Two investigators (W.R., L.C.) independently searched, PubMed, Web of Science from their inceptions (1966 and 1947, respectively) to June 1, 2020, for cohort studies related to family SEP in childhood and lung cancer in adulthood, without language restrictions. Keywords used were “socioeconomic position (SEP)”, “cohort”, “lung cancer”, “risk”, as well as their Medical Subject Headings (MeSH) terms. Only papers published in English were included. Details of the search terms and inclusion/exclusion criteria are shown in Figure 1.

The inclusion criteria were: (i) cohort study design; (ii) all cases were lung cancer patients, and data were classified according to different SEP, regardless of subtypes of lung cancer; (iii) In addition, the exclusion criteria were: (i) small sample sizes; (ii) non-English publications; (iii) lack of data and necessary information for meta analysis or dose response analysis.

All the included studies were divided into two groups: (I) studies reported adjusted estimates of the hazard ratios (HR) (including relative risk and odds ratios) and 95% confidence interval (95% CI); (II) studies missed data mentioned above. We also included studies that graded data based on parents' financial status and family life in childhood, and overall impact. The quintile or quartile method was incorporated and the first level was used as the reference.

Data were extracted by two investigators (W.R., H.Z.) independently, and any disagreements came to consensus after discussion. Basic data were recorded from all eligible studies, including the first author's name, publication year, country, study period, period of follow-up, number of lung cancer patients and HR with their 95% Cl. The results were reviewed by two senior investigators (L.W., H.J.).

The quality of the studies was evaluated using a score system that was designed with reference to the Newcastle-Ottawa Scale (NOS) tool (38–40). The system is based on a 0-9 points, with 9 reflecting the highest quality and 0 the lowest. Each point was allocated for the following: (I) representativeness of exposure arm(s), (II) selection of the comparative arm(s), (III) origin of exposure source, (IV) demonstration that outcome of interest was not present at start of study, (V) studies controlling the most important factors, (VI) studies controlling the other main factors, (VII) assessment of outcome with independency, (VIII) adequacy of follow-up length (to assess outcome), (IX) lost to follow-up acceptable (< 10% and reported).

Two researchers (W.R. and G.F.) independently evaluated the methodological quality of each included published cohort study. The results of the quality assessment were used for descriptive purposes, to provide an overall assessment of the quality of the included studies (38, 41–44).

A meta-analysis was performed following the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines and the study protocol of this study is 1,59,082 (45, 46). We collected study-specific HR with 95% Cl for lung cancer to combine the data. Through Cochran's I² statistic, we examined the heterogeneity across studies and statistical heterogeneity was considered if an I² statistic ≥ 50% (47). A random effect model was employed to synthesize hazard ratios (HRs) for lung cancer-specific mortality if high heterogeneity existed (p < 0.5, I² > 50%), otherwise a fixed effect model was conducted.

We utilized Funnel plot tests, Begg's test and Egger's test to evaluate the publication bias (48–50). In addition, subgroup analysis was performed according to gender. Sensitivity analysis was conducted by excluding each study in turn to access the stability of results and potential sources of heterogeneity. All statistical manipulation was employed by Stata software (version 12, StataCorp, TX, USA). All P-values were 2-tailed; statistical significance was set at P-value < 0.05 (51).

The dose-response study was analyzed by Benchmark Dose Software 3.1.2(BMDS 3.1.2) (52). All steps were established under the BMDS guidelines (53–56). Multiple models were selected for analysis, including extra risk assumptions for background and a benchmark response of 5% (39, 40). According to the included studies, the dose response was divided into two categories: family economic status in childhood and childhood housing conditions. At the same time, according to the original data with or without adjustment, it was divided into adjusted and unadjusted group for subgroup analysis and publication bias analysis. In all dose-response models, we used Exponential models and Hill models as the main dose-response observation results, (57) while using other results for verification. Duplicate studies were excluded.

The MR method was based on the following three assumptions: (i) the instrumental variables are strongly associated with the family SEP in childhood; (ii) the instrumental variables affect cancer only through their effect on family SEP in childhood and not through any alternative causal pathway; and (iii) the instrumental variables are independent of any confounders (58, 59). To avoid the potential violation of the first assumption, we selected SNPs that meet the threshold of genome-wide significance (P < 5 × 10⁻⁸). Besides, we applied some methods including weighted median and MR–Egger to test the second assumption. Then we evaluated the directional pleiotropy based on the intercept obtained from the MR Egger analysis to satisfy the third assumption (60). We also performed a leave-one-out analysis in which we sequentially omitted one SNP at a time, to evaluate whether the MR estimate was driven or biased by a single SNP.

The analysis was conducted to estimate the effect of family SEP in childhood (X) on the risk of lung cancer (Y) using genetic variants (g), and the causal estimate is equal to Yg/Xg (61). For the association between genetic variants and family SEP in childhood (Xg), summary data were utilized from published Genome-Wide Association Studies (GWASs), (62–77) including Social Science Genetic Association Consortium (SSGAC) (1,060,068 individuals), MRC Integrative Epidemiology Unit (MRC-IEU) (2,49,790 individuals), UK Biobank (75,244 individuals) and Neale Lab (4,55,571 individuals) (Supplementary Table 1) (78–80). Summary statistics for the association between genetic variants and lung cancer (Yg) are from the International Lung Cancer Consortium (ILCCO) (27,209 individuals) (Supplementary Table 2) (69).

