Power and sample size calculations in the presence of phenotype errors for case/control genetic association studies
- Brian J Edwards^{1},
- Chad Haynes^{1},
- Mark A Levenstien^{1},
- Stephen J Finch^{2} and
- Derek Gordon^{1}Email author
DOI: 10.1186/1471-2156-6-18
© Edwards et al; licensee BioMed Central Ltd. 2005
Received: 12 October 2004
Accepted: 08 April 2005
Published: 08 April 2005
Abstract
Background
Phenotype error causes reduction in power to detect genetic association. We present a quantification of phenotype error, also known as diagnostic error, on power and sample size calculations for case-control genetic association studies between a marker locus and a disease phenotype. We consider the classic Pearson chi-square test for independence as our test of genetic association. To determine asymptotic power analytically, we compute the distribution's non-centrality parameter, which is a function of the case and control sample sizes, genotype frequencies, disease prevalence, and phenotype misclassification probabilities. We derive the non-centrality parameter in the presence of phenotype errors and equivalent formulas for misclassification cost (the percentage increase in minimum sample size needed to maintain constant asymptotic power at a fixed significance level for each percentage increase in a given misclassification parameter). We use a linear Taylor Series approximation for the cost of phenotype misclassification to determine lower bounds for the relative costs of misclassifying a true affected (respectively, unaffected) as a control (respectively, case). Power is verified by computer simulation.
Results
Our major findings are that: (i) the median absolute difference between analytic power with our method and simulation power was 0.001 and the absolute difference was no larger than 0.011; (ii) as the disease prevalence approaches 0, the cost of misclassifying a unaffected as a case becomes infinitely large while the cost of misclassifying an affected as a control approaches 0.
Conclusion
Our work enables researchers to specifically quantify power loss and minimum sample size requirements in the presence of phenotype errors, thereby allowing for more realistic study design. For most diseases of current interest, verifying that cases are correctly classified is of paramount importance.
Background
One technique used in gene localization is the case-control genetic association study [1]. In this method, genotype and phenotype data are collected for case and control individuals [2]. Both genotype and phenotype data often contain misclassification errors [3, 4], adversely affecting statistical tests used to locate disease genes [5–9]. Though phenotype misclassification has been widely studied in conjunction with disease (e.g. cancer, depression, heart disease), such studies have primarily focused on environmental association, not genetic association [10–13]. We are aware of only one recent publication considering phenotype misclassification for a test of genetic association [14].
Page et al. [3] emphasize the importance of studying phenotype errors in the context of genetic studies. They note that more than 1300 National Institutes of Health (NIH)-funded studies of complex genetic diseases have yielded fewer than 50 causative polymorphisms in humans [15, 16]. More importantly, only 16%–30% of initially reported associations are confirmed without evidence of between-study heterogeneity or bias [15, 17, 18].
The problem of phenotype misclassification is particularly important, given the high error rates encountered in some studies. Lansbury [19] reports that postmortem pathological studies estimate that greater than 15% of Alzheimer's Disease and Parkinson's Disease cases are misdiagnosed in the clinic. Duffy et al. [12] report that in a breast cancer study conducted by Press et al. [20], nearly half (34 out of 69) of the individuals containing over expression of the immunohistochemical marker c-erbB-2 were misclassified. Burd et al. [21] found that 5%–12% of individuals previously diagnosed with Tourette syndrome were misdiagnosed. They further note that in their three-step model for linkage analysis, a 5% misclassification rate in the first step leads to a 20% error rate by the third step.
In the presence of random errors that are non-differential with respect to trait status (case or control), the type I error rate is constant [5]. That is, there is no change in significance of the classic chi-square test of independence on 2 × n contingency tables (the statistic of interest in this work). Here and elsewhere, n is the number of observed genotypes at the marker locus. However, there is a reduction in the power of the chi-square test and an increase in the minimum sample size needed to maintain constant asymptotic power at a fixed significance level [5, 22, 23]. A key issue that arises then is a quantification of power loss in the presence of phenotype errors.
