Taller height as a risk factor for venous thromboembolism: a Mendelian randomization meta-analysis

Summary

Background

Taller height is associated with greater risk of venous thromboembolism (VTE).

Objectives

We used instrumental variable (IV) techniques (Mendelian randomization) to further explore this relationship.

Methods

Participants of European ancestry were included from two cohort studies [Atherosclerosis Risk in Communities (ARIC) study and Cardiovascular Health Study (CHS)] and one case-control study [Mayo Clinic VTE Study (Mayo)]. We created two weighted genetic risk scores (GRS) for height; the full GRS included 668 single nucleotide polymorphisms (SNPs) from a previously published meta-analysis and the restricted GRS included a subset of 362 SNPs not associated with weight independently of height. Standard logistic regression and IV models were used to estimate odds ratios (ORs) for VTE per 10 cm increment in height. ORs were pooled across the three studies using inverse variance weighted random effects meta-analysis.

Results

Among 9143 ARIC and 3180 CHS participants free of VTE at baseline, there were 367 and 109 incident VTE events. There were 1143 VTE cases and 1292 controls included from Mayo. The pooled ORs from non-IV models and models using the full and restricted GRSs as IVs were 1.27 (95% CI: 1.11, 1.46), 1.34 (95% CI: 1.04, 1.73), and 1.45 (95% CI: 1.04, 2.01) per 10 cm greater height, respectively.

Conclusions

Taller height is associated with an increased risk of VTE in adults of European ancestry. Possible explanations for this association, including that taller people may have greater venous surface area, greater number of venous valves, or greater hydrostatic pressure, need to be explored further.

Introduction

Venous thromboembolism (VTE), comprising deep vein thrombosis (DVT) and pulmonary embolism (PE), occurs in 350 000 to 600 000 people and contributes to over 100 000 deaths in the U.S. each year [1]. Taller adult height is associated with greater risk of VTE [2–11]. For example, the Longitudinal Investigation of Thromboembolism Etiology (LITE) found that each 6.3 cm increment in height was independently associated with a 1.14 (95% confidence interval 1.05, 1.24) greater risk of VTE [9]. The physiologic explanation underpinning this increased risk is not yet fully understood; some proposed mechanisms include greater venous surface area, greater number of venous valves, and greater hydrostatic pressure in taller people [9].

It is possible that uncontrolled confounding by other risk factors explains the association of height with VTE, so additional techniques in causal inference would be useful to pursue. Genome-wide association studies have found many genetic variants related to height [12–20]. The most recent meta-analysis by Wood et al. found that 697 single nucleotide polymorphisms (SNPs) explained 20% of the heritability (and 16% of the overall variance) of height [20]. The high amount of variance in height explained by these SNPs, which should not be influenced by traditional confounding factors or by reverse causality, provides a unique opportunity to apply Mendelian randomization methods to explore the relationship between height and VTE. We used SNPs from the Wood et al. meta-analysis [20] as an instrumental variable (IV) for height, allowing us to better estimate the association between height and risk of VTE. We implemented these techniques in three separate studies (Atherosclerosis Risk in Communities study, the Cardiovascular Health Study, and the Mayo Clinic VTE Study) and then combined these results using meta-analytic techniques.

Methods

Study Population

The Longitudinal Investigation of Thromboembolism Etiology (LITE) is a combination of two population-based cohort studies: the Atherosclerosis Risk in Communities (ARIC) study [21] and the Cardiovascular Health Study (CHS) [22]. The study design and methods of LITE were described previously [23,24]. Briefly, ARIC began in 1987–89 with enrollment of 15 792 men and women aged 45 to 64 years, who were predominately black or white, and held follow-up examinations in 1990–92, 1993–95, 1996–98, and 2011–13, in addition to yearly telephone contact. CHS began in 1989–1990 with enrollment of 5201 mostly white men and women aged ≥65 years; an additional 687 African Americans were enrolled in CHS in 1992–93. Over the first 10 years of the study, participants in CHS were contacted every 6 months with either a telephone interview or clinic visit, and then by telephone interview every 6 months thereafter. The Mayo Clinic VTE Study (Mayo) is a case-control study of outpatients, including VTE cases referred to Mayo Clinic (n=1503) and controls undergoing general medical examinations at Mayo Clinic (n=1459). Cases and controls were selected for study participation from 1994–2009 and 2004–2009, respectively, using frequency matching on age, sex, state of residence, and prior myocardial infarction and stroke status [25]. Each study was approved by the institutional review board of its participating study center, and all participants provided written informed consent.

