Parity and pancreatic cancer risk: evidence from a meta-analysis of twenty epidemiologic studies.

Multiple studies have hypothesized parity is associated with pancreatic cancer risk but obtained conflicting results. We conducted a meta-analysis (including a dose-response approach) of current available epidemiologic studies to investigate the association between parity and risk of pancreatic cancer. Ten cohort studies and ten case-control studies including 8205 cases were eligible for inclusion. The combined RR (relative risk) of pancreatic cancer for the parous vs. nulliparous was 0.91 (95% CI, confidence interval = 0.85-0.97, I(2) = 39.0%, P(h) = 0.01). We observed an inverse association between giving birth to two children pancreatic cancer risk with RR of 0.86 (95% CI = 0.80-0.93, I(2) = 8.7%, P(h) = 0.36). And no evidence supported there was non-linear (P = 0.33) or linear relationship (P = 0.14) between number of parity and risk of pancreatic cancer. Findings from this meta-analysis indicate that giving birth to two children has the lowest pancreatic cancer risk, mechanism of this protective effect needs further investigation.

Pancreatic cancer remains an aggressive malignancy that caused the eighth leading cancer-related mortality worldwide1. Currently, the only possibility for cure is surgery and only 20% patients have such chance due to late detection and diagnosis2. The extremely low 5-year survival rate (<5%) highlighted the urgency to find out the etiology of pancreatic cancer. However, in the past few decades, smoking has been confirmed as the greatest risk factor for pancreatic cancer and accounts for approximately 25% of cases34. The other suspected propensities such as alcohol consuming, BMI, physical activities, chronic pancreatitis and dietary factors yield disaccord results and need to be further validated5. Notwithstanding the recent genome-wide association studies identified several genetic susceptibility loci of pancreatic cancer67, only explaining ~4% of the cases8.

The differences in sex distribution9 which could not entirely be explained by cigarette smoking and the fact that steroid hormone receptors exist in normal and neoplastic human pancreatic tissue101112 plus the findings that anti-estrogenic agents inhibit the growth of pancreatic cancer in both animal and human models13 prompted investigators to hypothesize that the reproductive factors may contribute to the etiology of pancreatic cancer. Up to now, numerous studies14151617 have focused on the roles of reproductive factors such as parity, age at first birth, age of menarche, age of menopause played in the development of pancreatic cancer, but obtained inconsistent findings. Of these factors the parity was the most extensively investigated and yielded conflicting results. The association ranged from protective effects1819 to substantial increased risk with the increasing number of birth20.

The efforts of making a conclusion across these studies hampered by small sample size rendering limited power to detect the association, and inconsistent referent groups making non comparability between studies. Therefore, we conducted a meta-analysis (including a dose-response approach) of current available epidemiologic studies to precisely evaluate the association between parity and risk of pancreatic cancer. And to extensively control the misclassification, we divided the number of parity into three groups and combined the corresponding data in each group separately.

Results

Study characteristics

Using the search strategy described in method paragraph we identified 107 articles, of which 80 articles were excluded after review of the title or abstract. 27 full-texts were reviewed and one article was excluded due to insufficient data21. Other two studies2223 were drived out due to employ the measure of standard incidence rate ratios (SIRs) and standard mortality ratios (SMRs) to estimate the risk. In addition, two studies2425 were removed for overlapping population with another longer duration follow-up report26 and larger sample size27. Two studies2829 were removed for focusing on the number of pregnancy rather than the number of parity. Finally, 10 cohort studies16171819202630313233 and 10 case-control studies14152734353637383940 were eligible for this meta-analysis yielding a combined number of cases of 8205 (Figure 1). Characteristics of the 20 included studies are shown in Table 1. Among these studies, seven were conducted in USA15183032333540, two in Canada2039, seven in Europe16172627343638, one in Japan31, and two in China1937, one in Egypt14. The average of following- up duration of cohort studies was 18.65 years, ranged from 7.1 to 38 years. Cohort size ranged from 37459 to 1292462, and the number of cases varied from 52 to 1959. Almost all studies provided relative risk estimates adjusted for age (n = 19), most studies for smoking (n = 14), BMI (n = 8) and few controlled for history of diabetes (n = 5) (Table 1).

