and government bodies for the development of interventions targeting behaviorrelated risk-profiles based on multilevel clusters of correlates that are simple to measure (Lakerveld et al., 2017;Yoon et al., 2015). Therefore, the main purpose of this investigation was to comprehensively describe the hierarchy of correlates from distinct domains that best differentiate between low and high levels of objectively measured PA, ST, and PF in older adults. Since similar behavioural and health-related correlates, such as regular sports practice and presence of disease, have been identified as important determinants of PA and ST (Peralta et al., 2018), we hypothesise that these correlates may best differentiate between low and high levels of these health-related behaviours.

Sample participants
Data were collected between March 2017 and November from a representative sample of non-institutionalized Portuguese older adults (≥65 years-old) selected by means of a proportionate stratified random sampling considering the number of citizens by age, sex and region from the mainland of Portugal. A total of 2666 older adults were assessed, of which 1461 had complete data on all correlates. Of those, 1109 participants had complete data on correlates plus PF and 634 participants had complete data on correlates plus PA/ST and were included in the present investigation. Written informed consents were obtained from participants. The present study was approved by the Ethics Committee of the Faculty of Human Kinetics, University of Lisbon (number: 25/2020).

Body composition
Body weight and height were measured using an electronic scale (model 799 SECA, Hamburg, Germany) and a stadiometer (model 220 SECA, Hamburg, Germany) according to standardised procedures (Lohman et al., 1988). Body mass index (BMI) was calculated by dividing body weight by height squared (kg/m 2 ).

Physical activity and sedentary time assessment
Physical activity and ST were objectively assessed using a tri-axial accelerometer (ActiGraph GT3x model, Fort Walton Beach, Florida, USA). Participants were asked to wear the accelerometer on the right side of the hip for 7 consecutive days and to only remove the monitor during sleep and water-based activities. The accelerometers were initialised on the first assessment day (i.e. 100 Hz frequency) and data were recorded in 15-s epochs and reintegrated into 60-s epochs. Periods of at least 60 consecutive minutes of zero counts (with 2-min of spike tolerance) were considered as non-wear time.
Only participants with at least three days (with at least one weekend-day) with more than 600 min/day of monitor wear were included in the analysis. Activity intensity thresholds were defined as: < 100 counts/min (ST), 100-2019 counts/min (light PA), and ≥2020 counts/min (moderate-to-vigorous PA, MVPA) (Troiano et al., 2008).
Participants were classified as inactive (<21.4 min/day of MVPA) or active (≥21.4 min/day of MVPA) based on World Health Organization PA guidelines , and dichotomised as sedentary (≥8 h/day of ST) or non-sedentary (<8 h/day of ST), according to the Canadian 24-Hour Movement Guidelines (Ross et al., 2020).

Physical fitness assessment
Physical fitness was measured using the Senior Fitness Test Battery (Rikli & Jones, 1999), which includes a set of six validated functional fitness tests to determine lower and upper body strength and flexibility, agility/ dynamic balance and aerobic endurance. The result of each test was converted into a Z-score, and the mean of the six Z-scores was used to compute a continuous measure of PF. Overall PF was categorised into tertiles, with the lowest (≤33.3th percentile) corresponding to low PF, and the two highest (>33.3th percentile) corresponding to medium-high PF (Santos et al., 2012).

Behavioral correlates
Personal (current sports practice, present and past sports federation) and the behavioural characteristics of relatives (family and friends sports practice) were questioned and dichotomised as no (0) or yes (1) (Commission, 2013).

Environmental and household correlates
Urbanity level was classified as urban area (1), semiurban area (2) and rural area (3). Neighbourhood characteristics (i.e. safety, access to sports facilities, existence of sidewalks, public transportation, cycling paths and attractive places to go, lack of green spaces, far distance to the nearest stores, and presence of people exercising) were answered and classified as agree (1) or disagree (2) (Saelens et al., 2003). The number of TVs at home was categorised as not having a TV (0), having 1 TV (1), having 2 TVs (2), and having ≥3 TVs (3) (Commission, 2013). Having a TV in the bedroom, at least one computer at home, and space to exercise at home, as well as owning a pet with street walking routines were questioned, and answers were dichotomised as no (0) or yes (1) (Commission, 2013).

