As part of AP&T's peer-review process, a technical check of this meta-analysis was performed by Mr M. Sajid.
Meta-analysis: the risk of venous thromboembolism in patients with inflammatory bowel disease
Version of Record online: 2 APR 2013
© 2013 Blackwell Publishing Ltd
Alimentary Pharmacology & Therapeutics
Volume 37, Issue 10, pages 953–962, May 2013
How to Cite
Aliment Pharmacol Ther 2013; 37: 953–962
- Issue online: 17 APR 2013
- Version of Record online: 2 APR 2013
- Manuscript Accepted: 11 MAR 2013
- Manuscript Revised: 20 FEB 2013
- Manuscript Revised: 26 JAN 2013
- Manuscript Received: 30 DEC 2012
Inflammatory bowel disease (IBD), which includes Crohn's disease (CD) and ulcerative colitis (UC), is a systemic disorder that predominantly affects the bowels but is also associated with venous thromboembolism (VTE).
To provide a quantitative assessment of the association of IBD with venous thromboembolism risk and to explore the possible sources of heterogeneity in the current literature, a meta-analysis of case–control and cohort studies was conducted.
Studies were identified by a literature search of the PubMed and Scopus databases (from inception inclusive 31 December 2012) for English language studies. Summary relative risks (RRs) with 95% confidence intervals (CIs) were calculated with fixed- and random-effects models. Several subgroup analyses were performed to explore potential study heterogeneity and bias.
Eleven studies met our inclusion criteria. The summary RR for deep venous thromboembolism (DVT) and pulmonary embolism (PE) comparing subjects both with and without IBD was 2.20 (95% CI 1.83–2.65). After adjusting for obesity and smoking, summary relative risks near 2.0 were seen for venous thromboembolism in both UC and CD patients.
This meta-analysis showed that inflammatory bowel disease is associated with an approximately two-fold increase in the risk of venous thromboembolism.
Venous thromboembolism (VTE) affects between 0.5 and 2 persons per thousand in the general community and is a major cause of death in hospitalised patients.[1-3] Thrombosis can start with the formation of a venous clot (thrombosis), often occurring in the deep veins of the legs, thighs, or pelvis, and can often be asymptomatic. The risk factors for venous thrombosis are primarily related to hypercoagulability, which can be either genetic (i.e. procoagulant proteins, protein C anticoagulant pathway, fibrinolytic proteins, homocysteine) or acquired as a result of immobilisation and venous stasis. Acquired risk factors for venous thromboembolism include surgery, trauma, pregnancy, puerperium, lupus anticoagulants, malignant disease, oral contraceptives, long-haul travel, obesity, myeloproliferative disorders and polycythemia vera.[4, 5] Deep vein thrombosis (DVT) and pulmonary embolism (PE) are the most common types of thromboembolism, but thromboses are also reported in unusual sites such as cerebral, innominate, retinal, hepatic and mesenteric veins.[6-9]
Inflammatory bowel disease (IBD), which includes Crohn' s disease (CD) and ulcerative colitis (UC), is a systemic disorder that predominantly affects the bowels, but is also associated with a number of extraintestinal manifestations, some of which may be more debilitating than the bowel symptoms. The possible association between IBD and VTE was first reported in 1936 by Bargen and Barker, who described 18 patients with thromboembolic disease (predominantly venous) from among more than 1000 patients treated for IBD at the Mayo Clinic. Since that time, several recent publications have suggested that patients with IBD have an increased risk of VTE, although some heterogeneity can be seen in the literature.[12-28] For example, while Nguyen et al. reported that IBD patients may have a more than sevenfold increase in the risk of VTE, Saleh et al. identified a much lower risk.
We conducted a meta-analysis of case-control and cohort studies to provide a quantitative assessment of the association of IBD with VTE risk, and to explore the possible sources of heterogeneity in the current literature. This topic is important as VTE could be a major cause of morbidity and mortality in people with IBD, and new information on the possible association between IBD and VTE could help raise awareness of this issue and assist in establishing guidelines for improving the health of patients with IBD.
Studies were identified by a literature search of the PubMed (from1961 through 31 December 2012) and Scopus databases (from 1978 inclusive 31 December 2012) for English language studies with the following medical subject heading terms and/or text words: venous thromboembolism, deep vein thrombosis, pulmonary embolism, inflammatory bowel disease, ulcerative colitis and Crohn's disease. We also reviewed the reference lists of the identified publications and relevant review articles for additional pertinent studies.