Figure 1

Flowchart of study selection procedure of this meta-analysis.

Table 1

Characteristics of eligible studies included in this meta-analysis of parity and pancreatic cancer risk

Author (Publication year)	Design and Study name	Country	Study period (follow-up years)	Age	N(case)	No. of participants/controls	Parity	OR (95%CI)	Factors investigated
Mesquita34 (1992)	PCC	Netherlands	1984–1988	35–79	176	487	1	Reference	Age, response status and life-time smoking of cigarettes
							2	0.62 (0.23–1.69)
							3	0.62 (0.22–1.74)
							4+	0.51 (0.21–1.25)
Cantor35 (1993)	PCC	USA	1984–1987	40–85	169	821	1	Reference	Age, having a first degree relative with the same cancer, and age at first birth
							2	0.7 (0.4–1.3)
							3	1.1 (0.6–2.1)
							4	0.6 (0.3–1.3)
							5+	0.5 (0.3–1.2)
Kalapothaki36 (1993)	HCC	Athens	1991–1992	-	66	66	Nulliparous	Reference	Age, hospital, past residence, years of schooling, cigarette smoking, diabetes melitus, as well as mutually (age at menarche, menopausal status, and age at menopause)
							Parous	0.39 (0.11–1.40)
Ji37 (1996)	PCC	China	1990–1993	30–74	451	1552	0–2	Reference	Age, income, education, smoking, green tea drinking, respondent status and BMI
							3–4	1.33 (0.8–2.23)
							≥5	1.88 (1.09–3.24)
Karlson38 (1998)	Nested CC	Swedish	-	Mean age, Case 52.7	1015	5073	1	Reference	Age
							2	0.91 (0.75–1.1)
							3	1.04 (0.84–1.29)
							4	1.20 (0.91–1.59)
							5+	1.16 (0.80–1.70)
Kreiger39 (2001)	PCC	Canada	1995–1996	20–75	52	233	0	Reference	Age, smoking status, BMI, tofu, and dietary fat, age at menopause, oral contraceptive use, estrogen replacement therapy, and age at menarche
							1–2	0.69 (0.24–2.00)
							3+	0.22 (0.07–0.65)
Skinner30 (2003)	Cohort Nurses' Health Study	USA	1976–1998 (22)	30–55	243	115474	0	Reference	Age, time period, cigarette smoking, diabetes, BMI and height
							1–2	0.86 (0.55–1.36)
							3–4	0.75 (0.48–1.17)
							5+	0.58 (0.34–0.98)
Duell40 (2004)	PCC	USA	1995–1999	21–85	241	818	0	Reference	Age, education, and smoking
							1	1.3 (0.75–2.2)
							2	1.0 (0.63–1.6)
							3	1.3 (0.81–2.1)
							4+	1.0 (0.63–1.7)
Teras18 (2005)	Cohort Cancer Prevention Study (CPS)-II	USA	1982–2000 (18)	60.5 ± 9.0	1959	387981	0	Reference	Age, race, smoking history, education, personal history of diabetes, BMI, height, exercise, and family history of pancreatic cancer
							1	0.99 (0.84–1.17)
							2	0.96 (0.83–1.11)
							3	1.10 (0.95–1.28)
							4	0.96 (0.80–1.15)
							5	0.78 (0.61–0.99)
							6	0.88 (0.65–1.19)
							7+	0.76 (0.57–1.03)
Silvera20 (2005)	Cohort National Breast Screening Study (NBSS)	Canada	1980–2000 (16.4)	40–59	187	89835	0	Reference	Age, cigarette smoking intensity, cigarette smoking duration, BMI, height, study left, and randomization group (intervention vs control)
							1–2	0.67 (0.43–1.06)
