Trajectories of health-related quality of life among people with a physical disability and/or chronic disease during and after rehabilitation: a longitudinal cohort study

The current study is part of the multicentre longitudinal cohort study Rehabilitation, Sports and Active lifestyle (ReSpAct) that was initiated to evaluate the nationwide programme Rehabilitation, Sports and Exercise (RSE; Dutch: ‘Revalidatie, Sport en Bewegen’) [23, 24]. The RSE programme has been implemented in eighteen rehabilitation institutions in the Netherlands (twelve rehabilitation centres and six rehabilitation departments of hospitals). The RSE programme aims to stimulate an active lifestyle during the rehabilitation period and to guide people with a physical disability and/or chronic disease in maintaining a physically active lifestyle in the home setting after discharge from rehabilitation [23, 24]. Participants of the RSE programme were referred to a sports counselling counter 3 to 6 weeks before discharge from rehabilitation for a face-to-face consultation with a sports counsellor, followed by four telephone-based counselling sessions up to 13 weeks after discharge from rehabilitation [23, 24]. All sessions were based on motivational interviewing [25] (see Online Resource 2 for a schematic overview of the RSE programme and the ReSpAct study).

Participants were included in the ReSpAct study from May 2013 to August 2015. Participants were monitored with questionnaires at given regular measurement times: at baseline (T0: 3–6 weeks before discharge) and 14 (T1), 33 (T2) and 52 (T3) weeks after discharge from rehabilitation (Online Resource 2). The study was approved by the ethics committee of the Center for Human Movement Sciences of the University Medical Center Groningen (reference: ECB/2013.02.28_1). All participants voluntarily participated after signing an informed consent.

Inclusion criteria were: (1) being at least 18 years of age, (2) having a chronic disease or physical disability (e.g. stroke, heart failure, Parkinson’s disease, spinal cord injury), (3) receiving inpatient or outpatient rehabilitation care or treatment at one of the participating rehabilitation departments or institutions, (4) participating in the RSE programme [24] and (5) filling in the RAND-12 Health Status Inventory (RAND-12) at two or more measurement occasions. Participants were excluded if they were not able to complete the questionnaires, even with help, or were participating in another physical activity stimulation programme.

HR-QoL was assessed by using the self-reported RAND-12 questionnaire [26], an adapted, abbreviated version of the RAND-36 Health Status Inventory (RAND-36) [27]. The RAND-12 contains at least one item from each of the eight subscales of the RAND-36, so that it adequately represents the wide range of relevant aspects of health status [28]. Six items of the RAND-12 contribute to the physical health composite (how health limits a person in activities, or how a person’s physical health causes problems with work or other activities) and six other items contribute to the mental health composite (how a person feels and how a person’s mental health causes problems with work or other activities) [27, 28]. All twelve items contribute to the general health composite, which represents all relevant aspects of health status [28]. We used an age-corrected general health composite score for this study [27]. A higher score on the RAND-12 indicated better HR-QoL. Because the RAND-12 only contains twelve items of the RAND-36 (range 0–100), scores on the RAND-12 range from 0 to 65. We found good reliability (internal consistency) of the RAND-12 based on the study sample at T0 (Cronbach’s α = 0.85, N = 974), at T1 (Cronbach’s α = 0.87, N = 957), at T2 (Cronbach’s α  = 0.88, N = 861) and at T3 (Cronbach’s α = 0.88, N = 780). Previous literature supports acceptable construct validity and test–retest reliability of the RAND-12 in among others clinical populations [28, 29].

All independent variables were measured at baseline (T0: 3–6 weeks before discharge). Person-related factors included gender, age, body mass index (BMI) and level of education, which was dichotomized into low (up to completed secondary education) and high (completed applied University or higher) to make it internationally comparable.

