Although the relationship between diet, physical activity (PA), and sedentary behavior on the one hand, and nutritional status on the other hand, is based on strong scientific grounds in both theoretical and empirical terms, quantifying it in a causal framework remains a challenge, especially in observational studies. While randomized controlled trials (RCTs) continue as a gold standard for causal inference, observational studies are more suitable for monitoring and tracking, both of which allow rapid adjustments in policy recommendations, regulations, and other interventions to improve nutritional health. With the advent of computerized evaluations of diet and PA by questionnaires, the cost-effectiveness of these methods has made them popular worldwide [1]. Reproducibility, internal, and (more rarely) external validity for some of these instruments have been reported, but to the best of my knowledge, their validity for causal inference on nutritional status has not been evaluated. The reason for that probably has a foothold in a firm belief that only RCTs, preferably case-control longitudinal studies, are entitled to causal interpretation [2].

The absence of random allocation of treatment (intervention, exposure, or any putative causal agent) in observational studies makes it difficult—but not impossible—to attribute the differences between the groups under comparison to the treatment. The effect of confounding and other omitted variables cancels out between the groups under random assignment but observational studies cannot control this aspect of the study design. Even so, under certain conditions, it is possible to balance the differences between the groups by estimating the likely values of the units (e.g., subjects) had they received the treatment applied to another group. Therefore, for each group both observed and hypothetical (“potential”) outcomes are presented, in addition to the observed treatments and baseline covariates, so from a statistical perspective it becomes a missing data issue. As long as the missing outcomes can be consistently estimated based on their association with observed data—a scenario known as “missing at random” [3]—unbiased differences between treatment groups can be calculated in observational studies.

To call these differences causal effects, a substantial theoretical basis must be present. For example, there are sufficient endocrinology grounds to claim that the pathogenesis of metabolic diseases is influenced by the quantity of sugar intake due to its contribution to advanced glycation end products, which in turn cause liver and muscle tissue damage [4]. Another point worth mentioning is the comparison of the differences between the end-of-study versus baseline performance of cases and controls in longitudinal studies. This design is known as difference-in-differences [5, 6] and has been widely applied to draw causal inferences in longitudinal studies with a control cohort.

The aforementioned reasoning rests on a key concept known as a potential outcome [7, 8], which can be translated as the unobserved outcome value under a particular scenario, such as receiving another treatment. In the context of cohort studies with intervention at a certain point in time, hypothetical values that would have been observed had there been no intervention can be estimated using so-called synthetic cohorts. For example, time trends in food consumption before and after relevant policy regulations may use this method to evaluate the policy impact.

A propensity score is another key concept for balancing the probability of receiving a particular treatment between groups under comparison in observational studies. The probability is estimated by binary logistic or multinomial regression for two or more treatment groups, respectively, using independent predictor variables such as baseline characteristics, and applied to matching units with similar probabilities, thus mimicking random allocation of treatment in RCTs. Various matching methods with software implementation, and their advantages and disadvantages, have been the subject of intense discussions [6].

Causal analysis is technically an application of marginal structural models [6]. An accessible introduction to causal methods in nutritional observational studies can be found from Mazzocchi et al. [5]. A special type of instrumental variable method to avoid bias in these studies, known as Mendelian randomization, is nicely explained in Bennet and Du [9], whereas the statistical basis for causal effect estimation is didactically provided by Cunningham [6].

The study aims to estimate the causal impact of diet and sedentary behavior on overweight in schoolchildren based on observational data. The emphasis is on the implementation and interpretation of causal analysis rather than on substantive theory.

The describes the relationship between sociodemographic characteristics, diet, and PA on the one hand, and two measures of excess weight on the other hand are described in the Table 1. The latter showed a clear-cut decline since the age of 11 years. Extreme categories of METs showed a difference in overweight including obesity, but it was reduced when only obesity was focused on. Higher totals of SSB, food items, and sedentary activities per day corresponded to higher proportions of overweight/obesity. The children who reported consuming lunch or dinner were more often overweight than those who did not report consumption of these meals.

No statistical significance is reported for the between-category differences concerning the outcomes as the data practically reached population coverage where the significance is assigned by definition, i.e. axiomatically.