Formulas allowing researchers to perform realistic power and sample size calculations in the presence of errors benefit researchers in the design of case-control studies by saving them the cost of excessive genotyping and phenotyping due to underpowered initial conditions. Mote and Anderson [22] computed power in the presence of what we call genotype error (although in a more general statistical setting) and proved that the power of the chi-square test of independence on r × c contingency tables (r = number of rows; c = number of columns) is always less than or equal to the power of the test when data are perfectly classified. Carroll et al. [24] developed methods for estimating the parameters of a prospective logistic model given a binary response variable and arbitrary covariates with case/control data when the covariates have measurement error. Gordon et al. [6, 7] developed formulas for power and sample size calculations for the specific situation of genotype error. They used Mitra's equation for the non-centrality parameter [6, 7, 25] to compute the power and minimum sample size both for data with and without genotype errors. Gordon et al. [6, 7] showed that a one percent increase in the sum of genotypic error rates typically results in a two to eight percent increase in the minimum sample size for the parameters and error models considered and that the increase in minimum sample size is larger when the allele frequencies are more extreme [7]. Kang et al. [8] extended this work by determining a linear approximation for the sample size increase needed to maintain constant asymptotic power at a fixed significance level. Kang et al. [8] found that (i) the cost of genotype misclassifications is a function of the true genotype frequencies and the ratio of controls to cases; (ii) in general, misclassifying a more common genotype as a less common genotype is more costly than the reverse error; and (iii) certain types of misclassification have costs that approach infinity as the minor SNP allele frequency approaches 0.
Our goal in this research is therefore two-fold: (i) to quantify power and sample size for the chi-square test of independence on 2 × n contingency tables in the presence of phenotype errors; and (ii) to quantify the cost of each type of phenotype error.
We present formulas to facilitate accurate power and sample size calculations in the presence of phenotype errors. We perform a genotypic test of association using the Pearson chi-square test statistic on 2 × n contingency tables. The degrees of freedom (in our case n-1) and the non-centrality parameter completely describe the power of the chi-square test. We express the non-centrality parameter in terms of the case and control sample sizes, genotype frequencies, and phenotype error model parameters. Rearranging the equation for the non-centrality parameter gives an equation for the minimum sample size. Additionally, this work extends Kang et al.'s [8] findings to the cost of phenotype errors.
Results
As noted in the Methods section (Distinguishing case from affected and control from unaffected), we use the term case to refer to an individual who has been diagnosed as being affected with a given disease, whether or not that individual is truly affected. Similarly, we use the term control to refer to an individual who has been diagnosed as being unaffected with a given disease, whether or not that individual is truly unaffected. We use the term affected (respectively, unaffected) to refer to an individual who is truly affected (respectively, unaffected) with the disease of interest.
All notation in the Results section is defined in the Methods section (Notation).
Design of simulation program – null and power calculations for a fixed sample size
Parameter settings for null and power simulations with di-allelic and tetra-allelic loci
Low | High | |
---|---|---|
True case and control genotype frequencies | p = 0.05 | p = 0.15 |
Pr(affected misclassified as a control) (θ) | 0.05 | 0.15 |
Pr(unaffected misclassified as a case) (φ) | 0.05 | 0.15 |
Disease prevalence (K) | 0.005 | 0.05 |
Number of cases ( ) | 500 | 1000 |
Number of controls ( ) | 500 | 1000 |
Significance level | 5% | 1% |
Genotype frequency parameter for tetra-allelic loci (power simulations) | ||
d | 1 | 2 |
Percentiles for absolute difference between asymptotic power and simulation power
5% significance level | 1% significance level | |
---|---|---|
Di-allelic locus | ||
Minimum | 0.0000 | 0.0000 |
10^{th} percentile | 0.0002 | 0.0002 |
25^{th} percentile | 0.0005 | 0.0004 |
50^{th} percentile | 0.0010 | 0.0011 |
75^{th} percentile | 0.0028 | 0.0026 |
90^{th} percentile | 0.0065 | 0.0057 |
Maximum | 0.0099 | 0.0119 |
Tetra-allelic locus | ||
Minimum | 0.0000 | 0.0000 |
10^{th} percentile | 0.0000 | 0.0000 |
25^{th} percentile | 0.0007 | 0.0008 |
50^{th} percentile | 0.0012 | 0.0014 |
75^{th} percentile | 0.0028 | 0.0032 |
90^{th} percentile | 0.0072 | 0.0081 |
Maximum | 0.0102 | 0.0111 |
Although the asymptotic power is a good enough approximation to the simulation power so that it can be used for design purposes, this difference is somewhat larger than would be expected in the event that the simulated power followed a binomial variation with probability equal to the asymptotic power (based on computation of 95% confidence intervals – results not shown). We discuss this issue below (see Discussion).
