Sample/Participants
There were 7,452 participants (Mage=59.06 years, SDage=8.94 years) recruited and followed up over 3 waves of data collection spanning 4 years (2011, 2013, to 2015; waves 1-3) from the China Health and Retirement Study (CHARLS) [21]. In this sample, 139 participants (Mage=57.81 years, SDage=9.43 years) self-reported a cancer diagnosis during the 4-year period and were alive at wave 3 data collection.
Data Collection
The original CHARLS was a sister study of the Health and Retirement Study (HRS) in the U.S. with aims to understand Chinese community-dwelling adults’ social, economic, and health status using a nationally representative sample of Chinese adults aged 45 years and older with multistage probability sampling methods. Data were collected via one-to-one interviews by trained interviewers or healthcare professionals to increase the response rate. The overall response rate was 80.51% in the first wave. A detailed description of CHARLS data collection methods has been published elsewhere [21].
Measures
Outcome variable.
Functional limitations were assessed via self-reported difficulty with seven tasks (1=yes, 0=no), including walking 100m, climbing stairs, chair stand, stooping/crouching/kneeling, lifting 11 pounds, extending arms up, and picking up a coin (range 0-7). Scores were summed such that higher scores indicated greater limitations. Cronbach’s alphas for each wave were 0.79, 0.82, and 0.82, respectively.
Cancer diagnosis.
The CHARLS survey asked participants to self-report any cancer diagnosis by a physician at each measurement occasion (wave). We recoded a between-person binary indicator for cancer diagnosis at each wave (1=yes, 0=no). We also coded a time-varying cancer diagnosis timing variable (i.e., at the within-person level, the change from having no cancer=0, to having cancer=1).
Time to/from diagnosis.
We centered each person at time 0 on the measurement wave where cancer diagnoses were first reported. Negative time scores indicate occasions prior to cancer diagnosis (pre-diagnosis), and positive scores indicate post-diagnosis occasions. The time to/from diagnosis variable for participants first reporting diagnosis at wave 1 were therefore coded as 0, 1, 2, whereas the those first reporting diagnosis at wave 2 and 3 were coded as -1, 0, 1, and -2, -1, 0, respectively.
Contributing factors.
Four sets of contributing factors for disability were assessed as, depressive symptoms (psychological factor), pain and falls (physical factor), self-reported memory problems (cognitive factor), and social contact, and availability of support (environmental). All measures are described below. Except for pain (which was only measured in wave 2), all variables were measured over 3 waves, and were time-varying predictors. To align person-level differences in time-varying predictors to person-level differences in functional trajectories, observations were summarized across each phase (pre-diagnosis, onset, and post-diagnosis). For example, repeated scores of subjective memory problems obtained prior to cancer diagnosis were averaged to obtain a person-level pre-diagnosis memory score predicting pre-diagnosis functional change, the memory score obtained at cancer diagnosis onset was used as the predictor of the intercept, and repeated scores obtained after diagnosis were averaged to a person-level score predicting post-diagnosis functional change (for the binary measures on falls, contacts, and participation in social activities, we used the maximum rather than the average). Phase-specific parsing made it possible to accommodate the time-varying nature of moderating factors within the multiphase modeling framework [10].
Depressive symptoms were assessed using 10-items of Center for Epidemiologic Studies Depression scale (CESD-10, range=0-24) [22], with a higher score indicating more depressive symptoms. Cronbach’s alphas of CESD-10 for each wave were 0.81, 0.76, and 0.80, respectively. Pain was assessed using the question “Do you feel any pain? (1=none, 2=a little, 3=some, 4=quite a bit, and 5=a lot)” in wave 2. Falls was assessed using the question “Have you fallen down in the last two years? (1=yes, 0=no).” Self-reported memory problems were assessed using the question “How would you rate your memory at the present time (1=excellent to 5=poor)?”, coded with a higher score indicating poorer self-rated memory. Social contact was measured by any weekly contact with children, including in-person meet, email, and phone or text (1=yes, 0=no). Availability of support was measured by number of people living in the same household (range=1-16), and participation in any social groups or activities (1=yes, 0=no).
Demographic covariates.
Demographic variables were assessed at baseline and grand mean-centered for participant age (in years), sex (0=female; 1=male), education (0=none, 1=less than lower secondary, 2=upper secondary and vocational training, 3=tertiary). Marital status was coded as a time-varying continuous variable to accommodate changes in status over time (1=married, 3=partnered, 4=separated, 5=divorced, 7=widowed, and 8=never married).
Data analysis
Preliminary analysis.
Descriptive statistics at baseline were assessed for all participants, and separately for those with and without cancer, with comparisons made using t-test or chi-square statistics. We further examined functional limitations over time by fitting an empty model with linear time as the only predictor.
Research question 1.
