The Mobility of Income-related Inequality in Food Preferences Among Chinese Adults: a Population-based Longitudinal Study

This study aimed to investigate the income-related inequality of food preference and its mobility 53 in the long run, and to quantify the determinants’ contributions of socioeconomic inequality 54 mobility in food preference in China. The data were sourced from the China Health and Nutrition Survey (CHNS) conducted in 2006, 57 2009, 2011 and 2015, respectively. A study sample of 3940 adults were included for analysis. Five 58 preselected questions were used to construct a summary food preference index. Cross-sectional 59 and longitudinal concentration indices were used to measure the income-related inequality in food 60 preference at a certain time and for a period, respectively. Health-related income mobility index 61 was used to measure the gap between cross-sectional and longitudinal income-related health 62 inequality concentration indices. Decomposition analysis was used to decompose income mobility 63 index into its determinants. run. Our study suggests that people with decreased income could be taken as targeted population 79 in future public intervention measures, and narrowing the regional gap between Eastern and 80 Western China could be strengthened in future Chinese policies. 81

Secondly, to find determinants of food preference; and lastly, to quantify each determinants' 131 contribution to the mobility of food preference inequalities. The study complements existing 132 literature by providing the first longitudinal study on inequality in food preference. Our study will 133 provide evidence-based findings for policy makers and public health workers to develop public 134 health policies and strategies to help people in making a healthy food preference and eliminating 135 inequality between the poor and the rich. 136 137

Study design 139
A retrospective cohort study was performed to track the change of food preference score and its 140 inequality over time as well as to analyze the contribution of each determinants on this mobility. 141

Data source 142
The data were sourced from the China Health and Nutrition Survey (CHNS). CHNS is a 143 multidisciplinary and national panel dataset on nutrition, health status, socioeconomic status, 144 community information and family relations of Chinese. The first wave of CHNS was conducted in 145 1989 and the following waves of CHNS were conducted every 2-4 years. A multistage, random 146 cluster sample design was used to draw the surveyed samples in 15 provinces and municipal cities 147 that substantially vary in geography, economic development, public resources, and health indicators. 148 CHNS was initially designed to estimate the impact of the health, nutrition, and family planning 149 policies and programs implemented by national and local governments. It was also designed to 150 investigate how the social and economic transformation of Chinese society was affecting the health 151 and nutritional status of the Chinese. Overall, the gross survey sample included 30,000 individuals 152 in 7200 households. Detailed information on procedure of data collection and quality assurance 153 measures could be found in CHNS official website (https://www.cpc.unc.edu/projects/china). This participants who were followed up in each survey were included. Since our study focused on adult 156 samples defined as those aged ≥18 years old (n=3978), respondents whose key information (food 157 preference/income) was missing were further excluded in this analysis, leaving a study sample of 158 3940 adult respondents. 159 Variables 160

Food preference 161
Five questions on preferences for five kind of food were preselected in the survey: fast foods, salty 162 snack foods, fruit, vegetables, and soft drinks/sugary fruit drinks. A five-point Likert scale (dislike 163 very much, dislike, neutral, like, like very much) was employed to estimate the degree of liking for 164 each question. A score for each category ranging from 1 to 5 was generated to measure the 165 preference degree, with a higher score indicating a healthier preference. A summary of food 166 preference index was created in our study to measure the overall degree of healthy preferences with 167 the sum of scores on five questions. The maximum score for food preference index is 25. 168

