Page 39 - IJPS-1-1

Danan Gu

ployed survival analysis techniques with time between any two adjacent surveys (in terms
of days) as the exposure length, and survival status (surviving vs. dead) at the end of the
period of the two adjacent surveys as an event. The dates of the interviews were recorded
for each wave. For those who died between two adjacent waves, the dates of death were
collected from officially issued death certificates whenever available; the next-of-kin and
the local residential committee were consulted when a death certificate was not available.
The quality of mortality data in the CLHLS is high (Gu and Dupre, 2008).
2.2.5 Controls
To obtain robust results, we controlled several sets of covariates that are shown in the lite-
rature to be associated with mortality (Wen and Gu, 2011); the covariates included demo-
graphics, socioeconomic status, and health practice. Demographics included age, residence
(urban vs. rural), and ethnicity (Han vs. non-Han). Sex was not considered as a covariate
since all analyses were stratified by sex. Socioeconomic status included years of schooling
(0, 1–6, and 7+), primary lifetime occupation (professional/administration (white collar) vs.
others), and economic independence measured by whether the respondent has retirement
wage/pension or own earnings (yes vs. no). Family support was measured by co-residence
with children (yes vs. no) and marital status (currently married vs. unmarried). Health
practice was measured by currently smoking (yes vs. no), currently consuming alcohol
(yes vs. no), and regularly exercising (yes vs. no). The sample characteristics are presented
in Table 1.
2.2.6 Analytical Strategy
The Weibull parametric survival function was applied because some of the variables vi-
olated the proportionality assumption required by the Cox proportional hazard model. As
noted early, respondents who were lost to follow-up were excluded because we did not
know their survival information. Other alternative approaches such as imputing the sur-
vival information, treating them as right-censored, or treating them as a specific category
yielded similar results. In the original design, several sequential models were conceived to
examine how the association between concordance/discordance of OSA and SSA was
changed in the presence of different covariates that included demographics, socioeconom-
ic status and family support, and health practice. However, because the results of the se-
quential models are very close to each other, we only presented the results from the final
model that includes all covariates (the results from the sequential models are available
upon request). The proportion of missing values for all variables in the analysis was less
than 2%. To reduce possible bias due to missing values in the analysis and inferences, we
employed multiple imputation techniques for all variables. We did not apply sampling
weights to the regression models because the CLHLS weight variable was unable to reflect
the national population distributions with respect to variables other than age, sex, and ur-
ban/rural residence (Wen and Gu, 2011)
We also performed additional tests to examine improved predictive power of concor-
dance and discordance of OSA and SSA for mortality risk (controlling for all covariates in
the analysis) by performing two alternative models that treated OSA and SSA as two in-
dependent variables. All analyses were performed using STATA 13.0.

3. Results
Table 1 presents the sample distribution for study variables and covariates by sex. One
interesting result in Table 1 is that it is more common to be in objectively poor health, but
subjectively good health (Type II) (18.7% for women and 21.4% for men), rather than the
other way around (Type III) (3.9% for women and 7.9% for men); it is much more com-
mon to be in both objectively and subjectively poor health (75.0% for women and 64.4%

International Journal of Population Studies | 2015, Volume 1, Issue 1 33