Height and Covariate Measurement

Baseline standing height was measured by trained technicians with participants standing as straight as possible without shoes and with back and heels against a vertical mounted centimeter ruler, to the nearest centimeter (cm) in ARIC and nearest half-cm in CHS. Weight was measured using a balance scale. Body mass index (BMI) was calculated as weight (in kilograms) divided by height (in meters) squared. Waist circumference was measured at the level of the umbilicus to the nearest cm. In order to confirm that potential confounders were not associated with the IV for height (a necessary assumption of a valid IV), we additionally consider other measured risk factors for VTE. Blood samples were drawn by staff and blood measurements were performed in study-specific central laboratories. Diabetes was defined as fasting blood glucose ≥7 mmol/L (≥126 mg/dL), non-fasting blood glucose ≥11.1 mmol/L (≥200 mg/dL), a self-reported physician’s diagnosis of diabetes, or reported use of a diabetes medication. Glomerular filtration rate was estimated from creatinine using an existing algorithm [26]. Factor VIII and activated partial thromboplastin time (available in ARIC only) were assessed as previously described [27,28]. C-reactive protein was measured via high-sensitivity immunoassay using Visit 2 (1990–92) plasma samples in ARIC [29] and baseline plasma samples in CHS [30]. In Mayo, height and weight were ascertained by self-report at the time of study enrollment for both cases and controls, and BMI was calculated as weight (in kilograms) divided by height (in meters) squared.

VTE Occurrence

Hospitalizations were identified by participants or proxy reports during annual contact by phone in ARIC and semi-annual contact by phone or clinic visit in CHS. Additional hospitalizations were found in ARIC through 2011 and CHS through 2001 by surveillance of local hospital discharge lists and Medicare records, respectively. Copies of hospital records were obtained for hospitalizations with International Classification of Disease (ICD) discharge codes indicating possible VTE events [23]. Validation of VTE events was performed by two physicians using standardized criteria requiring positive imaging tests for diagnosis of DVT and PE [23]. VTE events in Mayo were initially identified by a physician in the hematology clinic and validated by presence of imaging tests or by pathological examination of a surgically removed thrombus, as described previously [31]. DVT events in ARIC and CHS were restricted to those occurring in the lower extremity or vena cava. DVT events in Mayo also included those occurring in the arm.

Genotyping and Imputation

Details on genotyping, quality control, and imputation have been described previously [25,32,33]. Briefly, genotyping was performed in ARIC, CHS, and Mayo using the Affymetrix 6.0, Illumina 370CNV, and Illumina Human 660W Quad genome-wide SNP arrays, respectively. Before imputation, samples were filtered based on call rates, and SNPs were filtered based on call rates, allele frequencies, and Hardy Weinberg equilibrium P-values (see Table S1). Imputation to ≈2.5 million autosomal SNPs in the HapMap Phase II sample of European ancestry (CEU) was performed using MACH [34] v1.0.16 in ARIC and Mayo and BIMBAM [35] v0.99 in CHS.