Disease-related factors included the type of disease divided into eight categories: musculoskeletal disease, amputation, brain disorder (e.g. stroke or other non-congenital brain defects), spinal cord injury, other neurologic disease, organ disease, chronic pain and other diseases. Also, disease-related factors included the number of comorbidities dichotomized into no comorbidities and one or more comorbidities, because this variable included all diseases and disabilities reported by a participant. The level of acceptance of the disability or disease was assessed on a four-point Likert scale (1–4, no acceptance to complete acceptance), with a higher score indicating better acceptance of the disability or disease. The level of acceptance was dichotomized into no (no or little acceptance) and yes (acceptance to a large extent or completely), because when entering the level of acceptance as categorical variable in the logistic regression, we found that the odds ratios (ORs) did not linearly increased/decreased. Perceived fatigue was assessed with the 9-item Fatigue Severity Scale (FSS) [30], which is a valid and reliable questionnaire to determine the impact of perceived fatigue in clinical populations (in people with systematic lupus erythematosus r_validity = 0.81 and r_reliability = 0.89, and in people with multiple sclerosis r_validity = 0.47 and r_reliability = 0.81) [30‐32]. The FSS score ranges from 1 to 7, with a higher score indicating more perceived fatigue [30]. We found good reliability (internal consistency) of the FSS based on the study sample at T0 (Cronbach’s α = 0.91, N = 1044). The FSS includes items like “Exercise brings on my fatigue.” and “I am easily fatigued” [30]. The level of perceived pain was assessed on a six-point Likert scale (1–6, from no pain to severe pain), with a higher score indicating more perceived pain. The level of pain was dichotomized into no (no to light pain: score 1–3) and yes (moderate to severe pain: score: 4–6), because when entering perceived pain as categorical variable in the logistic regression, we found that the ORs did not linearly increased/decreased. Also, too few people reported severe pain (perceived pain = 6).

Lifestyle-related factors included the dichotomous variables smoking and alcohol use (“Do you smoke currently?” and “Do you consume alcohol currently?”: yes or no). In addition, the total minutes of physical activity per week was assessed by using the Adapted Short Questionnaire to Assess Health-enhancing physical activity (Adapted-SQUASH), a 19-item self-reported recall questionnaire. In a previous study, the Adapted-SQUASH has been shown to be a sufficiently reliable (intraclass correlation coefficient = 0.76, p < 0.001) and valid—compared to the Actiheart activity monitor—(intraclass correlation coefficient = 0.22, p = 0.027) questionnaire to determine self-reported physical activity in a similar sample (people with a physical disability and/or chronic disease) [33]. The Adapted-SQUASH is pre-structured in four main domains outlining types and settings of activity: ‘commuting traffic’, ‘activities at work and school’, ‘household activities’ and ‘leisure time activities’ including ‘sports activities’ [34]. The SQUASH [34] was adapted to make the questionnaire more applicable for this population (Adapted-SQUASH), as described in the study protocol of the ReSpAct study [24]. First, the items ‘wheeling in a wheelchair’ and ‘handcycling’ were added in the domains ‘commuting activities and leisure time’ and ‘sports activities’. Second, the self-reported intensity of the activity was categorised in ‘light’, ‘moderate’ and ‘vigorous’, instead of ‘slow’, ‘moderate’ and ‘fast’. Third, a large range of adapted sports (e.g. wheelchair basketball/rugby/tennis) were included for the item ‘sports activities’. Lastly, in the examples of different sports ‘tennis’ was replaced by ‘(wheelchair) tennis’. Information on sports participation (yes/no) was obtained from the Adapted-SQUASH. If the participant reported to perform at least one sports activity per week, than they were coded as ‘yes’, if not as ‘no’.

