Main findings
In this study, we presented a g-computation approach to evaluate the potential impact of hypothetical weight reduction scenarios on the burden of four NCDs in Belgium. We examined what would be the risk of suffering from diabetes, hypertension, CVD, and MSK disease if we could manipulate the BMI or the WC of Belgian adults and set them to values determined by hypothetical scenarios. The predicted risk was then compared to the risk under the “status quo” scenario, where no intervention would be implemented to the population. This is in contrast with the estimates we would have obtained using traditional regression models which produced stratum-specific odds ratios.
Our findings suggest that implementing weight reduction scenarios among individuals with excess weight could lead to a substantial and statistically significant decrease in the prevalence of diabetes, hypertension, CVD, and MSK diseases in Belgium. A major benefit was found for the fourth scenario, where the WC was lowered to half of the height for all Belgians with a ratio WC:height ratio above 0.5. Under this scenario, the prevalence of diabetes and hypertension would be drastically reduced, with respectively 36% and 25% of avoidable cases. The reduction was less pronounced for CVD and MSK diseases with a PIF of respectively 11% and 7%. A recent guideline report from the National Institute for Health and Care Excellence (NICE) mentioned that a waist measurement of more than half of a person’s height was a better indicator of increased fat in the abdomen compared to BMI and could better predict the risk of developing NCDs such as type 2 diabetes or CVD [53]. BMI remains however a useful practical measure to define overweight and obesity but should be interpreted with caution especially among older people and adults with high muscle mass, since it is less accurate to determine body fatness in these groups [58].
High PIFs were also observed under the first scenario, where the distribution of the BMI of all people with overweight would be shifted to the distribution of the BMI of people fallen in the “normal” BMI category. While this scenario may not be highly realistic, it is nonetheless valuable in defining the boundaries within which realist policy interventions could have an impact. This very theoretical scenario has the advantage to estimate the global burden of excess weight on NCDs and is closest to traditional PAF which estimate the risk of disease with a complete removal of the risk factor in the population. Under this first scenario, PIF for diabetes, hypertension, CVD, and MSK disease were 32%, 23%, 9%, and 6%, respectively. Those estimates were however lower in comparison to the PAF estimates obtained from the last Global Burden of Disease (GBD) study where the PAFs attributable to high BMI in Belgium were respectively of 50% for diabetes, 20% for ischemic heart diseases, 25% for stroke, 7% for back pain, and 13% for osteoarthritis [59].
It must be noted that those estimates cannot directly be compared to the estimates presented in this article. The g-computation approach is tailored to our data by estimating, for each individual, the conditional probability of developing a chronic disease given the variables included in the model, and subsequently averaging it at the population level.
In contrast, the GBD study's PAF estimates consider the overall contribution of high BMI to diseases across the entire population. They are not calculated directly from the specific population but often rely on relative risk estimates from external studies. These differences in data sources, methodologies, and the underlying framework for estimating population-level burden versus individual causal effects make direct comparisons between the two sets of estimates complex.
In addition, the variables for CVD and MSK used in this study were constructed based on a group of diseases (Table 1), which is difficult to compare with the GBD estimates, where PAFs are calculated for each disease separately.
The second and third scenario, where the BMI was respectively reduced by one unit and modified based on a ten percent reduction of the person’s weight, represent more realistic scenarios but had a smaller impact on the prevalence of the four diseases. A weight loss of 5–10% is considered by guidelines from the UK and the USA a the minimum weight loss to be achieved to have a clinical impact on health outcomes [58, 60]. To achieve this goal, evidence-based interventions include dietary modifications, physical activity, psychological interventions, pharmacotherapy, and bariatric surgery, for individuals with severe obesity [61, 62]. There is substantial evidence demonstrating that these interventions not only contribute to weight loss but also have a statistically significant impact on reducing the risk of obesity-related outcomes [63]. A one-unit reduction in BMI within the Belgian population would result in a reduction of 4.5% of the cases of diabetes, 3% of the cases of hypertension, 1.5% of the cases of CVD, and 1% of the cases of MSK disorders.
