Patterns and predictive factors of loss of the independence trajectory among community-dwelling older adults

Participants were members of the UPSAV prospective and longitudinal study of 221 conducted from July 2011 to November 2013 among community- living persons, aged over 75 years in Limousin, France. Each participant was followed for 2 years and was assessed by a geriatrician at 0, 6, 12, and 24 months. Our time variable (T0, T1, T2, T2, T3) corresponds to the four visits. The inclusion criteria were: age ≥ 75 years, registration with the social security system, complementary health or 100% coverage by social security, and the intellectual capacity to understand the protocol and submit to the interventions or mild to moderate dementia (Mini Mental Test Statement [MMSE] score ≥ 10). Also, the patient or their legal representative must have provided written informed consent. After inclusion, the study patients have benefited a comprehensive geriatric assessment. An intervention plan was established and coordinated by the UPSAV. Thus, the study participants benefited from a prevention strategy during follow-up time. The data were collected through questionnaires.

The SMAF is an indicator used to predict the transitions of Iso-SMAF profiles over 4 years with 1500 people over the age of 75 followed annually cohort PRISMA [39]) and examined covariates related to transitions of autonomy [40]. The PRISMA is the research group established to address the problem of lack of continuity to care experienced by older adults with chronic conditions in Quebec. Its objective was to evaluate the implementation of an Integrated Service Delivery Network (ISD French acronym) to improve the health, empowerment and satisfaction of frail older people and to change health and social service utilization without increasing caregiver burden [24]. SMAF is a quantitative variable ranging from 0 to 87 points [21‐23]. We chose this variable because it has never been the participant of a study of older adults autonomy trajectories and it is quite complete compared to other tools [10]. Based on epidemiological data and on the observation of the distribution of ISO-SMAF Profiles [10], a SMAF score between 0 and 7 indicates complete autonomy, between 8 and 14 we speak of average autonomy. A SMAF score ≥ 15 was determined to be the best descriptor of moderate to severe loss of autonomy.

The explanatory variables considered as potential predictive factors of independence trajectory included sociodemographic: age, sex, occupation, educational level, place of housing, type of housing, place of residence, marital status, lifestyle, monthly income, financial, human and technical assistance, hobbies, association membership. The health-related variables were: comorbidities [9], daily medications [37], urinary incontinence, anal incontinence, visual disorder, auditory disorder, and communication disorder. Cognitive ability measures consisted the Mini-Mental State Examination (MMSE) [6, 27, 37]. The total MMSE score is 30 points; a score of < 18 is defined as moderate or severe cognitive impairment [6]. We used the Cognitive Evaluation Reflection Group (GRECO) standards to dichotomize the MMSE scores; suspected dementia was defined as an MMSE score of < 24 [27]. Depressive state was evaluated using the 30-points Geriatric Depression Scale (GDS). The GDS scores were classified as: 0–9 no depression, 10–19 mild depression, 20–30 severe depression [5]. Nutritional status included the Mini Nutritional Assessment (MNA) [49], serum albumin level (Guigoz, 1997), body mass index (BMI). An MNA score of ≥24 is defined as an adequate nutritional status; an MNA score of 17–23.5 as risk of malnutrition, and an MNA score of < 17 is regarded as indicative of protein malnutrition [49]. In this study we defined a good and poor nutritional status as an MNA score of ≥24 and < 24, respectively. A serum albumin level of < 30 g /L was defined as a poor nutritional status. The body mass index (BMI) (kg/m²) was calculated by dividing the weight by the square of the height in meters. There are no standards for the interpretation of the BMI of older persons [31]. Nevertheless, obesity is generally defined as a body mass index (BMI) of 30 kg/m2 and higher. Overweight is defined as a BMI between 25 and 30 kg/m2 [42]. We categorized the subjects’ BMI as < 20 (abnormal weight), 21–24 (normal weight), or > 25 (excess weight). Fragility variables included Fried test [12, 13], 12-point Physical Performance Battery (SPPB) scale [18], fall during the previous year and unipodal support test [49]. A score of 0–6 indicates low physical performance, 7–9 average performance, and a score of 10–12 indicates good physical performance [18]. A unipodal support test result of < 5 s was regarded as indicative of an equilibrium disorder. Table 1 provides detailed overview of those variables.

Group-Based Trajectory Model (GBTM) [36] was used to identify latent trajectory groups for SMAF from scores between 0 to 87. GBTM is a particularity of finite mixture modeling. The method consists to cluster individuals into meaningful subgroups that show statistically similar trajectories [34, 35]. A statistical method is used to identify groups of distinctive trajectories which are summarized by a finite set of different polynomial functions of time. In our case, time is equal to visits. The complexity of estimating the parameters of the GBTM model requires maximization by the quasi-Newton procedure. The nature of the dependent variable SMAF (normal distribution) brought us to use the censored normal model [36]. Group’s trajectory, the form of each trajectory, are predicted. The probability for each individual of group membership is estimates. Which allow to assign them to the group for which they have the highest probability. Bayesian information criterion (BIC) criterion was used to select model [36]. We estimated seven models and selected the best model using the BIC. Missing data is a common drawback that appears in many real-world situations as in surveys. In our study, the lack of data was completely random and independent of the variable itself and any other external influences. For example, for the main variable SMAF, it was approximately 24% missing data in T1, 28% in T2 and 22% in T3. We used the multiple imputation method to manage missing data. Missing data were managed utilizing multiple imputation, which identifies missing values by performing repeated simulations [30]. We used PROC MI “multiple imputation procedure” in SAS to manage them. A multinomial logistic regression analysis was performed to analyze the dependence of the explanatory variable and to identify predictive factors. The final model was selected bases on the Bayesian information criterion [36]. The alpha level was set at 0.05.

