Origin of the data

Led by the Centers for Disease Control and Prevention in the United States, NHANES is a cross-sectional epidemiological study that applies a layered multi-phase probability sampling method.^{(Reference Zipf, Chiappa, Porter, Ostchega, Lewis and Dostal15)} The sampling process involves multiple stages: primary sampling units (counties or contiguous county groups) are first selected, stratified by geographic and socioeconomic criteria. Segments – such as census blocks or household clusters – are then chosen within these units, followed by random household selection. Eligible individuals within households are subsequently invited to participate. To enhance generalizability to the U.S. population, specific subgroups (e.g., racial/ethnic minorities, older adults, and low-income populations) are oversampled. This complex survey design ensures representativeness but requires the application of sampling weights in all analyses to account for differential selection probabilities, non-response, and oversampling. Started in 1999, NHANES runs on a two-year cycle, amassing comprehensive information about the health and nutrition of individuals within the U.S. demographic. This includes demographic details, physical examination outcomes, laboratory results, and dietary habits.⁽¹⁶⁾ An ethics review committee has granted approval for the data collection of NHANES, which is managed by the National Center for Health Statistics.⁽¹⁷⁾ As of September 2024, data analysis for the current research was conducted. Secondary data analyses do not require additional institutional review board approval, due to written informed consent is mandatory for all NHANES participants.

NHANES data can be accessed by the public through their official website (https://www.cdc.gov/nchs/nhanes/) (accessed on 4 September 2024). We based our study on adult participant data from the NHANES cycles spanning 1999 to 2004, selected for their inclusion of severe headache or migraine data within the miscellaneous pain questionnaire. This study focused on participants aged 20 years and above who had completed the interview process. To ensure the reliability of key outcome variables (severe headache or migraine status) and confounding factors (dietary carbohydrate consumption and baseline health covariates), individuals meeting any of the following exclusion criteria were excluded from the analysis: first, those with unavailable self-reported severe headache or migraine data (n = 11), as these primary outcome measures were essential for the study’s objectives; second, pregnant participants (n = 836), because oestrogen secretion levels during pregnancy could alter participants’ baseline headache status;^{(Reference Phillips, Koonalintip and Wakerley18)} third, individuals lacking dietary data (including carbohydrate consumption, n = 1783), since carbohydrate intake in dietary records was a critical variable under investigation in this study; fourth, participants with missing data for important baseline covariates (n = 2278), where “important covariates” specifically included age, sex, body mass index (BMI), blood pressure, and comorbid conditions identified in prior headache epidemiological studies as strongly associated with migraine prevalence or chronic headache outcomes, necessitating their inclusion to adjust potential confounding effects in causal inference;^{(Reference Liu, Wang, Wu, Dong and Yu19,Reference Liu, Wang, Dong and Yu20)} and finally, those with missing weight measurements (n = 11), as weight data were required for BMI calculation, a key covariate in analysing the relationship between headaches and dietary factors. The final analytical sample comprised 10,413 individuals, as detailed in Figure 1.

Figure 1.

Inclusion and exclusion process of the study was based on the 1999–2004 National Health and Nutrition Examination Survey.

Severe headache or migraine and carbohydrate consumption

This study identified individuals with a history of severe headache or migraine by analysing their responses to a particular inquiry located in the miscellaneous pain section of the NHANES survey: “Have you had a severe headache or migraine in the past three months?”^{(Reference Liu, Wang, Wu, Dong and Yu19,Reference Peterlin, Rosso, Rapoport and Scher21,Reference Slavin, Li, Khatri and Frankenfeld22)} In dietary evaluations, NHANES relies on the Automated Multiple Pass Method during the execution of 24-hour food recall surveys. This system, with computer assistance, thoroughly records the food and beverage intake of participants for the preceding 24-hour period, starting at midnight and ending at midnight.