Genetic Instrument for Height

The 697 height-related SNPs and their respective weights used for creating the weighted genetic risk score (GRS) IV were obtained from a genome-wide association meta-analysis study of height in 253 288 people of European ancestry [20]. The SNP weights are external to Mayo but are not external to ARIC and CHS since the two latter studies were included in the original meta-analysis [together ARIC and CHS comprised 11 338 (4.5%) of the total sample of 253 288]. In order to minimize possible linkage disequilibrium among SNPs selected for the GRS, we excluded one of two SNPs from any correlated pair of SNPs with r² > 0.1 as listed in Supplementary Table 1 from Wood et al. [20] (the first of the two correlated SNPs was excluded based on the listed order). We found that 29 SNPs were in possible linkage disequilibrium, leaving a total of 668 SNPs with which to create the “full GRS” (see Table S2, Column A for the list of SNPs used in the full GRS). Exploratory analyses indicated that the full GRS was associated with weight, independently of height, in ARIC and CHS (violating a key IV assumption); therefore, as a secondary analysis we created a “restricted GRS” that was uncorrelated with weight independently of height. The set of SNPs used to create the restricted GRS was determined using the ARIC cohort; in linear regression models (separately for each SNP), weight was regressed on each SNP and height, and the 306 SNPs showing the most statistically significant associations with weight (P < 0.4) independently of height were excluded. This alpha level was chosen because it was the smallest cut-off for which the restricted GRS was no longer associated with weight independently of height in ARIC. This left a subset of 362 SNPs (see Table S2, Column B for the list of SNPs used in the restricted GRS). Each GRS was created by taking Σ(beta value * estimated # of allele copies) over all of the SNPs, using the beta values listed in Column G of Supplementary Table 1 from Wood et al. [20].

Statistical Analysis

Only persons of European ancestry who provided written informed consent for DNA testing were included in this study. In ARIC and CHS respectively, there were 9349 and 3362 eligible participants with genotyping data available; we excluded those who had missing height information (n = 5 and 6), reported a VTE prior to baseline (n = 160 and 165), or were taking anticoagulants at baseline (n = 41 and 11), leaving a maximum of 9143 and 3180 participants for the present analyses, respectively. In Mayo, there were 2552 eligible participants with genotyping data available; after excluding those who had missing height information (n = 117) there was a maximum of 2435 participants for the present analyses.

To assess the strength of the GRS IVs, we used the F-statistic from a linear model regressing height on the GRS. If the GRS is associated with risk factors for VTE other than height, it would violate assumptions necessary for the IV analysis. To assess this, we estimated proportion of variance (R²) of each covariate explained by the GRS using linear regression. For Mayo, we only used controls to assess the association between covariates and the GRS because VTE is a relatively rare disease and we considered the controls to be a better representation of the general population compared to the combined group of cases and controls [36].

To examine the possibility of a nonlinear relationship between height and VTE, we initially fit logistic regression models with VTE as the outcome and modeling height using restricted cubic splines. Since the association between VTE and height was approximately linear in the logit scale (see Figure S1), we did not transform height in subsequent analyses. We estimated odds ratios (OR) and 95% confidence intervals (CI) for VTE per 10 cm increment in height using three separate logistic models. The first model was non-IV logistic regression adjusted for age, sex, body mass index, and study site in ARIC and CHS and for age, sex, body mass index, Minnesota residence, and history of stroke and myocardial infarction in Mayo (to obtain valid effect estimates, models using matched case-control data should be adjusted for the matching variables). The other two models were IV logistic structural mean models via a generalized method of moments estimator [37], using the full GRS and restricted GRS as the IVs for height in each respective model. Each of these models was fit separately to each of the three datasets. IV models using the full GRS and restricted GRS as the IVs were also run for DVT and PE separately. IV models were unadjusted for ARIC and CHS and were adjusted for the study matching variables for Mayo. As an alternative to using the restricted GRS IV model, we fit a model using the full GRS as the IV and adjusted for weight as a covariate in order to assess the possibility that weight mediates the height—VTE association.