We selected uncorrelated variants to construct the instrumental variables in the two-sample MR analysis. Using both summary statistics for Yg and Xg, an Inverse Variance-Weighted (IVW) meta-analysis was employed to estimate the effect of genetically determined family SEP in childhood on the risk of lung cancer using the method of Burgess et al. (81):

where Xg is the beta estimate for the association between the SNP and family SEP in childhood, Yg is the beta estimate for the association between the SNP and lung cancer, and σYg is the standard error for Yg. Corresponding HR and 95% CIs were calculated using β^IVW and se(β^IVW).

In addition, MR-Egger and Weighted Median were also conducted to identify the causality. Ward ratio would only be observed when the first three models cannot be used due to lack of SNPs (82). Leave-one-out analysis was conducted to estimate whether the result was driven by a single SNP. MR-Egger regression was also performed to access the pleiotropy.

A total of eight cohort studies consisting of 2,779,242 cases were included in the final meta-analysis, (16, 17, 20, 23–28) and seven studies were included in the final dose-response analysis (16–22). The age of the included participants ranged from 14 to 74, the majority of the respondents were between the age of 30 to 64 with 2,413 respondents recorded to have died from lung cancer, details about patients characteristics of these studies can be seen in Supplementary Table 3. Supplementary Table 4 presented the result of the quality assessment. Results and study characteristics including name of first author, publication year, country or region, follow-up period, sample size and number of cases or deaths, type of outcome, gender, age range, HRs and 95% CIs were extracted and can be seen in Supplementary Table 5.

The results of quality assessment did not suggest any evidence of publication bias (Supplementary Figures 1, 2) [Unadjusted group: (Begg: 0.673; Egger: 0.182)]; [Adjusted group: (Begg:1; Egger:0.228)]. And sensitivity analysis was performed by sequentially excluding individual studies (Supplementary Figures 3, 4).

In unadjusted group, the pooled HR for family economy conditions with best family SEP in childhood group was 1.42 (95% CI, 1.21, 1.66) (I2 = 73.4%, P < 0.001) (Table 1, Figure 2). The risk was slightly lower than adjusted group: 1.25 (95% CI, 1.10, 1.42) (I2 = 48.9%, P = 0.04) (Table 1, Figure 3). Subgroup analysis based on sex using the provided adjusted original data resulted in a pooled HR of 1.09 (95% CI 0.95, 1.26) (I2 = 0.0%, p = 0.78) for male, a pooled HR of 1.10 (95%CI 0.88, 1.35) (I2 = 0.0%, p = 0.57) for female, and a pooled HR of 1.39 (95%CI 1.20, 1.62) (I2 = 48.9%, p = 0.04) for both sexes. Subgroup analysis based on sex that provided unadjusted original data resulted in a pooled HR of 1.52 (95% CI 1.13, 2.06) (I2 = 87.1%, p = 0.006) for male and a pooled HR of 1.28 (95%CI 1.07, 1.54) (I2 = 19.4%, p = 0.007) for female and a pooled HR of 1.63 (95%CI 1.15, 2.32) (I2 = 58.7%, p = 0.006) for both sexes. The above results was consistent in dose response used the model summary with BMR of 1 Std. Dev. including the family economic conditions of parents and childhood housing conditions (Supplementary Figure 5).

Genetically predicted lower family SEP in childhood was associated with significantly higher odds of lung cancer (Table 2). The causal relationship between length of education and lung cancer was verified first. Using conventional MR analysis, one SD longer education was associated with a 50% lower risk of lung cancer (SSGAC 1: OR 0.50, 95% CI: 0.39, 0.63; P: 0.001). Another gene pool with the same exposure factor was also included for verification in our study, showing one SD longer education (genetic predisposition by 4 SNPs) was associated with a 60% lower risk of lung cancer (SSGAC 2: OR 0.40, 95% CI 0.32, 0.50; p: 0.001). With a sample size of 27 209, our sample provided sufficient statistical power (> 80%) to detect a causal effect of educational attainment on lung cancer. Power calculations on MR analysis were performed according to Brion et al. (83).

Meanwhile, associations were consistent in sensitivity analyses using weighted median (OR 0.57, 95% CI 0.45, 0.71; P: 0.001; Okbay et al.) (OR 0.58, 95% CI 0.48, 0.70; P < 0.001; Lee et al.) and MR-Egger method (OR 0.56, 95% CI 0.17, 1.88; P: 0.35; Okbay et al.) (OR 0.75, 95% CI 0.60, 0.94; P: 0.01; Lee et al.), but provided less precise estimates than with conventional MR (IVW method). Nonetheless, their causal estimates were similar in terms of direction and magnitude, and they were unlikely to occur by chance alone. These results were consistent with the hypothesis that genetic pleiotropy does not drive the result.

We also analyzed and observed the relationship between other exposure factors related to family SEP in childhood and lung cancer, and we found that the P-values for the intercept were large and the estimates adjusted for pleiotropy suggested null effects in four exposure factors: number of children fathered (Neale Lab): (OR 0.92, 95% CI: 0.29, 2.84; p: 0.88); mother's age at death (Neale Lab): (OR 1.00, 95% CI: 0.19, 5.26; p: 1.00); and average total household income before tax (MRC-IEU): (OR 0.90, 95% CI: 0.77, 1.04; p: 0.156). Nine out of the 15 independent exposure factor gene pools were unable to be analyzed due to the lack of SNPs corresponding to lung cancer.