Statistical analysis
Descriptive characteristics of the sample were examined using means and standard deviations (SD) (i.e. continuous variables), as well as frequencies and proportions (i.e. categorical and ordinal data). In addition, since a substantial number of participants were excluded due to missing data in at least one of the independent variables, secondary analyzes using the non-parametric Mann-Whitney U-test and Chi-square test of independence were performed.
A Chi-squared Automatic Interaction Detection (CHAID) analysis was used to identify and hierarchise correlates with most impact on distinct levels of PA, ST, and PF. CHAID is a decision tree analysis presented as a flow chart-like structure that identifies subgroups of independent variables according to their relevance in the model (Kass, 1980). Since this algorithm uses a nonparametric procedure (Kass, 1980), complex combinations between categorical, ordinal, and continuous predictors may be formed to express the likelihood of predicting different outcomes (Kass, 1980). Given that the analysis of one dependent variable did not consider the remaining dependent variables as independent variables, participants without any information on dependent variables or with missing data in at least one of the 36 correlates were excluded.
Regarding data analysis, the decision tree depth was limited to a maximum of 3 layers to form subgroups with a maximum of 3 correlates. According to the pruning criteria for our sample size (Bellazzi & Zupan, 2008), the minimum number of participants in the parent and child nodes was set at 100 and 50, respectively. Due to the hierarchical structure of the tree, it was possible to compare the response rate of the whole sample and the response index (RI) of the terminal nodes. Considering that the RI represents the percentage of subgroup response relative to the percentage of whole sample response, this indicator enables understanding the characteristics of the association between the dependent and independent variables (Lakerveld et al., 2017).
Statistical significance was set at 5%. All statistical analyses were performed using SPSS software (Version 25.0, IBM Corp., Armonk, NY, USA).

Participants
The analytical sample for the PA and ST outcomes compromised of 634 older adults (433 women; 76.6 ± 7.6 years), while the sample used for the PF analysis compromised of 1109 older adults (710 women; 77.7 ± 8.0 years). Most of the participants were retired, living in an urban area, and had at least one chronic disease. Similarly, in both analytical groups, approximately half of the participants had an education level less than or equal to the 4th grade, while approximately 44% of participants indicated having a financial income <500€ (Table 1).
Participants with missing data were older (i.e. group with missing PA/ST, 79.7 ± 7.9 years; group with missing PF, 80.5 ± 7.4 years), mostly widowed, and had lower education and financial income levels compared to the participants with complete data (p < 0.05). No differences were found between included and excluded participants in terms of BMI and presence of chronic diseases (p ≥ 0.05) (see Supplementary Table 1).

Hierarchy of correlates for high vs. low MVPA
As presented in Figure 1, marital status was the most relevant correlate to discriminate between PA levels, with 29.4% (n = 101; RI = 0.72) of all participants in a married/de facto union, divorced and single condition being physically active. The cluster of older adults with higher odds (45.7%; n = 42; RI = 0.57) of being physically active were those who reported to not be widowed, to have close family members that exercise regularly (2 nd level), and to have at least one computer at home (3 rd level). In contrast, only 11.3% (n = 33; RI = 1.86) of older adults who reported to be widowed were physically active. This percentage was even lower (5.3%; n = 8; RI = 3.88) when considering the subgroup of individuals less active, which consisted of widowers reporting not having a computer at home (2nd level), nor having sports facilities in their neighbourhood (3 rd level).

Hierarchy of correlates for high versus low ST
Similar to the correlates found for PA, marital status was the most important factor to distinguish between levels of ST (Figure 1). Although being married/de facto union was shown to increase the likelihood (38.6%; n = 90; RI = 0.70) of complying with ST recommendations, when considering the least sedentary cluster of older adults (i.e. being married/de facto union, having space to exercise at home and being a female) this percentage substantially increased (60%; n = 30; RI = 0.45). On the other hand, while 17.5% (n = 51; RI = 1.26) of all widowed older adults demonstrated having less than 8 h/day of ST, when considering the highest sedentary subgroup (i.e. older adults who reported to be widowed and not have sidewalks in the neighbourhood), the number non-sedentary individuals dropped to less than 10% (n = 11; RI = 3.00).

Hierarchy of correlates for high versus low PF
The factor that most contributed to differentiate between low and medium-high levels of PF was the financial income (Figure 1). A negative relationship between financial income and PF was found, with older adults reporting a financial income <500€, 500-999€, and ≥1000€, having a corresponding percentage of medium-high PF levels of 53% (n = 257; RI = 1.26), 69.8% (n = 289; RI = 0.96) and 92.9% (n = 195; RI = 0.72), respectively. When considering sample subgroups, the highest level of PF was found in older adults with a financial income ≥1000€ and who claimed to have at least one computer at home. In contrast, older adults with the lowest financial income, reporting to have physical problems frequently (2nd level), and being a man (3rd level), was the subgroup with the lowest indices of PF (29.7%; n = 22; RI = 2.25).

