IJE Advance Access originally published online on November 23, 2007
International Journal of Epidemiology 2008 37(1):136-146; doi:10.1093/ije/dym234
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
How should we use information about HWE in the meta-analyses of genetic association studies?
1Respiratory Epidemiology and Public Health Group, National Heart and Lung Institute, Imperial College London, SW3 6LR, London, UK.
2 Department of Health Sciences, Centre for Biostatistics and Genetic Epidemiology, University of Leicester, LE 1 7 RH Leicester, UK.
3 Clinical Epidemiology Unit, Mahidol University, 10400 Bangkok, Thailand.
4 Centre for Clinical Epidemiology and Biostatistics, University of Newcastle, NSW 2300 Newcastle, Australia.
* Corresponding author. Respiratory Epidemiology and Public Health Group, NHLI, Imperial College London, Emmanuel Kaye Building, Manresa Road, London, SW3 6LR, UK. E-mail: cosetta.minelli{at}imperial.ac.uk
 |
Abstract |
|---|
 Top Abstract MethodsResultsDiscussionSupplementary dataAcknowledgementsReferences |
|---|
Background: It is often recommended that control groups in meta-analyses of genetic association studies are checked for Hardy-Weinberg equilibrium (HWE) as a surrogate for assessing study quality. However, tests for HWE have low power and there is currently no consensus about how to handle studies that deviate significantly from HWE.
Methods: We identified 72 papers describing 114 meta-analyses of 1603 primary gene–disease comparisons. Based on these studies and on related simulations, we evaluated four different strategies for handling studies that appear not to be in HWE: (i) include them in the meta-analysis; (ii) exclude them if the test for HWE results in P < 0.05; (iii) exclude them if a measure of the size of departure from HWE is large and (iv) exclude them if (ii) and (iii).
Results: Of the 72 papers, 26 did not report information on HWE, with a trend toward increased reporting with time. HWE was evaluated through testing, with only three papers assessing the size of departure. On re-analysis, 9% of the 1603 primary comparisons showed significant deviation from HWE. The chance of an extreme departure from HWE was inversely related to the sample size of the study. Simulations suggest that there is no advantage in excluding studies that appear not to be in HWE.
Conclusions: Meta-analyses should report both the magnitude and the statistical significance of departures from HWE. Studies that appear to deviate from HWE should be investigated further for weaknesses in their design, but these studies should not be excluded unless there are other grounds for doubting the quality of the study.
Keywords Hardy-Weinberg equilibrium, genetic association studies, meta-analysis
Accepted 26 September 2007
The Hardy-Weinberg equilibrium (HWE) law states that if two alleles, G and g, with frequencies p and q = 1 – p, are in equilibrium in a population, then the proportion of people with genotypes GG, Gg and gg will be p2, 2pq and q2.1 Genuine departures from HWE in populations can be due to non-random mating, selection or migration, but usually the overall impact of these is limited, so that, although real populations are never in exact HWE, the extent of genuine deviations from HWE tends to be minimal.1,2 On the other hand, observed departures from HWE in genetic association studies are used in two main ways: (i) among cases, as a method for ‘rapid gene hunting’, when searching for possible polymorphisms associated with the disease3,4 and (ii) among controls, as a proxy for study quality.1,2
Departures from HWE in controls have been associated with problems in the design and conduct of genetic association studies particularly due to population stratification, genotyping error or selection bias. Population stratification is confounding caused by a varying mix of different ethnic groups in the study population when the frequency of the polymorphism and the incidence of disease varies between ethnic groups.5,6 It is theoretically important, although the impact of population stratification on real studies is a matter of debate.7,8 Genotyping error is a mistake in the laboratory identification of a subject's genotype due to sample contamination, observer variability or problems related to the specific genotyping technique, and it can lead to loss of precision and bias.9–12 However, although HWE testing is frequently justified on the grounds of genotyping error, recent theoretical work has demonstrated that HWE testing is not a reliable way of detecting genotyping error and the two-stage testing of HWE and then association can alter the type I error rate of the association test.13 Selection bias in the recruitment of controls in genetic case-control studies can also show up as a deviation from HWE.14 In addition to its role as a proxy for study quality, HWE is a fundamental assumption for the validity of the commonly used per-allele analysis.15,16
It is generally agreed that checking HWE among controls is a good idea, and it has been included in guides to the critical appraisal of genetic association studies.6,17–20 Despite this recommendation, there are still a number of unresolved issues:
- Often published primary genetic association studies do not report information on HWE; reporting rates ranging from 20 to 69% have been described for non-genetics journals.12,21–26 Even recent studies published in three high-profile genetics journals only had a 29% reporting rate for HWE.27 The same problem is apparent in the meta-analyses of genetic association studies; a review of 37 meta-analyses by Attia and colleagues, found that information on HWE was reported in only 24% of the meta-analyses.20
- With the modest sample size common in primary genetic association studies, tests have very low statistical power to detect any departure from HWE, and hence lack of evidence for departure from HWE does not necessarily imply adherence to HWE. Conversely, the few very large primary studies allow detection of minor degrees of deviation from HWE that may not be of practical importance.
