The considerable rise in the prevalence of overweight (including obesity) in India, where over a billion people reside [1,2,3,4], presents a serious public health concern given the association of overweight with increased non-communicable disease (NCD) risk [5].
In the early stages of economic development and urbanisation, overweight and obesity prevalence tends to be higher among individuals of a higher socioeconomic position (SEP), arguably due to an increased financial capability to meet and exceed nutritional requirement [6,7,8,9]. As societies develop economically, the prevalence of overweight increases among the poor and rural population [6,7,8,9,10,11,12,13,14].
Since India’s economic liberalisation in the early 1990 [15], economic growth has not been uniformly distributed across the country. In addition to considerable heterogeneity in culture, customs and diet, the current levels of economic development between India’s states varies substantially. For example, the Gross Domestic Product of Delhi is eight times greater than that of the state of Bihar [16]. Consequently, the prevalence of overweight, and the extent of the increase in its prevalence in recent decades, varies considerably sub-nationally [1,2,3]. For instance, in Bihar, the prevalence of overweight among women increased from 3.7 to 11.7% (an absolute increase of 8%) between 1998 and 2016, whereas in Delhi, the prevalence increased from 12 to 33.5% over the same period (an absolute increase of 21.5%) [1]. However, little is known about variation in the sub-national socioeconomic patterning of overweight.
In this paper, we aimed to understand how recent trends in the association between overweight and SEP differ between India’s most and least economically developed states between 1998 and 2016, a period in which India’s Gross Domestic Product per capita quadrupled from US$432 to US$1750 [17]. The main rationale for this study was to unmask subnational heterogeneity in trends in the association of overweight and SEP in India not observed when analysing national trends. Demonstrating this would imply that national-level trends may not be generalisable at a subnational level [18]. A study of this nature is of importance as health policy is dictated at the state level; therefore, estimating the prevalence by state development and urban and rural areas may highlight different immediate health policy priorities between less and more developed states.
We conducted secondary analysis, using repeated cross-sections from state-representative data from 1998 to 2016 to estimate the prevalence of overweight in India by SEP in the five most and least economically developed states in India. In more economically developed societies, there is usually higher prevalence of overweight among poorer individuals where, for instance, there is a higher exposure to relatively cheaper fatty foods [6, 9, 19]. This is more likely to be the case in urban areas, where risk factors for overweight are usually much greater. We therefore hypothesise that in India’s most developed states, we will observe a considerable increase in the prevalence of overweight among lower SEP individuals and relatively smaller increases among higher SEP individuals. On the other hand, in India’s least developed states, we expected to find larger increases among higher SEP individuals, compared to lower SEP individuals. This is supported by the fact that poorer individuals in societies with lower levels of economic development are more likely to be unable to afford to meet nutritional requirements, whereas the relatively rich may be more exposed to overweight due to a greater access to excess food [6, 9].
Data
We used the National Family Health Survey (NFHS) Surveys 2 (1998–99), 3 (2005–06) and 4 (2015–16). All three surveys collected health and demographic data on women aged 15–49, whereas surveys 3 and 4 collected data on men aged 15–54. The sampling method was designed to include a nationally-representative sample of individuals within a nationally-representative sample of households. Additionally, in India, the NFHS surveys are also representative at the level of the state.
The NFHS surveys select rural and urban samples separately. Specifically, in rural areas in all three waves analysed, rural samples were selected using two-stage sampling, whereby the first stage involved selecting primary sampling units (PSUs), or villages, with a probability proportional to size (PPS), and the second stage involved selecting random households from each village. In urban areas, NFHS 2 and 3 used a slightly different sampling procedure to the one in NFHS 4. In NFHS 2 and 3, three-stage sampling was adopted whereby in the first stage wards were selected with a PPS, in the second random census enumeration blocks (CEB) were chosen in each ward and, in the third, random households were chosen from each CEB [2, 3] On the other hand, NFHS 4 adopted a two-stage approach in urban areas, whereby CEBs served as the PSU, selected using a PPS, and households from each PSU randomly selected. Were a PSU to contain fewer than 40 households, the PSU was joined to the nearest PSU. The 2011 census helped determine the sampling frame in NFHS 4 [1].
In all three surveys Interviews used a uniform questionnaire and were conducted by survey teams. A woman’s eligibility for the survey was determined by whether they were between ages 15–49 and, for the NFHS 3 and 4, whether they spent the previous night in the selected households. Men aged 15–54 in the households were eligible for the Men’s survey in NFHS 3. Of the selected households in NFHS 4, a random sample of households were selected to determine eligibility for the men’s survey [1].
In India there are currently 36 States/Union Territories. We restricted our analysis to states that have been in existence since the collection of the NFHS 2 survey. States created between the surveys were not considered in the analysis. We selected five states to indicate the most and least developed states as the study aimed to demonstrate a divergence in the trends in their socioeconomic patterning. Our primary objective was to highlight variation in trends in the socioeconomic patterning of overweight within India. We therefore chose not to include all the states in India as this would lead to the inclusion of states that are closer to the average level of per capita net state domestic product for India. As a result, we would risk placing states at similar levels of economic development in the Most and Least developed states categories, consequently underestimating the extent of the variation in trends.
