Birth Order and Risk of Non-Hodgkin Lymphoma—True Association or Bias?

Abstract

There is inconsistent evidence that increasing birth order may be associated with risk of non-Hodgkin lymphoma (NHL). The authors examined the association between birth order and related variables and NHL risk in a pooled analysis (1983–2005) of 13,535 cases and 16,427 controls from 18 case-control studies within the International Lymphoma Epidemiology Consortium (InterLymph). Overall, the authors found no significant association between increasing birth order and risk of NHL (P-trend = 0.082) and significant heterogeneity. However, a significant association was present for a number of B- and T-cell NHL subtypes. There was considerable variation in the study-specific risks which was partly explained by study design and participant characteristics. In particular, a significant positive association was present in population-based studies, which had lower response rates in cases and controls, but not in hospital-based studies. A significant positive association was present in higher-socioeconomic-status (SES) participants only. Results were very similar for the related variable of sibship size. The known correlation of high birth order with low SES suggests that selection bias related to SES may be responsible for the association between birth order and NHL.

Non-Hodgkin lymphoma (NHL) is a common cancer in industrialized nations, and incidence has increased dramatically over the last 50 years (1). Lymphomas arise from cells of the human immune system, which is known to be critically shaped by early-life environmental exposures, including infections. Infection very early in life may fundamentally alter immune function and in this way may influence lymphoma risk (1). Increasing birth order has been widely studied as a proxy for increased likelihood of early-life exposure to infection and as a risk factor for lymphoma. Although an inverse relation between increasing birth order and risk of Hodgkin lymphoma has been reported in most case-control (2) and cohort (3) studies, results from studies of NHL have been inconsistent. Recently, investigators in several case-control studies reported a positive relation between increasing birth order and risk of NHL (4–7), but others have found no association (8–12). In addition, researchers in other case-control studies have reported that NHL risk increases with sibship size (13), while others have reported no association (8, 11, 14). In no studies have investigators reported a significant inverse association. Thus, there are inconsistent data from case-control studies that increasing birth order may be associated with increasing NHL risk (15). Until recently, cohort studies have been too small to provide informative data on the association between birth order and NHL incidence. In 2006, in a large-scale population-based Swedish cohort study of over 7,000 cases of incident NHL, Altieri et al. (3) reported no association between NHL and birth order (P-trend = 0.464). However, this was a study of predominantly young persons, with a mean age of NHL occurrence of 40 years (3).

Recently, it has been hypothesized that selection bias may explain the inconsistency in the results of case-control studies of birth order and NHL risk (11). To further investigate the association between birth order and related variables indicative of childhood crowding and NHL risk, we conducted a pooled analysis of case-control studies from the International Lymphoma Epidemiology Consortium (InterLymph).

MATERIALS AND METHODS

We performed a pooled analysis of data from 18 case-control studies included in InterLymph (www.epi.grants.cancer.gov/InterLymph). Participating studies (Table 1) met the following eligibility criteria: cases diagnosed with incident, histologically confirmed NHL as adults (ages 16–96 years); collection of data on birth order or sibship size; and the availability of an electronic data set in March 2007. Organ transplant recipients and persons with human immunodeficiency virus infection were excluded. Because of variations in study design among the 6 EpiLymph study centers (Table 1), we treated this study statistically as 6 separate studies, to allow for assessment of center-specific effects and heterogeneity.

Table 1.