							3–4	0.81 (0.53–1.24)
							5+	1.05 (0.64–1.74)
Lin31 (2006)	Cohort The Japan Collaborative Cohort Study for Evaluation of Cancer Risk (the JACC study)	Japan	1988–1999 (10)	40–79	154	63273	0–1	Reference	Age, pack-years of smoking, history of diabetes, history of gallstone/cholecystitis, and parity or age at first birth
							2–3	0.74 (0.38–1.43)
							4–5	0.84 (0.40–1.75)
							≥6	0.80 (0.31–2.11)
Lo14 (2007)	HCC	Egypt	2001–2004	15–90	194	194	0–3	Reference	Age, residence, and active smoking
							4–6	1.3 (0.6–2.8)
							7+	0.5 (0.2–1.3)
Prizment32 (2007)	Cohort Iowa Women's Health Study	USA	1986–2003 (18)	55–69	228	37459	0	Reference	Age
							1–2	1.14 (0.69–1.88)
							3–4	1.10 (0.67–1.81)
							>4	1.09 (0.63–1.87)
Heuch26 (2008)	Cohort	Norway	1956–1998 (38)	32–74	449	63090	0	0.68 (0.45–1.04)	Age, birth cohort and breastfeeding
							1	Reference
							2	0.97 (0.69–1.35)
							3	1.47 (1.03–2.09)
							4	1.49 (0.98–2.26)
							≥5	1.63 (1.01–2.64)
Steven16 (2009)	Cohort Million Women Study	UK	7.1	-	1182	995192	0	Reference	Age, region, socioeconomic status, smoking, body mass index, and diabetes
							1	0.88 (0.72–1.04)
							2	0.78 (0.70–0.85)
							3	0.81 (0.71–0.91)
							4+	0.84 (0.72–0.96)
Zhang15 (2010)	HCC	USA	1976–2006	23–78	284	1096	0	Reference	Age, study left, year of interview, and race
							1–2	1.5 (1.0–3.2)
							3–4	1.4 (0.9–2.2)
							4+	1.5 (0.8–2.6)
Chang19 (2010)	Cohort	China	1978–2007 (29)	Age at first birth	243	1292462	1	Reference	Age at first childbirth, Marital status (married or unmarried), years of schooling (≤9 or ≥9 years), and birth place
							2	0.69 (0.49–0.98)
							3+	0.64 (0.44–0.93)
Lucenteforte26 (2011)	HCC	Italy	1983–1992 and 1991–2008	Median age Case (61) Control (58)	285	713	0	Reference	Study/left, age, education, area of residence, year of interview, history of diabetes, and tobacco smoking
							1	0.72 (0.4–1.16)
							2	0.86 (0.55–1.33)
							3	0.88 (0.53–1.46)
							4+	0.46 (0.26–0.85)
Duell17 (2012)	Cohort European prospective investigation into cancer and nutrition (EPIC) cohort	Europen	1992–2006 (14)	35–70	304	328610	0	Reference	Age, left, smoking, education, alcohol consumption at baseline, red meat intake, BMI, and diabetes history
							1	1.01 (0.64–1.59)
							2	1.03 (0.69–1.53)
							3	0.92 (0.59–1.43)
							4+	1.15 (0.71–1.85)
Lee33 (2013)	Cohort California Teachers Study	USA	1995–2009 (14)	Mean age 51.9	323	118164	1	Reference	Race/ethnicity, BMI, menopausal status HT use, age at menarche, smoking status and history of diabetes
							2	0.98 (0.69–1.41)
							3	0.88 (0.60–1.30)
							4+	0.83 (0.55–1.29)

Abbreviations: OR, odds ratio; CI, confidence interval; BMI, body mass index; HT, hormone replacement treatment; PCC, population-based case–control; HCC, hospital-based case–control.

