General comments
This review has many of the properties of a systematic review (e.g., specification of search terms and databases, evaluation of heterogeneity), and even of a meta-analysis (summary measures), without conforming fully with accepted methodology for the former (see Methods: Literature search and data extraction). Such reviews, following published guidelines, are usually made to evaluate effects of interventions. While many of the same principles remain relevant to other types of data, guidance and reporting standards for systematic reviews of prevalences are lacking [14]. Hence, we conceived this as a narrative rather than a systematic review of publications reporting headache prevalences, updating our previous review of 2007 [5].
In comparison with our previous review, we found an apparent increase in prevalence of migraine but not those of other headache types. We also found wide geographical variations, as did the 2007 review. Estimates for migraine were very high in populations with very good knowledge and interest in the disorder (e.g., neurologists, and other health personnel). While estimates were influenced to varying degrees by several methodological factors (nature of the screening question, number of conditions investigated, sampling method, number of participants, how patients were engaged and not least – for migraine and TTH – how ICHD criteria were applied with regard to definite and/or probable diagnoses), many of these influences were not significant. Since the review included studies of all populations of interest other than clinical populations, and extracted many data on methodological and other aspects of each study, we believe it provides a strong basis for examination of these influences, which we discuss in the following sections. Meanwhile we note that our prevalence estimates in this review derive from studies varying with regard to these factors, so that it remains in doubt whether the observed differences over time and between geographical regions are real.
Estimated medians for the different headache types were approximately 20% lower than the means, whereas, for headache overall, the median estimate was 3% higher. It could be argued that medians, uninfluenced by outliers with very high prevalences, were more correct and conservative summary measures. On the other hand, we had no a priori reason to believe that studies with high estimates were less correct or representative for the population at large than those with low estimates. In many ways, studies finding high prevalences might have been methodologically better (neutral screening question, considering both probable and definite diagnoses, face-to face interviews, etc). Therefore, we based estimates on the means.
The sensitivity analysis showed that the bivariate analyses were little influenced by statistical method (parametric or non-parametric).
Relation to case definition: “active headache disorder”
Case definition is of over-riding importance, and likely to be the single most influential factor in any enquiry dependent on prevalence estimation (including burden estimation) [4, 15]. For headache disorders, case definition revolves around timeframe, diagnostic criteria and how the latter are applied. Since 1988, use of ICHD criteria has become mandatory in the absence of acceptable alternatives [4], but, from a pragmatic perspective, some flexibility in their application to epidemiological enquiry is invariably necessary. Empirically, we found different approaches towards this flexibility among the reviewed studies, and these were not always described.
As to timeframe, we might have restricted this review to studies reporting 1-year prevalence, but we believe the ICHD definition of “active headache disorder” [9] does not exclude studies reporting “current”, 6- or 3-month prevalences (see Methods: Timeframe of headache). These latter timeframes were more restrictive, and therefore given to underestimation, as appears from Tables 1 and 3 (although only of episodic headache, by exclusion of those with low-frequency episodes). On the other hand, those reporting lifetime prevalences would tend to give higher estimates through inclusion of “non-active” headache. Those in which timeframe was unknown (not specified) might go either way. Overall, our estimates may be somewhat conservative because we subsumed shorter timeframes into our case definition of “active headache disorder”, but this approach is validated by the fact that, for all timeframes subsumed (1-year, 6-month, 3-month, “current” and unknown), 95% CIs were widely overlapping (except for H15+, with only 3 studies).
As to ICHD criteria and their application, we might also have restricted our case definitions for migraine and TTH to definite diagnoses (meeting all ICHD criteria). This would have had a large influence: probable diagnoses accounted for > 40% of total migraine and > 30% of total TTH. The persuasive arguments for including probable diagnoses have been set out before [4]. In clinic, probable diagnoses are, or should be, confirmed or refuted during follow-up, and have utility in allowing management plans to be set meanwhile. This utility is lacking in epidemiological studies, with subsequent enquiry rarely possible, but on the one hand excluding probable cases would clearly provide a partial account of headache prevalence while, on the other, probable cases are probably what they appear to be [4]. Most studies did not explicitly state how ICHD criteria were applied, and, while some gave estimates for both definite and probable cases, a few reported only the sum of these. We would, therefore, have excluded the majority of studies had we limited the review in this way, while grossly underestimating true prevalences, as evidenced for migraine in particular in Tables 4 and 6. In the MLR analyses (Table 7), this was the factor with the highest SPCsqr, alone explaining 6.4% (Model 1) to 7.8% (Model 2) of the variation in migraine prevalence.
