3.1 Literature search
A flow diagram of the literature selection process is shown in Fig. 1. Initially, 706 articles were identified by searching three databases and manually screening the references and grey literature. Then, 179 articles were found to be duplicates and deleted. After reviewing the unique titles and abstracts, 74 records appeared to be potentially relevant for inclusion in the meta-analysis. After further full-text review, 67 studies were excluded because 15 of them did not report schizophrenia patients as the study population, and 52 lacked of data on the mortality rate of any site-specific cancer. Ultimately, seven studies were included in our meta-analysis. The kappa statistic indicated satisfactory agreement between two raters (kappa = 0.82).
3.2 Study characteristics and quality assessment
The characteristics of the included studies are depicted in Table 1. A total of 1,162,971 patients with schizophrenia were enrolled in the seven included studies. Four common cancer types (i.e., breast, colon, lung and prostate cancer) were investigated, considering that a minimum of three data values of mortality rate of a certain cancer were needed for our analysis. Most studies compared cancer mortality in the general population as a control to calculate the RRs[7-10, 22, 23], one study used the cancer mortality rates in a population without schizophrenia to determine RRs[24]. The mortality risk of four selected site-specific cancers was individually reported as SMRs with 95%CIs in major studies and documented as HRs with 95%CIs in only one article[24]. One study did not provide data of 95%CIs corresponding to the SMRs of different cancers[9]. One study reported the data of P values and RRs[23]. Thus, we calculated the value of corresponding 95%CIs.
Four of the studies provided mortaliry rates of site-specific cancers according to gender[7, 8, 22, 24]. One article reported only on breast cancer mortality in female schizophrenic patients and lung cancer mortality in all participants with schizophrenia[9], one study merely reported mortality data for lung cancer[23], and one study did not provide the mortality data for prostate cancer[10]. The follow-up periods ranged from 7 years to 38 years. Four studies were conducted in European countries[8, 9, 22, 24], two in North America[10, 23], and one in Japan[7]. Most studies used the ICD-9 or ICD-10 and hospital records to assess schizophrenia. Confirmation of cancer death was mainly based on death registry data or death certificate.
All seven cohort studies scored no less than six stars according to the NOS scale, indicating a high level of methodological quality in this meta-analysis (Table 2).
3.3 Mortality risk of site-specific cancer in schizophrenia
Based on each included study’s cancer classifications, we calculated the pooled RRs of four common types of cancer, including prostate (men), breast (women), lung (all patients, men, women), and colon (women, men). The mortality risk of lung and colon cancer was reported in men or women with schizophrenia. We applied the random-effects model to calculate the pooled RRs of site-specific cancers due to the presence of substantial heterogeneity among the included studies (Fig. 2). The pooled RR for breast cancer was 1.97 (95%CI 1.38–2.83, P < 0.001), that for lung cancer was 1.93 (95%CI 1.46–2.54, P < 0.001), that for prostate cancer was 1.58 (95%CI 0.79–3.15, P = 0.195), and that for colon cancer was 1.69 (95%CI 1.60–1.80, P < 0.001). Thus, patients with schizophrenia showed significantly increased mortality risk rates of breast, lung, and colon cancer, but not prostate cancer, compared with the general population or control group.
3.4 Site-specific cancer mortality analyzed according to gender
Substantial heterogeneity was observed among the mortality rates of lung cancer in all participants with schizophrenia (I2 > 50%). We therefore reassessed the results according to gender, which is regarded as a potential confounder (Fig. 3). We also calculated the mortality rate of colon cancer by gender (Fig. 4). For women, the pooled RRs of lung and colon cancer were 2.49 (95%CI 2.40–2.59, P < 0.001, Fig. 3) and 2.42 (95%CI 1.39–4.22, P = 0.002, Fig. 4), respectively. For men, the pooled RRs of lung and colon cancer were 2.40 (95%CI 2.30–2.50, P < 0.001, Fig. 3) and 1.90 (95%CI 1.71–2.11, P < 0.001, Fig. 4), respectively. Thus, the mortality rate of lung and colon cancer significantly increased in patients with schizophrenia compared with that of the general population or control group regardless of gender.
3.5 Heterogeneity and sensitivity analysis
The χ2 test and I-squared statistics were used to evaluate the heterogeneity among studies. The calculated χ2 revealed no significant heterogeneity in the mortality rates of site-specific cancer among the included studies (P > 0.05, Figs. 2–4), except for lung cancer in all patients (P < 0.001, Fig. 2) and colon cancer in female patients (P = 0.039, Fig. 4). The I2 statistic indicated high levels of heterogeneity among the cohort studies for breast cancer (I2 = 54.2%, Fig. 2) and lung cancer (I2 = 85.1%, Fig. 2), as well as low levels of heterogeneity for prostate cancer (I2 = 9.1%, Fig. 2). During analysis of cancer mortality rates by gender, significant heterogeneity was observed among the included studies for colon cancer in women (I2 = 64.0%, Fig. 4). And statistical heterogeneity did not exist among the studies for colon cancer in men (I2 = 0.0%, Fig. 4), and lung cancer in women and men (I2 = 0.0%, Fig. 3).
Sensitivity analysis was performed to evaluate the stability of our pooled results in the random-effects model. During multivariate analysis of breast cancer mortality, the results did not significantly flip with a new RR (1.74, 95%CI: 1.24–2.44), which was calculated by the trim-and-fill method (Fig. S1, Additional file 1). Each study included in the meta-analysis of breast cancer made no difference to our pooled results in the random-effects model (Fig. S2, Additional file 1). During multivariate analysis of lung cancer mortality, the trim-and-fill method was used to obtain a new RR (1.93, 95%CI: 1.46–2.54), which was roughly equivalent to the previous RR (1.97, 95%CI: 1.38–2.83) (Fig. S1). Moreover, the pooled results remained stable regardless of which study in the meta-analysis of lung cancer was omitted (Fig. S2). During multivariate analysis of prostate cancer mortality, although the recalculated RR was similar to the previous value, with no significant difference (Fig. S1), the pooled result became unstable when the study by Crump et al. was excluded (Fig. S2).
3.6 Publication bias
According to the Cochrane handbook, at least 10 studies are required to perform the tests for funnel plot asymmetry efficaciously. Thus, Egger’s test was used to assess the potential publication bias. No evidence of potential publication bias was found among the included studies for breast cancer (P = 0.472), lung cancer (P = 0.120), and prostate cancer (P = 0.299). There were not enough studies for colon cancer to conduct the Egger’s test.