Inclusion and exclusion criteria
Published reports that met the following criteria were included: (i) evaluated risk of first venous thromboembolism, deep vein thrombosis, or pulmonary embolism among patients with Crohn' s disease and/or ulcerative colitis; (ii) had a case–control or cohort design which included a comparison group that lacked inflammatory bowel disease; (iii) reported an estimate of relative risk (RR), either an odds ratio (for case-control studies) or a rate or risk ratio (for cohort studies), accompanied by a corresponding measure of uncertainty [i.e. 95% confidence interval (CI), standard error, variance, or P value]; (iv) evaluated adult patients (>20 years old); and (v) evaluated age and gender as potential confounders. For multiple reports on the same population or subpopulation, we used either the relative risk estimates from the most recent report or the one containing the most cases. These criteria specifically excluded studies that evaluated cerebral thromboembolism or recurrent VTE and those that had a cross-sectional, ecological, or prevalence design.
The data extracted included the year of publication, country of the population studied, study design, number of cases, number of exposed and unexposed subjects (cohort studies), number of controls, source of the controls (case-control studies), follow-up period (for cohort studies), type of IBD (UC or CD), RR estimates with their corresponding confidence intervals (CIs) and variables controlled for by matching or in the multivariable models. The data were abstracted separately by two authors (HY and CS); discordant results were resolved by consensus.
Summary RR estimates were calculated using both the fixed-effects (inverse-variance weighted) method and the random-effects method. Odds ratios were assumed to provide estimates of RR. Statistical heterogeneity between studies was evaluated with Cochran's Q-test and the I2-statistic. I2 describes the percentage of total variation across studies due to heterogeneity rather than chance. I2 can be readily calculated from basic results obtained from a typical meta-analysis as follows:
where Q is Cochran's heterogeneity statistic and df is the degrees of freedom. Publication bias was assessed by constructing a funnel plot and using Egger and Begg tests.[33, 34] In addition, to further evaluate the possibility of publication bias, and due to the fact that publication bias primarily affects smaller, less precise studies, reports were divided into two groups based on their sample size and precision. Owing to the fact that the funnel plot showed a clear distinction in precision at this point, the five least precise studies were able to be compared to the other studies. All statistical analyses were carried out with STATA, version 11.0 (Stata Corp, College Station, TX, USA). P values < 0.05 were considered statistically significant. All statistical tests were two-sided. The analysis and reporting is consistent with the guidelines recommended for meta-analyses of observational studies.
Detailed search steps are described in Figure 1. The electronic search identified 403 abstracts for review. From among these, 38 articles potentially met inclusion criteria and were retrieved for detailed review. The bibliographies of all 38 studies were reviewed and three additional studies were identified, providing 41 articles for full review. We excluded one candidate study due to overlapping publications and one study because it reported relative risks that were not adjusted by age. Eleven studies[12-22] met our predefined inclusion criteria and were thus included in the final analyses, the details of which can be seen in Table 1.