In summary, with regard to case definition, we believe that best estimates from available data are made by including studies with various timeframes, and including definite and probable (or unspecified) diagnoses. These inclusions allow many more studies to be taken into account, providing data from countries and regions that otherwise would have few or none. This is not to say that future studies should not give due regard to these very important methodological issues.
Relation to publication year and geography
Unlike case definition, there were no obvious a priori reasons for expecting these factors to be significantly influential. Although geographical variation of course implies genetic, environmental, cultural, economic, lifestyle and general health variations, among these only dwelling altitude has clearly demonstrated impact, and this only in Nepal [16]. Urban/rural divide and relative wealth or poverty are weakly related to headache prevalence in some studies, not always in the same direction. Population age may be a factor: age itself is strongly associated with prevalence of all headache types, and some national populations are relatively young in comparison with the global mean (but, it should be noted, this factor is not independent of population poverty or general health).
Compared to our previous summary of prevalence data from 2007 [5], this review has found increased prevalences of active headache (from 46% to 52%), of migraine (from 11% to 14%) and of H15+ (from 3% to 4.6%) but decreased prevalence of TTH (from 42% to 26%). However, in both bivariate and MLR analyses, association between prevalence and publication year was significant only for migraine. An increase in prevalence was also found in some [17, 18] but not all [19, 20] studies performed in the same populations at different time points. As to geography, mean reported prevalences of all headache types varied considerably, albeit with wide and mostly overlapping CIs. Compared with HI superregion, lower prevalences were estimated for all headache and migraine in SSA and SEEAO, and for TTH in NAME.
However, it is uncertain whether or to what extent these differences over time and place are real: overall, the MLR analyses show that the present models explain relatively little of the large variations in prevalence estimates between studies (for migraine less than 30%, and even less for other headache types, possibly because of fewer studies). There are many other aspects of methodology (see below), some that are difficult to identify and impossible to quantify but which are likely or certain to have varied between studies in different times and places, and perhaps been influential. Standardisation according to accepted guidelines will help in future studies [4], but these factors, and the uncertainties they generate, will persist in the large corpus of historical studies from mostly HI countries.
The relatively few studies from the other GBD superregions are vulnerable to incidental findings. Therefore, with the large and unexplained variations, it may be risky (less accurate) to adjust the estimate of global prevalence for proportions of the global population living in each region. For example, SA and SEEAO have half the world’s population but only 9% (17/180) of headache prevalence studies. Better estimates of global prevalences may be obtained by considering all studies as a way of cluster sampling the world, although adjustment could and should still be made for the huge oversampling from HI countries. Arguably, the rich variety of methods used make the overall estimates more resistant to yet unknown sources of variation.
Relation to population of interest
This review included studies of all populations of interest other than clinical (patient) populations. In contrast, the review from 2007 [5] included only studies with > 500 participants on whole populations or representative samples of these within specified age ranges in communities, towns or countries, together with studies performed in schools. This means we took account here of studies on smaller and more selected populations (e.g., among employees of a company or factory, hospital staff, university students, neurologists, ethnic minorities, inhabitants of a slum area, a tribe, a monastery, etc). Such populations were likely to differ from the general population in the country in question (i.e., more or less healthy, poorer or richer, more or less educated, etc) but, by including these, we gathered data from more diverse settings and many more countries. Again, we believe this approach results in a more representative sample of the world’s population, while, with one exception (see below), neither bivariate nor MLR analyses found significant differences between estimates from selected or more general populations.
The exception was studies on health-care personnel (medical students, hospital staff, neurologists). These yielded significantly higher prevalence estimates for migraine in both bivariate and MLR analyses. The studies on neurologists are an interesting example: three studies from USA [21], France [22] and Norway [23], considering either 1-year prevalence or “current” headache (unknown timeframe), found a mean prevalence of 42.2%. It has been argued that this very high prevalence is a true finding because neurologists are expected to diagnose themselves with very high sensitivity and specificity [23]. This is not to say it reflects the true population prevalence: neurologists are in many ways a highly selected and unrepresentative group.