Cost functions
Using the mathematics presented in the Methods section (Cost functions), we compute the following formulas:
Cost coefficients for different types of misclassification
K | R* | p | C _{ θ } | C _{ φ } |
---|---|---|---|---|
0.005 | 0.5 | 0.05 | 0.01 | 540.29 |
0.15 | 0.01 | 458.99 | ||
1 | 0.05 | 0.01 | 478.32 | |
0.15 | 0.01 | 432.67 | ||
2 | 0.05 | 0.01 | 440.18 | |
0.15 | 0.01 | 415.60 | ||
0.05 | 0.5 | 0.05 | 0.09 | 51.59 |
0.15 | 0.10 | 43.82 | ||
1 | 0.05 | 0.08 | 45.67 | |
0.15 | 0.10 | 41.31 | ||
2 | 0.05 | 0.08 | 42.03 | |
0.15 | 0.10 | 39.68 |
When the prevalence K = 0.005, the cost coefficient C_{ φ }becomes larger by an order of magnitude. The minimum value of C_{ φ }is 415, occurring as above when R* = 2 and p = 0.15. That means that a 1% increase in the value of φ requires at least a 415% increase in cases and controls to maintain the same power at any significance level.
A second finding that becomes clear from studying equation (1) is that the cost coefficient C_{ φ }has an infinite limit as the prevalence K approaches 0 (for any set of fixed values of the other parameters), while the cost coefficient C_{ θ }has a limit of 0. This results comes from the observation that the dominating terms for the cost coefficients C_{ φ }and C_{ θ }in equation (1) are (1 - K)/K and K/(1 - K), respectively.
It should be noted that the linear Taylor approximation is not very accurate for even small values of φ. The linear Taylor approximation is useful, though, in that it serves as a lower bound for the percentage sample size increase. That is, percent increase in sample size is at least C_{ φ }for any value of φ. We illustrate this point in the next section.
Minimum sample size requirements in presence of phenotype misclassification – Alzheimer's disease ApoE example
Our results for the cost functions are consistent with the findings here. For values of φ less than 0.02, sample size increase appears to be constant in the parameterθ. That is, misclassification of an affected as a control does not affect the sample size estimates at all. However, even a 1% misclassification of an unaffected as a case requires a sample size increase from 486 to 921 (φ = 0.01, θ = 0.0 in figure 1; exact results not shown) to maintain constant power, an approximately 90% increase. As the probability of misclassifying an unaffected as a case φ increases, there appears to be an interaction between the two misclassification parameters, requiring even larger sample size increases than would be expected if the sample size increase were linear in each misclassification parameter (figure 1).
Comparison of power loss for fixed sample size when only one misclassification parameter is non-zero
The results of figure 2 further illustrate the importance of distinguishing between the two types of misclassification. When the φ parameter is 0, the asymptotic power is virtually independent of the value of the φ parameter and the disease prevalence K. Power values for all settings of φ and K are approximately 99%. When the θ parameter is 0, the asymptotic power reduces to 91% when φ = 0.01, K = 0.05 and to 33% when φ = 0.01, K = 0.01. When φ = 0.02, power reduces to 76% when K = 0.05 and to 11% when K = 0.01. These examples further document the dominating effect that disease prevalence has on power and/or sample size requirements in the presence of phenotype misclassification error.
