Heterosexual AI is common among young people worldwide, although patterns appear to vary substantially both within and between groups and regions. While it is clear that many young people experience AI, it is unclear how regularly it is practised. The available data suggest that condoms are used less frequently during heterosexual AI than during VI.

Lifetime AI prevalence increases with age among all young people (including some sexually inactive), but not among the sexually active, which may suggest that those who are sexually active at younger ages (<16 years) engage disproportionally in AI (Fig. S1b). This finding is corroborated by a study in Zambia, which found that AI was the first sex act of 9 % of primary school girls, and 0 % of secondary school girls [50].

AI prevalence did not vary by recall period, which may indicate that individuals who initiate AI continue to practise it. Alternatively, differences may be obscured by reporting bias, with more accurate reporting of behaviours over shorter recall periods. The latter conjecture is supported by a meta-analysis examining reported sexual behaviour over different recall periods, which found greater accuracy in reporting of AI over 30 days compared to 6 months [51].

AI is a highly stigmatised behaviour in many populations and thus its reporting is likely subject to social desirability bias. Therefore, it may be more accurately reported using more confidential interviewing methods [52]. Our review found significantly higher prevalence reported using ACASI, followed by SAQ and FTFI, although as articles using ACASI tended to be more recent (all after 2002); this finding may be confounded by increasing AI prevalence over time. Studies conducted in South Africa provide a good illustration of the substantial heterogeneity found in reported AI prevalence, some of which is likely a result of bias in reporting this stigmatised behaviour. In Cape Town alone, the two estimates of lifetime AI prevalence among sexually active school students (14–15 years) vary widely: from 56 % in one study of randomly selected young people throughout the city using ACASI [53], to 6 and 15 % using SAQ and ACASI methods respectively, in a smaller study of a single school year of the same age [54]. Studies employing FTFI reported the lowest lifetime AI prevalence in the country, with a national survey reporting 5 % among the sexually active [55], while two vaginal microbicide trials found <2 % prevalence [56, 57].

Only one study in our review directly compared AI prevalence using different interview methods, but their findings of higher reported prevalence using more confidential methods are supported by other studies [54]. For example, 3.5 % of married men in Cotonou, Benin reported lifetime AI in a FTFI, but 17.5 % using the more anonymous polling booth survey (PBS) method [58]. Discrepancies in reporting between more and less confidential interview methods imply that effort should be made to develop and utilise more reliable tools to gather data on stigmatised behaviours.

There is a popular opinion that heterosexual AI is on the increase [59]. Anecdotally, general practitioners at US universities have reported an increasing number of female students presenting with anal fissures caused by AI [60]. Some authors have linked recorded increases in AI practice to increased exposure to pornography at young ages, arguing that it causes a de-stigmatisation of anal sexual behaviour [28, 61]. Higher AI prevalence has been found among Swedish and US adolescents exposed to online pornography [62, 63]. Participants in a qualitative study on AI among 16–18 year olds in England frequently cited pornography as a main reason for young people practising AI, although the authors argue that this explanation is simplistic [64].

We found some evidence in this review to support the argument that AI prevalence is increasing, but it is difficult to separate an actual change in prevalence from a possible lessening in social stigma and thus a reduction in social desirability bias. Although our meta-analysis found a significant increase in AI prevalence over time only in males and among European youth, an increase was reported by series cross-sectional studies. Prevalence among Swedish female university students was found to increase by 12.1 percentage points over 10 years, and national surveys from the US and Croatia reported increases of 2.2 % points over 4 years and 8.3 % points over 5 years respectively among sexually active females, with similar increases among males [28, 33, 34, 61]. This discrepancy between our meta-analysis findings and the findings of the series cross-sectional studies may be explained by the comparatively greater diversity in study populations and survey methods seen across the articles in this review, introducing greater heterogeneity and making it more difficult to conclusively identify trends.

This study has a number of limitations. We searched for published studies through established databases and through reference scanning and, did not include non-English language articles, and thus may have missed some eligible articles. This criteria, however, is unlikely to have influenced results much given the large number of articles included and the small number of eligible articles that were excluded on the basis of language (N = 3). Where the survey year and mean age of the sample was not reported and attempts to contact authors were unsuccessful, we approximated it from available information in order to carry out the analysis. Our use of mean age, rather than maximum age as the upper cut-off, meant that small numbers of older adults are also included in some of the articles in this review, particularly from samples of university students. However, given that lifetime prevalence among the sexually active did not differ significantly by study sample (data not shown) or by age, it is unlikely that this has affected our findings. As a significant amount of heterogeneity remains unexplained, it is possible that we may have failed to identify possible explanatory variables due to inconsistency of reporting.

Other than the previously discussed social-desirability bias, other biases could have affected the results of this meta-analysis. Selection bias may have been introduced if study populations were chosen a priori for their perceived higher risk. Our use of engagement in VI as the definition of sexual activity may mask the practice of AI by those who hadn’t initiated VI, however this may be small since two US studies indicate may be 1 % and a study in Zambia with small sample size indicates may be approximately 4 % [50, 65, 66].

Many articles reported incompletely on sexual behaviour, which in turn limited the scope of this review. Data on males from several articles were excluded for failing to report homosexual and heterosexual AI separately, while other articles were excluded for compiling AI practice together with other sexual activities. Of the 136 included articles, 30 failed to report separately by gender and 15 failed to report the recall period of AI prevalence. The dearth of data from Asia and Latin America hindered examination of trends by continent. Most of the included articles had small sample sizes, with a paucity of data from larger population-based studies. Our estimates for AI frequency are based only on the handful of articles which reported it. We focused our analysis on lifetime prevalence of AI as this was overwhelmingly the most common recall period. Shorter recall periods are, however, more epidemiologically relevant and useful.

This review has a number of strengths and makes a valuable contribution to understanding this neglected sexual risk behaviour. We have included a large number of studies, including also those which did not report AI data in the abstract, thus minimising reporting bias. Had we searched for and included only articles which referred to AI in the title or abstract our summary estimates would likely have been higher (Fig. S3).