- Separate testing for HWE of each primary study in a meta-analysis leads to inflated type I errors unless adjustment is made for multiple testing.
- Even when a statistically significant departure from HWE is detected, information on its magnitude is rarely if ever reported.20,27 If departures from HWE can bias the study results, then the influence is likely to depend on the extent, and possibly direction, of the departure. Different measures of the extent of departure from HWE have been proposed, including the inbreeding coefficient,28 the disequilibrium parameter29 and
,30 but they are rarely used in either primary studies or meta-analyses.
- In the meta-analysis of genetic association studies there is no consensus on what to do with studies that are not in HWE. Some authors suggest performing sensitivity analyses, pooling both with and without the studies that appear not to be in HWE and assessing whether studies classified as not being in HWE provide a different estimate of the genetic effect.20,27 Others suggest total exclusion from the meta-analysis of studies that show a significant deviation from HWE.19
This article explores these issues by re-analysing 114 published meta-analyses that include 1603 primary gene–disease comparisons and by using related simulations. The specific questions addressed are: (i) what is the extent of departure from HWE in published studies? (ii) should meta-analyses test for HWE in primary studies, or measure the size of the departures, or both? and (iii) what will be the best strategy for dealing with departures from HWE in a meta-analysis? In particular, should studies showing departures from HWE be excluded?
|
Methods |
|---|
|
TopAbstractMethodsResultsDiscussionSupplementary dataAcknowledgementsReferences |
|---|
Empirical evidence of departures from HWE in meta-analysis
The Human Genome Epidemiology Network (HuGENet) maintains a list of published genetic association meta-analyses (http://www.cdc.gov/genomics/hugenet/reviews_arch.htm). All 434 published articles listed on the HuGENet website on February 13, 2007 were examined and accepted for inclusion in this review if they: (i) reported the raw data on genotype frequencies for cases and controls; (ii) considered a binary disease outcome; (iii) were written in English and (iv) included >5 primary gene–disease comparisons. Because of the associated problems of assessing HWE and estimating the odds ratio (OR),31 comparisons based on <10 cases and/or <10 controls or with zero cells for both cases and controls in any specific genotype group were excluded. The selection process and numbers of exclusions are shown in Figure 1. The 72 papers eventually reviewed included reports on 114 meta-analyses describing 1603 primary comparisons. References for the 72 articles are reported in Supplementary Table.
|
We tested for HWE in controls in each of the 1603 primary comparisons using the exact test, as implemented by the genhwcci command in Stata,32 with statistical significance set at the 5% level.
The parameter 30 was used for estimating the magnitude of departures from HWE. This is defined as
|
|
= 0 when the proportions are in HWE, is positive when there are excess homozygotes and negative when there are excess heterozygotes. Although can theoretically range between plus and minus infinity, absolute values over 1 are relatively uncommon. was estimated by 
with approximate SE 
where n1, n2 and n3 are the numbers of subjects with each of the three genotypes. When one of these numbers was zero, an empirical adjustment was made by adding 
to each number. For comparison, we also calculated two other measures of departure from HWE, the inbreeding coefficient28 and the disequilibrium parameter.29
The inbreeding coefficient is defined as
|
|
|
|
Comparison of strategies to deal with HWE in meta-analysis
Four strategies of dealing with HWE were considered:
- Do nothing—all studies are included in the meta-analysis regardless of any evidence of departure from HWE.
- Exclude any study for which P < 0.05 in the exact test.
- Exclude any study for which the absolute value of the estimated > 0.5.
- Exclude any study for which P < 0.05 and the absolute value of the estimated > 0.5.
Since there is no reference threshold for departures from HWE in genetic association studies, our choice was based on the distribution of observed in the 1603 comparisons of the empirical study, where 
11% had an estimated outside the range –0.5 and 0.5 (see ‘Results’ section).