Discussion
As we noted above (Results – Design of simulation program – power calculations for a fixed sample size), the asymptotic power is a good enough approximation to the simulation power so that it can be used for design purposes. However, the difference is somewhat larger than would be expected in the event that the simulated power followed a binomial variation with probability equal to the asymptotic power. One possible explanation may be that our simulation studies were "under-powered" so that the asymptotic theory did not hold. Indeed, the median power value at the 5% significance level for our simulation studies (table 1) was 13% (full results not shown). Given such low overall power levels and also the fact that, for the SNP minor allele frequency of 0.05, Cochran's condition of a minimal expected cell count of 5 is not achieved [26], it is conceivable that effective sample sizes are not sufficient for power values based on asymptotic theory to hold. Other authors studying misclassification error have also observed this phenomenon [27].
While we have considered a genetic model-free framework here, we note that our work easily extends to a genetic model-based framework as well [6, 7]. We will implement calculations using a genetic model-based framework in our web tool (next paragraph).
Given the accuracy of our method (absolute errors no larger than 0.012, based on simulations), we conclude that researchers may use our method to accurately determine power and sample size calculations for case/control genetic association studies in the presence of phenotype misclassification. We have developed a web tool that performs these calculations online. The URL for this tool is: http://linkage.rockefeller.edu/pawe/paweph.htm.
Conclusion
In this work, we developed a method for performing realistic power and sample size calculations in the presence of phenotype errors. Simulation results suggest that our formulas (equations (2) and (3)) may be used to design case/control genetic association studies incorporating phenotype misclassification. We confirmed that phenotype misclassification always reduces the power of the chi-square test of association (as was first shown by Bross [5]), and consequently, increases the minimum sample size needed to maintain constant asymptotic power.
Our cost calculations reveal two significant findings. The first is that power and/or sample size is most significantly altered by a change in disease prevalence. Specifically, the cost coefficient for misclassifying an affected as a control is of the order of magnitude K/(1 - K) and the cost coefficient for misclassifying an unaffected as a case is of the order of magnitude (1 - K)/K, where K is the disease prevalence (equation (1)). This finding suggests that, for many diseases of current interest, where prevalence is usually less than or equal to 0.10, it is much more important to insure that cases are truly cases rather than controls being truly controls. Zheng and Tian [14] made this same observation (without the explicit computation of cost coefficients) for the linear test of trend applied to cases and controls genotyped at a SNP marker.
Methods
Distinguishing case from affected and control from unaffected
Throughout this work, we use the term case to refer to an individual who has been diagnosed as being affected with a given disease, whether or not that individual is truly affected. Similarly, we use the term control to refer to an individual who has been diagnosed as being unaffected with a given disease, whether or not that individual is truly unaffected. We use the term affected (respectively, unaffected) to refer to an individual who is truly affected (respectively, unaffected) with the disease of interest. A key assumption we make through the paper is that we collect only cases and controls for our test of genetic association.
Notation
We use the following notation:
Count parameters
a = Number of alleles at the marker locus. The number of genotypes at the marker locus is always a(a + 1)/2 = n.
Probability parameters
K = Prevalence of disease.
p_{0j}= Frequency of genotype j at the marker locus for the affected group, 1 ≤ j ≤ a(a+1)/2.
p_{1j}= Frequency of genotype j at the marker locus for the unaffected group, 1 ≤ j ≤ a(a+1)/2.
Error model parameters
θ = Pr (affected individual classified as control) = 1 - Se, where Se is the sensitivity of the phenotype measurement instrument.
φ = Pr (unaffected individual classified as case) = 1 - Sp, where Sp is the specificity of the phenotype measurement instrument. This notation was used by Bross [5].
A key assumption we make here is that these errors are random and independent. Furthermore, they are non-differential with respect to a particular genotype [14].
Cost parameters
C_{ θ }= Cost of misclassifying an affected individual as a control. This value is the percent increase in minimum sample size necessary to maintain constant power for every one percent increase in the value of θ.
C_{ φ }= Cost of misclassifying an unaffected individual as a case. This value is the percent increase in minimum sample size necessary to maintain constant power for every one percent increase in the value of φ.