Parous vs. Nulliparous

Eleven studies including 38 reports reported the results using nulliparous (number of parity = 0) as reference. The summary multivariable-adjusted RR (95% CI) of pancreatic cancer associated with parity for parous versus nulliparous was 0.91 (95%CI = 0.85–0.97) (Table 2, Figure 2), indicating inverse association between parity and pancreatic cancer risk. Moderate between-study heterogeneity was observed among studies (P = 0.01, I2 = 39.0%). The 38 report-specific RRs of parous versus nulliparous ranged from a low of 0.89 (95% CI = 0.84–0.95) after omission of report by Teras et al18 to a high of 0.92 (95% CI = 0.86–0.98) after omission of report by Steven et al16. The power for the sample size of meta-analysis to detect an RR of 0.91 is more than 95%.

Table 2

Summary risk estimates of the association between parous and pancreatic cancer.

	No. of reports	RR (95%CI)	P	I2	Ph*
Overall	38	0.91 (0.85–0.97)	<0.01	39.0	0.01
Subgroup analysis
Study design
Cohort	24	0.89 (0.84–0.95)	<0.01	28.7	0.10
Case-control	14	0.95 (0.76–1.18)	0.62	50.6	0.02
Number of cases
<100	27	0.90 (0.81–1.00)	0.05	26.8	0.10
>100	11	0.81 (0.84–0.99)	0.03	60.2	0.01
Location
Europe	15	0.82 (0.76–0.89)	<0.01	16.3	0.27
North America	23	0.97 (0.89–1.04)	0.37	20.8	0.18

*P value for heterogeneity.

Abbreviations: OR, odds ratio; CI, confidence interval.

Figure 2

Forest plot of parous and pancreatic cancer risk.

Different number of parity

The results of effects of different number of parity on pancreatic cancer risk were presented in Table 3. To explore the effects of different number of parity, we divided the number of parity into three groups and included the studies using number of parity of 0 or 1 as reference. 11 studies were assigned to the first group (number of parity = 2), and the summary multivariable-adjusted RR (95%CI) of pancreatic cancer associated with give birth to two children versus number of parity of 0 or 1 was 0.86 (95% CI = 0.80–0.93) with I2 = 8.7% (Ph = 0.36) (Figure 3). The second group (>2 and <5) contained 16 studies including 28 reports, and our analysis of these reports yielded the combined risk estimate of 0.98 (95%CI = 0.89–1.09) with I2 = 53.7% (Ph < 0.01). 9 studies including 11 reports in the third group (≥5) yielded the combined risk estimate of 0.95 (95% CI = 0.75–1.20) with I2 = 66.3% (Ph < 0.01). The sensitivity analysis revealed that the pooled RRs for the second and third groups were similar before and after elimination of individual report. Considering the first group, the 11 report-specific RRs ranged from a low of 0.82 (95% CI = 0.76–0.89) after omission of report by Teras et al18 to a high of 0.92 (95% CI = 0.84–1.01) after omission of report by Steven et al16.

Table 3

Summary risk estimates of the association between parity number and pancreatic cancer risk

