We used data from the 2002–2019 Korea National Health Insurance Service Health Screening Cohort (NHIS-HEALS). All Korean citizens aged ≥ 40 years are eligible for the biennial general health screening program. The NHIS-HEALS comprised 514,866 general health screening participants aged between 40 and 79 years in 2002, equating to a 10% simple random sample of the target population [18]. The data were provided by the National Health Insurance Service (NHIS), which covers 97% of the Korean population. Since the NHIS also manages the healthcare claims of the remaining 3% of the Korean population, the medical aid program beneficiaries, the NHIS database contains the medical records of the entire Korean population. The NHIS-HEALS includes anonymized participant information (sex, age, health insurance premium decile determined by household income, residential areas, medical records, and health screening database). All participants were followed up until their loss of eligibility due to death or emigration [18].

The migraine cohort was constructed as follows: First, individuals with at least two diagnoses of migraine (International Statistical Classification of Diseases and Related Health Problems, 10th revision [ICD-10] G43) during the study period were assigned to the migraine cohort. Next, patients with medical records of migraine from January 1, 2002, to December 31, 2003 (a two-year washout period), were excluded. The migraine cohort was further subdivided into the MA group (ICD-10 code G43.0), MO group (ICD-10 code G43.1), and unspecified group (none of the above).

The primary outcome was the incidence of all-cause dementia. The secondary outcomes were each cause of dementia, including AD (ICD-10 codes G30 or F00), VaD (ICD-10 code F01), other specified dementias (ICD-10 codes F02, G31.00, G31.82), and unspecified dementia (ICD-10 code F03). Individuals were classified as dementia patients if they had at least two ambulatory visits or one hospital admission for dementia and did not have any medical records of dementia before December 31, 2003.

This study tried to mimic the prospective study design and overcome the inherent limitations of the retrospectively constructed NHIS-HEALS through the risk-set matching method using time-dependent propensity score [19, 20]. The control group was selected from individuals at risk of migraine, considering other potential confounders such as age, sex, socioeconomic status, comorbidities and lifestyle factors.

First, we calculated the hazard component for being patients with migraine represented as a time-dependent propensity score, using a Cox proportional-hazard model with January 1, 2004 as baseline (after the 2002–2003 washout period) and migraine diagnosis as an event [21]. The baseline characteristics of the participants used for propensity score matching were collected over two years before the baseline (2002–2003). Age (continuous variable), sex, household income level (medical aid beneficiary or decile for NHIS enrollees), residential area (urban or rural), registered disability, past medical history (stroke, diabetes mellitus, hypertension, and depression; individuals with more than two outpatient visits or one admission based on ICD-10 codes), smoking status (never smoker, ex-smoker, current smoker, or unspecified), BMI (< 25 or ≥ 25), and alcohol consumption (none, ≤ 7, 7–14, or > 14 units per week) were included as covariates.

Second, each patient with migraine was matched to individuals of the same age, sex, baseline household income level, and at risk of migraine at the index date (the date of criteria for migraine cohort were fulfilled). The index dates of the control individuals were set the same as the index dates of their matched patients with migraine. We repeated this risk-set matching sequentially for all migraine patients [22]. To mimic the design of a prospective study, risk-set matching was performed independently of the future diagnosis of migraine. In other words, because the control individuals were those who had not yet developed migraine (at risk of migraine) at the time of matching, they had the possibility of developing migraine during follow-up. Therefore, most individuals who had been assigned to the migraine cohort before matching were included as patients with migraine in the main analysis, but a minority were included as controls for other patients who developed migraine before them.

Subsequently, a 1:1 matching on time-dependent propensity score was performed for each risk set using a nearest-neighbor matching algorithm with a maximum difference of hazard components between patients with migraine and control individuals of < 0.1 [19, 23]. The matched patient-control set was removed from the subsequent risk set to generate a non-overlapping sample. The 1:1 propensity score matching was repeated until no more patients with migraine remained in the risk set.

The balance of baseline characteristics between the migraine and control groups was assessed with a standardized difference; if the absolute value of the standardized difference was less than 0.1, the distribution of covariates was considered balanced [24]. We used the Kaplan–Meier method and stratified log-rank test to evaluate the cumulative incidence of dementia in patients with migraine and matched controls. We calculated the incidence rate (IR, the number of dementia cases per 10,000 person-years) of dementia and the 95% confidence interval (CI) using a generalized estimating equation with a Poisson distribution. The effect size was estimated as the hazard ratio (HR) using the Cox proportional hazards model. Subgroup analyses according to age at migraine diagnosis, follow-up duration, sex, and presence of aura were also performed. In some cases, the first date of migraine diagnosis during study period might be indicative of an active period of migraine (instead of new migraine onset), particularly among patients who were initially diagnosed with migraine at an older age. Therefore, we performed additional sensitivity analyses that investigated the association between migraine and dementia (all-cause, AD, VaD, mixed or other specified, and unspecified dementia) among individuals with first migraine diagnosis at age < 60 years. Moreover, same analyses were performed with five years of washout period (instead of two years) to exclude more active period of previously diagnosed migraine. Statistical significance was defined as a two-tailed p-value of < 0.05. All analyses were performed using the SAS Enterprise Guide software (version 7.1; SAS Institute, Cary, NC, USA) and R (version 4.1.3; Vienna, Austria; Rproject.org/).