This meta-analysis was registered in PROSPERO (ID: CRD4202365834). Databases including PubMed, Web of Science, Embase, and Cochrane Library were searched until October 1, 2022. The searches were conducted using the following medical subject heading (MESH) terms: “education,” “oral cavity,” and “pharyngeal cancer;” these terms and their variants were linked using Boolean operators.

Studies meeting the following criteria were included: (1) focus on education and OCPC (including broader definitions such as head and neck cancer) that reported odds ratios (ORs), with 95% confidence intervals (95% CIs), or adequate data to calculate these measures, (2) inclusion of participants aged 18 years or older, (3) availability of the full text of the study and published in English, and (4) sample size of ≥200. Review articles, meta-analyses, case reports, and editorials were excluded. Two researchers (GC and JX) independently assessed the full texts after screening all the titles and abstracts.

Data on the full title of a study, first author, year of publication, study design, follow-up period, sources of cases and participants, number of cases and controls, and ORs were independently extracted by two investigators (GC and JX). Since the studies included education levels assessed using different methods and scales, we redefined education level to facilitate data collection and comparison. Therefore, for each study, we defined the lowest category reported in the study as the lower education level and the highest category as the higher education level. The Newcastle Ottawa Scale (NOS) tool was used to evaluate the quality of each study (low, medium, and high quality). Any disagreements between the researchers were resolved through consensus.

Statistical analyses were conducted using Stata 14.0. I² analysis was used to assess heterogeneity, and an I²-value >50% indicated heterogeneity between the studies. Data synthesis based on different research populations and methods was conducted using a random-effects model. Subgroup and meta-regression analyses were performed. The robustness of the literature results was evaluated by sensitivity analysis. Funnel plots derived from the Egger and Berger bias tests were used to check for publication bias.

A genome-wide association study (GWAS) dataset, which included 766,345 individuals of European descent who had received an education, was selected as the exposure set (Supplementary Table 1) (16). These individuals were aged older than 30 years, and EA was measured as the number of completed schooling years. Years of schooling were converted and standardized, where each unit represented 4.2 years of schooling (1 standard deviation [SD] = 4.2 years). Based on published MR studies (17, 18), a total of 268 single nucleotide polymorphisms (SNPs) were selected in our study for MR analysis (Supplementary Table 2). According to previous studies, these SNPs accounted for 12% of the variation in EA among individuals (16). The tools used in these studies can be considered to have a strong predictive effect on education level since the F-statistic was 41.76 (19).

The Integrative Epidemiology Unit GWAS database, which included 6,034 cases of OCPC and 6,585 controls, was selected as the outcome set (Supplementary Table 1) (20). The cases included were in line with the International Classification of Diseases 10 (ICD-10) standard. We focused on participants of European descent, including patients with OCPC, oral cavity cancer (OCC), and OPC.

To explore the potential mediators of the EA–OCPC pathway, we used the inverse-variance weighted (IVW) method to assess the potential causal relationship between EA and common risk factors for OCPC. Based on the current database, we analyzed the following common risk factors of OCPC, including HPV, number of sexual partners in lifetime, hypertension, type 2 diabetes, smoking, and alcohol consumption. Supplementary Table 1 presents the GWAS summary data of the above risk factors. Multivariable MR (MVMR) was mainly used to evaluate the impact of multiple potential exposures on the results as a whole and to identify potential risk factors.

According to the guidelines for MR, several MR approaches, including the IVW method, weighted median method, MR-Egger method, simple mode, and weighted mode, were used to estimate the relationship between EA and OCPC. The IVW method is primarily used to evaluate causal relationships since it possesses the highest statistical power. ORs were transformed with effect estimates (equivalent to beta coefficients), and the results were represented with 95% CIs. Since many instrumental variables (IVs) were associated with multiple traits (pleiotropy), it was necessary to apply sensitivity analysis to the results. Therefore, we applied MR-Egger regression and the leave-one-out sensitivity test. If the regression intercept was close to 0, it indicated the absence of horizontal pleiotropy (21). The leave-one-out sensitivity test was mainly used to calculate the MR results of the remaining IVs after removing the IVs one by one. No difference between the estimated MR result and the result after removing an IV indicated that the MR result was robust. All MR analyses in this study were conducted using the TwoSampleMR package in R (22).