Characteristics of Participants From 18 Non-Hodgkin Lymphoma Case-Control Studies Included in a Pooled Analysis of Birth Order/Sibship Size and Risk of Non-Hodgkin Lymphoma, 1983–2005

First Author, Year (Reference No.)	Location of Study	Years of Study	Age Range, years	Matching Variables	Cases		Controls
					No.	Participation Rate, %	No.	Participation Rate, %	Source of Controls
Talamini, 2004 (28)	Aviano and Napoli, Italy	1999–2002	18–84	None	225	97	504	91	Hospital patients
Spinelli, 2007 (21)	British Columbia, Canada	2000–2003	20–79	Age, sex, and region	828	85	845	50	Random selection from client registry of Ministry of Health
Zhang, 2005 (29)	Connecticut, United States	1995–2001	21–84	Age	597	72	716	47–69	<65 years: RDD; ≥65 years: random selection from CMS
EpiLymph (multicenter case-control study of lymphoma)	Europe
de Sanjose, 2004 (22)	Spain	1998–2003	17–96	Age, sex, and region	435	82	630	96	Hospital patients
Becker, 2004 (10)	Germany	1999–2002	18–82	Age, sex, and region	518	87	518	44	Random selection from population registries
Besson, 2006 (23)	Ireland	1998–2004	19–85	Age, sex, and center	144	90	208	75	Hospital patients
Becker, 2007 (12)	Czech Republic	2001–2003	19–82	Age, sex, and region	199	90	199	60	Hospital patients
Becker, 2007 (12)	France	2000–2003	18–82	Age, sex, and region	217	91	272	74	Hospital patients
Cocco, 2008 (24)	Italy (Sardinia)	1998–2004	25–81	Age, sex, and region	219	93	336	66	Random selection from population registries
Vineis, 2000 (14)	11 cities in Italy	1990–1993	20–74	Age, sex, and region	1,640	82	1,771	74	Random selection from demographic or National Health Service files
Cerhan, 2007 (25) (Mayo Clinic)	Minnesota, Iowa, and Wisconsin, United States	2002–2005	20–87	Age, sex, and region	500	65	499	69	Patients attending a prescheduled general medical examination
Cozen, 2007 (6) (NCI-SEER)	Detroit, Michigan; Iowa; Los Angeles, California; and Seattle, Washington, United States	1998–2001	20–74	Age, sex, region, and race/ethnicity	1,316	76	1,055	52	<65 years: RDD; ≥65 years: random selection from CMS
Chiu, 2005 (30)	Nebraska, United States	1999–2002	20–75	Age and sex	386	74	535	78	RDD
Tavani, 2000 (26)	Aviano and Milan, Italy	1983–1992	17–85	None	429	>97	1,157	>97	Hospital patients
Grulich, 2005 (4)	New South Wales and Australian Capital Territory, Australia	2000–2001	20–74	Age, sex, and region	694	85	694	61	Random selection from electoral rolls
Smedby, 2007 (5) (SCALE)	Denmark and Sweden	1999–2002	18–74	Age, sex, and country	3,055	81	3,187	71	Random selection from population registries
Bracci, 2006 (7)	San Francisco, California, United States	1988–1993	21–74	Age, sex, and region	1,305	72	2,404	78	<65 years: RDD; ≥65 years: random selection from CMS (population-based)
Willett, 2005 (27)	Parts of northern and southwestern England, United Kingdom	1998–2003	16–69	Age, sex, and region	828	75	897	71	Random selection from general practice lists

Abbreviations: CMS, Centers for Medicare and Medicaid Services; NCI, National Cancer Institute; RDD, random digit dialing; SCALE, Scandinavian Lymphoma Epidemiology Study; SEER, Surveillance, Epidemiology, and End Results [Program].

Exposure assessment

Investigators in all studies collected self-reported data on birth order and/or sibship size. Age differences with the nearest older and younger siblings were collected in 10 and 9 studies, respectively. For the birth-order analyses, the referent category was firstborn children, including only children. Analyses were also performed after separating only children from the referent category. Because results were very similar, they are not presented here (data not shown). Socioeconomic status (SES) groups in each study were based on the tertile distribution of years of education (11 studies) or area-based deprivation levels obtained from census data (11) (2 studies) in the controls within each individual study.