Strengths and limitations
An important strength of this study lies in the didactic application of the g-computation approach and the description of the steps required to estimate the population effect of a potential intervention in cross-sectional data. The methodological tool used in this present study, based on a g-computation approach and a random-forest multiple imputation method, allows the assessment of the potential effects of any well-defined intervention and targeting of any subgroup of interest, while also addressing the bias related to self-reported data and the missing data issue in health interview surveys. This paper contributes to familiarizing a public health audience with the g-computation approach enabling them to estimate policy-relevant effects of hypothetical health interventions. Compared to standard analytic techniques, the g-computation approach has the advantage to provide flexibility in simulating real world interventions. It enables modeling the impact of dynamic interventions, where different subjects can receive varying levels of the exposure under study, as well as joint interventions, where the values of multiple exposures can be modified simultaneously. Another additional benefit of the g-computation approach, lies in its ability to handle time-varying confounders (i.e., confounders whose value changes over time), especially in situations where there's treatment confounder feedback (i.e., when the confounder is affected by the exposure) [64]. However, the cross-sectional nature of the data in this study did not allow us to take full advantage of this benefit.
This study also represents the first application of the random-forest multiple imputation method to address the bias related to self-reported health and anthropometric data in the BHIS. This method has been recently identified as a more adequate approach for valid measurement error correction in comparison to regression calibration [37]. Whenever feasible, self-reported information from health interview surveys should be combined with objective information from health examination surveys to address the bias related to self-reported anthropometric data and therefore provide more accurate PIF. A second important strength of the present study is the consideration of the potential confounding role of the environmental factors in the association between excess weight and chronic diseases. In particular, the linkage of the BHIS data with objective environmental factors at the residential address of the participants provides a significant improvement on the state of the art, as most studies do not consider environmental factors in the link between BMI and chronic diseases. Also, environmental factors are often assessed on a broad scale, using exposure e.g. in administrative units. Our study used the residential address, thus considerably refining the spatial scale. The limits of this approach are discussed further in the section measurement error.
Findings of this study must nevertheless be seen in the light of some limitations. If the g-computation approach allowed to evaluate the PIF of several weight reduction scenarios, the obtained estimates should however be treated with caution and several assumptions need to be met to interpret them causally. The first assumption is the “temporal ordering assumption” where we assume that the exposure precedes the outcome and the confounding factors precede the exposure. Unfortunately, this required assumption is not met by the cross-sectional structure of the data and is undoubtedly the most questionable assumption in this present study. While we can reasonably assume that fixed variables such as age, sex or education are causes rather than effects of the excess weight risk factor, it is not that obvious that the excess weight risk factor precedes chronic disease or that lifestyle factors precede the weight status. Making the distinction between unintentional weight loss, which may result from chronic disease, and intentional weight loss can be challenging [65]. People suffering from chronic disease could also be physically less active and therefore be at greater risk of gaining weight. For instance, individuals with CVD, MSK disorders or diabetes may exhibit weight gain due to factors like reduced mobility (leading to a decrease in calorie expenditure), medications, or fluctuations in blood sugar levels. Another challenge with cross-sectional data is the inability to differentiate whether covariates function as mediators or confounders. In this study, physical activity was considered as a confounding factor but it cannot be ruled out that excess weight may impact physical activity and indirectly the risk of chronic disease. One possible consequence could be underestimation of the true causal effect because the PAF would not incorporate all burden for the disease that is attributable to the excess weight risk factor. Physical activity could also function as a collider variable (a variable that is a common effect of both the exposure and the outcome) and adjusting for it may have introduced collider bias, potentially generating a spurious association between excess weight and chronic diseases.
The second assumption is the “exchangeability” assumption which assumes that there are no unmeasured confounding factors in the exposure-outcome association. Indeed, the exposure may only be considered as randomized within each stratum of the confounders if all confounders are considered in the model. This assumption is also very difficult to meet in the available cross-sectional study. Although we included in our analyses all the confounders identified in the literature that were available in our data, there remain several potential unmeasured confounding factors, such as genetic factors or nutritional habits which can both play an important role in the association between excess weight and chronic disease. Even though the variables related to nutritional habits were available in the BHIS, it was decided to not include them in the model because they were highly prone to a reverse causation effect.