A total of 36 studies on OCPC (including 13 studies that delineated their focus on OCC) were finally included in the meta-analysis (Figure 1). All included studies had a case-control design, and their characteristics have been summarized (Supplementary Tables 3, 4). In terms of geographical location, six studies were conducted in Europe, 21 studies in the Americas, and nine in Asia. There were 67,326 cases (OCPC) and 37,903 controls (non-OCPC) in the 36 studies. NOS assessment revealed a moderate risk of bias.

The meta-analysis of the studies on the relationship between EA and OCPC revealed significant heterogeneity (I² = 92.7%, P < 0.001). Random-effects model analysis showed a significant negative association between EA and OCPC, and the pooled OR was 0.439 (95% CI: 0.383–0.503, P < 0.001). The forest plots showed 36 studies (Figure 2). In the subgroup analysis, no heterogeneity was found in region, quality, year of publication, or sample size (Supplementary Table 5). The meta-regression analysis showed no significant heterogeneity in the year of publication, sample size, and quality score (Supplementary Table 6). The combined results of the remaining studies did not change significantly after the sequential elimination of each study in sensitivity analysis. The funnel plots, Egger's test results (P = 0.775), and Berger's test results (P = 0.505) confirmed that there was no publication bias (Supplementary Figure 1). We further explored the relationship between education and OCC (including 13 studies that clearly defined OCC): our random-effects model analysis showed a significant negative association between EA and OCC (pooled OR = 0.425, 95% CI: 0.345–0.549; P < 0.001) (Supplementary Figure 2). Subgroup analysis indicated that the existing heterogeneity may have been caused by region and quality (Supplementary Table 7). Further meta-regression analysis showed no significant heterogeneity in year of publication, sample size, or quality (Supplementary Table 8). After the sequential elimination of each study, the combined results of the remaining studies did not change significantly. The funnel plot, Egger's test results (P = 0.913), and Berger's test results (P = 0.054) showed no evidence of publication bias.

There was a significant causal effect of EA on OCPC in the IVW method (OR: 0.349, 95% CI: 0.222–0.548, P < 0.001). Additionally, a positive result was obtained by applying the weighted median method (OR: 0.303, 95% CI: 0.155–0.592, P < 0.001). Although no statistical significance was found, similar results were found with the MR-Egger, simple mode, and weighted mode methods (Table 1). The slope of the MR scatter diagram represents the impact of the exposure on the outcome. The MR fitting results showed that risks of OCPC, OCC, and OPC decreased with the increase in EA (Supplementary Figure 3). Since the intercept in MR-Egger regression was close to 0 (intercept β = 0.003; SE = 0.012, P = 0.800), it was unlikely to have horizontal pleiotropy (Supplementary Table 9). In addition, the forest plot showed that the MR fitting line was completely on the left side of 0, indicating that the increase in EA level reduced the risk of OCPC (Supplementary Figure 4). The funnel plots in this study were symmetrical; therefore, no pleiotropic effects were observed (Supplementary Figure 5). Sensitivity analysis yielded robust evidence, and the results were not affected by any SNP (Supplementary Figure 6).

Further exploration of the relationship between EA, OPC, and OCC showed that EA was significantly associated with OPC and OCC (Table 1). MR-Egger regression showed the absence of horizontal pleiotropy (Supplementary Table 9). Similarly, the funnel plot and leave-one-out sensitivity analysis showed that MR results were reliable.

Assessing the causal relationship between EA and the common risk factors of OCPC was conducive to exploring the interference factors that may mediate the association between EA and OCPC. The preliminary results showed that 4.2 years of additional education was significantly related to a reduced risk of common risk factors (including BMI, hypertension, type 2 diabetes, and smoking) (Table 2). To control the pleiotropic pathway, when we further applied the MVMR model after correcting for the above mediating factors, education level still had a protective effect against OCPC (OR: 0.361, 95% CI: 0.281–0.463, P < 0.001) (Table 3).