Statistical methods

Odds ratios and 95% confidence intervals were computed from unconditional logistic regression models, using a 2-stage random-effects model to estimate relative risk (hereafter called “risk”) of NHL and a joint fixed-effects model to estimate risk by NHL subtype (16). All World Health Organization classification (17) subtypes of NHL except multiple myeloma were included in the analysis, as recommended for epidemiologic analyses (18). Results in all models were adjusted for the matching variables age, sex, and region/study center. Tests for a linear trend in odds ratios were performed using a generalized least-squares trend estimation procedure that assumed a linear relation in exposure for increasing birth order and sibship size (19).

Heterogeneity among study centers was assessed using Cochran's Q statistic and the I² statistic (20). In the presence of significant heterogeneity (P < 0.10), forest plots were used to identify outlying studies, and sensitivity analyses were performed with and without the outlying studies; no individual study was consistently identified as outlying. Because of a predominance of study participants of Caucasian origin, stratification by race/ethnicity was not meaningful. Restriction of analyses to Caucasians gave results similar to those from analyses that included all participants. All statistical tests were 2-sided and assumed an α error level of 0.05. Analyses were performed using STATA software, version 10.0 (Stata Corporation, College Station, Texas).

To examine the possible effect of bias, we conducted sensitivity analyses by stratifying by study design factors, including published response rates in cases and controls (tertiles) and source of controls (population vs. hospital). In addition, results were stratified by sex and SES.

The pooled analyses were approved by the University of New South Wales Human Research Ethics Committee.

RESULTS

This analysis included data on 13,535 cases and 16,427 controls. Data on both birth order and sibship size were collected in 14 studies (4–6, 10, 12, 21–27); in 4 additional studies (14, 28–30), data were available on sibship size only (Table 1). Demographic characteristics of participants are summarized in Table 2.

Table 2.

Demographic Characteristics of Cases and Controls From 18 Non-Hodgkin Lymphoma Case-Control Studies Included in a Pooled Analysis of Birth Order/Sibship Size and Risk of Non-Hodgkin Lymphoma, 1983–2005

Demographic Factor	Cases		Controls
	No.	%	No.	%
Pooled total	13,535		16,427
Sex
Male	7,329	54.1	8,865	54.0
Female	6,206	45.9	7,562	46.0
Age, yearsa
<20	25	0.2	56	0.3
20–29	381	2.8	842	5.1
30–39	924	6.8	1,602	9.8
40–49	1,823	13.5	2,386	14.5
50–59	3,472	25.7	3,729	22.7
60–69	4,313	31.9	4,759	29.0
70–79	2,448	18.1	2,885	17.6
≥80	147	1.1	168	1.0
Education/SES (tertile)
Low	5,205	38.5	5,802	35.3
Medium	4,643	34.3	5,883	35.8
High	3,577	26.4	4,642	28.2
Unknown/other	110	0.8	100	0.6
Race/ethnicity
White	5,672	41.9	6,757	41.1
Black	202	1.5	327	2.0
Other/mixed	539	4.0	558	3.4
Missing data	7,122	52.6	8,785	53.5

Abbreviation: SES, socioeconomic status.

^aThe median age of cases was 58 years (range, 16–96); the median age of controls was 60 years (range, 17–89). Data on age were missing for 2 cases.

Overall, there was no significant association between increasing birth order and increasing risk of NHL (Table 3). Among participants whose birth order was fifth or higher, the pooled odds ratio was 1.12 (95% confidence interval: 0.95, 1.32) as compared with those who were firstborn, and the test for trend with increasing birth order was not significant (P-trend = 0.082). The direction of the association was the same for sibship size and was also nonsignificant (P-trend = 0.313). Within individual strata of sibship size, there was no consistent association of NHL risk with increasing birth order (data not shown). Study-specific odds ratios for all birth orders and most sibship sizes were heterogeneous, with the greatest variation being seen at higher birth orders and higher sibship sizes (see Web Figure 1 (http://aje.oxfordjournals.org/) and Table 3). A statistically significant positive association with birth order and sibship size was present in Australian and North American studies, but in Europe an association was present only in the Scandinavian study (5). All of these studies were population-based.