Compared to not consuming sugary drinks, two or more of these per day significantly increased the odds of obesity by 3% (95% CI 0–6%), and overweight including obesity by 6% (2–10%) when unreliable diet records were excluded (Table 2). With regression adjustment instead of exclusion of these records, corresponding impacts were 6% (3–10%) and 9% (5–13%), in the same order.

Consuming 10–19 foods per day increased the odds of obesity by 6% (4–8%), and overweight including obesity by 13% (9–17%), compared to the children who consumed up to 9 food items per day, with the exclusion of unreliable dietary records. With the latter included and adjusted for as regression covariates, the corresponding figures climbed to 11% (8–15%) and 24% (16–32%), respectively. Compared to the same baseline, consuming 20 or more food items significantly increased the odds of overweight/obesity only when unreliable dietary records were excluded, by 17% (6–29%) and 9% (2–16%), respectively.

Using up to two sedentary activities as a baseline and the covariate adjustment, 3–5 such activities increased the odds of overweight including obesity by 12% (8–15%) and of obesity alone by 8% (4–12%). Only the former effect was statistically significant with the exclusion adjustment, increasing the odds by 7% (2–12%). Six or more sedentary activities significantly increased average odds in the range of 7–12% for adjustment methods of dietary records quality.

Consuming one or none of the ultra-processed foods per day increased the odds of being overweight/obese by 3% (0–6%) with the exclusion, but not with the covariate, adjustment. Eating 2 or more ultra-processed foods showed no significant effects on weight status.

Eating breakfast reduced the odds of overweight/obesity by 3% (0–7%) under covariate adjustment, whereas eating dinner increased these odds by 5% (1–10%) but only with the covariate adjustment method (Table 3).

No other effects of consuming specific meals/snacks were statistically significant.

To facilitate the understanding of the snacks, their top five food/beverage items are provided. In the morning, these consisted of bread/biscuits (15%), fruits (12%), yogurt (10%), cream biscuits (9%), and fruit juice (6%); in the afternoon, the preferred items were bread/biscuits (28%), cream biscuits (13%), fruits (12%), yogurt (12%), chocolate drinks (10%), and coffee with milk (10%); in the evening, the preference was for fruits (9%), sweets (8%), cream biscuits (8%), sodas (7%), fruit juice (7%), and bread/biscuits (7%).

To the best of the author’s knowledge, the present study is the first to estimate the causal effects of diet and sedentary behavior in nutritional epidemiology based on observational data. Daily consumption of sugary drinks, eating ten or more foods, and engaging in three or more sedentary behaviors per day significantly increased the odds of being overweight/obese. Most of these odds followed a dose-response relationship, i.e. they increased with higher levels of exposure.

Causal analysis with observational study data has a strong foothold in econometrics and public health, especially for evaluating the impact of policy regulations on an area level (county, district, state), and in epidemiological analyses of the impact of public health interventions such as vaccination, using synthetic cohort method [6, 29–32]. It usually requires a large sample size to guarantee a sufficient number of comparable case and control units, particularly if the effect size is small, such as in the present study. However, even a small risk increase may translate into a large disease burden if the cause is frequent in the population. This is certainly true for consuming sugar-added drinks and ultra-processed foods, frequent eating (“snacking”), and sedentary activities, as shown in Table 1. Also, all of these are modifiable risk factors as opposed to immutable genetic factors, so the weight of the former for health interventions in schoolchildren is of utmost value.

Lunch is the main meal in Brazil, with rice and beans traditionally accompanying meat or chicken. Brazilian schoolchildren who consumed such a lunch which concentrated most of the daily energy intake had lower obesity risk compared to other dietary patterns based on the time of day of eating events [33]. This is in line with the present study finding that frequent eating (“snacking”) increases the risk of overweight/obesity and some reviews shared this concern, especially in the context of social isolation during the coronavirus disease (COVID) epidemic [32–34].

The consumption of ultra-processed foods increased worldwide, including in the two largest countries in the Americas, the USA [35] and Brazil [36], and was associated with lower-quality diets in children and adults [37]. The present study did not find consistent evidence of a causal link between these foods and overweight/obesity, so further research is needed to verify if a longer follow-up to late adolescence and adulthood may confirm this link.