Each of the 114 meta-analyses was re-analysed to find how many studies would be excluded by each strategy and whether applying different strategies made a difference to the pooled estimate of the genetic effect and its 95% confidence interval (CI). We then used simulations to investigate which strategy provided the best trade-off between bias and precision. The genetic effect was estimated using both a genotype-based analysis in which we estimated the OR between the two extreme genotypes, i.e. GG vs gg (referred to as ORGG), and a per-allele analysis in which we estimated the OR of allele G vs g (referred to as ORallele G). By comparing the two extreme genotypes, in the first analysis we avoid having to specify a genetic model, which, in practice, is rarely known a priori. On the other hand, the allele-based analysis implicitly assumes an additive genetic model.
In the simulations, the values for the parameters were chosen to reflect the empirical study combined with different assumptions about the association between the departure from HWE and the size of the OR. For each scenario 10 000 meta-analyses, each of 10 primary studies, were randomly simulated, and the four strategies for dealing with HWE were applied in turn. The pooled genetic effect was estimated using a random effects meta-analysis model, implemented in Stata using the metan command. The differences between strategies were compared in terms of: (i) the number of excluded studies; (ii) the mean of the estimates of the pooled logOR; (iii) the percentage bias, that is the percentage difference between the mean estimate of logOR and its true value; (iv) the average SE of logOR estimates, calculated as the square root of the average variance; (v) the root mean square error (RMSE), that is the square root of the mean squared errors and (vi) the coverage of the 95% CIs, that is the percentage of intervals that included the true value.
|
Results |
|---|
|
TopAbstractMethodsResultsDiscussionSupplementary dataAcknowledgementsReferences |
|---|
Empirical evidence of departures from HWE in meta-analysis
Details of the 114 meta-analyses are presented in Supplementary Table. HWE was reported in only 46 of the 72 papers. In 43 articles, HWE was only assessed through hypothesis testing, while in three papers33–35 a measure of departure from HWE was also calculated and used to adjust the estimate of the genetic effect.36 The
-squared test was used in 16 papers, the exact test in 12 papers, and a permutation test in 2 papers; 15 papers did not specify which test was used, while 1 paper did not perform any tests but did report the results of the tests as evaluated in the primary studies. A trend is noticeable toward increased reporting on HWE in more recent articles and toward use of the exact test instead of the -squared test. Prior to 2005, 12 of 26 (46%) papers reported on HWE compared with 34 of 46 (74%) from 2005 onward (P = 0.01). No statistically significant departures from HWE were found in 30 of the 82 meta-analyses covered in the 46 papers that reported on HWE. In the 52 meta-analyses with evidence of departure from HWE, 3 different strategies were used; in 12, all studies were included regardless of the departures from HWE; in 29, sensitivity analysis was used, and in 11, studies showing departure from HWE were excluded from the main analysis.
Overall 9.0% (145/1603) of the primary comparisons showed statistically significant deviation from HWE (P < 0.05). However, because the nominal level chosen for statistical significance was 5%, we would expect 80 significant departures by chance alone. This suggests 65 true departures and an approximate false discovery rate of 55% (80/145).
There was high correlation between the three measures of departure from HWE as shown in Figure 2. The estimates of in the 1603 comparisons showed an approximately symmetrical distribution centred close to zero (mean 0.04 and SD 0.36). Only 0.5% of comparisons had an estimated < –1, and 1.4% had over 1. For limits of –0.5 and 0.5, the corresponding figures were 4.2 and 7.2%, respectively. The test for HWE was significant in 71% of the comparisons when was outside the range of ±1, and 43% of those with outside ±0.5. However, in almost all primary comparisons the estimates of were very imprecise because of small sample size. The average SE for was 0.26 with 90% of the individual SE lying between 0.08 and 0.69.
|
Given that the magnitude of departures from HWE are thought to reflect the extent of methodological problems with the study, it might be expected that large studies, which are usually associated with higher quality, would tend to show lower values of
. In order to investigate this, the estimates of observed in the 1603 associations were plotted against the corresponding total number of controls, as shown in Figure 3. Although we focus on sample size, allele frequencies and, in a cohort study, disease frequencies are also important determinants of the precision of a study. In Figure 3 the estimates of from small studies are subject to more sampling error, but even allowing for this, there is good evidence that the absolute magnitude of the underlying true deviation from HWE is inversely related to sample size.