The third assumption, known as the “no-interference assumption”, asserts that the outcome of each individual is not affected by the exposures and outcomes of the other individuals. We can reasonably expect that this assumption is fully met in our study for the reason that chronic diseases are not contagious. This, however, may vary depending on the intervention and study group. For instance, the implementation of a dietary intervention to reduce BMI of participants, such as changing the cooking style in the family, could potentially influence members of the same family similarly.
The fourth assumption, the “experimental treatment assignment” assumption, also called the positivity assumption [66], assumes that the exposure to the risk factor is possible for all individuals in each stratum of the covariates. In the context of this study, it means that the BMI values generated under the considered scenarios must be attainable for all individuals in which the scenario took place. This assumption is closely related the realism of the scenario and is therefore more likely violated for the first and fourth scenarios, which requires changes in the BMI or in the WC that are rarely observed in the population (e.g. a drop in the BMI from 35 to 25). In concrete terms, this means that each stratum of the covariates that contains overweight individuals should also contain individuals with a normal BMI. To evaluate the positivity assumption, we compared the probability of individuals being overweight among the two populations groups under study (individuals with overweight and individuals with a “normal” BMI). We built a model for BMI based on all confounders, and predicted, for each individual with overweight, the probability of being overweight. This process was repeated for individuals with a normal BMI. The observed overlap between the two probability distributions suggests that this assumption is plausible (Additional file 12).
The fifth assumption is the “consistency” assumption, which assumes that “an individual’s potential outcome under his observed exposure history is precisely his observed outcome” [19]. While consistency is plausible for medical treatments, because it is easy to manipulate hypothetically an individual's treatment status, consistency may however be problematic when the exposure is a biologic feature and the manipulation difficult to conceive [67]. Violations of consistency assumption often occur when there is ambiguity in the definition of interventions to change exposure. In the context of this study, BMI interventions remain vague because they specify attributes rather than specific behaviors. The main limitation of our approach lies in the highly theoretical nature of the hypothetical scenarios considered, which do not accurately mirror real-world interventions. Ambiguity arises from the fact that there are many competing approaches to decrease an individual’s BMI and each of these approaches may have a different causal effect on the outcome [68]. By presenting an estimate for the effect of a “BMI reduction”, we implicitly assume that all interventions on BMI have the same effect on the risk of suffering from a chronic disease, which is unlikely to hold. Another difficulty arising from ill-defined interventions is the challenge of selecting the confounding factors required to achieve conditional exchangeability. Firstly, the set of confounding factors to be considered may vary for different versions of the intervention. Secondly, because BMI is not an intervention in itself but rather a physiological risk factor, identifying all the confounders becomes a practically impossible task due to the necessity of also considering genetic factors. Even if we manage to account for all potential confounding factors including genetic factors, there is a high likelihood that the positivity assumption will be violated. Certain genetic traits could exert such a strong influence on body weight that all subjects possessing them automatically become obese [68]. Another issue with interventions on BMI is that the better we adjust for confounders that determine both excess weight and chronic diseases, the more we narrow our focus to the remaining factors that have a direct effect on BMI (such as genetic predispositions). Consequently we isolate a potential intervention that changes the remaining determinants of BMI. In this study, we compared the risk of suffering from a chronic disease of overweight vs non overweight individuals conditional on their physical activity level, smoking status, environmental and alcohol consumption. This means that our estimates correspond to the effect of other versions of the intervention “BMI reduction”, such as healthy diet or genes. However, other versions of the intervention may not be manipulable and not be of primary interest for policymakers. Successful interventions with evidence for effective weight reduction are multifactorial and it is unrealistic to assume that BMI in the population could be modified without considerable changes to all other aspects of lifestyle. Our findings may therefore be underestimated, since our analyses adjusted for possible confounding by physical activity or alcohol consumption and thereby do not entirely take into account the co-benefits of weight reduction intervention via changes in physical activity or alcohol consumption.