Although using a GRS as an IV typically avoids the problem of weak instrument bias, there may still be inconsistent effect estimation by the individual variants comprising the GRS [38]. Thus, we performed sensitivity analyses to test for the influence of bias from potential directional pleiotropy or heterogeneity of individual SNP IVs using three robust IV methods: inverse variance weighted (IVW), MR-Egger regression, and weighted median estimation [38,39]. Summary IV estimates and standard errors for each SNP (using the ratio estimator [40]) were analyzed using the Stata package mrrobust to obtain robust IV OR estimates for the full and restricted set of SNPs for each of the three studies. Evidence for directional pleiotropy or heterogeneity of individual IV estimates was assessed by evaluating the I² index and Q test [41] from the IVW analysis and the significance of the intercept term from MR-Egger regression, as well as by visual inspection of funnel plots for asymmetry. The standard errors used in the funnel plots were estimated using the delta method [42].

For each of the three sets of results for the main VTE analysis, four sets of results for the DVT and PE analysis, and six sets of results for the sensitivity analysis, inverse variance weighted random-effects meta-analysis was performed to obtain a pooled OR and 95% CI across the three studies. Analyses were performed in Stata, version 12.1 (ARIC and CHS) and R, version 3.1.1 (Mayo) [43,44].

Results

Baseline characteristics are summarized for ARIC and CHS in Table 1 and for Mayo in Table 2. Notably, participants in CHS were on average older and shorter compared with those in ARIC and Mayo, and mean BMI was higher in Mayo compared with ARIC and CHS. In each study, height was strongly associated with the IVs (all F-statistics >105 and >63 for the full and restricted GRS, respectively, where F<10 is indicative of a weak IV [45]; Tables 1 and 2). Other anthropometric factors, including BMI, weight, and waist circumference, were also associated with the full or restricted GRS IVs (Tables 1 and 2). However, these anthropometric covariates were not associated with the restricted GRS IV after adjustment for height (all P-values > 0.2; however, adjustment for height could potentially create bias). The GRS IVs also were not generally associated with other common VTE risk factors in the three different studies (Tables 1 and 2). Additionally, a separate GRS for VTE, consisting of 5 SNPs in F5 (Leiden), F2, F11, FGG, and ABO genes [46], was not associated with the full or restricted GRS IVs in ARIC (R²=0.00; P>0.5).

Table 1.

Baseline characteristics and relationship with GRS for height in ARIC and CHS participants of European ancestry.

Characteristic	ARIC (1987–99) N=9143					CHS (1989–90) N=3180
		Full GRS (668 SNP)		Restricted GRS (362 SNP)			Full GRS (668 SNP)		Restricted GRS (362 SNP)
	Mean (SD) or %	R2 of GRS	P	R2 of GRS	P	Mean (SD) or %	R2 of GRS	P	R2 of GRS	P
Standing height, cm	169 (9)	0.080	<.001	0.045	<.001	165 (9)	0.068	<.001	0.039	<.001
Age, years	54 (6)	0.000	0.81	0.000	0.77	72 (5)	0.000	0.48	0.000	0.98
Female sex	53%	0.000	0.67	0.000	0.91	60%	0.000	0.35	0.001	0.13
Hormone replace therapy in females	21%	0.001	0.10	0.000	0.97	14%	0.000	0.72	0.000	0.49
Body mass index, kg/m2	27 (5)	0.002	<.001	0.000	0.93	26 (4)	0.002	0.008	0.000	0.43
Weight, kg	77 (16)	0.014	<.001	0.013	<.001	71 (14)	0.012	<.001	0.010	<.001
Waist circumference, cm	96 (13)	0.001	0.002	0.003	<.001	93 (13)	0.000	0.29	0.001	0.12
Diabetes	9%	0.000	0.62	0.000	0.45	12%	0.000	0.73	0.000	0.46
Fasting glucose, mmol/L	5.8 (1.7)	0.000	0.78	0.000	0.46	6.0 (1.6)	0.000	0.51	0.000	0.93
eGFR, mL/min/1.73 m2	100 (12)	0.000	0.69	0.000	0.85	66 (15)	0.000	0.57	0.000	0.24
C-reactive protein, mg/L	3.9 (7.0)	0.001	0.03	0.001	0.03	3.1 (5.5)	0.000	0.84	0.000	0.76
Factor VIII, %	125 (34)	0.000	0.67	0.000	0.35	119 (36)	0.000	0.62	0.000	0.91
aPTT, seconds	29 (3)	0.001	0.002	0.000	0.42	--	--	--	--	--

GRS: genetic risk score, ARIC: Atherosclerosis Risk in Communities, CHS: Cardiovascular Health Study, SNP: single nucleotide polymorphism, SD: standard deviation, eGFR: estimated glomerular filtration rate, aPTT: activated partial thromboplastin time

Table 2.