|
Empirical study comparing strategies for HW deviation
The results from the empirical study of the four strategies are presented in Supplementary Figure. In 31 of the 114 meta-analyses (27%) no studies were excluded by any strategy, while in most of the remaining meta-analyses the strategies made little difference to the estimate of the pooled OR. When using a genotype-based analysis (ORGG), in meta-analyses where studies were excluded the average percentage difference in the pooled estimate for all strategies 2–4 compared with do nothing (strategy 1) was between 7 and 8%. The percentage difference between the pooled estimates for strategy 2 (P-value only), 3 (value of
only) and 4 (combination of P-value and ) compared with the estimate for strategy 1 was >10% in 14, 16 and 7 meta-analyses, respectively. The corresponding number of meta-analyses showing a difference >20% was 7, 9 and 5, respectively. While there were relatively small differences in estimated ORs, strategy 2 showed a greatest loss in precision. Compared with strategy 1, the percentage difference in the width of the 95% CI of the pooled logOR in those meta-analyses where studies were excluded was 11% for strategy 2, 9% for strategy 3 and 6% for strategy 4. A strategy based on significance has greater power to exclude large studies, and it is large studies that have the greatest influence on the precision of the estimated OR. When using an allele-based analysis, the number of meta-analyses with a difference >10% and >20% between the OR of any of the strategies of exclusion compared with ‘do nothing’ was only 5 (4%) and 1 (1%), respectively.
For the three strategies that were compared with ‘do nothing’, the chance of excluding any study in a meta-analysis increased with the number of studies in the meta-analysis; the average number of studies in meta-analyses that showed any exclusion under the P-value strategy was 16 compared with 11 in those meta-analyses in which there was no exclusion. The corresponding average number of studies per meta-analysis for the strategy based on and the strategy based on a combination of P-value and were 17 vs 10 and 18 vs 11, respectively. Although not conclusive, these findings seem to support an important role of chance in determining apparent departures from HWE.
Simulation study comparing strategies for HW deviation
The simulations were designed to reflect the values found in the empirical study. Ten studies were simulated in each meta-analysis, corresponding to the median in the empirical study. The total sample sizes for the 10 studies were fixed at; 70, 150, 200, 250, 310, 390, 470, 630, 940 and 2070, corresponding to 5th, 15th, 25th, ... up to 95th percentiles of the empirical study, with equal numbers of cases and controls in each study. The number of cases and controls in each of the three genotype groups for a simulated study was generated assuming:
- ORGG of 1.4, corresponding to the mean ORGG observed in the 1603 published associations.
- An additive genetic model, which assumes that logORGg = 1/2 logORGG, so that ORGg = ORallele G = 1.18.
- An allele frequency of 0.28, chosen to reflect the mean allele frequency observed in the 1603 published associations.
- randomly generated, at each simulation, by sampling from a normal distribution with mean 0 and standard deviation depending on the size of the study.
Estimates of in the empirical studies were distributed symmetrically around zero, with the absolute size of depending on the size of the study, where larger studies tend to show smaller values of (Figure 3). To mirror this pattern, the true value of for each of the 10 simulated studies (i, with i = 1, ..., 10) was sampled from:
|
|
Since the underlying value of simulated for each study was expected to play a major role in determining which strategy might perform better, a sensitivity analysis to assess the assumption of dependence of on sample size was also performed. In this analysis, a mixture model was used, in which the possibility that the value of depended on sample size as described above was given a 90% probability, while a 10% probability was given to the possibility that might not depend on the sample size at all.
The most crucial assumption in these simulations concerns the presence and magnitude of the possible effect of the departure from HWE on the size of the genetic association and the shape of this relationship. For the empirical study, the difference of each study-specific logOR from the pooled logOR of the relative meta-analysis was plotted against the observed . As Figure 4 shows, there is a slight indication of a negative trend but this is to be expected as the estimates of and the logOR will be negatively correlated when based on data from the same study. The empirical evidence does not support a strong association between genetic effect and departure from HWE, so, in our simulations we considered three different scenarios; the first had no association between and the logOR; the second assumed a small association, specifically a doubling of the estimate of the logOR per unit change in ; the third assumed the same degree of association, but in opposite direction (negative association).
|
Table 1 presents the results of the simulations for both the genotype- and allele-based analyses. All four strategies seem to perform well in terms of bias even when the OR depends on departure from HWE. Although studies with positive deviation from HWE overestimate the logOR, this is balanced by studies with negative deviation that underestimate it, so, on average, the bias tends to be small.