Analyses were conducted in a two-step approach. First, trajectories of HR-QoL during and after rehabilitation among participants with two or more valid measurements over time were identified using Latent Class Growth Mixture (LCGM) modelling with quadratic (assuming non-linear change over time), linear (assuming linear change over time) and latent class analyses (lca) models [35], using the Mplus software program 7.11. The choice for linear and quadratic models was made based on previous research [14], showing trajectories of HR-QoL to be both linear as well as quadratic (non-linear). Additionally, latent class analyses were conducted for descriptive purposes. These analyses gave us insight in the (heterogeneity of) patterns of change in HR-QoL without a priori assuming a trajectory shape. LCGM models are regression-based models that assume that individuals in the sample do not necessarily come from one underlying population but might come from multiple underlying (or latent) subpopulations. LCGM modelling aims to find the optimal number and characteristics of these subpopulations. Common, stepwise modelling strategies were applied [35], using the Guidelines for Reporting on Latent Trajectory Studies (GRoLTS) as well [36]. A one-class model was first determined, thus assuming one underlying population, and subsequently more classes were added one at a time and model fit indices were inspected. The optimal number of classes was determined according to the following model fit criteria: (1) a lower Bayesian Information Criterion (BIC), where a difference of 10 points lower is usually regarded as sufficient improvement [37], (2) a higher entropy (range from 0 to 1), a standardised measure of how accurately individuals’ trajectories are classified, where higher values indicate better classification [38, 39] and (3) average posterior probabilities of ≥ 0.80 [35]. The choice for the optimal number of classes was additionally made considering clinical interpretation (rejecting solutions that do not make clinical sense) and class size. Finally, individuals were classified into their most likely class based on their posterior probability.

Second, multiple binomial multivariable logistic regression analyses were performed to assess associations between the previously described person-, disease- and lifestyle-related factors and trajectory membership using version 24 of the Statistical Package for the Social Science (SPSS). The outcome of the LCGM modelling, the nominal variable of trajectory membership, was used as dependent variable.

Independent variables at baseline were all entered block wise (block 1: person-related factors, block 2: disease-related factors and block 3: lifestyle-related factors) in multivariable models. Descriptive statistics of these variables were analysed at baseline. Assumptions of normality and linearity were checked. The continuous independent variables age, BMI, fatigue, and physical activity/week were standardised. Results of the multiple binomial multivariable logistic regression analyses are presented as odds ratio (OR) and corresponding 95% confidence interval (CI). Because three comparisons between two trajectories were needed to compare all HR-QoL trajectories, a Bonferroni-corrected p-value, to correct for multiple testing, of 0.017 (0.05/3 = 0.017) was used to give a 95% probability of correctly concluding not to reject the null hypothesis [40].

To facilitate transparency and reproducibility, additional information is available on: (a) the dataset of the HR-QoL (Online Resource 1) and (b) the Mplus syntax of the LCGM modelling and the SPSS syntax of the multiple binomial multivariable logistic regression analyses (Online Resource 2).

In total 1100 participants were included in this study. Participants had an average age of 51.0 ± 13.5 years and 52.0% were female. The three most common disease groups were brain disorder (26.0%, N = 286), musculoskeletal disease (18.1%, N = 199) and chronic pain (15.6%, N = 172) (Table 1).

Based on descriptive characteristics at baseline (Table 1), participants excluded for the LCGM modelling analyses were on average more often female, younger, lower educated, lived less independently, had worse acceptance of their disease, perceived more fatigue, smoked less, received less counselling moments and had lower levels of HR-QoL. Descriptive characteristics at baseline were missing of around 250 excluded participants, which might give skewed descriptive characteristics.

The results of the fit indices for quadratic, linear and lca models with one to six trajectories of HR-QoL are presented in Table 2. Comparing these models with the model fit criteria alone proved to be complicated, as the model fit criteria were not always in agreement, which is a common finding in LCGM modelling [41]. After careful consideration, we chose the three-class quadratic model as the optimal model in this sample, although the average posterior probabilities were slightly below 0.80, indicating possibly less distinct trajectories and subsequent fuzzy classification, yet it avoids inclusion of an extremely small class, as is the case in the four-class and five-class quadratic models. The three-trajectory model consisted of two large and stable, but distinctly different trajectories: moderate (N = 635, 55.1%) and high (N = 429; 40.9%) trajectory. In addition, one smaller intermediate trajectory is provided, which increases between 3 and 6 weeks before discharge from rehabilitation and 33 weeks post rehabilitation and then stabilises (i.e. recovery) (N = 36; 4.0%) (Fig. 1).

Descriptive statistics of the mental, physical and general health composites for the three trajectories at each measurement time are presented in Table 3. Overall, mental health followed the same but higher course and physical health followed the same but lower course compared to general health. Supplementary figures are given in Online Resource 2, including estimated mean trajectories for each model, estimated means with individual trajectories for each latent class and the estimated with observed means for the final model. Although the plots with estimated means with individual trajectories for each latent class show large heterogeneity in individual trajectories of HR-QoL, all individual trajectories follow the same growth pattern over time for each latent class.