Discussion
In this investigation, we sought to hierarchically classify the multidimensional correlates that were most related with different levels of objectively measured PA, ST, and PF, while considering a large sample of Portuguese older adults. Even though some of the most important factors seemed to repeat amongst different domains, our findings support that the hierarchical sequence of correlates should be expected to change depending on the specific domain under investigation.
In general, we found that the most important factor to distinguish between attaining or not MVPA recommendations was marital status, with widowers fitting into the less active group. In line with our findings, it has been shown that non-married older adults were less likely to be physically active compared with married/ cohabiting older adults (Koyanagi et al., 2017), which can be explained by the expected higher rates of loneliness in the non-married group (Schrempft et al., 2019). Conflicting results may reflect the different groups within the non-married classification (i.e. divorced, widowed, and single), where distinct levels of PA may be expected. Although we identified widowed as a priority group to promote PA during late adulthood, few other studies have integrated this variable as a sociodemographic correlate for PA (Koyanagi et al., 2017).
In addition to not being widowed, an older adult's family PA habits were identified as the second most important factor to positively influence PA levels. Even though previous investigations suggested that rather than family, the social support of friends has a higher influence on PA levels during late adulthood (Resnick et al., 2002), current evidence shows that the influence of social support is expected to change depending on several sociodemographic factors, including marital status (Cobb et al., 2015). Married older adults, for example, generally have an increased likelihood to comply with PA recommendations if their spouse is also physically active (Cobb et al., 2015). Therefore, gathering multilevel information on family and social support is imperative when developing interventions targeting PA. Having at least one computer at home was the third most powerful correlate related with high levels of compliance to PA recommendations, while not having a computer was found to be the second most important correlate for an individual to be classified as physically inactive. Most of the previous investigations have only focused on the relationship between computer use and cognitive capacity (Wagner et al., 2010), However, since the ability to use a computer relies on a wide range of factors, such as marital status, level of education, and financial income (Wagner et al., 2010), owning a computer should be assumed as a protective factor regarding PA and targeted in future interventions to enhance or maintain PA.
The absence of sport facilities in the neighbourhood was found to be the third most impactful correlate for an older adult to not fulfil the PA recommendations, which is in line with evidence showing that recreational or sport centres are the places where older adults most engage in structured PA (Hoekman et al., 2017). This confirms recent findings reporting that built environment characteristics, such as the presence of recreational and sport-related facilities, are directly associated with the promotion of PA (Bonaccorsi et al., 2020;Júdice et al., 2021). In this regard, there is a further need to provide user-friendly facilities that encourage and provide the opportunity for older adults to be physically active (Hoekman et al., 2017).
As with the findings related to PA, we found that marital status was the most important factor for differentiating between high and low levels of ST, with married older adults being more likely to spend less time in sedentary pursuits. Since there is strong body of evidence suggesting an independent time and healthrelated relationship between ST and PA levels (Patterson et al., 2018;Thorp et al., 2011), it is expected that similar correlates, as well as their hierarchical arrangement will stand out for these two behaviours. In this perspective, we found that being widowed emerged as the keyfactor that most related with low PA and high ST during late adulthood, whereas not being widowed and being married were found to be associated with favourable levels of PA and ST, respectively. As previously stated, compared with married older adults, widowers are particularly exposed to reduced social networks and isolated environments, which accentuates the deterioration of physical and mental health (Li et al., 2016). In line with this, a large body of evidence supports the idea that feelings of loneliness and depression negatively impact ST and levels of PA (Li et al., 2016).
Notwithstanding, in the present investigation, widowers reporting to not have sidewalks in their neighbourhood were identified as the risk group more likely to be classified as sedentary. There is a robust body of literature highlighting the importance of the presence of sidewalks on behavioural outcomes such as PA and ST (Bonaccorsi et al., 2020;Mooney et al., 2017), with the low presence of sidewalks being found to be unfavourably associated with overall levels of PA and positively associated with ST (Cleland et al., 2019;van den Berg et al., 2016), which support our findings. Indeed, it has been shown that the presence of sidewalks not only enriched environmental walkability (Bonaccorsi et al., 2020), but it also enhanced social connectedness, which directly contributed to less social isolation (van den Berg et al., 2016). Thus, when designing interventions to positively influence ST, public health authorities should not only target individuals according to their sociodemographic characteristics, but also consider the development of an appropriate built environment that offers various sport-related and walkability opportunities.