In line with the present study, schoolchildren’s sedentary time, especially their screen time, was associated with overweight/obesity in other studies [38, 39]. Other studies also found an impact of consuming SSB on excess weight in children and adults [40–43]. In Brazil, the consumption of SSB increased over the weekend by more than a third compared to the weekday [36]. The problem is compounded by the increasing tendency of both consuming SSB and sedentary activities worldwide [40–50].

The strengths of the present study include a large sample size and consequently a high statistical power, the application of robust statistical analysis of causal effects, with a probe into their sensitivity to broader versus the narrower definition of excess weight and to misreporting of dietary intake and PA. Also, the covariates controlling for other variables associated with the outcomes covered principal risk factors established in the literature. The best available statistical methods were applied for causal analysis. The survey questionnaire has been thoroughly tested for over fifteen years, and its reliability and validity are at least as good as other similar instruments, thus providing confidence in its capacity to evaluate schoolchildren’s diet and PA. Intraclass correlations between survey reports from the children in the same classroom were also taken into account, thus making statistical inference more robust.

The present study’s limitations include memory error, inevitable in self-reports, only one day of the week being asked about for each child, and not knowing the quantities of the foods/beverages consumed [51]. The survey application across different days of the week attenuated the issue of the representativeness of the results, as did a virtually total coverage of the target population. These limitations stem from the trade-off between the feasibility of the questionnaire designed for monitoring and surveillance of schoolchildren’s diet and PA on the one hand, and its accuracy and precision on the other hand. Moreover, statistical adjustment for multiple comparisons was not made, so the interpretation of isolated P-values should be cautious. For example, the effect of consuming dinner (Table 3) was significant in only 1 (P = 0.016) of 4 probes that combined 2 outcomes and 2 data quality adjustment methods. Also, despite a wide coverage of the principal factors known to be associated with excess weight, residual confounding cannot be ruled out. For example, genetic and early-life influences [52, 53] were not accounted for in the present study.

It is beyond the scope of this work to discuss the strengths and limitations of a variety of causal analysis designs and methods. However, it is worth noticing that the application of Mendelian randomization has been growing exponentially in medical research, including the field of nutrition [54, 55]. When the genetic analysis is unavailable, other designs that improve similarities between the baseline characteristics of the groups under comparison are still available [5, 6, 9, 31, 32]. The best design to evaluate the effectiveness of the interventions to prevent overweight/obesity—a matter of ongoing debate [56–58]—may benefit from a causal analysis, despite some skeptical views held by those who think confounding is ubiquitous in observational studies [2]. Cost-benefit analysis of these interventions is also affected by causal analysis. The aforementioned issues are bound to dominate future research in nutrition epidemiology and related disciplines.

The lack of a standardized definition of implausible dietary records has hampered more rigorous between-study comparisons, so the present study followed the recommendation to use both covariate adjustment and the exclusion of such records [59]. Some nationally representative surveys found that under-reporters were more likely to be overweight and conceal consumption of dietary sugar, including SSB [59]. Web-CAAFE validation studies showed that more than 30% of schoolchildren misreport dietary intake to some degree [15], especially for rarely and frequently consumed items [33]. Despite examining two extreme scenarios for Web-CAAFE misreporting, residual confounding cannot be ruled out, although its impact on the present study conclusions is likely small.

It is worth emphasizing that “causal” here means “unbiased” or “not confounded” plausible effect whose mechanism is often only partially known. In the last two decades, statistical advances and software development have brought causal analysis within the reach of researchers who are not specialists in this area. Contemporary information technology allows fast retrieval of big observational data where confounding is a primary concern because often the researchers did not have a say in the design of the data collection. Contrary to the widespread opinion, the statistical theory has firmly established the unbiasedness of causal analysis, even for cross-sectional studies, given the conditions similar to those required for traditional regression methods [6]. Accurate estimates of effect size provide a huge advantage for better preventive actions and evaluating their costs and benefits.

In conclusion, this work has pointed to the causal impact of consuming the SSB, eating ten or more foods per day, skipping breakfast, and sedentary behavior on increasing the odds of overweight/obesity. Under the right conditions, this type of analysis can be used in observational studies in nutrition sciences.