Carbohydrate intake density in the diet was assessed using data from the USDA’s Food and Nutrient Database for Dietary Studies. The calculation of carbohydrate energy percentage, defined as the proportion of energy derived from carbohydrates (expressed as % of kcal), is based on the assumption that carbohydrates provide 4 kcal/g. This percentage is calculated using the formula: carbohydrate energy percentage (% of kcal) = carbohydrates intake (g/day) × 4 (kcal/g)/energy intake (kcal/day).^{(Reference Lieberman, Fulgoni, Agarwal, Pasiakos and Berryman23)}

Covariates

A range of possible covariates were considered in our study, informed by existing research literature. These encompassed age, BMI, gender, spousal status, racial/ethnic background, level of education, dietary intake, household financial status, smoking behaviour, alcohol consumption, physical activity, diabetes, cardiovascular condition (confirmed by a doctor), and metabolic syndrome.^{(Reference Liu, Wang, Wu, Dong and Yu19,Reference Liu, Wang, Dong and Yu20,Reference Slavin, Li, Khatri and Frankenfeld22,Reference Meng, Zhou, Li, Wang, Kang and Fu24–Reference Gan, Estus and Smith26)} The poverty income ratio (PIR) was used as the basis for classifying family income into low (PIR ≤ 1.3), medium (PIR 1.3–3.5), and high tiers (PIR > 3.5), according to a US government publication.⁽²⁷⁾ Following the guidelines from previous research, participants were categorized as never, former, or current smokers.^{(Reference Liu, Wang, Chen, Dong and Yu25)} Alcohol consumption was similarly classified into never, former, or current drinker. Engaging in moderate to vigorous exercise for over 10 minutes during the last 30 days was considered physical activity; participants who did not meet this criterion were labelled as physically inactive. A 24-hour dietary recall interview, conducted at the mobile examination centre, provided detailed information on the participants’ total energy, protein, total fat, and dietary fibre intake. The diagnosis of diabetes was confirmed if any of the following criteria were met: a doctor’s diagnosis, Plasma glucose levels of ≥11.1 mmol/L in a 2-hour oral glucose challenge, ≥11.1 mmol/L in a random test, ≥7.0 mmol/L when fasting, a glycosylated haemoglobin (HbA1c) greater than 6.5%,⁽²⁸⁾ or being treated with insulin or diabetes medication. Participants were classified as having diabetes if they met at least one of the above criteria. To resolve potential conflicts in classification (e.g., a participant on medication with a normal HbA1c), a pre-defined hierarchy was used: treatment with glucose-lowering medication was given the highest priority, followed by a self-reported doctor’s diagnosis, and then any abnormal biochemical value. This hierarchy ensures that individuals receiving treatment for diabetes are appropriately classified regardless of their current biochemical status. Participants identified solely by elevated biochemical markers, without a prior diagnosis or treatment, were defined as having undiagnosed diabetes. Individuals with missing data for all diabetes-related variables were excluded from the analysis. In this study, metabolic syndrome was described based on the 2005 Adult Treatment Panel III criteria.^{(Reference Grundy, Cleeman, Daniels, Donato, Eckel and Franklin29)}

Statistical analyses

This research involved a reanalysis of an open-access dataset, following the analytical guidelines established by NHANES.^{(Reference Johnson, Paulose-Ram, Ogden, Carroll, Kruszon-Moran and Dohrmann30)} Our analysis meticulously accounted for the complex sampling design and associated sampling weights. We utilized dietary weights in our weighted analysis. Specifically, NHANES data from the 1999–2000 and 2001–2002 cycles were analysed with the four-year dietary weighting system (WTDR4YR), whereas the first-day dietary weighting system (WTDRD1) was used for the 2003–2004 data. The determination of sample weight values over the 1999–2004 timeframe was based on the following scheme: for the years 1999–2002, weights were calculated as 2/3 × WTDR4YR, and for the subsequent years, as 1/3 × WTDRD1. This weighting scheme is consistent with the NHANES analytical guidelines for combining multiple survey cycles. The periods 1999–2000 and 2001–2002 constitute 4 years of data and were therefore assigned a four-year dietary weight (WTDR4YR). The 2003–2004 period constitutes 2 years of data and was assigned a first-day dietary weight (WTDRD1). To create a composite weight that accurately represents the U.S. population over the entire 6-year study period (1999–2004), the original weights were proportionally scaled. The weight for the 1999–2002 data was multiplied by 2/3 (4 years/6 years), and the weight for the 2003–2004 data was multiplied by 1/3 (2 years/6 years). This adjustment ensures that each sub-period contributes to the analysis in proportion to its sample size and time span, preventing over- or under-representation of any single cycle and providing unbiased national estimates for our study window.