|Case-control studies of IBD and DVT and/or PE incidence|
|Author, year (Ref. No), country||No. of case patients by subsites||No. of control subjects (source of controls)||Type of IBD||RR (95% CI) thrombosis subsite||Controlled variables|
|Kappelman 2011 Denmark|| |
125 963 VTE
(71 367 DVT
54 108 PE)
|Population controls||UC and CD|| |
a1.5 (1.2–1.8) VTE
1.5 (1.2–1.9) DVT
1.5 (1.1–2.1) PE
Age, gender, malignancy, surgery, fracture, pregnancy, congestive heart failure, diabetes,
Myocardial infarction, stroke hormone replacement, antipsycholic
|Cohort studies of IBD and DVT and/or PE incidence based on incidence rate ratios|
|Author, year (Ref.No), country, (follow-up-period)||Study population||RR (95% CI),thrombosis subsite (number of cases)||Controlled variables|
|Bernstein 2001 Canada (1984–1997)|| |
Manitoba Health administrative
Exposed group: 6027 persons with IBD Comparison group: 5529 matched controls
3.47 (2.94–4.09) VTE (187)
4.10 (3.21–5.25) CD VTE (89)
3.04 (2.43–3.81) UC VTE (98)
3.54 (2.89–4.34) DVT (126)
4.71 (3.53–6.29) CD DVT (67)
2.77 (2.07–3.69) UC DVT (59)
3.32 (2.49–4.32) PE (61)
2.94 (1.83–4.73) CD PE (22)
3.59 (2.49–5.17 UC PE (39)
|Age, gender, postal area of residence|
|Mieshler 2004 Austria (N/A)|| |
Three out-patient clinics of Division of Gastroenterology and Hepatology Exposed group: 618 patients with IBD
Comparison group: 618 controls matched by age and gender
|3.6 (1.7–7.8) VTE (48)||Age, gender, operation, injuries, oral contraceptive use, pregnancy, BMI, smoking|
|Bernstein 2007 Canada(1994–2004)||The Statistics Canada's Health Person Oriented Information database|| |
50 years and older
a1.3 (1.23–1.37) VTE (N/A)
1.32 (1.24–1.41) DVT (N/A)
1.26 (1.13–1.39) PE (N/A)
Younger than 50 years
a1.57 (1.42–1.72) VTE (N/A)
1.53 (1.37–1.71) DVT (N/A)
1.67 (1.39–2.01) PE (N/A)
|Nguyen 2008 USA (1998–2004)|| |
Nationwide Inpatients Sample (NIS) Exposed group: 73197 CD discharges
43645 UC discharges
Comparison group: 1% random sample of non-IBD hospital discharges
a1.66 (1.33–2.06) (N/A)
1.48 (1.35–1.62)CD VTE (N/A)
1.85 (1.70–2.01) UC VTE (N/A)
|Age, gender, health insurance carrier, neighbourhood income, respective bowel surgery, calendar year, hospital characteristics|
|Nguyen 2009 USA (2005)|| |
Nationwide Inpatients Sample (NIS)
Exposed group: 694 CD patients and
387 UC patients in obstetric hospitalisation
Comparison group:964392 non-IBD patients in obstetric hospitalisations
a7.07 (4.07–12.3) (N/A)
6.12 (2.91–12.9) CD VTE (N/A)
8.44 (3.71–19.2) UC VTE (N/A)
|Maternal age, race/ethnicity, median neighbourhood income, comorbidity, health insurance, geographical region, hospital location and teaching status, Caesarean delivery|
|Grainge 2010 UK (1987–2001)|| |
General Practice Research Database(GPRD) Exposed group: 13 756 IBD patients
Comparison group: 71 672 matched controls
|3.4 (2.7–4.3) UC VTE (784)||Age, gender, BMI, smoking, cancer diagnosis, history of PE or DVT|
|Rothberg 2011 USA (2004–2005)|| |
Patients discharged from 374 acute care facilities in the United States that participated in Premier's Perspective
Exposed group: 814 IBD patients
Comparison group: 241924 non-IBD patients
|3.11 (1.59–6.08) VTE (1052)||Age, gender, VTE prophylaxis, length of stay> 6 days, pneumonia, chronic obstructive pulmonary disease, stroke, congestive heart failure, urinary tract infection, respiratory failure, septicaemia, obesity, inherited thrombophilia, cancer, central venous catheter, mechanical ventilation, urinary catheter, chemotherapy, steroids|
|Merrill 2011 USA (2008)|| |
National Surgical Quality Improvement Program Participant Use Data File
Exposed group:2249 IBD patients Comparison group: 269 119 patients without IBD
|2.03 (1.52–2.7) VTE (2665)||Age, gender, race/ethnicity, admitted from home, smoker, BMI>30, medical historyb, clinical factorc|
|Saleh 2011 USA (1979–2005)||The National Hospital Discharge Survey|| |
1.64 (1.62–1.66) UC VTE (21 000)
1.08 (1.06–1.09) CD VTE
1.77 (1.74–1.80) UC DVT
1.24 (1.22–1.26) CD DVT
1.40 (1.37–1.44) UC PE
0.74 (0.72–0.76) CD PE
Bröms 2012 Sweden
The Medical Birth, Patient, and Prescribed Drug Registers of all residents in Sweden
Exposed group:1209 UC patients and 787 CD patients
Comparison group: 10773 patients without IBD
a2.61 (1.24–5.49) VTE
3.78 (1.52–9.38) UC VTE(8)
1.26 (0.35–4.53) CD VTE(3)
|Age, parity, year and month of birth,|
In total, one case-control and 10 cohort studies. No study received greater than 13% of the total weight in the meta-analysis.