Subject characteristics and their relationship with the GRS for height in the Mayo VTE study (1994–2009)

Characteristic	Mean (SD) or %		Full GRS (668 SNP)*		Restricted GRS (362 SNP)*
	Cases N=1143	Controls N=1292	R2 of GRS	P	R2 of GRS	P
Standing height, cm	172 (11)	171 (10)	0.075	<.001	0.046	<.001
Age, years	55 (16)	56 (16)	0.002	0.09	0.003	0.04
Female sex	50%	52%	0.001	0.19	0.001	0.41
Body mass index, kg/m2	31 (8)	29 (6)	0.000	0.80	0.001	0.30
Weight, kg	94 (25)	84 (20)	0.018	<.001	0.007	0.003
Prior stroke/myocardial infarction	20%	11%	0.000	0.78	0.000	0.76
Minnesota residency	45%	55%	0.001	0.23	0.003	0.05

^*Among controls only

GRS: genetic risk score, VTE: venous thromboembolism, SNP: single nucleotide polymorphism, SD: standard deviation

Among 9143 ARIC and 3180 CHS participants, there were 367 and 109 incident VTE events after 23 and 12 years of follow-up, respectively. There were 1143 VTE cases and 1292 controls included from Mayo. Combining the study-specific non-IV logistic regression model results, the summary OR for VTE per 10 cm increment in height was 1.27 (95% CI: 1.11, 1.46) with moderate heterogeneity between studies (I²=42%, P=0.18; Figure 1). Sex-stratified point estimates were similar in men and women in each study and by meta-analysis (all P for interaction >0.36, data not shown). The summary ORs obtained by combining study-specific IV model results were similar whether using the full GRS IV [OR: 1.34 (95% CI: 1.04, 1.73)] or the restricted GRS IV [OR: 1.45 (95% CI: 1.04, 2.01)] with little to no evidence of heterogeneity between studies [I²=22% (P=0.28) and I²=0% (P=0.98), respectively; Figure 1]. There was slightly less precision in the estimate of the pooled OR using the restricted GRS IV potentially because inclusion of fewer SNPs resulted in a weaker instrument. The summary OR obtained from the full GRS IV model additionally adjusted for weight was slightly attenuated [OR: 1.24 (95% CI: 0.96, 1.61)]. Combining study-specific IV model results using the restricted GRS IV, height was associated slightly more strongly with DVT [OR: 1.46 (95% CI: 0.98, 2.16)] than with PE [1.20 (0.79, 1.82); see Figure S2).

Figure 1.

Meta-analysis of non-IV and IV logistic model odds ratios (OR) and 95% confidence intervals (CI) of VTE per 10 cm increment in height in ARIC, CHS, and Mayo.

ARIC: Atherosclerosis Risk in Communities, CHS: Cardiovascular Health Study, VTE: venous thromboembolism, IV: instrumental variable, GRS: genetic risk score, SNP: single nucleotide polymorphism.

†Fit using traditional logistic regression with adjustment for age, sex, body mass index, and study site.

‡Same as † except adjusted for age, sex, body mass index, Minnesota residence, and prior stroke and myocardial infarction.

§Fit using IV logistic structural mean model via a generalized method of moments estimator with the GRS as the IV. Mayo IV estimates are adjusted for the study matching variables (age, sex, Minnesota residence, and prior stroke and myocardial infarction); ARIC and CHS IV estimates are unadjusted.