Other variables 169
Socioeconomic characteristics considered in this study were comprised of age, gender, household 170 size, education (illiterate, elementary, middle school, high school, and university), marital status 171 (unmarried, married, and others), working status and income. Income was measured by net per 172 capita household expenditure. Net per capita household expenditure was calculated by subtracting 173 health expenditure from total household expenditure and dividing by household size. of this indicator includes the 12 aspects of urbanization, which are population density, economic 179 activity, traditional markets, modern markets, transportation, health infrastructure, sanitation, 180 communication, social services, diversity, education, and housing[9]. Each aspect was given a score 181 from 0 to 10, and then weighted equally in the overall index and added together for an overall 182 maximum possible score of 120. A higher urbanization score indicates greater urbanization. 183 Compared to urban/rural dichotomy measure, urbanization index encompasses the underlying 184 complexity and heterogeneity of the community that would otherwise be missed in binary measure. 185 Detailed construction procedure could be found on official CHNS website 186 (https://www.cpc.unc.edu/projects/china). Lastly, since there was a potential association between 187 dietary knowledge and food preference, dietary knowledge was also considered as a covariate in 188 our study [10,11]. Dietary knowledge score was a summary variable constructed from the 12 189 questions surveyed in CHNS. Respondents were asked to choose "strongly disagree", "disagree", 190 "neutral", "agree" and "strongly agree" for each question. For each positive question, we assigned 191 a score ranging from 0 to 2, with 2 for strongly agree, 1 for agree, and 0 for other answers, and vice 192 versa. The maximum dietary knowledge score is 24 with a higher score indicates better knowledge. where 00 is the grand mean of y, and 0 is the difference between the ith average y and grand 214 The random intercept model with one independent variable can be written as follows: 216 Concentration index is nowadays the most widely used analytic tool to measure the relative Where =̄̄̄， is the relative rank of individual i in the distribution of average incomes 237 after T periods, is the average health of individual i after T periods. 238 Health-related income mobility could be calculated from following formula: 239 Where is the mobility index of y after T periods. 241 The mobility index could also be decomposed into the contributions of different determinates with 242 the use of following expression: 243 Where ̂i s the coefficient for kth variable, ̄ is mean of kth variable at time t, is 245 concentration index of kth variable at time t. ̄ is the mean of y at time t.
is the 246 is the mobility index of kth independent variable after T 247 period.
is the residual error. 248 All statistical analyses were carried out on SAS 9.4. A p value of 0.05 or less is considered 249 statistically significant in the study. 250 251

252
Summary statistics of respondents' characteristics at baseline are presented in Table 1  However, residents with university degree were evidently higher than others, and the illiterate had 267 the lowest scores. Although there was little difference between married and unmarred, residents 268 with other marital status (including widowed, separated, and divorced) had absolutely the lowest 269 scores among all groups. 270 - Figure 1 The mean score of food preference by characteristics groups over years-271 Table 2 shows the factors associated with food preference identified from multi-level random  Table 2. Adjusted association between food preference and determinants-278 Table 3 shows the concentration and mobility indices for food preference by year. The 279 concentration indices of food preference score on income in each year were presented in CI t 280 column. The indices were all positive, indicating that there were pro-rich inequalities in food 281 preference with the rich had the higher food preference score in each period. The longitudinal 282 concentration indices were presented in CI T column. In a long run, the degree of pro-rich 283 inequality in food preference were larger than that at the baseline.

-Table 3 Concentration and mobility indices for food preference by year-295
The results of decomposition analysis of food preference mobility index are showed in Table 4. 296 The first column presents the longitudinal concentration index of independent variables. In a long 297 run prospective, higher income, educational attainment, urbanization index, and married were 298 concentrated in the rich, which means there was a pro-rich inequality in these characteristics. 299