|
In terms of precision, the strategies also seem to perform similarly. The strategy with exclusion based on P-value alone tends to produce a loss in precision compared with ‘do nothing’, as shown by a small increase in the average SE and the RMSE in Table 1. As observed in the empirical study, large studies were more likely to show statistically significant deviation from HWE and their exclusion from the meta-analysis has most impact on the precision. With this strategy, 11% of studies were excluded on average. In comparison, the strategy with exclusion based on a combination of P-value and magnitude of
excluded only 6% of studies. With the latter strategy, the excluded studies are not necessarily the largest ones, since large studies, although more likely to overcome the problem of a lack of power of HWE tests, will also tend to have smaller values of . Average performance, as shown in Table 1, is only partially re-assuring. The real issue for a researcher is whether the strategy will make a difference to their specific meta-analysis? Figure 5 shows box plots of the individual logOR estimates from the 10 000 simulated data sets for both the genotype- and the allele-based analyses. Outlying estimates are relatively rare. For the genotype-based analyses, they represent between 0.7 and 0.8% of the 10 000 simulations for all strategies when the OR is unrelated to departure from HWE, between 0.8 and 0.9% when the OR is positively related to departure and between 0.7 and 0.8% when the OR is negatively related to departure.
|
Sensitivity analyses
Sensitivity analyses were performed to assess the robustness of the results to the choice of the thresholds for P-value and
. We evaluated the impact of using: (i) a significance level for the HWE test of P < 0.1, to compensate for the lack of power of statistical tests for HWE, by analogy with the argument for testing heterogeneity at the 10% level in a meta-analysis;37 (ii) a more conservative absolute value of , > 1 and (iii) a combination of these two, i.e. P < 0.1 and absolute value of > 1. The results of these sensitivity analyses showed similar patterns to the main analyses. It is interesting to note that, as expected, the use of a higher threshold for the P-value showed greater loss in precision, with an average of 17% of studies excluded compared with 11% in the main analyses. On the other hand, when using a strategy which combines P-value and magnitude of , the difference in the exclusion rate using a threshold of 0.1 compared with the main analysis is less marked (8% vs 6%).
We also performed sensitivity analyses aimed at assessing the impact of: (i) varying the effect of departures from HWE on the size of the genetic effect; (ii) assuming a linear relationship between the logOR and the absolute value of ; (iii) assuming that in 10% of studies, does not decrease with sample size, that is large studies could also show large deviations and (iv) using a fixed effect meta-analysis model instead of random effect model. Once again the overall findings were unchanged. If a stronger association is assumed between the logOR and the departure from HWE, then the patterns in Table 1 remain similar but the differences become more noticeable.
|
Discussion |
|---|
|
TopAbstractMethodsResultsDiscussionSupplementary dataAcknowledgementsReferences |
|---|
Our review of 1603 gene–disease comparisons from 114 published meta-analyses illustrates the inconsistent reporting of HWE in meta-analyses, with increasing use of testing but extremely limited use of measures of the magnitude of the departure from HWE. Nine per cent of the comparisons showed a statistically significant deviation from HWE, which is within the range of figures reported by other authors.27
The empirical study shows that the chance of an extreme deviation from HWE decreases with increasing sample size. This is crucial since large studies are more likely to have the power needed to be able to detect a statistically significant departure from HWE, if such a departure exists, but are less likely to show major departures from HWE. In fact, this would not be surprising if departures from HWE are a proxy for poor quality, and if it is true that quality tends to be higher in large studies. Although study quality cannot be predicted based on sample size alone, on average an association has been shown between quality and sample size.38 The recommended exclusion of studies based on hypothesis testing of HWE at the level of 0.05 is questionable as larger studies will produce smaller P-values for the same sized departure from HWE.
The empirical comparison of different strategies for handling HWE suggests that in most instances the impact of strategy choice on the estimate and precision of the pooled genetic effect is small. Even in those few instances (10 and 1% of the meta-analyses for the genotype- and allele-based analyses, respectively) where there is a substantial impact, we cannot say from empirical evidence alone which estimate is best.