Table 3.

Odds Ratios for Non-Hodgkin Lymphoma According to Birth Order/Sibship Size in a Pooled Analysis of Results From 18 Non-Hodgkin Lymphoma Case-Control Studies Within the International Lymphoma Epidemiology Consortium (InterLymph), 1983–2005

Familial Structure Variable	No. of Cases	No. of Controls	Odds Ratioa	95% Confidence Interval	Heterogeneity
					No. of Studies	P Value	I2, %
Birth order
Firstborn	3,610	4,579	1.00
Second-born	2,595	3,311	1.01	0.92, 1.10	14	0.19	24.5
Third-born	1,520	1,841	1.05	0.93, 1.20	14	0.02	48.1
Fourth-born	816	977	1.12	0.96, 1.30	14	0.05	42.0
Fifth-born or higher	1,228	1,401	1.12	0.95, 1.32	14	<0.01	60.9
P-trend			0.082
Sibship size
Only child	1,283	1,581	1.00
No. of other siblings
1	2,820	3,591	1.02	0.93, 1.12	18	0.83	0.0
2	2,668	3,290	1.06	0.96, 1.16	18	0.72	0.0
3	1,902	2,314	1.04	0.88, 1.21	18	0.02	46.7
4	1,248	1,591	1.01	0.87, 1.18	18	0.09	32.8
≥5	2,720	3,235	1.10	0.93, 1.29	18	0.01	50.9
P-trend			0.313

^aOdds ratios and 95% confidence intervals were computed using a 2-stage random-effects model, adjusted for age (in 5-year intervals), sex, and study center.

Pooled odds ratios varied by case and control response rate and by study design. Positive associations with birth order and sibship size of 5 or more were confined to studies in the lower 2 tertiles of case response rate (Table 4). In contrast, among studies in the top tertile (90%–97%), the directions of the associations were reversed, in that cases were more likely rather than less likely to be firstborn and were less likely to come from large sibships. Similarly, the positive associations in the highest category of birth order and sibship size were confined to studies with the lowest control response rate tertile (Table 5). Reported response rates tended to be lower in population-based studies than in hospital-based studies, and positive trends with birth order and sibship size occurred in population-based studies (birth order: P-trend < 0.001; sibship size: P-trend < 0.001) but not in hospital-based studies (birth order: P-trend = 0.391; sibship size: P-trend = 0.118). Because of this overlap, it was impossible to separate the impacts of study design and response rates on the study-specific odds ratios.

Table 4.

Odds Ratios for Non-Hodgkin Lymphoma According to Birth Order/Sibship Size, by Response Rate Among Cases, in a Pooled Analysis of Results From 18 Non-Hodgkin Lymphoma Case-Control Studies Within the International Lymphoma Epidemiology Consortium (InterLymph), 1983–2005

Familial Structure Variable	Tertile of Response Rate in Cases (Range)
	Low (65%–76%)				Medium (81%–87%)				High (90%–97%)
	No. of Cases	No. of Controls	ORa	95% CI	No. of Cases	No. of Controls	ORa	95% CI	No. of Cases	No. of Controls	ORa	95% CI
Birth order
Firstborn	1,196	1,641	1.00		2,300	2,001	1.00		413	638	1.00
Second-born	838	1,156	1.02	0.91, 1.15	1,615	1,469	1.05	0.90, 1.24	288	540	0.84	0.68, 1.04
Third-born	474	634	1.09	0.94, 1.26	853	872	1.20	1.02, 1.40	174	354	0.75	0.54, 1.05
Fourth-born	275	318	1.29	1.07, 1.55	452	431	1.18	0.89, 1.57	110	207	0.78	0.59, 1.03
Fifth-born or higher	388	423	1.34	1.14, 1.58	569	635	1.26	0.97, 1.63	205	409	0.77	0.61, 0.96
P-trend			<0.001				0.011				0.018
No. of studies			4				5				5
Sibship size
Only child	358	484	1.00		828	974	1.00		97	123	1.00
No. of other siblings
1	928	1,288	1.03	0.87, 1.21	1,624	1,858	1.05	0.93, 1.19	268	445	0.78	0.53, 1.15
2	930	1,207	1.11	0.94, 1.31	1,473	1,612	1.08	0.96, 1.23	265	471	0.75	0.54, 1.04
3	636	877	1.07	0.90, 1.28	1,072	1,039	1.23	1.07, 1.40	194	398	0.62	0.36, 1.06
4	450	592	1.14	0.94, 1.38	651	690	1.12	0.96, 1.30	147	309	0.61	0.36, 1.04
≥5	960	1,069	1.31	1.11, 1.56	1,321	1,270	1.21	1.02, 1.44	439	896	0.58	0.41, 0.82
P-trend			<0.001				0.004				0.001
No. of studies			6				6				6