The overall results of the meta-analysis are shown in Figure 2 and Table 2. All individual studies had RR estimates above 1.0 with statistical significance. The summary RR for DVT and PE comparing subjects with and without IBD was 2.20 (95% CI, 1.83–2.65) for the 11 studies combined. Evidence of heterogeneity was observed (Q = 243.0 P < 0.01, I2 = 95.9%). A sensitivity analysis identified the study by Saleh et al. as the largest contributor to this heterogeneity. In an analysis performed that excluded this study, the summary RR between IBD and VTE was somewhat higher [summary RR = 2.53, 95% CI (1.79–3.58)], although the test for heterogeneity was still statistically significant (Q = 219.7 P < 0.01, I2 = 95.9%).
|Subgroup||No. of studies||Summary RR (95% CI) Fixed effect model||Q||P-heterogeneity||I 2 Statistic%||Summary RR (95% CI) Random effect model|
|All studies||11||1.37 (1.36–1.38)||243.0||<0.01||95.9||2.20 (1.83–2.65)|
|Smaller size||5||4.16 (2.98–5.80)||16.1||<0.01||75.2||3.29 (1.98–5.47)|
|Larger size||6||1.37 (1.36–1.38)||187.5||<0.01||97.3||1.90 (1.55–2.31)|
|DVT||4||1.45 (1.37–1.54)||82.4||<0.01||96.4||1.80 (1.11–2.91)|
|PE||4||1.42 (1.29–1.55)||42.8||<0.01||93.0||1.63 (0.98–2.72)|
|UC||6||1.65 (1.63–1.67)||91.9||<0.01||94.6||2.57 (2.02–3.28)|
|CD||5||1.09 (1.08–1.11)||176.9||<0.01||97.7||2.12 (1.40–3.20)|
|Adjust for smoking and BMI|
|Yes||3||2.81 (2.36–3.35)||7.91||0.02||74.7||2.20 (1.83–2.65)|
|No||8||1.37 (1.36–1.38)||171.4||<0.01||95.9||1.99 (1.65–2.41)|
Evaluation of Heterogeneity
We used various subgroup analyses to further evaluate several potential sources of heterogeneity. Table 2 shows the results of subgroup meta-analysis by study size, thrombosis sites, IBD type and adjustment for confounding factors [smoking and body mass index (BMI)].
The association between IBD and VTE incidence was significant for deep vein thrombosis [summary RR = 1.80, 95% CI (1.11–2.91); n = 4 studies] but not for PE [summary RR = 1.63, 95% CI (0.98–2.72); n = 4 studies]. The evaluation of differences in the effect size by disease type (i.e. Crohn's disease vs. ulcerative colitis) demonstrated that the association between IBD and VTE incidence was similar in patients with UC [summary RR = 2.57, 95% CI (2.02–3.28); n = 6 studies] and CD [summary RR = 2.12, 95% (CI 1.40–3.200; n = 5 studies]. In the subgroup analysis that was restricted to the three publications that controlled for smoking and obesity, the positive association of IBD with VTE [summary RR = 2.20, 95% CI (1.83–2.65); n = 3 studies] remained.
Visual inspection of the Begg funnel plot demonstrated the asymmetry typically associated with publication bias (Figure 3). That is, smaller, less precise studies [those with the larger standard errors (S.E.)] appear to have higher RRs than the large, more precise studies. Evidence of publication bias was also seen with the Egger or Begg tests (Egger P = 0.02). On visual inspection of the Begg funnel plot, five outlying studies appeared to be the cause of the asymmetry. After removing these five studies, no evidence of publication bias could be found (Egger P = 0.16).
This meta-analysis showed increased risk of VTE in patients with IBD. Overall, this study had several key findings. First, the elevated summary RRs for all studies combined, as well as those in the subgroup analyses of both venous thrombosis and PE, provide strong evidence that patients with IBD are at increased risk (approximately twofold) for both DVT and PE. Second, the increased RRs are similar, regardless of whether or not the studies were adjusted for major potential confounders like smoking or BMI. The similarity of adjusted and unadjusted RRs suggests that while smoking and BMI may increase the risk of DVT or PE, they did not confound the relationship between IBD and DVT/PE.