The sixth underlying assumption of g-computation approach is the “no model misspecification” assumption. A necessary condition (but not sufficient) for the absence of model misspecification is that the model should be able to accurately predict the outcome under no intervention. Variables from the model were selected based on their theoretical relevance and guided by a DAG that reflects the hypothesized causal structure. Non-linear relationships were assessed by testing the quadratic terms, while interactions were examined using the StepAIC algorithm (a variable selection method that iteratively adds or removes variables from a model based on their impact on the Akaike Information Criterion, aiming to find the most parsimonious model with a good fit). The AUC demonstrated a good predictive performance for the four NCDs models.
Lastly, like other studies based on observational data, the validity of our results relies on the key assumption of no measurement error. It can however be challenging to accurately assess the exposure to risk factors of NCDs through observational studies, such as abdominal obesity or environment. Although we applied a correction method to address the bias of self-reported anthropometric data and used both BMI and waist circumference separately to approximate abdominal obesity, another measure that could have been used is the Body Shape Index (ABSI), a comprehensive indicator of body shape integrating both waist circumference and BMI [69]. For the environment also, it is important to keep in mind that air pollution exposure is extrapolated from the mean annual concentration of a given area to individual exposure, and does not take into account the time spent in this area. Personal mobility could be integrated in dynamic exposure assessments, but determining individual buffer values to delimit a person’s neighborhood is still an active field of research. Other methods to determine environmental exposure are human biomonitoring or deploying wearable sensors, but this is unfortunately impossible to apply for large samples, over long time periods or for past studies. There was also a time lag between health data collection and environmental data. However, as environmental change is slow, we do not expect a strong impact on our results. A certain degree of measurement error also applies for the diseases. While the bias related to self-reported diabetes and hypertension could be addressed based on clinical information from the BELHES, the same correction could not be applied for self-reported CVD and MSK diseases, as the relevant clinical information was not available in the BELHES.
Furthermore, our estimates apply to the Belgian population and may not be generalizable to other populations characterized by different NCDs risk factor distributions. For example, we estimated the risk of diabetes in the Belgian population for a distribution shift of the BMI of individuals with overweight to the distribution of individuals with a normal BMI, but the BMI distribution may be very different in other populations. Our PIF estimates may also vary a lot for different diseases within the same CVD or MSK group, limiting the possibility of comparing our results with the GBD estimates.
A final limitation of our study lies the lack of detailed analysis regarding the differential effects of BMI on different types of diabetes. While our findings demonstrate a significant association between BMI and diabetes, it must be recognized that the impact of BMI may vary between type 1 and type 2 diabetes. While the link between obesity and type 2 diabetes is well-established, emerging evidence suggests a link between obesity and type 1 diabetes as well [70]. Future research could explore this aspect further to elucidate whether BMI affects both types of diabetes similarly.
Whilst obesity is widely considered as a major modifiable risk factor for many chronic diseases, nevertheless, a rigorous examination of the mentioned assumptions underscores the challenge in determining its causes and consequences. Addressing this is however important, as the prevention of any disease requires that interventions focus on causal risk factors. Although all the required assumptions of the g-computation approach may not be fully met, based on the literature knowledge regarding the relationship between excess weight and NCDs, the evidence from literature supports the direction of causality investigated in this study.
https://news.google.com/rss/articles/CBMiT2h0dHBzOi8vYm1jbWVkcmVzbWV0aG9kb2wuYmlvbWVkY2VudHJhbC5jb20vYXJ0aWNsZXMvMTAuMTE4Ni9zMTI4NzQtMDI0LTAyMjEyLTfSAQA?oc=5
2024-04-14 09:41:07Z
CBMiT2h0dHBzOi8vYm1jbWVkcmVzbWV0aG9kb2wuYmlvbWVkY2VudHJhbC5jb20vYXJ0aWNsZXMvMTAuMTE4Ni9zMTI4NzQtMDI0LTAyMjEyLTfSAQA
Bagikan Berita Ini
0 Response to "The potential impact fraction of population weight reduction scenarios on non-communicable diseases in Belgium: application of the g-computation approach - BMC Medical Research Methodology - BMC Medical Research Methodology"
Post a Comment