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

^aOdds ratios and 95% confidence intervals were computed using a 2-stage random-effects model, adjusted for age (in 5-year intervals), sex, and study center.

Table 5.

Odds Ratios for Non-Hodgkin Lymphoma According to Birth Order/Sibship Size, by Response Rate Among Controls, in a Pooled Analysis of Results From 18 Non-Hodgkin Lymphoma Case-Control Studies Within the International Lymphoma Epidemiology Consortium (InterLymph), 1983–2005

Familial Structure Variable	Tertile of Response Rate in Controls (Range)
	Low (44%–61%)				Medium (66%–74%)				High (75%–99%)
	No. of Cases	No. of Controls	ORa	95% CI	No. of Cases	No. of Controls	ORa	95% CI	No. of Cases	No. of Controls	ORa	95% CI
Birth order
Firstborn	1,047	1,106	1.00		1,794	2,002	1.00		769	1,471	1.00
Second-born	746	760	1.02	0.82, 1.26	1,282	1,430	1.01	0.89, 1.14	567	1,121	1.00	0.87, 1.14
Third-born	435	380	1.27	1.08, 1.50	730	751	1.00	0.78, 1.27	355	710	0.89	0.69, 1.15
Fourth-born	240	183	1.45	1.17, 1.80	371	400	1.06	0.83, 1.36	205	394	0.88	0.64, 1.22
Fifth-born or higher	267	215	1.34	1.00, 1.79	572	528	1.08	0.79, 1.48	389	658	0.96	0.70, 1.31
P-trend			<0.001				0.577				0.511
No. of studies			5				5				4
Sibship size
Only child	299	337	1.00		797	926	1.00		187	318	1.00
No. of other siblings
1	836	890	1.05	0.87, 1.27	1,442	1,651	1.03	0.89, 1.18	542	1,050	0.94	0.76, 1.16
2	788	774	1.14	0.95, 1.38	1,319	1,445	1.06	0.93, 1.20	561	1,071	0.93	0.73, 1.19
3	565	534	1.20	0.98, 1.47	915	910	1.13	0.92, 1.38	422	870	0.72	0.46, 1.13
4	342	360	1.06	0.85, 1.32	595	586	1.15	0.93, 1.42	311	645	0.73	0.49, 1.09
≥5	641	573	1.28	1.04, 1.56	1,210	1,234	1.06	0.81, 1.40	869	1,428	0.85	0.54, 1.33
P-trend			0.012				0.206				0.351
No. of studies			6				6				6

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

^aOdds ratios and 95% confidence intervals were computed using a 2-stage random-effects model, adjusted for age (in 5-year intervals), sex, and study center.

There was also variation in the pooled odds ratios when data were examined according to the demographic features of the respondents. When results were stratified by SES, the positive trend with birth order was not found in the lowest tertile, was of borderline significance in the middle tertile, and was significant in the top tertile (Table 6). For increasing sibship size, a positive trend was also present only in the top tertile of SES. Positive trends were closer to statistical significance among women (birth order: P-trend = 0.008, sibship size: P-trend = 0.052) than among men (birth order: P-trend = 0.088; sibship size: P-trend = 0.167).