	2					> 2 and <5					≥5
	No. of reports	RR (95%CI)	P	I2	Ph*	No. of reports	RR (95%CI)	P	I2	Ph*	No. of reports	RR (95%CI)	P	I2	Ph*
Overall	11	0.86 (0.80–0.93)	<0.01	8.7	0.36	28	0.98 (0.89–1.09)	0.75	53.7	<0.01	11	0.95 (0.75–1.20)	0.64	66.3	<0.01
Subgroup analysis
Study design
Cohort	6	0.87 (0.77–0.99)	0.04	45.3	0.10	15	0.99 (0.88–1.11)	0.83	58.0	<0.01	7	0.93 (0.72–1.20)	0.56	63.6	0.01
Case-control	5	0.89 (0.76–1.04)	0.14	0.0	0.83	13	0.95 (0.78–1.16)	0.60	49.3	0.02	4	0.93 (0.52–1.66)	0.80	73.8	0.01
Number of cases
<100	5	0.89 (0.73–1.10)	0.29	0.0	0.80	21	1.05 (0.92–1.21)	0.46	38.8	0.04	11	0.95 (0.75–1.20)	0.64	66.3	<0.01
>100	6	0.87 (0.78–0.97)	0.02	44.9	0.11	7	0.90 (0.79–1.02)	0.08	64.0	0.01	0	-	-	-	-
Location
Europe	5	0.84 (0.76-0.93)	<0.01	14.5	0.33	12	0.96 (0.83–1.12)	0.64	63.8	<0.01	2	1.33 (0.96–1.84)	0.09	16.1	0.28
NorthAmerica	5	0.94 (0.84–1.07)	0.35	0.0	0.86	15	1.04 (0.93–1.17)	0.47	20.8	0.22	7	0.81 (0.65–1.01)	0.06	47.0	0.08

*P value for heterogeneity

Abbreviations: OR, odds ratio; CI, confidence interval

Figure 3

Forest plot of giving birth to two children and pancreatic cancer risk.

Dose response meta-analysis

The dose-response analysis on the number of parity and risk of pancreatic cancer involved 19 studies, we did not find a curve linear association between number of parity and risk of pancreatic cancer (P = 0.33 for non-linearity, I2 = 25%, Ph = 0.05).

The combined relative risk of pancreatic cancer for per live birth was 0.98 (95% CI = 0.95 to 1.01; P = 0.14 for linear trend) with I2 = 66.2% (Ph < 0.01), indicating no linear relationship between number of parity and risk of pancreatic cancer (Figure 4).

Figure 4

Forest plot of parity number (per 1 live birth) and pancreatic cancer risk.

Meta-regression

We conducted a meta-regression to comprehensively explore the source of heterogeneity of parous versus nulliparous. Six factors such as study design, location, whether adjusted for smoking, BMI and diabetes, number of cases were included in the meta-regression model. In this model, the Adj R-squared was 100.00%, and Prob > F was 0.0024, which indicated that the model was significant. After 100 times permutation, only study location appeared to be significant (adjusted P < 0.01) to explain the between-study heterogeneity.

Subgroup analyses

In the subgroup analysis of parous versus nulliparous, significant protective effects of parous on pancreatic cancer were observed in cohort studies (RR = 0.89, 95% CI = 0.84–0.95, I2 = 28.7%, Ph = 0.10), number of cases of more than 100 (RR = 0.89, 95% CI = 0.82–0.99, I2 = 60.2%, Ph = 0.01) and European population (RR = 0.82, 95% CI = 0.76–0.89, I2 = 16.3%, Ph = 0.27). In the subgroup analysis of effects of different number of parity on pancreatic cancer, the effects still turned out to be null in each stratum in both second and third groups. However, significant protective effects were observed in the subgroups of cohort studies (RR = 0.87, 95% CI = 0.77–0.99, I2 = 45.3%, Ph = 0.10), number of cases of more than 100 (RR = 0.87, 95% CI = 0.78–0.97, I2 = 44.9%, Ph = 0.11) and European population (RR = 0.84, 95% CI = 0.76–0.93, I2 = 14.5%, Ph = 0.33) of giving birth to two children (Table 3).

Publication bias

The result of Egger's test indicated no evidence of substantial publication bias for parous versus nulliparous (P = 0.52), and in the first group (P = 0.54), second group (P = 0.68) and third group (P = 0.70). And the of Begg's funnel plot of parous versus nulliparous was presented in Figure 5.

Figure 5

Beeg's funnel plot of parous versus nulliparous.