A recent study by Trikalinos and coworkers36 has assessed the empirical evidence of the impact of departures from HWE on the pooled OR in a data set of 42 meta-analyses. They re-analysed the meta-analyses by excluding studies with statistically significant departures from HWE and by adjusting the analyses for the magnitude of the departures. It is interesting to note that, although their results are consistent with our findings for the exclusion based on a P < 0.05, their conclusions are quite different. They interpret the fact that the exclusion of studies with departures from HWE may sometimes result in the loss of statistical significance and in changes in the estimate of the pooled OR, as evidence of bias associated with the inclusion of studies with departures from HWE. They advocate that studies with departures from HWE should be excluded (or their results adjusted for the magnitude of the departure) even when there is no evidence of underlying methodological problems. Our view is that the change in statistical significance and in the magnitude of the genetic effect might well be due to the exclusion of large studies with small departures from HWE and that the impact of these exclusions are just as likely to be detrimental as beneficial.
As an empirical investigation cannot say which strategy is best because the true OR is unknown, simulations studies are needed alongside the empirical study. Our simulations suggest that only under extreme amounts of association between the departure from HWE and the size of the OR will the exclusion of studies that appear to deviate from HWE reduce the amount of bias in the estimate of the OR, and there is no empirical evidence that the OR is associated with departure from HWE. In fact, the ‘do nothing’ strategy of including all studies regardless of departure from HWE performed well in all of our simulations. Marginally the worst-performing strategy was based on statistical significance alone, which can result in a small loss in precision.
Overall, our results suggest that the real issue regarding HWE is not just the lack of power of the tests for HWE as previously suggested,27 but also the failure to consider the magnitude of the departure. The belief that including studies that appear to show a departure from HWE will bias the estimate of the genetic effect might well be unjustified. Although the ‘do nothing’ strategy performs well, we would not advocate that HWE should be ignored in either primary studies or meta-analyses. Instead, evidence of departures from HWE, based on both the result of a hypothesis test and the estimation of the magnitude of the departure, should be considered as a warning signal for the possible presence of methodological problems requiring further investigation. For instance, in the context of a primary study, this might lead to an exploration of the possibility of population stratification. In a similar manner, evidence of departure from HWE in studies included in a meta-analysis should lead to further investigation of the possibility of methodological problems in those studies. Failure to find any reasonable grounds for doubting the quality of a study should, however, make us question the sense of excluding that study. The data alone are unlikely to provide sufficiently strong evidence to justify exclusion. It is hard to argue against a sensitivity analysis that compares exclusion with inclusion, but perhaps we should tend toward accepting the analysis that includes all studies.
Our simulations were based on typical published gene–disease comparisons. Further investigation is required to show whether our conclusions hold under other circumstances, such as a very low allele frequency or a non-additive genetic model.
Similar concerns about the reliability of excluding from a meta-analysis studies with statistically significant departures from HWE have also been voiced by Salanti and coworkers.39 In a recent paper, they proposed an alternative approach for dealing with the problem, which adjusts the pooled estimate of the genetic effect for departure from HWE (measured by the fixation coefficient), using a meta-regression framework. Such an approach represents a compromise between the two opposite strategies of either excluding studies with statistically significant departures from HWE, as currently recommended, or including all studies by ignoring departures from HWE, as indicated by the simulation work presented in this article. However, the potential usefulness of this method in practice requires further evaluation, with the major limitation being that it can only be applied to very large meta-analyses. In fact, the adjustment tends to result in pooled ORs with wider 95%CIs, thus decreasing the power to detect an association. Many studies need to be included in the meta-analysis in order to estimate the parameters of the meta-regression. Moreover, this approach requires a Bayesian analysis, which might be difficult to implement by most researchers carrying out meta-analyses of genetic association studies.
More practical is the approach of Zou and Donner13 who advocated a simple adjustment to the -squared test of association to account for the impact of departure from HWE. On theoretical grounds they concluded that in a single primary study, it is ‘not necessary, and possibly even harmful to test the HWE assumption’. While we agree with most of their conclusions, we would still investigate departure from HWE but not make decisions about the usefulness of the study based on HWE alone.
|
Supplementary data |
|---|
|
TopAbstractMethodsResultsDiscussionSupplementary dataAcknowledgementsReferences |
|---|
Supplementary data are available at IJE online.
|
Acknowledgements |
|---|
|
TopAbstractMethodsResultsDiscussionSupplementary dataAcknowledgementsReferences |
|---|
Cosetta Minelli would like to thank the Department of Health, UK, for supporting this research via a National Research Scientist in Evidence Synthesis Award.
Conflict of interest: None declared.
KEY MESSAGES
|
|
References |
|---|
|
TopAbstractMethodsResultsDiscussionSupplementary dataAcknowledgementsReferences |
|---|
1 Sham P. Statistics in Human Genetics. (2001) London, UK: Arnold Publishers.
2 Khoury MJ, Little J, Burke W. Human Genome Epidemiology: A Scientific Foundation for Using Genetic Information to Improve Health and Prevent Disease. (2004) New York, USA: Oxford University Press.