Two other classification methods were used to identify trajectories. It is about k-means for longitudinal datasets (Kml) [14, 15] and hierarchical ascending classification (HAC). For k-means method, we used the Calinski-Harabasz criterion [7] to identify the optimal number of trajectory groups. Calinski-Harabasz criterion combines the within and between matrices to evaluate clustering quality. We used the “Kml” package in R software (v. 3.4.1; Core Team (2014) R Foundation for Statistical Computing, Vienna, Austria; http://​www.​R-project.​org/​) [15]. We used Ward’s aggregation criteria [51] to identify the optimal number of groups for the hierarchical ascending classification. Ward’s criteria consist to minimize intragroup inertia and maximize intergroup inertia. The method was implemented in R software.

Despite the application of those three methods, in this proposal, GBTM is the principal method because it is simple to implement, useful for describing the heterogeneity of SMAF scores evolution, identifying the risk factors, and potentially for informing clinicians about patients’ subgroups who would need more attention to maintain their functional autonomy. According to Twisk [48], GBTM was shown to be superior for identifying underlying longitudinal trajectories. The k-means and hierarchical ascending classification were performed to compare the optimal number of trajectory groups with the GBTM. Thus, for k-means and hierarchical ascending classification, we presented only the results of the trajectory groups. The results of Baseline characteristics and the logistic regression are based on the GBTM method. These methods are more detailed Bimou and colleagues’ study [2].

Table 1 summarizes the description of the study sample at baseline. Variables including occupation, educational level and monthly income had rare modalities that were grouped together. The participants mean age were 86.1 ± 5.0 years old; About 64% of the participants were > 80 years old. Most study participants were female, resided in an urban area, had hobbies, no cognitive disorders and not depressive symptoms, whereas a relatively small minority had significant loss of weight and low monthly income.

The BIC’s values, Calinski-Harabasz’s and Ward’s criteria are listed in Table 2. GBTM results showed a fairly significant decrease between the first model (k = 2, BIC = − 3229) and the second model (k = 3, BIC = − 2424); 14.5% of participants were classified into the smallest subgroup in first model, compared to 8.9% in second model. Calinski-Harabasz’s criterion decreased from 378 (k = 3) to 317 (k = 4), subsequently increased rapidly from k = 4, and thereafter decreased. Ward’s criterion provided a large jump of inertia between k = 2 and k = 3. Inertia value begins to stabilize when the group number exceeds three. Thus, the best-adapted models included three groups of independence trajectories.

Figures 1, 2 and 3 show the three trajectory groups formed by the three methods. Among the seven models performed, only the two- and three-group models converged for GBTM method (Fig. 1). Therefore, we selected the three-group model for further analysis. Similarly, in the k-means and hierarchical ascending classifications, the model comprising three independence trajectory groups best fit the data.

The GBTM model comprising three groups showed a posterior probability of 0.73 ± 0.14 to 0.98 ± 0.17. The three groups were: Stable functional autonomy trajectory (SFA) (n = 117, average SMAF score between 5.8 and 6.7, 53%, highly independent older adults), Stable then decline functional autonomy decline trajectory (SDFA) (n = 72, 33%, average SMAF score 16.7–21.8, older adults with moderate-to-severe dependence), and Constantly functional autonomy decline trajectory (CFAD) (n = 32, 14%, average SMAF score 33.7–42.5, dependent older adults). The three trajectory groups obtained by k-means and hierarchical ascending classification showed the similar groups those found by the GBTM and described in the same way. Thus, we obtained for k-means: SFA (n = 150, SMAF average 6.4–7.3, 67.9%), SDFA (n = 53, 24%, SMAF average 18.8–25.6), and CFAD (n = 18, 8%, SMAF average 37.2–42.4). Hierarchical ascending classification showed following groups: SFA (n = 136, 62%, average SMAF average 5.1–5.9), SDFA (n = 60, 27%, SMAF average 14.2–19.1), and CFAD (n = 25, 11%, SMAF average 31.4–37.9).

Table 3 provides the adjusted values of OR from multinomial logistic regression. Multinomial logistic regression revealed that specific baseline characteristics predicted membership within each of the three functional autonomy trajectory groups as compared to the Stable functional autonomy trajectory group. The predictive factors of Stable then decline functional autonomy decline trajectory were as follows: farmer (OR = 10.7, 95% CI = 1.09–14.44, p = 0.041), non-membership of an association (OR = 2.67, 95% CI = 1.02–7.00, p = 0.005), and a fall in the previous year (OR = 2.72, 95% CI = 1.28–5.77, p = 0.009). The predictive factors of a Constantly functional autonomy decline trajectory were: worker (OR = 10.33%, CI = 0.74–15.60, p = 0.081), lack of financial assistance (OR = 2.35, 95% CI = 0.09–7.56, p = 0.009), lack of human assistance (OR = 3.30, 95% CI = 0.03–8.26, p = 0.001), lack of hobbies (OR = 22.21, 95% CI = 1.44–34.25, p = 0.026), and cognitive disorder (OR = 2.12, 95% CI = 0.95–10.05, p < 0.0001). The previous occupation and multiple pathologies were predictive factors for both above trajectories.