Expressing case and control genotype frequencies in terms of affected and unaffected genotype frequencies
For a derivation, see the Appendix.
It is interesting to note that determination of case and control genotype frequencies in the presence of only phenotype error differs from determination of the same frequencies in the presence of only genotype error in that one needs to specify disease prevalence for phenotype error (in addition to specifying the respective misclassification probabilities for phenotype and genotype) [7, 14].
Test statistic for genotypic association
The test statistic considered in this work is Pearson's chi-square statistic on 2 × n contingency tables. Here, the two rows refer to the two possible classifications (case or control) and the n columns correspond to the n different genotypes, where n = a(a + 1)/2. Using this statistic on 2 × n contingency tables, we test for association between genotype and disease status. We selected the genotypic test of association because the null distribution of the allelic test of association cannot be determined when either the case or control group genotype frequencies deviate from Hardy-Weinberg Equilibrium (HWE) [28, 29]. Let G_{ rc }equal the observed count of the c^{th} genotype in the r^{th} group, where 1 ≤ c ≤ n and r = 0 for the case population and r = 1 for the control population. Then, the chi-square statistic is given by the formula .
In this expression, the expected cell count of the c^{th} genotype in the r^{th} group, E_{ rc }, is determined by the equation E_{ rc }= S_{ r }D_{ c }/N, where is the row total for the r^{th} group, is the column total for the c^{th} genotype, and is the total sample size.
Under the null hypothesis of no association between the marker locus and the disease (p_{0j}= p_{1j}for all j), the statistic X^{2} is asymptotically distributed as a central χ^{2} with n - 1 degrees of freedom. We verify this statement in our simulations (see Results).
Asymptotic power calculations
In this section, we describe our method for computing asymptotic power in the presence of errors. The asymptotic power is summarized by a non-centrality parameter λ, which is a function of the case and control sample sizes and the respective genotype frequencies.
Asymptotic non-centrality parameter
Mitra [25] derived the asymptotic power function for the chi-square test for unmatched cases and controls. Under the alternative hypothesis, the distribution is a non-central χ^{2} with n -1 degrees of freedom and non-centrality parameter λ*. Mitra [25] showed that for perfectly classified data (i.e., θ = φ = 0)), the non-centrality parameter is given by
Increase in minimum sample size
Design of simulation program – null and power calculations for a fixed sample size
We perform simulations using 100,000 iterations to verify (i) the nominal significance levels under the null hypothesis; and (ii) the asymptotic power calculations provided by equation (2). We use a 2^{7} factorial design [30] in which we set lower and upper bounds for each set of parameters. In the simulations, we consider both di-allelic and tetra-allelic loci. For each simulation, both the affected and unaffected genotype frequencies are in HWE. For the power simulations using di-allelic loci, the genotype frequencies are specified as follows using a parameter p: for the affected group, p_{01} = (1 - p)^{2}, p_{02} = 2p(1 - p), p_{03} = p^{2}, and for the unaffected group, p_{11} = (1 - p - 0.1)^{2}, p_{12} = 2(p + 0.1)(1 - p - 0.1), p_{13} = (p + 0.1)^{2}. That is, the SNP minor allele frequency in the unaffected population is equal to the sum of the SNP minor allele frequency in the affected population (p) and 0.1. For the null simulations, both the affected and unaffected groups have genotype frequencies as specified above for p_{0j}, j ∈ {1,2,3}. Our parameter settings for the factorial design are shown in table 1.
For the tetra-allelic loci, the parameter settings are the same as for the di-allelic loci with the exception of the affected and unaffected genotype frequencies. For the tetra-allelic loci, we let p = 0.25 and specify the genotype frequencies for power simulations as follows using a parameter d. For the affected population, the probability of a homozygous genotype is p^{2}+d(0.03) and the probability of a heterozygous genotype is 2p^{2} - d(0.02), where d = 1,2. For the control group, the probability of a homozygous genotype is 0.0625 and the probability of a heterozygous genotype is 0.125. For null simulations, we set d = 0.