Table 6.

Odds Ratios for Non-Hodgkin Lymphoma According to Birth Order/Sibship Size, by Socioeconomic Status of Participants, in a Pooled Analysis of Results From 18 Non-Hodgkin Lymphoma Case-Control Studies Within the International Lymphoma Epidemiology Consortium (InterLymph), 1983–2005

Familial Structure Variable	Tertile of Participant Socioeconomic Statusa
	Low				Medium				High
	No. of Cases	No. of Controls	ORb	95% CIb	No. of Cases	No. of Controls	OR	95% CI	No. of Cases	No. of Controls	OR	95% CI
Birth order
Firstborn	1,151	1,262	1.00		1,292	1,754	1.00		1,138	1,525	1.00
Second-born	885	1,001	1.00	0.86, 1.17	934	1,273	1.02	0.91, 1.15	755	1,009	1.06	0.92, 1.23
Third-born	573	670	0.93	0.77, 1.14	544	662	1.21	1.00, 1.46	385	496	1.11	0.93, 1.32
Fourth-born	345	389	1.03	0.87, 1.23	288	362	1.24	0.98, 1.58	180	225	1.14	0.86, 1.49
Fifth-born or higher	608	654	1.07	0.87, 1.32	377	516	1.11	0.90, 1.37	236	225	1.49	1.14, 1.95
P-trend			0.548				0.067				0.007
Sibship size
Only child	439	436	1.00		408	575	1.00		432	561	1.00
No. of other siblings
1	884	978	0.91	0.77, 1.08	1,005	1,306	1.15	0.97, 1.35	910	1,283	1.00	0.84, 1.18
2	915	1,038	0.91	0.77, 1.08	935	1,178	1.17	0.98, 1.39	797	1,058	1.11	0.93, 1.32
3	751	814	0.99	0.83, 1.18	655	848	1.15	0.88, 1.51	482	637	1.06	0.87, 1.30
4	563	685	0.89	0.73, 1.07	404	524	1.14	0.86, 1.52	273	368	1.08	0.86, 1.37
≥5	1,521	1,705	1.03	0.87, 1.22	768	1,054	1.05	0.78, 1.40	416	467	1.31	1.06, 1.62
P-trend			0.146				0.803				0.008

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

^aSocioeconomic status groups in each study were based on the tertile distribution of years of education (11 studies) or area-based deprivation levels obtained from census data (11).

^bOdds ratios and 95% confidence intervals were computed using a 2-stage random-effects model, with adjustment for age (in 5-year intervals), sex, and study center.

The associations with birth order and sibship size were not specific to any particular NHL subtype. Statistically significant positive trends with increasing birth order were found for B-cell NHL overall (P-trend < 0.001) and for follicular NHL (P-trend = 0.001), diffuse large B-cell NHL (P-trend = 0.002), and precursor B-cell NHL (P-trend = 0.025). Trends were also significant for T-cell NHL (P-trend = 0.002) and for the mycosis fungoides subtype (P-trend = 0.039). For increasing sibship size, significant positive trends were seen for B-cell NHL (P-trend < 0.001), follicular NHL (P-trend = 0.012), and B-cell lymphoma not otherwise specified (P-trend = 0.004) and were of borderline significance for diffuse large-cell NHL (P-trend = 0.062). However, the majority of lymphoma subtypes, including the category of NHL not otherwise specified, were not associated with either birth order or sibship size.

There was no association with number of younger siblings, age difference with the nearest older or younger sibling, or sharing a bed or bedroom as a child (data not shown).