Third, both the funnel plots and the Egger test suggest that publication bias in the literature may be artificially inflating the apparent association between VTE and IBD. However, publication bias typically affects studies with smaller sample sizes. Consistent with this finding, the summary RR of the five smallest studies [RR = 3.29, 95% CI (1.83–2.65)] was substantially higher than that of the six largest studies [RR = 1.90, 95% CI (1.55–2.31)]. However, analyses limited to only the six largest studies provided RRs that were only moderately smaller than analyses that contained all the studies (summary RR of 1.90 vs. 2.20, respectively). This suggests that publication bias was not the sole cause of the approximately twofold increase in risk identified here.
Fourth, although our results showed that patients with UC may be at a somewhat greater risk for VTE than those with CD, the difference is small; furthermore, elevated RRs were seen in both groups. The possibility of a greater risk among patients with UC, in which rates of DVT or PE were 50/1000 person-years in UC and 40/1000 person-years in CD, was suggested by Bernstein et al. Bernstein's results and this meta-analysis both suggest that health care providers should be aware of the increased risk for both CD and UC patients.
Data from several sources suggest that the association between IBD and the risk of VTE is biologically plausible. IBD creates localised and systemic inflammation, and activation of the coagulation cascade occurs during inflammation. Proinflammatory cytokines such as IL-6 have been shown to stimulate coagulation without concomitant fibrinolysis. Fibrinolysis is depressed in patients with IBD, with decreased plasma levels of tPA activity and increased plasminogen activator inhibitors reported. Thus, inflammation represents a state of relative hypofibrinolysis, promoting a hypercoagulable state.[38-40] Local and systemic inflammation, together with other required risk factors such as surgery, prolonged immobilisation, central venous catheters, fluid depletion, steroid therapy, smoking, oral contraceptives, high levels of antiphospholipid antibodies and hyperhomocysteinaemia (because of vitamin deficiencies), may induce a hypercoagulable state and prothrombotic conditions in patients with IBD.[14, 41]
Several limitations of this study must be considered. All of the studies that assess the association between IBD and VTE are observational studies, and are thus subject to bias. Although adjustment for smoking or BMI actually increased the strength of the association between IBD and VTE, we were unable to evaluate variables such as diet, alcohol and other health conditions. However, despite the inherent limitations of observational studies, the association we found between IBD and an increased risk of VTE may still be valid for several reasons. First, the strength of the associations identified decrease the possibility that they are due to chance alone. Second, all 11 studies included in this meta-analysis reported a RR above 1.0 with statistically significant results. Although some statistical heterogeneity was identified, which makes the exact magnitude of the association less certain, the consistency of the RRs among the 11 studies, despite different study designs, populations and methods, suggests that the positive direction of the association is also not due to chance alone.
We found little evidence that these findings were due to confounding factors. Several of the studies were adjusted for the most likely confounders, including smoking and BMI. The positive association of IBD with VTE risk increased when the meta-analysis was limited to studies that controlled for smoking and BMI [summary RR of studies adjusting for the confounders 2.58, 95% CI (1.89–3.51) and summary RR of studies unadjusted for the confounders 2.10, 95% CI (1.38–3.19)]. This suggests that obesity and smoking do not explain the association between IBD and VTE.
Disease activity seems to be highly associated with thromboembolic risk, as it has been suggested that in IBD, activation of the coagulation cascade is more pronounced in active disease. However, in this meta-analysis, we could not do subgroup analyses based on IBD disease activity. A study by Grainge et al. was the only one to evaluate the association between IBD disease activity and the occurrence of VTE in their population-based study in the United Kingdom. They found a 4.5-fold higher rate of VTE during acute disease flares when compared with periods of remission among IBD patients. The Grainge study selected IBD patients experiencing moderate-to-severe flares, given that a disease flare was defined by the need for corticosteroids. It is possible that corticosteroid therapy may have contributed to an additional risk for the development of VTE in that study. Two studies reported that between 60% and 80% of IBD patients have some component of active disease at the time of VTE diagnosis.[13, 14]
There are several potential reasons for the residual heterogeneity. First, there may have been differences in the anatomical location of IBD between patients. We were not able to do a subgroup analysis based on the anatomical location of IBD because only two studies, one by Miesher et al. and the other by Nguyen et al.[14, 17], in this meta-analysis specified the location of IBD. A study by Solem et al., which was excluded from this meta-analysis due to the RRs of VTE risk among IBD patients being unavailable, showed that CD patients with VTE typically have colonic disease involvement (ileocolonic in 56% and colonic in 23%), and that most UC patients with VTE (76%) have pancolonic disease. This implies that the extent of colonic disease in IBD may correlate with a patient's thromboembolic risk.