Here, we briefly describe the algorithm used to simulate our phenotype and genotype data for each replicate of a particular simulation. Note that a simulation is completely described by the each of the 7 parameter settings provided in table 1. For each individual in each replicate, we first randomly assign the individual an affection status (affected or unaffected) using the disease prevalence K. We then randomly assign the individual a genotype conditional on the affection status using the conditional probabilities p_{0j}and p_{1j}. Once affection status and genotype are determined, we then randomly assign case or control status using the individual's affection status and the phenotype misclassification probabilities. Within each replicate, we repeat this procedure until we have the specified number of cases and controls. Because of the low prevalence, we invariably reach our required number of controls much more quickly than we reach our required number of cases. In such situations, we simply ignore all assigned control individuals after reaching our required number, and keep collecting cases until we achieve that required number.
Cost functions
We demonstrate how to compute the sample size cost coefficient of phenotype misclassification to gain insight into which type of misclassification requires the greater increase in sample size for fixed power. Let λ equal the non-centrality parameter when there is no phenotype misclassification and let λ* equal the non-centrality parameter in the presence of phenotype errors. To find the sample size adjustment needed to maintain constant power, we set λ = λ*. We considered this condition previously when studying the cost of genotype error [8]. Let and . Then the condition λ = λ* may be rewritten as or . Though the cost of misclassification for cases is mathematically defined as the ratio /N_{ A }, we instead consider the reciprocal ratio N_{ A }/ because the latter allows for more straightforward computation. We approximate N_{ A }/ using a first-order Taylor Series expansion centered at (θ, φ) = (0,0). We obtain . Here, (∂/∂θ)[f]|_{(0,0)} is the partial differential operator (with respect to θ) acting on the function f and evaluated at the point (0,0). An identical definition holds for (∂/∂φ)[f]|_{(0,0)}.
Minimum sample size requirements in presence of phenotype misclassification – Alzheimer's disease ApoE example
We determine the minimum sample size necessary to maintain a constant power of 95% at the 5% significance level using formula (3) and considering estimated genotype frequencies from a recently published genetic association analysis of Alzheimer's Disease (AD) cases and controls genotyped at the ApoE marker locus [9]. In most populations there are three alleles at the ApoE locus. Conventionally, they are denoted ε_{2}, ε_{3}, and ε_{4} and we label them 2, 3, and 4 respectively in this work. In a well known and often replicated association finding, every copy of the 4 allele in a person's genotype increases that person's risk of getting late-onset AD by a factor of 2.5–3 [31]. Furthermore, recently published estimates of prevalence for Alzheimer's Disease in the US hover around the 2% range [32]. Thus, for our sample size calculations, we assume a prevalence K = 0.02.
If we index the six genotypes as 1 = 22, 2 = 23, 3 = 24, 4 = 33, 5 = 34, 6 = 44, then the genotype frequency values we use for our sample size calculations (taken from our previous work [9]) are:
p_{01} = 0.019, p_{11} = 0.000, p_{02} = 0.057, p_{12} = 0.118, p_{03} = 0.019, p_{13} = 0.024, p_{04} = 0.465, p_{14} = 0.699, p_{05} = 0.344, p_{15} = 0.159, p_{06} = 0.096, p_{16} = 0.000.
As it has been documented that phenotype misclassification in Alzheimer's Disease may run as high as 15% or more [19], we consider phenotype misclassification values 0 ≤ θ, φ ≤ 0.15, in increments of 0.01. It is assumed that there are equal numbers of cases and controls (R* = 1).
Appendix
= [Pr(genotype = j, case, affected) + Pr(genotype = j, case, unaffected)]/Pr(case)
= [Pr(genotype = j | case, affected) Pr(case | affected) Pr(affected) + Pr(genotype = j | case, unaffected) Pr(case | unaffected) Pr(unaffected)]/[Pr(case | affected) Pr(affected) + Pr(case | unaffected) Pr(unaffected)]
= [p_{0j}(1 - θ)K + p_{1j}φ(1 - K)]/[(1 - θ)K + φ(1 - K)].