The second and third most powerful correlates to positively influence levels of ST were having a space to exercise at home and being a woman, respectively. Previous studies have reported a reduction in ST among older adults who live in houses with more available space (i.e. detached house) (Cerin et al., 2020). This may be explained by the fact that the ownership of houses with more available space provide more opportunities to be active (i.e. housework and gardening) and, hence, reduce ST. Furthermore, our finding that women were particularly more likely to be less sedentary than men is in line with data from studies using similar population-based samples (Santos et al., 2018;van Ballegooijen et al., 2019). Nevertheless, given that many of the secondary factors related with ST observed in our investigation did not stand out in previous reports within the scope of ST correlates, more research is needed to further understand the impact of these sociodemographic and household characteristics on objectively measured ST.
Regarding PF, the most relevant correlate to distinguish between low and medium-high levels was related to financial income, with older adults reporting higher financial incomes (≥1000€) showing higher levels of PF. Since socioeconomic status is positively associated with education levels, older adults with higher financial income are expected to have an increased literacy regarding health and sport-related outcomes (Beenackers et al., 2012). In addition, individuals with higher socioeconomic status are more likely to have access to a wider range of sports opportunities, in particular leisure-time PA (Beenackers et al., 2012). This finding demonstrates the negative impact that socioeconomic inequities may have on different domains of PA, and consequently on different levels of PF. Furthermore, we found that owning a computer was the second most impactful factor related with higher levels of PF. Interestingly, owning a computer during late adulthood has been described in the literature as an important indicator of a high socioeconomic position due to its positive association with a large number of the variables previously mentioned (Wagner et al., 2010). Despite that the effect of owning a computer on PF has been poorly studied in older adults, our findings highlight the protective nature that owning a computer may have on both PF and PA.
The second most relevant factor to contribute to low PF was related with the presence of physical problems, with older adults reporting to be physically impaired or having physical problems frequently being at the lowest level of PF. The natural deterioration of physical health during late adulthood is expected to decrease the levels of functional fitness (Gomez-Bruton et al., 2020), which increases the risk for losing physical independence (Sardinha et al., 2016). Considering the interrelation between PF, PA and ST, this situation may be particularly critical in more sedentary older adults (Santos et al., 2018), especially men, who, were identified as being at risk for low levels of PF in our investigation.
Even though this investigation has made a significant contribution to the understanding of PA, ST, and PF correlates during the late adulthood, it is not without limitations. First, due to the cross-sectional nature of our investigation, it was not possible to establish causality between correlates and main outcomes. Secondly, even though we used objective measures of ST and PA, it is important to note that seasonal effects, as well as the selected algorithms and cut points (i.e. non-wear time and MVPA) to process raw accelerometer data may influence overall PA outcomes (Migueles et al., 2017). Since there are no clear guidelines discriminating the best cut-points to analyze accelerometer data in older adults, it is important to take the cut-points chosen into consideration when interpreting and comparing study results. Third, even though our study included 36 correlates from four different domains, there may be missing important discriminators, such as physical limitations, disease severity, and social isolation (Dugan et al., 2018;Schrempft et al., 2019). Fourth, our investigation was mostly based on Caucasian older adults, thus limiting the extension of our results to other ethnic groups. In addition, there was a high number of participants without information on at least one of our variables of interest. Since participants from the group under analysis slightly differed from the group of participants with missing data and considering that the data-driven models are specific for the correlates included in the analysis, this investigation should be framed as an exploratory investigation. Although the CHAID analysis allowed us to identify the most powerful correlates of each dependent variable, it was not possible to adjust our decision trees for the influence of objectively measured PA, ST and PF on each other. Lastly, considering that decision trees composed of deep layers are prone to overfitting (Bellazzi & Zupan, 2008), a higher minimum number of participants per parent and child nodes was set to avoid this modelling error.
In summary, the present investigation highlighted how multilevel correlates hierarchically impacted distinct levels of PA, ST, and PF using a large sample of Portuguese older adults. Public health agencies and organisations targeting to enhance PA and PF levels and counteract ST during older adulthood should prioritise widowed individuals with low-income levels. In addition, developing strategies toward the optimisation of computer use and ownership, improvement in access to sport facilities and enhancement of street walkability may help to attain better healthrelated behaviours in older adults. Nevertheless, future research is needed to understand how the hierarchical relation between multilevel correlates and distinct levels of PA, ST and PF may change over time and according to the season, and to identify additional health-related and social correlates that best differentiate between low and high levels of PA, ST, and PF.

Disclosure statement
No potential conflict of interest was reported by the author(s).