Second, because VTE is difficult to diagnose, some degree of differential misclassification is likely to have occurred in the studies included in this meta-analysis. VTE is often clinically silent and in many cases, the first sign of the disease is a sudden fatal PE.[44, 45] Despite modest increases in the antemortem diagnosis of PE over the years, PE diagnosis is confirmed by objective testing in only about 20% of patients.
Third, in this meta-analysis we included a number of very different types of studies of different populations and settings. Some studies are population-based cohorts, some are from referral centres, and others look only at hospitalised patients through discharge databases. This variety could have contributed to the heterogeneity among the studies (>90%).
Fourth, detection bias could lead to artificially increased RRs in this meta-analysis. For example, due to more rigorous medical follow-up of IBD patients, subclinical events may be detected more often in IBD patients than in non-IBD patients. However, given the seriousness of the outcomes assessed, especially PE, it seems unlikely that this bias would account for the entire twofold increased risks identified here.
Our results have important clinical and public health implications. This study showed an increased risk of VTE among IBD patients, which we hope will help increase awareness of this subject. Based on this emerging evidence, several practice guidelines now include recommendations on VTE risk for patients with IBD. For example, the American College of Gastroenterology (ACG), UC practice guidelines state that physicians should consider prophylaxis with heparin in patients hospitalised for a severe disease flare. However, no such recommendations are contained within the current ACG CD guidelines. Guidelines for the treatment of IBD by the British Society of Gastroenterology recommend pharmacological prophylaxis for VTE in hospitalised patients with severe UC. Similar recommendations are included in the latest European Crohn's and Colitis Organization consensus statements on the current management of UC and CD.[50, 51] However, a recent study by Tinsley et al. showed that a total of 29.1% of US gastroenterologists were unaware of any recommendations addressing pharmacological prophylaxis included in the ACG IBD guidelines, and 34.6% would give pharmacological VTE prophylaxis to a hospitalised patient with severe UC.
In conclusion, this study provides evidence of an association between IBD and VTE, and this evidence should be considered in practice guidelines. Future studies involving evaluations of the possible diagnostic bias discussed above, as well as randomised trials evaluating the risks and benefits of both VTE screening and prophylaxis in patients with IBD, could add further clarity to the discussion.
Guarantor of the article: Tetsuya Mine.
Author contributions: H. Yuhara contributed to the study concept and writing of manuscript. J. Koike and M. Igarashi contributed to the data acquisition. H. Yuhara, C. Steinmaus and D. Corley contributed to the analysis and interpretation of data and the drafting of the manuscript. C. Steinmaus, D. Corley, T. Suzuki and T. Mine contributed to the critical revision of the manuscript for important intellectual content. T. Mine supervised the study. All authors approved the final version of the manuscript.
Declaration of personal and funding interests: None.
- 4Harrison's Principles of Internal Medicine, 18th edn. McGraw-Hill Professional, pp. 986–7., , , et al.
- 9Homocysteine in inflammatory bowel disease: a risk factor for thromboembolic complications? Am J Gastroenterol 2000; 95: 2825–30., , , et al.Direct Link:
- 10Extensive arterial and venous thrombosis complicating chronic ulcerative colitis. Arch Intern Med 1936; 58: 17–31., .
- 16Rising prevalence of venous thromboembolism and its impact on mortality among hospitalized inflammatory bowel disease patients. Am J Gastroenterol 2008; 103: 2272–80., .Direct Link:
- 19Risk factor model to predict venous thromboembolism in hospitalized medical patients. J Hosp Med 2012; 147: 120–4., , , et al.
- 28Incidence and risk factors of venous thromboembolism in colorectal surgery: does laparoscopy impart an advantage? Arch Surg 2011; 146: 739–43., , , et al.
- 29Meta-analysis. In: Rothman K, Greenland S, eds. Modern Epidemiology. 2nd ed. Philadelphia: Lippincott Raven, 1998; 643–73..