Whereas the longitudinal concentration indices of age, other marital status, living in Central and 300
Western China were negative, indicating that these characteristics were more concentrated in the 301 poor. The second column shows the mobility index of independent variables, which measures how 302 much the longitudinal perspective alters the picture that would emerge from cross-sectional view. 303 A negative mobility index of independent variables indicates that the weighted sum of the cross-304 sectional concentration indices of independent variables underestimates the degree of long-run 305 inequality, and vice versa. In a long run, income and marital status were overestimated, however, 306 education, age, central and western china were underestimated using the cross-section perspective. 307 The third column shows the elasticity of food preference score with respective to each determinant 308 variable. Last column shows the contribution of independent variables on mobility indices for 309 food preference score. Married, and living in Western China made healthy food preference more 310 concentrated among the rich in the long run, whereas increasing age and dietary knowledge 311 contributed to making good food preference behavior less concentrated among the rich in the long 312 run. 313 - Table 4 Decomposition of the mobility index of food preference by determinants-314 315 Discussion 316 Among a set of health maintenance behavior, healthy eating is one of the best things people can do 317 to maintain their weight and prevent health problems such as heart disease, high blood pressure, 318 type 2 diabetes, and some types of cancer. Healthy eating starts with healthy food choices. In 319 today's fast-paced world, however, more and more individuals are choosing to remove certain 320 foods from their daily consumption habits based on personal preferences. Our study observed a 321 significant change in the increase of healthy food preference score over the 10-year period. There 322 are likely many reasons to explain this phenomenon. In macroscopic level, many initiatives were 323 conducted by Chinese government in this period. Firstly, Dietary Guidelines, which was published 324 to guide Chinese people to make healthy food and beverage choices, was updated by the Chinese 325 for unhealthy food on TV, internet, newspapers, and magazines, such as fast food, programs on 333 healthy food and nutrition also plays an important role in providing residents with numerous 334 dietary knowledge by mass media. In microscopic level, people have become increasingly more 335 health-conscious than they used to be. They have developed the idea that good health is above 336 wealth, and hence try to seize any chance to seek nutrition knowledge for improving their 337 Our study identified some factors to be the determinants of healthy food preference for Chinese 353 adults. Most determinants were similarly reported in previous studies[7, 23-25]. Firstly, our study 354 found that food preferences were strongly influenced by their sociodemographic characteristics. 355 As expected, increasing age, higher educational attainment, and being married were positively 356 associated with healthy food preference [23]. Our study confirmed the evidence from cross-357 sectional studies that there was a statistical difference in food preference between male and female 358 [26,27]. Since more than 25% of men were unwilling to increase their knowledge about food and 359 nutrition, men were less likely to form a healthy preference compared to women [28]. 360 Longitudinal analysis is more powerful than cross-sectional analysis in detecting gender 361 association with food preference. 362 Secondly, positive association between economic characteristics and food preference was also 363 confirmed in our study. A study on adolescents in China found that income was negatively 364 associated with healthy food preference with the controlling of covariates [29]. This difference 365 may be resulted by two reasons. First reason is that the study populations in the two studies were 366 different. The adolescents are more likely to have irrational food beliefs than adults. They prefer 367 taste to health reasons in making food choices. The other reason is that only cross-sectional data 368 was used in the former study, therefore more strong evidence could be obtained from more 369 powerful study design. Our study found that there was pro-rich inequality in food preference score across different income 390 groups. People with a higher income were more likely to have a higher food preference score. This 391 is especially true when we adopt a long-term perspective view. Weighted cross-sectional 392 concentration index underestimated this pro-rich inequality. The reason was that respondents with 393 income downwardly mobile tended to have below average levels of food preference score compared 394 to upwardly mobile respondents. The downward income mobile had a larger influence on food 395 preference than upward mobile. Future public intervention measures could take people with The idea that health is determined by factors outside the traditional health-care setting has become 398 an increasing recognized approach in improving public health and addressing health disparities. 399 Previous cross-sectional studies found out that income and education contributed a large proportion 400 to health sector variables, such as chronic disease incidence and health service use [35][36][37][38][39][40]. Our 401 study found income did not contribute as pro-rich inequality in food preference in a longitudinal 402 view as much as cross-section data obtained. Our results showed that in the long run living in 403 Western China was more concentrated in the poor, and it was more obvious when using the 404 longitudinal view than using cross-sectional approach. Since this characteristic had a negative 405 association with healthy preference, living in Western China contributed to increasing pro-rich 406 inequality in food preference in the long run. It is notable that urbanization index also contributed 407 to pro-rich inequality by 3% in the long run. Despite China launched western development strategy 408 in 2000, the regional gap between different areas is still very large. Chinese government could 409 strengthen the effort to reduce the regional gap between Eastern and Western China and the gap 410 between urban and rural areas. Our results showed that the unequal distribution of dietary score was 411 more concentrated among the rich in the long run, and the elasticity of food preference with respect 412 to this characteristic was positive, thus this characteristic contributes to making healthy food 413 preference more concentrated among the advantaged in the long run. 414 It is worth highlighting some limitations of our study. Firstly, all the data were self-reported, thus 415 recall bias may have affected the validity of our findings. Secondly, our study focuses on a sample 416 aged 18 years and over. This selected sample limits the ability to generalize the results to the 417 minors. Thirdly, potential determinants of the food preference considered in the study were 418 selected from the survey questions. Other unobserved factors that may affect dependent variables 419 were not included in multilevel regression model. 420

422
The results of this study showed that the mean of food preference score increased from 2006 to 423 2015. Some factors, such as age, income, dietary knowledge score, and urbanization index were 424 positively associated with food preference score, whereas living in central and western China were 425 the negative predictors of healthy food preference. There was pro-rich inequality in food 426 preference score with the rich had higher preference scores in each period. Longitudinal 427 concentration index indicated that the degree of pro-rich inequality in food preference was larger 428 than that at the baseline in the long run. The cross-sectional measure of food preference score was 429 underestimated in the long run. Further decomposition analysis on mobility index showed that 430 living in Western China and urbanization index contributed to the increase of pro-rich inequality 431 in food preference in the longitudinal perspective. Our study suggests that people with decreased 432 income could be taken as targeted population in future public intervention measures; and 433 narrowing the regional gap between eastern and western China and between urban and rural areas 434 could be strengthened in future Chinese policies.