To examine functional trajectories for those participants with and without cancer, we used an ordinary growth curve model with functional limitations as the outcome (Model 1). In the level-1 within-person model, we specified functional limitations as:
\({\text{Functional limitations}}_{ti}\)=\({\pi }_{0i}\)+\({\pi }_{1i}\)(Timeti)+\({\epsilon }_{ti}\)
|
(1)
|
where functional limitations for person i at time t was a function of an intercept \({(\pi }_{0i},\) baseline functional limitations), linear time (\({\pi }_{1i},\)within-person association between time and functional limitations), and the within-person residual, \({\epsilon }_{ti}\), whose variance was \({\sigma }_{\epsilon }^{2}\) and assumed to be homogeneous across persons. In the level-2 model, the individual specific intercepts and slopes were specified as:
\({\pi }_{0i}={\beta }_{00}+{\beta }_{01}\)(Diagnosisi)+\({\upsilon }_{0i}\)
\({\pi }_{1i}={\beta }_{10}\)+\({\beta }_{11}\)(Diagnosisi)
|
(2)
|
where βs were sample-level parameters. The person-specific intercept, \({\pi }_{0i}\), and the person-specific slope, \({\pi }_{1i}\), from equation (1) were each modeled as a function of time-invariant and between-person cancer diagnosisi (1=participants had cancer and 0=participants without cancer), while controlling for demographic variables (not shown in equation 2). \({\upsilon }_{0i}\) were between-person differences in the intercept with a variance, \({\sigma }_{\upsilon }^{2}\).
Research question 2.
Among participants with cancer, we applied multiphase growth curve models (Model 2) to examine whether levels and slopes of functional limitations post-diagnosis differed from pre-diagnoses. Model 2 had predictors of time to/from diagnosisit, time-varying cancer diagnosisit (1=had cancer and 0=no cancer), and the interaction between time to/from diagnosisit × cancer diagnosisit, while controlling for demographic variables. The level-1 within-person multiphase growth model was specified as:
\({\text{Functional limitations}}_{ti}\)=\({\pi }_{0i}\)+ \({\pi }_{1i}\)(Time to/from diagnosisti)+\({\pi }_{2i}\)(Diagnosisti) +\({\pi }_{3i}\)(Diagnosisti ×Time to/from diagnosisti)+\({\epsilon }_{ti}\)
|
(3)
|
where functional limitations for person i at time t was a function of an intercept \({(\pi }_{0i},\) functional limitations at the first report of cancer diagnosis), an individual specific slope parameter (\({\pi }_{1i},\)linear rate of cancer diagnosis-related change in functional limitations before cancer diagnosis, when diagnosisti=0), an individual specific parameter (\({\pi }_{2i},\) discrete differences in level of functional limitations between pre- and post-cancer diagnosis), a second individual specific slope parameter (\({\pi }_{3i},\) differences in the linear rate of cancer diagnosis-related change in functional limitations between the pre and post-diagnosis phases), and the within-person residual, \({\epsilon }_{ti}\), whose variance was \({\sigma }_{\epsilon }^{2}\) and assumed to be homogeneous across persons. In the level-2 model, the individual specific intercepts and slopes were specified as:
\({\pi }_{0i}={\beta }_{00}+{\beta }_{01}\)(Agei)+\({\beta }_{02}\)(Malei)+\({\beta }_{03}\)(Educationi)+\({\beta }_{04}\)(Marital statusi)+\({\upsilon }_{0i}\)
\({\pi }_{1i}={\beta }_{10}\)
\({\pi }_{2i}={\beta }_{20}\)
\({\pi }_{3i}={\beta }_{30}\)
|
(4)
|
where βs were sample-level parameters, representing the mean intercept (\({\beta }_{00}\)) and mean slopes (\({\beta }_{10}, {\beta }_{20}, {\beta }_{30}\)) of the functional limitation trajectory pooling over all participants with cancer in the sample. The person-specific intercept, \({\pi }_{0i}\), from equation (3) was further modeled as a function of participant age (\({\beta }_{01}\)), being male (\({\beta }_{02}\)), education level (\({\beta }_{03}\)), and marital status (\({\beta }_{04}\)). \({\upsilon }_{0i}\) were unexplained between-person differences in the intercept with a variance, \({\sigma }_{\upsilon }^{2}\), representing the degree of individual variability around the mean intercept.
Research questions 3.
We applied the full multiphase growth curve model (Model 3) by adding covariates to equation (4) of Model 2. More specifically, we explored whether contributing factors had effects on levels (as main effects, Model 3.1) and slopes by fitting additional interaction terms between the disability contributing factors × time to/from diagnosisit×cancer diagnosisit (Model 3.2). We trimmed non-significant variables, one at a time, to achieve model parsimony. All analyses were performed using SAS (version 9.4), and statistical significance was considered at p<.05 level (2-tailed).