3 Lee WC. Searching for disease-susceptibility loci by testing for Hardy-Weinberg disequilibrium in a gene bank of affected individuals. Am J Epidemiol (2003) 158:397–400.
4 Wittke-Thompson JK, Pluzhnikov A, Cox NJ. Rational inferences about departures from Hardy-Weinberg equilibrium. Am J Hum Genet (2005) 76:967–86.[CrossRef][Web of Science][Medline]
5 Freedman ML, Reich D, Penney KL, et al. Assessing the impact of population stratification on genetic association studies. Nat Genet (2004) 36:388–93.[CrossRef][Web of Science][Medline]
6 Cardon LR, Palmer LJ. Population stratification and spurious allelic association. Lancet (2003) 361:598–604.[CrossRef][Web of Science][Medline]
7 Thomas DC, Witte JS. Point: population stratification: a problem for case-control studies of candidate-gene associations? Cancer Epidemiol Biomarkers Prev (2002) 11:505–12.
8 Wacholder S, Rothman N, Caporaso N. Counterpoint: bias from population stratification is not a major threat to the validity of conclusions from epidemiological studies of common polymorphisms and cancer. Cancer Epidemiol Biomarkers Prev (2002) 11:513–20.
9 Gomes I, Collins A, Lonjou C, et al. Hardy-Weinberg quality control. Ann Hum Genet (1999) 63:535–38.[CrossRef][Web of Science][Medline]
10 Hosking L, Lumsden S, Lewis K, et al. Detection of genotyping errors by Hardy-Weinberg equilibrium testing. Eur J Hum Genet (2004) 12:395–99.[CrossRef][Web of Science][Medline]
11 Little J, Bradley L, Bray MS, et al. Reporting, appraising, and integrating data on genotype prevalence and gene–disease associations. Am J Epidemiol (2002) 156:300–10.
12 Xu J, Turner A, Little J, Bleecker ER, Meyers DA. Positive results in association studies are associated with departure from Hardy-Weinberg equilibrium: hint for genotyping error? Hum Genet (2002) 111:573–74.[CrossRef][Web of Science][Medline]
13 Zou GY, Donner A. The merits of testing Hardy-Weinberg equilibrium in the analysis of unmatched case-control data: a cautionary note. Ann Hum Genet (2006) 70:923–33.[CrossRef][Web of Science][Medline]
14 Edland SD, Slager S, Farrer M. Genetic association studies in Alzheimer's disease research: challenges and opportunities. Stat Med (2004) 23:169–78.[CrossRef][Web of Science][Medline]
15 Schaid DJ, Jacobsen SJ. Biased tests of association: comparisons of allele frequencies when departing from Hardy-Weinberg proportions. Am J Epidemiol (1999) 149:706–11.
16 Sasieni PD. From genotypes to genes: doubling the sample size. Biometrics (1997) 53:1253–61.[CrossRef][Web of Science][Medline]
17 Weiss ST, Silverman EK, Palmer LJ. Case-control association studies in pharmacogenetics. Pharmacogenomics J (2001) 1:157–58.[Medline]
18 Salanti G, Sanderson S, Higgins JP. Obstacles and opportunities in meta-analysis of genetic association studies. Genet Med (2005) 7:13–20.[Web of Science][Medline]
19 Munafo MR, Flint J. Meta-analysis of genetic association studies. Trends Genet (2004) 20:439–44.[CrossRef][Web of Science][Medline]
20 Attia J, Thakkinstian A, D’Este C. Meta-analyses of molecular association studies: methodologic lessons for genetic epidemiology. J Clin Epidemiol (2003) 56:297–303.[CrossRef][Web of Science][Medline]
21 Kocsis I, Gyorffy B, Nemeth E, Vasarhelyi B. Examination of Hardy-Weinberg equilibrium in papers of Kidney International: an underused tool. Kidney Int (2004) 65:1956–58.[CrossRef][Web of Science][Medline]
22 Kocsis I, Vasarhelyi B, Gyorffy A, Gyorffy B. Reanalysis of genotype distributions published in "Neurology" between 1999 and 2002. Neurology (2004) 63:357–58.