Our classification of states was based on the per capita net state domestic product (PCNSDP) in 2014–15 using the base year 2011–12. The most economically developed states were Goa, Maharashtra, Sikkim, Haryana and Kerala with a PCNSDP ranging from ₹112,444 to ₹241,081, compared to an all India average of ₹72,805. The least economically developed states included Bihar, Assam, Uttar Pradesh, Manipur, and Madhya Pradesh with NSDPPC ranging from ₹23,223 to ₹44,809 [20]. We limited our sample to non-pregnant women, whose inclusion could bias the associations we sought to identify. This left a total of 96,365 women and 18,729 men in the most developed states category, and 289,200 women and 54,669 men, respectively, in the least developed states category.
As NFHS-2 only sampled ever-married women, we restricted our samples in 2005–06 and 2015–16 to this population to allow the comparability of the study population across surveys. Additionally, respondents with missing height and weight data were also omitted from the sample, leaving 76,050 women (12,168 in 1998–99; 14,000 in 2005–06; 49,882 in 2015–16) and 18,729 men (8518 in 2005–06 and 10,211 in 2015–16) as the study population in the most economically developed states, and 213,195 women (22,266 in 1998–99; 20,459 in 2005–06; and 170,470 in 2015–16) and 54,669 men (19,377 in 1998–99; and 35,292 in 2015–16) in the least economically developed states. As multi-stage sampling approaches were adopted in the collection of the NFHS, we included the sampling weights included in the data set to account for unequal selection probabilities.
Outcome
We used the Body Mass Index (BMI) variable included in the surveys (measured as the respondent’s weight divided by the square of their height) to separate individuals into two groups: overweight (BMI over 24.99 kg/m2), and not overweight (BMI 24.99 kg/m2 or under). This categorisation is based on the WHO’s recommended cut-offs for BMI classification [5]. Rather than split the continuous BMI measure into multiple subcategories of overweight, we used this classification as the main aim of the paper was to analyse trends in excess adiposity, and research has found an elevated risk of NCDs and mortality beyond a BMI of 24.99 kg/m2 [21, 22]. We did not use a continuous measure of nutritional status, as observed population-level increases in BMI we would expect to observe over the study period could be driven by a both individuals moving into overweight categories, and individuals moving from underweight to normal weight; the latter of which does not capture increases in excess adiposity.
Height and weight information on women aged 15–49 in NFHS-2, 3 and 4, and men aged 15–54 in NFHS-3 and 4, were collected by specially trained investigators. A solar-powered SECA digital scale was used to measure the weight of respondents, with the NFHS-2 report claiming an accuracy of ±100 g. The height of respondents in NFHS-2 and 3 was measured using a measuring board designed for use in survey data collection. In NFHS-4, the Seca 213 stadiometer was used to collect respondent’s height information [1,2,3].
Independent variables
Exposure of interest
We used a measure of educational attainment as our primary indicator of SEP. This was based on the answer to a question regarding the number of completed years of schooling, and respondents were assigned to one of the following education categories: No Education (0 years); Primary Education (1–5 years); Secondary Education (6–12 years); and Higher Education (12+ years). Higher levels of education can increase earning capability, along with the accumulation of employable skills, both of which make it a suitable proxy for SEP.
For sensitivity analysis we verified our results using a standard of living (SoL) asset-based index as an alternative measure of SEP. In surveys, measures of SEP are seldom examined in isolation, as one measure cannot adequately describe all socioeconomic differences in a health outcome [23]. As education and SoL capture different aspects of SEP, the pathways through which it is associated with overweight may also differ. For example, those with high education may work in more sedentary jobs [6,7,8,9], increasing their risk of overweight, whereas SoL may be positively associated with overweight through determining the ability to afford excess food [6,7,8,9]. Some suggest that in low/middle income settings, where there is a substantial informal employment sector and earnings not in the form of monetary enumeration, household income may not be an appropriate measure of SEP. Rather, the stock of assets may be more reliable [24]. Data on household income to proxy SEP is likely to be very sensitive to seasonal fluctuations in repeated cross-sections and may not capture the true level of wealth of the household. Additionally, in transitioning societies, it may be more common to receive income ‘in-kind’ rather than monetary enumeration [25], and households may draw money from multiple sources [24], limiting the ability for respondents to adequately recall all income in a questionnaire.
We created our own SoL index using principal components analysis (PCA) after pooling the household surveys over time. The inputs we used into the PCA included information on the household’s stock of assets, their access to services, and other household characteristics. We completed this process for urban and rural areas separately due to differences in the importance of different assets between urban and rural residents. The percentage of respondent households in urban and rural areas by characteristics used to build the SoL index in each survey is presented in Additional file 1: Table S1. We then ranked households based on this new index and assigned the first, second and last third of the weighted sample a SoL classification of ‘Higher’, ‘Medium’ or ‘Lower’ Standard of Living (SoL).
We examined the validity of the SoL index we created by comparing the ranking of households using the index from the pooled data, within one survey, and the survey-specific wealth index already included in the data. The correlation coefficient in each of the three surveys used was greater than 0.95, suggesting a very strong agreement with our measure and the household rankings determined the survey-specific index.
Covariates
Our final models were adjusted for the respondent’s age (15–29; 30–39; and 40–49 (40–54 for men)) and marital status. Marital status was categorised as either ‘currently married’ or ‘not currently married’ and was included as married individuals have been found to be at higher risk of being overweight [26]. We would have also preferred to control for the respondent’s occupation. Higher prevalence of overweight may be expected to be observed among individuals in more sedentary jobs [6,7,8,9], and sedentary labour may be expected to be more prevalent among higher SEP individuals. However, it was not possible to control for occupation in our research due to the fact that it was collected on a very limited subsample of the respondents in NFHS 4 (approximately 5% of women in the NFHS-4 national sample).