Discussion

We systematically reviewed ten cohorts and ten case-control studies including 8205 cases on the association between number of parity and pancreatic cancer risk. The factor became interesting candidate exposure related to pancreatic cancer because the difference of gender distribution and several biologic researches suggested that estrogen had protective effect against pancreatic cancer414243 and pancreatic cancer might be in case another hormone-dependent cancer101112. The current meta-analysis indicated that parous was inversely associated with pancreatic cancer risk compared with nulliparous. Giving birth to two children has the lowest risk of pancreatic cancer. We did not observe a non-linear or linear relationship between number of parity and pancreatic cancer. To the best of our knowledge, it is the first meta-analysis deeply mined the effects of different number of parity on pancreatic cancer risk.

The meta-regression analysis revealed that the study location might be the major source of between-study heterogeneity. We further carried out subgroup analysis by study design, study location and number of cases. And the between-study heterogeneity was largely removed when stratified according to study location, which further indicated that it mainly contributed to the heterogeneity. In the subgroup analysis of parous versus nulliparous, significant protective effects of parous on pancreatic cancer were identified in cohort studies, number of cases of more than 100 and European population. Similar with the subgroup analysis of parous, the subgroup analysis of different number of parity revealed that giving birth to two children was inversely associated with pancreatic cancer risk in those previously mentioned subgroups. The sample size of cohort studies was larger than the case-control studies and most studies conducted in Europe also have larger sample size than the studies conducted in North-America. So the detection of significant inverse associations might be due to dramatically increased statistic power of the combined large sample size. And the underlie mechanisms of the geographic variation of the effects of parity on pancreatic cancer risk is largely unknown and needs further validation and exploration.

Our meta-analysis indicated parous women had lower risk of pancreatic cancer compared with nulliparous women. Parous women are likely to have had longer periods of exposure to high levels of circulating estrogens. And animal studies reported that the estrogen had the inhibitory effect on the growth of preneoplastic pancreatic lesions or transplanted pancreatic carcinoma in rats1112, indicating that estrogen may have effect against pancreatic cancer. However, the discrepancy between giving birth to two children and more than two is not fully understood, several potential biologic mechanisms might have been proposed. First, there was a progressive increase in insulin resistance during pregnancy44, and insulin resistance was suggested to be independently associated with pancreatic cancer risk45. Another explanation may be obesity which is known as a risk factor for pancreatic cancer46 and pregnant women are more tend to be obese. So the harmful effects of insulin resistance and obesity may counteract the protective effect of estrogen against pancreatic cancer.

Our current meta-analysis has several strengths. First, we included ten cohorts and ten case-control studies which enabled us significant statistical power to detect potential association. Second, to control misclassification, we combined different number of parity separately rather than used the highest versus the lowest number. Third, we applied the model adjusting for most established risk factors therefore largely controlled the confounders. In addition, no publication biases were detected, which suggested the entire pooled result may be unbiased.

Despite these advantages, some limitations of the current study should be acknowledged. First we did not have access to the primary data from the studies included in this meta-analysis, so could not perform additional adjustments for potential important covariates. Second, there is a relatively wide range of values for cutoff of highest level of number of parity so we could not accurately assign an exposure value to open-ended category, which might impact our current analysis.

In conclusion, the current meta-analysis indicates that parous is inversely associated with pancreatic cancer risk compared with nulliparous, and giving birth to two children has the lowest pancreatic cancer risk. The exact mechanism of this protective effect needs further investigation.

Methods

Literature search and eligibility criteria

We searched the Pubmed and Embase databases from inception to November 2013 targeting the studies investigating the relationship between parity and the risk of pancreatic cancer. To undertake a comprehensive literature search, we used the combined terms: the MeSH heading and keywords relating to pancreatic neoplasms, the MeSH heading “reproductive history” and the keywords regarding to the exposure (reproductive factors, parity, birth), without language restriction. Reference lists of eligible studies and relevant review articles were also scanned to identify further pertinent studies. Studies were considered eligible if they met the following criteria: (a) the design of the study is prospective cohort study or case-control study; (b) investigate the relationship of parity and pancreatic cancer risk; (c) provide or allow the calculation of odds ratio (OR), relative risk (RR), or hazard ratio (HR) estimates with 95% CI. We excluded non-human, case reports, comparative studies not using an analytical epidemiologic design, or studies not reporting analyses of primary data (e.g., letters, editorials, narrative reviews) and not providing sufficient data. When multiple studies pertained to the same or partially overlapping populations, we used the result with the longest follow-up time or the largest sample size.

