From a public health perspective, the first consideration we can draw from the analyses above is that the strong positive trend in BMI observed in the last decades in all countries of Southern Africa is still present in the South African population. Even considering the reduction in the rate of increase observed in men, this fact in itself calls for the urgent implementation of public health interventions to curb the obesity epidemic. In the absence of effective interventions, the overall proportion of adult South Africans who are overweight or obese is extremely likely to increase by a few percentage points by 2020, rather than to decrease by 10 % as per South African Government’s strategic plan [31]. This is likely to further increase rather than reduce the relative mortality from non-communicable diseases, which have become the largest broad cause of premature mortality since around 2009 [32].
A second consideration that can be drawn from analyses suggests that trends in obesity are not homogeneous across population strata defined by biological, behavioural and socioeconomic characteristics. The identification of risk factors in the latter two categories has immediate potential from a public health perspective, because these factors are potentially modifiable. The identification of biological factors is also of public health interest since this knowledge can help targeting high risk groups more effectively, avoiding the waste of resources associated with interventions excessively broad in scope.
The existence of socioeconomic inequalities in obesity prevalence is a well-established finding, and has been previously confirmed in the South African population [3, 8]. This study adds to those findings showing that socioeconomic inequalities exist also in the temporal trends of obesity. In particular, our analysis suggests that subjects at the extreme of the income distribution (fourth and fifth quintiles) and belonging to the White population group (a strong indicator of high socioeconomic status) are more at risk of increasing their BMI compared to those in the lower socioeconomic strata. This conclusion is supported by the fact that subjects with tertiary education (another strong indicator of high socieconomic status) also show higher rates of change, even though the relationship does not reach statistical significance.
Our results also indicate rural vs. urban dwelling as a risk factor for increasing BMI. This is an interesting finding, because the results from other studies show that the prevalence of obesity in South Africa is higher in urban that rural areas, especially among women [3]. Our analyses confirm that rural dwelling is associated with significantly lower baseline BMI, but also suggest that subjects in rural areas are ‘catching up’, possibly because of the rapid spread of urban lifestyles (high consumption of processed food, reduced physical activity) into rural areas. From a public health perspective, this indicates that targeting rural areas before the prevalence of overweight and obesity increases to ‘urban’ level, could have a significant impact on the future trends in the obesity epidemic in the country.
BMI in women is increasing more rapidly than in men. Moreover, while the rate of increase seems to be slowing down among men, among women there is no evidence of a similar reduction. Young women, in particular, seem to be the most vulnerable.
Among behavioural factors, our study offers evidence that smokers tend to increase their BMI less than non-smokers, that smoking cessation is an independent risk factor for faster BMI increase among subjects with normal weight, and that obese/overweight subjects who start smoking tend to lose weight.
However, these results (which add to the consistent findings of many population studies showing that smoking is associated with lower BMI, [33] and that smoking cessation is associated with weight gain [34]) should not be interpreted as a support of smoking as an efficient way of controlling body weight. The relationship between smoking and body weight is incompletely understood and, even though a direct effect of nicotine on body weight is plausible (because nicotine increases energy expenditure and could reduce appetite), many other factors are likely to be involved in explaining the observed relationships. Among those, the results of various studies suggest that smoking inception is more frequent among subjects with greater weight concern and previous attempts to lose weight, which may indicate that changes in body weight following smoke inception are at least partly determined by psychological and behavioural factors that precede the initiation [35]. These factors were not measured in our study.
Moreover, a growing literature shows that smoking is associated with increased insulin resistance and risk of type 2 diabetes, as well as greater waist circumference (an indicator of the amount of visceral adipose tissue) thus suggesting that, despite the lower BMI, smokers, and especially heavy smokers, have an increased cardiovascular risk compared to non smokers [35]. The finding of a greater waist circumference among smokers is supported by the results of a secondary cross-sectional analysis of our data, which show that heavy smoking (>20 cigarettes/day) is significantly associated with greater waist circumference compared to no smoking (linear regression coefficient c = 3.25 cm, 95 % CI: 0.25 ; 6.2)4.
In any case—regardless of the dubious underlying causal mechanism—our finding that smoking cessation is associated with weight gain might be of public health interest because of the observed downward trend in the number of smokers in South Africa, [16] which, besides its overwhelming benefits for general population health, could foster an increase in BMI. Subjects who decide to quit smoking should be considered for obesity prevention strategies.
Among normal/underweight subjects, high levels of physical exercise were associated with lower rates of increase of BMI during the study period, but with higher baseline BMI. A possible explanation of this incongruence between cross-sectional and longitudinal relationships is that they might be expression of reversed causal processes. That is, while the causal precedence in the longitudinal relationship between exercise frequency and subsequent BMI change is determined by the temporal sequence, it might be that the observed cross-sectional association is the result of a greater tendency of subjects with higher BMI to exercise more in order to decrease their weight.
Finally, special consideration of waist circumference is deserved. In our analysis, waist circumference (besides being a risk factor for cardiovascular disease per se) was directly associated with higher rates of increase in BMI, independently of the BMI class. These considerations strongly suggest centrally obese subjects as primary targets of obesity reduction campaigns.
Several limitations of this study need to be acknowledged. Low reliability of self-report data, including those on physical exercise, alcohol and tobacco use is a well-known problem in population-based surveys and the measurements used in this study are no exception. However, in absence of specific reasons to think of an association between measurement error and individual BMI, it is probable that this measurement unreliability resulted in observed associations biased towards the null [36]. More precise measurements are therefore likely to strengthen the result of our analyses rather than invalidate them.
Independent variables introduced in the models as possible predictors of different BMI trajectories were identified according to previous evidence of association with BMI in population studies in SubSaharan Africa and availability in the NIDS dataset. Other factors are likely to play a significant role in explaining inter-individual differences, and, among those it is worth mentioning dietary habits and metabolic disorders, especially diabetes. For both these variables, the NIDS dataset provides neither direct measures nor reliable proxies5.
Suboptimal response and greater attrition rates were observed in some social strata in the NIDS survey. Even though differences between respondents and non-respondents in observed characteristics have been taken into account through appropriate adjustment of sampling weights, we cannot exclude the possibility that unobserved differences might have biased the results of our study in an unpredictable way.
Finally, the availability of three successive measurements only allowed for the estimation of linear trends. The availability of measurements from the forthcoming waves of the NIDS study will allow the possibility of non-linear trends and to better forecast future scenarios.