Results from sensitivity analyses using robust IV estimators that take a weighted summary of the individual effects of the SNPs are shown in Figure S3. When using the set of 362 restricted SNPs as individual IVs, the summary OR using MR-Egger regression [OR: 1.34 (0.99, 1.82); I²=0] was similar to the estimate obtained when using the restricted GRS as a single IV. By comparison, the summary ORs using IVW [OR: 1.15 (95% CI: 0.96, 1.39); I²=0] and weighted median [OR: 1.15 (95% CI: 0.87, 1.54); I²=0] methods were attenuated. Within each study, there was little to no evidence of heterogeneity between SNP IVs in the full or restricted set of SNPs in IVW models [I² from 0% to 7%; all Q test P-values ≥ 0.16; see Figures S4–S9]. Funnel plots did not display obvious evidence of heterogeneity or directional pleiotropy using either set of SNPs, which was supported by formal testing of MR-Egger regression intercept terms in ARIC and Mayo (P-values > 0.2) but not in CHS (P-values ≤ 0.02; see Figures S4–S9). Pooled across studies, the summary MR-Egger regression intercept term did not show evidence of directional pleiotropy [intercept: −0.006 (95% CI: −0.020, 0.008) using full set of SNPs and −0.008 (95% CI: −0.024, 0.007) using restricted set of SNPs].

Discussion

In this meta-analysis we found a consistent positive relationship between height and VTE whether estimating the association of height with VTE using standard logistic regression and adjusting for relevant confounders or using IV analysis with a single weighted GRS, which suggested the risk of VTE increased by approximately 30–40% for each 10 cm increment in height. Sensitivity analyses using robust IV methods with multiple IVs (i.e., individual SNPs) generally supported our main findings; however the associations were attenuated.

Previous studies have found that taller height is associated with lower risk of most cardiovascular disease outcomes and total mortality [10], but taller height is associated with higher risk of VTE [2–11], atrial fibrillation [47], and abdominal aortic aneurysm [48]. Previous Mendelian randomization analyses provide additional support for an association of taller height with lower risk of coronary heart disease [49,50] but not with risk of stroke [49]. To our knowledge, this is the first study using Mendelian randomization to identify taller height as a risk factor for VTE.

Vessel wall damage, blood flow changes, and hypercoagulability are the three important factors for the pathogenesis of thrombosis [51]. A recent analysis in LITE found that the association of height with VTE risk was not mediated by markers of hemostasis or inflammation, suggesting that taller height more likely influences VTE risk through greater venous area, venous stasis, or vessel wall damage than through hypercoagulability [52]. Taller height is a risk factor for various types of cancer [53], another key risk factor for VTE. However, the non-IV OR for VTE was unchanged after excluding 112 cancer-provoked VTE cases in ARIC [OR: 1.46 (95% CI: 1.18, 1.79) per 10 cm in height], and active cancer was an exclusion criterion in Mayo. Thus, it is unlikely that cancer (and its effects on hypercoagulability) acts as a mediator of the association between height and VTE. Taller height is associated with higher resting venous pressure during quiet standing in healthy people [54]. Over time, this higher pressure could lead to more damage to venous walls and thereby greater VTE risk. Also, taller people have a greater venous surface area and possibly more venous valves, allowing for a greater area in which thrombi can form. Indeed, the majority of thrombi form in valve pockets of the deep veins in lower limbs where blood flow is stagnant [55]. This is supported by our finding of a stronger association of height with DVT than with PE.