23 Nemeth E, Vasarhelyi B, Gyorffy B, Kocsis I. Unreported deviations of genotype distributions from Hardy-Weinberg equilibrium in articles published in Critical Care Medicine between 1999 and 2003. Crit Care Med (2004) 32:1431–33.[Web of Science][Medline]
24 Bardoczy Z, Gyorffy B, Kocsis I, Vasarhelyi B. Re-calculated Hardy-Weinberg values in papers published in Atherosclerosis between 1995 and 2003. Atherosclerosis (2004) 173:141–43.[CrossRef][Web of Science][Medline]
25 Gyorffy B, Kocsis I, Vasarhelyi B. Missed calculations and new conclusions: re-calculation of genotype distribution data published in Journal of Investigative Dermatology, 1998–2003. J Invest Dermatol (2004) 122:644–46.[CrossRef][Web of Science][Medline]
26 Gyorffy B, Kocsis I, Vasarhelyi B. Biallelic genotype distributions in papers published in Gut between 1998 and 2003: altered conclusions after recalculating the Hardy-Weinberg equilibrium. Gut (2004) 53:614–15.
27 Salanti G, Amountza G, Ntzani EE, Ioannidis JP. Hardy-Weinberg equilibrium in genetic association studies: an empirical evaluation of reporting, deviations, and power. Eur J Hum Genet (2005) 13:840–48.[CrossRef][Web of Science][Medline]
28 Weir BS. Genetic Data Analysis II: Methods for Discrete Population Genetic Data. (1996) Sunderland, MA, USA: Sinauer Associates.
29 Hernandez JL, Weir BS. A disequilibrium coefficient approach to Hardy-Weinberg testing. Biometrics (1989) 45:53–70.[CrossRef][Web of Science][Medline]
30 Lindley D. Statistical inference concerning Hardy-Weinberg equilibrium. Bayesian Stat (1988) 3:307–26.
31 Sweeting MJ, Sutton AJ, Lambert PC. What to add to nothing? Use and avoidance of continuity corrections in meta-analysis of rare events. Stat Med (2004) 23:1351–75.[CrossRef][Web of Science][Medline]
32 Cui J. Hardy-Weinberg equilibrium test in case-control studies. Stata Tech Bull (2000) 57:17–19.
33 Thakkinstian A, Han P, McEvoy M, et al. Systematic review and meta-analysis of the association between complement factor H Y402H polymorphisms and age-related macular degeneration. Hum Mol Genet (2006) 15:2784–90.
34 Zintzaras E, Koufakis T, Ziakas PD, Rodopoulou P, Giannouli S, Voulgarelis M. A meta-analysis of genotypes and haplotypes of methylenetetrahydrofolate reductase gene polymorphisms in acute lymphoblastic leukemia. Eur J Epidemiol (2006) 21:501–10.[CrossRef][Web of Science][Medline]
35 Gilbody S, Lewis S, Lightfoot T. Methylenetetrahydrofolate reductase (MTHFR) genetic polymorphisms and psychiatric disorders: a HuGE review. Am J Epidemiol (2007) 165:1–13.
36 Trikalinos TA, Salanti G, Khoury MJ, Ioannidis JP. Impact of violations and deviations in Hardy-Weinberg equilibrium on postulated gene–disease associations. Am J Epidemiol (2006) 163:300–9.
37 Sutton AJ, Abrams KR, Jones DR, Sheldon TA, Song F. Methods for Meta-analysis in Medical Research. (2000) Chichester, UK: John Wiley & Sons Ltd.
38 Ioannidis JP, Trikalinos TA, Ntzani EE, Contopoulos-Ioannidis DG. Genetic associations in large versus small studies: an empirical assessment. Lancet (2003) 361:567–71.[CrossRef][Web of Science][Medline]
39 Salanti G, Higgins JP, Trikalinos TA, Ioannidis JP. Bayesian meta-analysis and meta-regression for gene–disease associations and deviations from Hardy-Weinberg equilibrium. Stat Med (2007) 26:553–67.[CrossRef][Web of Science][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  J. Little, J. P.T. Higgins, J. P.A. Ioannidis, D. Moher, F. Gagnon, E. von Elm, M. J. Khoury, B. Cohen, G. Davey-Smith, J. Grimshaw, et al. STrengthening the REporting of Genetic Association Studies (STREGA): An Extension of the STROBE Statement Ann Intern Med, February 3, 2009; 150(3): 206 - 215. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 G. Davey Smith Big business, big science? Int. J. Epidemiol., February 1, 2008; 37(1): 1 - 3. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||