DISCUSSION

In this large pooled analysis of data from case-control studies, we found no significant association between birth order and sibship size and risk of NHL overall. However, significant positive associations were present for a number of B- and T-cell lymphoma subtypes. Substantial heterogeneity in study-specific risks was observed, which may reflect variations in the association by response rate, study population, and/or SES. The positive associations were confined to population-based studies, which had lower response rates in cases and controls, and were absent in hospital-based studies, which had higher response rates. After stratifying results by SES, the positive associations were present among persons of upper SES but not those of lower SES. The fact that the birth order association was confined to studies with lower response rates and to upper-SES participants and was not specific to particular B- or T-cell subtypes suggests that the previously reported positive association of NHL risk with birth order and sibship size in case-control studies may have been due to selection bias mediated by SES.

Selection bias may have arisen because of differential response rates by SES. In almost all developed countries, lower SES is strongly associated with a higher total fertility rate, and therefore with birth order of the offspring (31–33). Although we did not have data on parental SES, the SES of parents correlates strongly with that of their children (34), so participant SES is likely to be correlated with the participant's own birth order. Among controls in this study, increasing participant SES was strongly inversely correlated with the participant's own birth order (P < 0.001). Low response rates increase the probability of selection bias. Thus, our finding that the positive association with birth order was confined to those studies with lower response rates is consistent with the association's being due to selection bias. However, low response rates do not in themselves cause selection bias (35). Selection bias arises only if participation differs between exposed (or nonexposed) cases and exposed (or nonexposed) controls. A SES-related selection bias leading to a positive association between NHL risk and birth order can arise in 2 main ways. First, among controls, it can arise if the participation rate is higher among persons with a low birth order. Second, among cases, it can arise if the participation rate is higher among persons of high birth order. Because SES and birth order are correlated, this equates to a higher participation rate among high-SES controls than among low-SES controls or a higher participation rate among low-SES cases than among high-SES cases.

There is consistent evidence that in population-based case-control studies, the response rate is higher in high-SES controls than in low-SES controls. This has been shown for controls recruited through a variety of methods, including random digit dialing (36–38), contact with their general practitioners (39, 40), and letter or telephone after random selection from health insurance (41), electoral (42), or census (43) rolls. This selection bias would tend to generate a positive association between birth order and disease risk. However, this does not explain why the effect of birth order was most prominent in the high-SES stratum (Table 6), in which response rates are likely to be higher.

Since response rates are generally substantially lower in controls than in cases in population-based studies, there is less potential for selection bias to arise from factors associated with case selection. In case-control studies of cancer, the reasons for case nonparticipation are clearly different from the reasons for control nonparticipation. The most common reasons include having died of cancer prior to interview or being too ill to participate (11). In a population-based Italian case-control study of cancer, Richiardi et al. (43) reported that nonparticipating cases were of high SES rather than low SES, and they hypothesized that this may have been because more educated and wealthy subjects retained more independence when hospitalized for a serious disease. However, other cancer case-control studies have identified a pattern similar to that of controls, with higher participation among cases of higher SES (38, 44, 45).

In addition to selection bias, there is at least 1 alternative explanation for the lack of an association between birth order and NHL risk in studies with the highest response rates included in our analysis. Most (but not all) of the studies with high response rates, particularly among controls, were hospital-based case-control studies. Even when response rates are high, studies of this design are prone to different selection biases (46). For example, if all patients with NHL are treated in a hospital but low SES predicts hospital attendance for other common conditions such as trauma or cardiovascular disease, then low-SES (high birth order) people may be overrepresented among hospital-based controls. Among hospital-based studies in our data set, birth order was not associated with NHL risk. Among population-based studies, which are generally regarded as methodologically superior to hospital-based studies, there was a significant association. However, because these were also the studies with the lowest response rates, it was not possible to conclude that this positive association was likely to reflect a causal association. Even among population-based studies, the positive association between birth order and NHL varied by case and control response rates (data not shown).