Our ability to estimate the effect of height on VTE risk depends on the validity of our IV. In the context of this study, the three assumptions of a valid IV are that it (1) is associated with height, (2) is not associated with any confounder of the height–VTE association, and (3) is conditionally independent of VTE given height and confounders. Although we found that assumption (1) was satisfied, it is not possible to explicitly confirm assumptions (2) and (3) since they involve unmeasured confounders and the function of many of the SNPs used to construct the GRS IVs is unknown. Nevertheless, we found that the GRS IVs (particularly the restricted GRS) were generally not associated (independently of height) with measured risk factors for VTE that could potentially confound the association between height and VTE, which provides support for the plausibility of assumption (2). For assumption (3), we assume that the GRS IVs are not related to VTE through any pathway other than height [i.e., there is no horizontal pleiotropy (also known as biological or type I pleiotropy)] [38,56]. Given the large number of SNPs used to construct the GRS IVs, it is difficult to know whether this assumption is reasonable. However, to our knowledge, none of the genes containing the 697 height SNPs from the Wood et al. meta-analysis [19] has previously been reported in genome-wide association studies of VTE [25,33,57–59], and we found no association between the GRS IVs for height and a previously reported 5-SNP GRS for VTE [46]. Beyond the three IV assumptions, two additional assumptions need to be met in order to identify the effect using the generalized method of moments estimator. First, we assume there is no effect modification (on the multiplicative scale) by the IV on the association between height and VTE [60]. There was no evidence that this assumption was violated in any of the three studies (all interaction P-values ≥ 0.52) except in ARIC when using the full GRS IV (interaction P-value = 0.04). Second, we must assume the effect of height on VTE is linear on the logit scale and independent of the IV [37], which appeared reasonable from restricted cubic spline modeling.

The main GRS IV and IVW analyses are unbiased only when all IVs are valid. However, the MR-Egger regression and weighted median analyses allow for consistent effect estimation with some relaxation of IV assumptions, albeit with some loss of precision compared to IVW. Specifically, MR-Egger regression allows all IVs to violate assumption (3) provided they meet the weaker assumption that any pleiotropic effects of an IV (effects of the IV on VTE through pathways other than height) are uncorrelated with the strength of the IV–height association [38], while weighted median estimation allows for up to 50% of the weight contributed by IVs to violate assumptions (2) and (3) [39]. Summary robust IV effect estimates were attenuated compared to estimates from the main IV analysis, which could have possibly resulted from using different estimation methods (ratio estimator vs generalized method of moments) and different IV weighting schemes. The fact that summary robust IV effect estimates were slightly larger when using the restricted set of SNPs may suggest reduced pleiotropy in comparison to using the full set of SNPs.

Although there was an association between height and risk of VTE in the main meta-analyzed results overall, in CHS the estimated effect was weak or non-existent and had wide confidence intervals. The apparent lack of association in CHS may be partially explained by the older age of this cohort, whereby the effect of height is masked in older age when there is a higher prevalence of other VTE risk factors, or by selection bias. Other observational studies have analogously reported that the effect of family history as a risk factor for VTE gets weaker with increasing age [61,62].

Strengths of this analysis include using data from three different study populations (two population-based prospective studies and one clinic-based case-control study) and having validated VTE cases in each study. However, there are some limitations of our study. We only included people of European ancestry in order to be consistent with the genome-wide meta-analysis study of height [20], which limits the generalizability of our results. Further, although we have shown that the genetic instrument was not associated with many common VTE risk factors (i.e., potential confounders), we were not able to explore the precise physiologic mechanisms that may mediate the relationship between height and VTE risk, including that taller adults may have greater venous surface area, a greater number of venous valves, and greater hydrostatic pressure. Future studies should examine the relationship of height and VTE risk in other racial/ethnic groups and further explore the mechanisms connecting taller height with VTE risk.

In summary, this analysis provides evidence that taller height is associated with increased risk of VTE in adults of European ancestry. Additionally, we supplement the results of many previous studies reporting associations between height and VTE by use of an IV technique that allows for effect estimation in a way that may be less influenced by traditional confounding factors than observational analyses. Future research should continue using IV techniques to better understand the relationship of other exposures with VTE risk.

Essentials.

• Observational data suggest taller people have a higher risk of venous thromboembolism (VTE).

• We used Mendelian randomization techniques to further explore this association in three studies.

• Risk of VTE increased by 30–40% for each 10 cm increment in height.

• Height was more strongly associated with deep vein thrombosis than with pulmonary embolism.