Relation to age and gender
This review mostly confirmed earlier studies on age and gender distributions (Tables 1, 4 and 5). Headache in general was most prevalent in the group with mid-range age values between 10 and 19 years. Overall, and in females, migraine was most prevalent between 20 and 64 years, whereas in males it was somewhat higher from 10 to 19, although with overlapping 95% CIs. TTH was most prevalent in the 20–64 years age group for both genders (except for the single study in those above 64 years), as was H15+. The age group 20–64 years is, of course, a very broad category, and this grouping only (and imprecisely) distinguishes adults from children and adolescents on the one hand and from the elderly on the other. However, the data did not allow a more refined analysis.
All headache types were more common in females, most markedly for migraine (17.0% versus 8.6%) and H15+ (6.0% versus 2.9%), the more disabling types, while the difference was small and not significant in TTH. Gender differences were least in the youngest age groups (Table 1).
The relationships between age and prevalences of dMig and pMig, and of dTTH and pTTH (Table 5), are worthy of comment. The proportion of pMig to total migraine was higher in the age group 10–19 years than in those who were older, and also tended to be higher among males. This suggests that ICHD migraine criteria are somewhat less fitting for children and adolescents than for adults, and perhaps for males than for females. The proportion of pTTH to total TTH was equal between genders, but again much higher in young people. In the latter group, a high prevalence has been noted in several studies of headaches unclassifiable within ICHD, meeting the criteria for neither definite nor probable migraine or TTH, and described as “undifferentiated headache” [24].
Comparisons with GBD2019
For migraine and TTH, our estimates of global prevalence and those of GBD2019 had widely overlapping CIs (Table 2), but for all headache our estimate (52.0% [48.9–55.4]) was markedly higher than that of GBD2019 (35.0% [32.3–37.7]). Possibly, this is because, in GBD, ‘“all headache’” is the sum of migraine and TTH, whereas our estimates are based on studies among which the majority asked about headache in general, therefore also including other headache types. However, it is puzzling that in GBD2019 [13] (and earlier versions [25, 26]) total headache prevalence is markedly lower than the sum of migraine and TTH prevalences. The explanation does not, apparently, lie in corrections for comorbidity of the two headache types.
GBD estimates are based on sophisticated mathematical modelling [2], which combines methods of meta-analysis with regression techniques (“meta-regression”). It not only adjusts for gender and age composition of populations, for comorbidity and for some methodological variables, but also takes account of multiple geographical levels (country, region, superregion, global) and year of study (“borrowing strength over space and time”) [27]. These are smoothing methods, whereby GBD is able to make estimates for all gender- and age-groups, for smaller regions, individual countries and even parts of countries. GBD estimates are more even across regions and time periods than ours: we can make reasonable estimates for global prevalences over a long time span but, for countries, smaller regions and defined time periods, in the absence of data specific to these, GBD methods achieve what we cannot. This is not a comment on which estimates might be truer in populations from whom data have been directly derived in one or more methodologically sound studies.
Influences of methodological factors
In the MLR analyses, publication year appeared important as a factor explaining variation in migraine prevalence estimates (6.4% of variation in Model 2, higher estimates associated with more recent publication), but it played no role in other headache types. The apparent increase in migraine prevalence over time may be real, perhaps related to environmental, physical, behavioural or psychological changes, but more probably it has to do with methodological developments over the years, leading to better techniques of access and engagement and improved diagnostic instruments, both likely to enhance case ascertainment. The positive correlations between prevalence estimates of all diagnoses indicate that case ascertainment is an important factor also for the other headache types, and for headache in general. Certainly they speak against a hypothesis that, if one diagnosis is made more often, it is at the expense of another. It may also be that better knowledge and greater awareness of a condition lower the thresholds for reporting relevant symptoms. That knowledge of migraine and its diagnosis can affect estimates is supported by the findings of higher migraine prevalence among medical personnel, although these are, largely, findings specific to the high end of the spectrum of knowledge. There has been no similar focus on other common headache types.
The negative association of prevalence estimates of all headache, migraine and H15+ with number of study participants (Model 2: 3.2%, 1.6% and 12.5% of variations respectively) may indicate that smaller studies can afford more sensitive methods (personal interview, face to face or by phone) to detect cases. Alternatively, it may be caused by greater selection bias in small studies. This would explain the absence of similar effect on estimates for TTH, since people whose lives are less impacted by a disease have less interest in surveys enquiring into it. That more random sampling – generally more methodologically rigorous – was negatively associated with migraine prevalence estimates may similarly be explained by reduced selection bias.