Descriptive statistics of possible determinants before discharge from rehabilitation for the HR-QoL trajectories are presented in Table 4. Multiple binomial multivariable logistic regression analyses were performed to determine associations among the personal-, disease- and lifestyle-related factors before discharge from rehabilitation and the HR-QoL trajectories (Table 5).

Compared with participants in the moderate HR-QoL trajectory (N = 635), participants with a higher BMI (OR 0.77, 95% CI 0.64–0.94), participants who perceive fatigue (OR 0.47, 95% CI 0.39–0.58) and/or participants who perceive pain (OR 0.22, 95% CI 0.15–0.33) are less likely to belong to the latent class with a high HR-QoL trajectory (N = 429), while participants who accept their physical disability and/or chronic disease (OR 3.25, 95% CI 2.25–4.68) are more likely to belong to the latent class with a high HR-QoL trajectory. Also compared to the moderate HR-QoL trajectory, based on the limits of the 95% CI which both lie above or below one (but not significant), participants who are older (OR 1.27, 95% CI 1.04–1.55), participants who drink alcohol (OR 1.44, 95% CI 1.01–2.05) and/or participants who are more physically active (OR 1.21, 95% CI 1.01–1.44) are more likely to belong to the latent class with a high HR-QoL trajectory, while participants who smoke (OR 0.58, 95% CI 0.35–0.94) are less likely to belong to this latent class.

There were no significant determinants before discharge to distinguish between the moderate HR-QoL (N = 635) and the recovery HR-QoL (N = 36) trajectories. But, based on the limits of the 95% CI which both lie above one (but not significant), participants who drink alcohol (OR 3.05, 95% CI 1.09–8.53) are more likely to belong to the latent class with a moderate HR-QoL trajectory, compared to the recovery HR-QoL trajectory.

A comparison of the recovery HR-QoL trajectory (N = 36) and the high HR-QoL trajectory (N = 429) showed that participants who are older (OR 1.97, 95% CI 1.18–3.29), participants who accept their physical disability and/or chronic disease (OR 5.09, 95% CI 2.04–12.69) and/or participants who drink alcohol (OR 4.60, 95% CI 1.53–13.83) are more likely to belong to the latent class with a high HR-QoL trajectory (N = 429).

Remarkably, gender, education level, type of disease, having comorbidities, level of physical activity and sports participation before discharge were not significant determinants to distinguish between trajectories of HR-QoL.

In addition, we checked whether the found significant determinants in the multiple binomial multivariable logistic regression analyses were still found after controlling for general HR-QoL scores at baseline (Table 5). HR-QoL scores at baseline were found to be significant determinants in the comparisons between the moderate and high HR-QoL trajectories (OR 5.86, 95% CI 4.14–8.30) and between the recovery and high HR-QoL trajectories (OR 45.24, 95% CI 10.26–199.47). When controlling for HR-QoL score at baseline, only perceived fatigue (OR 0.69, 95% CI 0.55–0.87) and perceived pain (OR 0.56, 95% CI 0.35–0.88) remain significant determinants when comparing the moderate and high HR-QoL trajectories (Table 5).

More than one third of our sample obtained a relatively stable high HR-QoL, but more than half obtained moderate HR-QoL after participating in a person-centred physical activity promotion programme; the RSE programme. We found several modifiable disease-related factors to be important in determining HR-QoL, which emphasises the importance for optimising person-centred advice in focusing on fatigue and pain management and on better acceptance of the disability during rehabilitation. Also, the identified HR-QoL trajectories are not disease specific, which might imply a disease-overarching mechanism.

Furthermore, to make the LCGM modelling more transparent, the data, syntax and results are available in electronic supplementary material. Especially in latent trajectory studies, open communication is important due to the data-driven aspect of the analyses and the difficult choices made to find the optimal model fit. We would like to encourage other researchers in the field of latent trajectory studies, to provide open communication of their analyses and results, and to use the GRoLTS checklist [36] in reporting the analysis of the latent trajectory study. This will benefit comparison of the results in different study populations.