Page 14 - IJPS-3-1

Exposure to urban life and mortality risk among older adults in China

self-reported activities: (a) visiting neigh bors, (b) shopping, (c) cooking, (d) washing
clothes, (e) walking one kilometer, (f) lifting a 5-kg bag, (g) crouching and standing up
three times, and (h) taking public transportation; dichotomous coding was similar to
that used for ADL disability. Cognitive function was measured using the Mini-mental
State Examination (MMSE) that includes six domains of cognition—orientation,
reaction, calculation, short-term memory, naming, and language—with a total score
of 30. The MMSE items were adopted from the Folstein MMSE scale (Zhang, Gu
and Hayward, 2008). Respondents were categorized as cognitively impaired if their
MMSE score was below 24 (Zhang, Gu and Hayward, 2008). Given the low level
of educational attainment among most older adults in China, we assessed alternative
criteria (e.g., score of 18) for those with no education to test the sensitivity of different
cut-points for defining cognitive impairment (available upon request from the authors);
we obtained similar results to those presented here. To account for possible difference
in mortality risk over time, we controlled for year of the CLHLS survey.

2.3 Analytical Strategy
We modeled the association between exposure to urban life and mortality under
different measurement schemes of urban life exposure: residential status at birth
and at older ages (the preliminary measurements; Table 3), change or stability in
residential status between these two life stages (four-type classification; Table 4),
change or stability in residential status plus mid-life exposure (i.e., PLO) (eight-type
classification; the upper panel of Table 5), and finally, we included migration (fourteen-
type classification; the lower panel of Table 5).
We used Weibull hazard regression models to examine the association between
urban exposure and mortality, with two sequential models. Model I (the partial model)
controlled for age (single year) and sex whenever appropriate; Model II (the full
model) additionally controlled for socioeconomic status, family/social support, health
behaviors, and health condition. We also designed other models that added only one
set of all covariates in Model II into Model I, but the results were similar to Models I
and II. To save space, we thus opted to present the simplest model and the full model.
Multicollinearity among variables was tested and found to be not a problem, with
all variance inflation factors less than 3 (Chatterjee and Hadi, 2012). We performed
analyses separately by sex and age group to investigate possible differences between
men and women and between the young-old and the oldest-old (ages 65–79, ages 80+).
However, we did not do so for models that included PLO because of the small sample
size of some categories. In all models by age group, the single year of age was still
controlled for.
In the analytical sample, all variables had a missing value of less than 2%. We used
multiple imputation for these missing values, assuming that the respondents who had
missing values would have the same value for a given variable as those who had no
missing values if the former had the same conditions on factors with non-missing
values.
For survival status and the length of exposure to death, we applied multiple
imputation to impute missing survival/mortality status, and it produced results close
to those we present here, in which we did not impute survival/mortality status. The
reasons that we did not use imputed results were because the survival status—and
the length of exposure to death—are dependent variables in the survival analysis
and because its proportion of missing is high (nearly 30%). Those who had at least
two interviews and were lost to follow-up afterwards in the subsequent waves were
included in the analysis; however, only information before lost to follow-up was
included. Those who were only had one interview were excluded from the analysis.
In all analytical models, we did not apply the sampling weight because the CLHLS
weights were constructed from population distributions of age, sex, and urban/rural
residence—variables that were controlled for in the models (Winship and Radbill,
1994). Furthermore, no longitudinal weight was attached in the released CLHLS
datasets. All analyses were performed using Stata version 13.1.

8 International Journal of Population Studies 2017, Volume 3, Issue 1