If the birth order association is due to selection bias, a similar effect is likely to occur in case-control studies of birth order as a risk factor for other diseases. Birth order has been studied in some detail as a risk factor for Hodgkin lymphoma. Investigators in population-based cohort studies (3, 47) and case-control studies with low refusal rates (9, 48–50) are consistent in reporting that increasing birth order and/or sibship size is associated with decreasing Hodgkin lymphoma risk, especially the young adult type. However, in some recent case-control studies which had lower response rates in controls, researchers reported no association with low birth order or other markers of higher SES (51–53). One possibility for the lack of association is a temporal change in early-life child-care arrangements, resulting in increased early-life exposure to infection that mutes the effect of birth order. Alternatively, in 1 of these studies, Glaser et al. (54) reported that the absence of an effect was probably related to a combination of the low response rate in controls and the fact that controls who participated were of higher SES than those who did not participate. This is consistent with the form and direction of selection bias which may have occurred in NHL research: The difference is that for Hodgkin lymphoma, selection bias has sometimes obscured a true inverse relation with birth order, whereas for NHL it may have created the impression of a positive relation. For other types of cancer, birth order has been less intensively studied, although patterns of increased and decreased risk have been described for certain cancer sites in cohort studies, perhaps reflecting the effect of SES on cancer risk (55).

In case-control studies that have observed an association of NHL risk with increasing birth order, researchers have interpreted the finding as suggesting that the immunologic consequences of early infection may be responsible for increased NHL risk in persons of late birth order (4, 6, 7). Our finding of no significant association between NHL risk and birth order does not rule out an effect of early life environment, but it does suggest that such an effect, if present, is unlikely to be reliably mediated by birth order.

The large size of this pooled analysis using individual data allowed stratification by study design and participant demographic characteristics, thus enabling exploration of factors related to the weak association observed for some NHL subtypes. In previous InterLymph pooled analyses we have carried out, for atopic diseases and for autoimmune disorders, there was very little difference in the pattern or significance of results when we stratified results by these design features or demographic variables (56, 57). A possible reason for the heterogeneity of the findings we have demonstrated for birth order is that the infection-related consequences of birth order are highly variable among the countries in which these studies were conducted. In terms of selection bias, unfortunately investigators in most of the participating studies did not collect data on nonresponders to allow direct investigation of the hypothesis presented here. However, in 1 population-based study, Mensah et al. (11) investigated nonresponse and reported that both the participating cases and controls were of higher-than-expected SES. Shen et al. (37) reported varying results on participation rate by SES according to individual study center. When response rates are substantially lower in controls than in cases, higher participation rates in persons of higher SES will lead to selection bias, causing an association between increasing birth order and NHL risk, as we have hypothesized may be the case. Because of variations in how investigators in participating studies reported nonresponse, it is likely that there was some misclassification of true nonresponse rates. The absence of universally accepted criteria for reporting response rates (58) and a lack of consistency between studies in recording of details of nonresponse precluded any recalculation of standardized response rates in this pooled analysis.

In summary, we did not find a significant overall association between birth order or sibship size and risk of NHL in this large pooled analysis. However, a positive association was found in population-based studies and among people in the highest SES stratum. Sensitivity analyses by study design factors and participant characteristics, a strength of pooled analyses, suggested that a likely explanation for the weak positive association between birth order and NHL risk reported in some case-control studies may be selection bias. The fact that low-SES controls in this pooled analysis were more likely to be of high birth order and were less likely to participate may have generated a weak positive association between NHL risk and birth order. More generally, our results show that because of the close association between birth order and SES and between SES and subject nonresponse, case-control studies of birth order as a risk factor for disease will only be valid if the response rate is very high in both cases and controls and if other sources of selection bias related to SES are eliminated. Confounding or selection bias due to SES should be ruled out in all studies which have identified birth order as a risk factor for disease. The association is best explored in population-based case-control studies with high response rates, or in large prospective or retrospective cohort studies that utilize linked cancer and birth registry records.

Glossary

Abbreviations

NHL

non-Hodgkin lymphoma

SES

socioeconomic status