In this analysis, means and standard errors were used to describe continuous variables, while unweighted counts and weighted proportions were applied to represent categorical variables. We utilized linear regression models for numeric variables and χ² tests for categorical data to assess variations between groups. A regression model with multiple variables was employed to investigate the relationship, and the findings are presented as odds ratios (OR) with 95% confidence intervals (95% CI) for the link between carbohydrate consumption and severe headache or migraine prevalence.

We constructed three multivariable regression models using a sequential adjustment strategy to comprehensively assess the association of carbohydrate consumption with self-assessed severe headache or migraine, and to evaluate the potential influence of different confounding pathways. This approach allows for the examination of how the effect estimate changes upon the addition of groups of covariates that represent different types of confounders, while avoiding over-adjustment for potential mediators.^{(Reference Böhnke31,Reference Szklo, Nieto and Miller32)} Model 1 (Minimal adjustment): This baseline model adjusted for demographic and basic socioeconomic factors, including age, gender, racial/ethnic background, marital status, level of education, and household financial status, as well as NHANES survey cycles. These factors are considered fundamental upstream confounders in health disparities research.^{(Reference Arab, Karimi, Garaulet and Scheer33)} Model 2 (Model 1 + Lifestyle factors): This model further adjusted for modifiable behavioural and anthropometric risk factors, including smoking status, BMI, alcohol use, and physical activity. This step was taken to test whether the observed association was independent of these major lifestyle-related confounders. Model 3 (Model 2 + Clinical comorbidities): This fully adjusted model additionally included history of clinical conditions: diabetes, metabolic syndrome, and cardiovascular disease. These factors were added last because they are potential mediators on the causal pathway between the exposure and the outcome. Adjusting for them in the final model allows us to examine if the exposure exerts an effect beyond or independent of these advanced disease states.^{(Reference Böhnke31,Reference Szklo, Nieto and Miller32)}

To investigate a possible non-linear association between the proportion of energy from carbohydrates and severe headache or migraine, we built a regression model adjusted for multiple variables, using a restricted cubic spline with three knots. After adjusting for covariates in Model 3, we used a two-piece logistic regression model with smoothing to evaluate the threshold relationship between dietary niacin intake and migraine, using the median percentage of carbohydrate intake in total energy as the reference point. The inflection points were identified using likelihood-ratio tests and the bootstrap resampling method.

Additionally, to assess the potential effect modification by key demographic and metabolic factors, we conducted stratified analyses based on a priori hypotheses. Sex was selected due to well-established differences in headache prevalence and pathophysiology linked to hormonal influences.^{(Reference Vetvik and MacGregor34,Reference van Casteren, Verhagen, Onderwater, MaassenVanDenBrink and Terwindt35)} Age was stratified at 50 years to approximate the menopausal transition in women and to explore age-related metabolic changes. Household financial status was chosen as a determinant of dietary quality and access to healthcare, which may confound or modify the diet-headache relationship.^{(Reference Hamdan, Tuot and Crews36)} Body Mass Index (BMI) was included because obesity is a pro-inflammatory state and a known risk factor for headache chronification, which may alter susceptibility to dietary triggers.^{(Reference Ulutaş, Bayır, Çetinkaya, Güler, Ağırcan and Özkan37,Reference Yan, Duan, Yin, Wang, Li and Li38)} Specifically, we stratified by sex, age (20–50 years and ≥50 years), household financial status (low vs. middle or high), and BMI (<25 kg/m² and ≥25 kg/m²). We separately applied logistic regression models and likelihood ratio tests to formally assess statistical heterogeneity and interaction across these subgroups.

Statistical power calculations were not conducted a priori due to the sample size being solely determined by the available data. We used R software version 4.2.1, in conjunction with the R survey package version 4.1-1 and Free Statistics software version 1.7.1, to perform statistical analyses.^{(Reference Yang, Zheng, Chen, Chen, Wen and Chen39)} Statistical significance was evaluated through a two-tailed P-value below 0.05.