The systematic analysis and modelling study complied with the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guideline [10], and the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER) [11]. The study protocol was prospectively registered on PROSPERO (number CRD42020215140).

We searched five databases (PubMed, Web of Science, Embase, China National Knowledge Infrastructure, and Wanfang Database) without a lower time limit and until 23 June, 2021, to identify observational studies that investigated cervical cancer screening adherence among general women. A combination of search terms relating to cervical cancer screening and adherence was adopted, without language restriction. Additionally, we searched ProQuest theses database and Google Web for potentially eligible grey literature. Reference lists of related systematic reviews and the included articles were screened to identify additional relevant studies. Full details of the search strategies are listed in the Appendix (Table S1).

Studies were retained if they reported data on adherence to cervical cancer screening, including sample size, the number of adherent subjects, and/or adherence rate. Studies that were conducted in hospitals or in special population (e.g. HIV-infected women) were excluded. Reviews, conference abstracts, commentaries, and case series were also excluded.

Against the selection criteria, two authors (WZ and KG) independently reviewed all titles, abstracts, and full-text articles that were potentially relevant. Inconsistencies during the review process were resolved through discussion with a third author if necessary (MJ).

The following data were independently extracted from the included articles by two authors (WZ and KG): author, year published, year investigated, country, region (WHO, World Bank [WB] region), place (rural, urban, both), Socio-demographic Index (SDI), data source (self-reporting, objective-recording), screening guideline information (guideline name, definition of adherence, screening methods and intervals), sample size, the number of participants who were adherent to cervical cancer screening guideline, age and education of participants. As cervical cancer screening strategies and programs were generally adapted and implemented by the local health care system, in full considerations of country- and area-specific contexts (e.g., differences in cervical cancer prevalence, economical level, social and cultural factors), it was anticipated that the recommended methods and intervals would not be completely consistent across studies. Therefore, we defined the adherence rate of cervical cancer screening as the proportion of participants who were adherent to local guidelines. The geographical regions of study were designated as African Region (AFR), Region of the Americas (AMR), Southeast Asia Region (SEAR), European Region (EUR), Eastern Mediterranean Region (EMR), and Western Pacific Region (WPR) using the WHO categorization, and the regions were also categorized into HICs and LMICs according to the WB criteria. Furthermore, SDI was also obtained from Global Burden of Disease Study 2019, which is a composite indicator of total fertility rate under the age of 25 years, mean education for those aged 15 years and older, and lag distributed income per capita, reflecting the overall socioeconomic-development of a country [12]. The data sources were defined as self-reporting (by survey to subjects) and objective-recording (by checking the records in health care systems and hospitals, etc.). Wherever applicable, data on age-specific or investigation year-specific adherence rates were separately extracted. For studies in which the year investigated was not provided, we imputed the investigation year by subtracting five years from the year published (Table S2). For studies where reported censored age ranges, we imputed the missing age band by taking the same width reported in other age ranges on the same study.

For studies, where associated factors of adherence to cervical cancer screening were explored using multivariable logistic regressions, the definitions of each factor and the corresponding effect size estimates (odds ratio [OR] and 95% confidence interval [CI]) were extracted.

We evaluated the risk of bias of included studies using a quality scale based on the Strengthening the Reporting of Observational Studies in Epidemiology statement (STROBE) [13]. This scale assesses sources of bias from five dimensions: sample population, sample size, participation, outcome assessment, and analytical methods (Table S3). The total score ranges from 0 to 10, and studies with a score greater than 6 points were considered high quality. Any discrepancies were resolved by consensus through group discussion.

For meta-analysis, the effects of associated factors on adherence to cervical cancer screening were synthesized by random-effects (DerSimonian) model [14]. Only factors that shared similar definitions and had at least three individual multivariate-adjusted effect sizes from different studies were included [15].

For modelling analysis, one study might provide two or more data points about cervical cancer screening adherence in different age groups or investigation years. To make full use of this information, we used a multilevel mixed-effects meta-regression approach. To take the country-specific context of screening guidelines and policies into account, the country identification number (μi) was employed as the random effect in models. Given that adherence rate (p) = number of participants adherent to guideline / number of participants, we stabilized the rate with the logit link as follows:

$$\textrm(p)=\ln \left(\frac\right)=\ln (odds)=\alpha +_1_1+_2_2+\cdots +_i$$

Where α is the intercept term, β is the coefficient, and x is the variable.

The effects of cluster-level variables, including age, place, year investigated, SDI (per 1%), and WHO region, were assessed by univariable and multivariable meta-regression (Table S4). Age, SDI, and year investigated were revealed to be significantly associated with cervical screening adherence based on those analyses. We only included age and SDI in our final model since there was considerable collinearity between SDI and year investigated, and SDI may have a broader interpretation on social health status than chronological years. Therefore,

$$\textrm(p)=\alpha +_1\ast \textrm+_2\ast \textrm+_i$$

then,

$$p=\frac_1\ast \textrm+_2\ast \textrm+_i\right)}_1\ast \textrm+_2\ast \textrm+_i\right)}$$

We only estimated the adherence rates for women aged 20–69 years, considering this is a proper age range for cervical cancer screening. To address both the features of geography and income, we classified the world into ten regions, namely AMR (HIC), EUR (HIC), EMR (HIC), WPR (HIC), AFR (LMIC), AMR (LMIC), SEAR (LMIC), EUR (LMIC), EMR (LMIC) and WPR (LMIC). The overall and age-specific adherence rates for these ten regions were generated for the year 2019 based on the above model. The SDI value for each region was calculated as the weighted-average of country-specific SDI in this region, and the weight was population size in 2019.

For meta-analysis of national adherence rate, we pooled the rates using random-effects model for countries with adequate adherence data.

A two-sided P value of less than 0.05 was considered statistically significant. All analyses were performed with R version 4.0.2 (R Foundation for Statistical Computing) and Stata version 14.0 (College Station, TX, USA).