Two population-based cross-sectional studies were conducted from June to August in 2001 and 2005 respectively in rural Western China. 41 counties of 8 provinces in 2001 and 40 counties of 9 provinces in 2005 were all determined by the Chinese Ministry of Health (MOH) and the United Nations Children’s Fund (UNICEF) rather than sampled randomly. Of them, there were the same 8 provinces and 29 counties in both 2001 and 2005. After a multi-stage probability proportion to size sampling (PPS) method was adopted, 5 townships in the chosen counties and then four villages from each township were selected randomly. As an example, how to select five townships from a chosen county is presented as follows. Firstly, we sort all townships in a county from smallest to largest according to the population size, and then calculate the accumulated population size. Suppose that there is a total population of 202000 in a township, and the sample interval can be calculated as 202000/5 = 40400. After that, a random number is determined based on the banknote serial number of Renminbi (Chinese currency). The random number consists of the last digits of this serial number, which has the same length as that of the sampling interval. For example, we have selected the serial number 98272809 from Renminbi banknote, and then the serial number 72809 is determined. Next, the first selection number is 32049 by taking the absolute value of difference between the random number and sample interval. If the first selection number is contained in the accumulated population size of this township, the first township is included in this study. The second selection number is the first selection number plus the sampling interval, and so the second township is selected. The third, fourth and the fifth townships are chosen with the same method. Similarly, four villages are selected from each chosen township. The sampling interval for village selection is calculated by dividing the total population size of a township by four. Finally, eight mothers with children under three years old were selected randomly from each chosen village. The main inclusion criteria were that mothers (>15 years old) were non-pregnant with a strong willingness of participate and without other diseases during the survey.
After this study was fully explained to participants, the informed consent was obtained from them. Socioeconomic and demographic characteristics were collected from the questionnaires. The protocol was also reviewed and approved by the Human Research Ethics Committee of the Xi’an Jiaotong University College of Medicine.
Capillary blood was collected from each participant using a finger-prick method to extract three drops of blood from the left middle and/or ring finger. After the first two drops of blood were wiped away, the third drop of blood was used immediately for testing hemoglobin concentration (HBC) by Hb photometer (B-Hemoglobin, precision of 0.1 grams/decilitre, Hemocue AB, Sweden). The anemia of nonpregnant women in this study was defined as HBC less than 12.0 g/dl. Besides, three levels of severity of anemia were also distinguished: mild anemia (10.0-11.9 g/dl), moderate anemia (7.0-9.9 g/dl), and severe anemia (less than 7.0 g/dl) . Many studies had shown that HBC increased rapidly with altitude especially at altitude of higher than 1000 m sea level [18–20]. Previous study had also found that Centers for Disease Control (CDC) method might be more suitable to adjust altitude in Western China compared with other methods . Thus, in this study CDC method  was solely utilized with the following expression:
Where ΔHb was the increment of HBC by increased altitude above sea level, and Alt was the altitude (m).
A pilot study was carried out to pretest all questionnaires and procedures and then the detailed interviewer guides were developed. The interviewers were trained to standardize questionnaires administration and HBC measurements at least one week before commencement of the survey. After signing the informed consent form, all participants were interviewed face-to-face by means of the pre-coded structured questionnaires to collect the information on socio-demographic characteristics of mothers. Meanwhile, blood sample collections of mothers were carried out to measure their HBC.
Nineteen investigation teams from Xi’an Jiaotong University College of Medicine were established for these counties. Each team consisted of four or five members and one supervisor. During the survey, all fieldworkers were closely monitored by their supervisors and randomly examined. Participants were re-interviewed immediately when errors and/or missing values were detected.
Mothers’ anemia was considered as a unique outcome variable in the study. We considered the certain individual characteristics of mothers as the proximate covariates of anemia. Individual-level covariates included ethnicity, which consisted of Han, Tibet, Uighur, Hui, Zhuang and Others; parity (1, 2, or >2); mothers’ age; and breastfeeding duration. The county-level covariate was altitude of county (<500 m, 500 ~ 1500 m, or >1500 m). Due to lack of income data of each household, wealth index was constructed through the principal component analysis for assessing economic status of the household . The principal component analysis synthesized information on a set of household assets and living conditions: the ownership of a car, television, bicycle and motorcycle; the availability of clean water; the resources of household income and so on. Based on the tertiles of the first principal component, the socioeconomic status of the households was classified into 3 levels indicating the poorest, middle and wealthiest households.
Two-level logistic regression
Considering the hierarchical structure of the study sample, a two-level logistic model was applied to assess the influences of the covariates on the outcome variable. The observed responses y
are proportions with the standard assumption that they are binomially distributed y
), where n
is the denominator for the proportion. The variance-components model correct for the problem of correlated observations in a cluster, by introducing a random effect at each cluster. Consider the 2-level logistic variance model where the expected proportion is modeled using a logit link function.
Where, was scale parameter. The model could also be expressed as follows,
In the study, level 1 represented the individual and level 2 was the county. P
was the proportion of mothers anemia from the jth county in the ith individual. β
m were the regression coefficients corresponding to the effects of fixed covariates, X
were the observed characteristics of the mothers and the county. The random effects were presented as random variance with standard error (SE), and fixed effects as an odds ratio (OR) with 95% confidence interval (CI). The Variance Partition Coefficient (VPC, %) was calculated as the proportion of variation that is attributable to the county-level sources of variation.
The data was entered into Epi Info Version 6.0 (CDC, Atlanta, GA, USA) by double entry. All analysis was performed using STATA Version 12.0 (Stata Corporation, College Station, TX, USA). Independent t-test was used to compare the means of two different populations, and the comparison between two independent proportions was made by means of χ
2 test. In addition, effect coding was used in the two-level logistic analysis to estimate the differences among 10 provinces with the grand mean of the 10 provinces as the reference category.