Data extraction

For each eligible study, two investigators (BBZ and LZ) independently carried out the eligibility evaluation, data abstraction, and quality assessment, and disagreements were resolved by consensus. The following data were extracted from eligible study: the first author, publication year, geographic region, demographics of participants, sample size (number of case and total participants), duration of follow-up or study period, covariates adjusted in the multivariable analysis, parity number categories, the corresponding OR or RR (with their 95% CI) and person-years for each category. For the RR, we chose the one represented the greatest degree control of potential confounders. The individual authors were contacted with E-mail if the data of interest were not shown in the publications.

Statistical analysis

Because the absolute risk of pancreatic cancer is low in human, the ORs and HRs were considered equivalent to RRs and we simply used RRs representing all of the three measures. To control confounding factors in the greatest extent, we extracted the maximally adjusted RRs (95%CI).

First of all, we evaluated the overall effect of parous compared with nulliparous, if the study considered nulliparous as reference, we summed up all the parous categories (number of parity >0) as ever parity in each study together and treated these different categories as different reports. For those studies considering number of parity of 0 or 1 as reference, we further assessed the effects of different number of parity. We first divided the number of parity into three groups (2, >2 and <5, ≥5) based on the pre-analysis of the data structure across these eligible articles, and then combined the corresponding data in each group separately. We pooled the RRs for the overall effects of parous and respective effects of different groups in a random effects model described by DerSimonian and Laird47 which takes into account both within- and between-study variability. Then we explored the possible non-linear or linear relationship between number of parity and risk of pancreatic cancer using a random-effects dose-response meta-analysis according to the method described by Greenland and coworkers4849. This method requires that the distribution of cases and person-years/number of cases and controls, and the RR with 95%CI for at least three quantitative exposure categories. For each study, the median or mean level of exposure category was assigned to each corresponding RR. When the median or mean exposure level per category was not presented in the study, we assigned the midpoint of the upper and lower boundaries in each category as value of exposure. If the highest category was open-ended, we assumed that it had the same amplitude as the closest category. Furthermore, a potential cure linear relationship between parity and pancreatic cancer risk was assessed using restricted cubic splines with 3 knots at fixed percentiles (10%, 50% and 90%) of the distribution50. A P value for curve linearity or non-linearity was calculated by testing the null hypothesis that the coefficient of the second spline is equal to zero.

We quantified the extent of heterogeneity using Q-test51 and I2 score52. We conducted a meta-regression analysis and subgroup analyses to explore source of heterogeneity and subgroup analyses were performed, if feasible, according to study design, geographic region, and number of cases. Sensitivity analysis was also performed to assess the influence of each individual report on overall estimates by sequential removal of individual studies. Funnel plots and the Egger's test53 were applied to examine the publication bias. All statistical analyses were carried out by Stata (version 10.0). The power of our meta-analysis was calculated using PowerV3.0 (http://www.mds.qmw.ac.uk/statgen/dcurtis/software.html).

Author Contributions

Conceived and designed the study strategy: X.P.M.; Acquisition of data: statistical analysis and interpretation of data B.B.Z., L.Z., W.C.; Drafting or revision of the manuscript: B.B.Z., J.H.; Reference collection and data management: N.S., J.Y.L., X.Q.C.; Wrote the manuscript: B.B.Z.; Prepared the tables and figures: C.L., Y.S., R.Z.; Study supervision: X.P.M.; All authors reviewed the manuscript.