It seems obvious that studies including pMig would report markedly higher estimates of migraine prevalence, since > 40% of total migraine prevalence in these studies was attributed to pMig. Including pTTH did not significantly affect estimates of TTH, perhaps because cases that might have been classified as TTH met criteria for pMig, counterbalancing the otherwise expected increase. That inclusion of pTTH was associated with a higher prevalence of all headache (Model 2) suggests this captured cases, probably of mild headache, that were otherwise overlooked.
It also seems obvious that a screening question setting some threshold of severity, and therefore excluding milder cases, would lead to lower prevalence estimates of all headache (9.0% and 7.2% of variation) and of TTH, a less severe headache type (8.0% and 5.4% of variation), but not of migraine or H15 + .
Lifetime and 1-day prevalences
Lifetime prevalences are higher than those of active headache disorders, except for H15+, but information about headache earlier in life is not relevant to estimations of population burden deriving from active headache. In addition, recall error is probably a bigger problem when asking people about their whole lifetimes rather than recent time periods.
The opposites apply to 1-day prevalence. Enquiry into headache yesterday may almost eliminate recall error. Although 1-day prevalence is substantially lower than 1-year prevalence, at least for episodic headache, it offers a very sound basis for measuring population burden, provided that samples are large enough. However, questions about headache yesterday provide reliable data only in “on-the-spot” interviews (face-to-face or by telephone): in internet surveys or those relying on mailed questionnaires, participants with headache on the day they receive it may postpone answering it until the first day without, then truthfully but spuriously report headache yesterday. The relatively few studies reporting 1-day prevalence suggest that 15.8% of the world’s population have headache on any day, and almost half of them migraine (7.0%).
Study quality
We found no significant association between study quality score [4] and prevalence. This does not mean that the quality criteria are meaningless: higher quality lends credibility to findings. While the criteria may need refinement, they are based on sound general epidemiological principles, to which all future studies should endeavour to adhere. However, higher scores on one criterion may tend to increase prevalence estimates (e.g., distinction between definite and probable migraine), and higher scores on another have the opposite effect (e.g., random sampling), with the net result being very small. This probably explains why quality score explains so little of the variability in our models (Table 7).
Burden estimates
While prevalence is a determining component of disease-attributed burden, it provides no estimate of it in the population. Headache-attributed burden not only has multiple and diverse components [15] but also is very unevenly distributed in populations with current headache, being quite low in the relatively large proportions who have infrequent attacks of mild or moderate intensity [28]. Of the two further factors that must be known to estimate burden – the level of burden associated with the ictal (symptomatic) state and the proportion of time spent in this state – both can be assessed as population averages. The former may be a constant at population level for each headache type (the view taken in GBD studies, with “disability weights” assessing lost health), so that a sound estimate of population burden for all headache disorders may be based on the mean Time in Ictal State (TIS) (or Time in Symptomatic State in the GBD studies) [29]. Estimates of disease-attributed burden based on mean TIS may be relatively insensitive to estimated prevalence if higher estimates of the latter are generated by more searching enquiry, identifying participants with very low TIS. But all studies are likely to capture those with high burden, so that burden estimates are highly sensitive to participation bias.
It is worth noting that most cases of pMig fail the attack duration criterion (< 4 h for adults, < 2 h for children) [30,31,32]. Since duration is a factor in TIS, inclusion or not of probable diagnoses has rather less impact on burden estimates than on prevalence estimates.
Study limitations
Including not only strictly population-based prevalence studies but also studies on all other non-clinical samples made a rules-based literature search strategy difficult. This might have introduced some subjectivity and unknown study-selection biases.
The fact that dependent variables were not normally distributed might have rendered the MLR analysis less reliable. However, we believe that, by also presenting a model with normalized variables, we have generated results that for explorative purposes are robust, although the exact size of the contribution of each variable remains uncertain.
The fact that < 30% of the variability could be explained by the variables included in this analysis (time and place of studies, age and gender of populations, and methodological variables) indicates that there are other influential variables that are not yet understood or measured. We can only speculate as to what these may be: population differences in genes, general health, quality of health services for headache, exposure to climatic factors, light or altitude, nutritional factors, chemicals and air pollution, level of stress, attitudes towards pain and knowledge about headache in the population. Even differences in connotations in the words used for headache or migraine may play a role.