= [Pr(genotype = j, control, affected) + Pr(genotype = j, control, unaffected)]/Pr(control)
= [Pr(genotype = j | control, affected) Pr(control | affected) Pr(affected) + Pr(genotype = j | control, unaffected) Pr(control | unaffected) Pr(unaffected)]/[Pr(control | affected) Pr(affected) + Pr(control | unaffected) Pr(unaffected)]
= [p_{0j}θK + p_{1j}(1 - φ)(1 - K)]/[θK + (1 - φ)(1 - K)].
Declarations
Acknowledgements
The authors gratefully acknowledge grants K01-HG00055 (DG) and HG00008 (to J. Ott) from the National Institutes of Health. BJE was supported by the Rockefeller University Science Outreach Program. The authors also gratefully acknowledge two anonymous reviewers whose comments led to significant improvements and simplifications in the research.
Authors’ Affiliations
References
- Breslow NE, Day NE: Statistical Methods in Cancer Research. The Analysis of Case-Control Studies. 1980, Lyon, International Agency for Research on Cancer, 1: 350-EighthGoogle Scholar
- Ott J: Analysis of Human Genetic Linkage. 1999, Baltimore, The Johns Hopkins University PressGoogle Scholar
- Page GP, George V, Go RC, Page PZ, Allison DB: "Are we there yet?": Deciding when one has demonstrated specific genetic causation in complex diseases and quantitative traits. Am J Hum Genet. 2003, 73: 711-719. 10.1086/378900.PubMed CentralView ArticlePubMedGoogle Scholar
- Rice JP, Saccone NL, Rasmussen E: Definition of the phenotype. Adv Genet. 2001, 42: 69-76.View ArticlePubMedGoogle Scholar
- Bross I: Misclassification in 2 x 2 tables. Biometrics. 1954, 10: 478-486.View ArticleGoogle Scholar
- Gordon D, Levenstien MA, Finch SJ, Ott J: Errors and linkage disequilibrium interact multiplicatively when computing sample sizes for genetic case-control association studies. Pac Symp Biocomput. 2003, 490-501.Google Scholar
- Gordon D, Finch SJ, Nothnagel M, Ott J: Power and sample size calculations for case-control genetic association tests when errors are present: application to single nucleotide polymorphisms. Hum Hered. 2002, 54: 22-33. 10.1159/000066696.View ArticlePubMedGoogle Scholar
- Kang SJ, Gordon D, Finch SJ: What SNP genotyping errors are most costly for genetic association studies?. Genet Epidemiol. 2004, 26: 132-141. 10.1002/gepi.10301.View ArticlePubMedGoogle Scholar
- Gordon D, Yang Y, Haynes C, Finch SJ, Mendell NR, Brown AM, Haroutunian V: Increasing power for tests of genetic association in the presence of phenotype and/or genotype error by use of double-sampling. Stat Appl Genet and Mol Biol. 2004, 3: Article 26-Google Scholar
- Brown RP, Sweeney J, Frances A, Kocsis JH, Loutsch E: Age as a predictor of treatment response in endogenous depression. J Clin Psychopharmacol. 1983, 3: 176-178.PubMedGoogle Scholar
- Appels A, Mulder P: Imminent myocardial infarction: a psychological study. J Human Stress. 1984, 10: 129-134.View ArticlePubMedGoogle Scholar
- Duffy SW, Rohan TE, Kandel R, Prevost TC, Rice K, Myles JP: Misclassification in a matched case-control study with variable matching ratio: application to a study of c-erbB-2 overexpression and breast cancer. Stat Med. 2003, 22: 2459-2468. 10.1002/sim.1477.View ArticlePubMedGoogle Scholar
- Jacobsen SJ, Roberts RO: Re: Effect of nonsteroidal anti-inflammatory agents and finasteride on prostate cancer risk. J Urol. 2003, 169: 1798-1799. 10.1097/01.ju.0000057804.01025.13.View ArticlePubMedGoogle Scholar
- Zheng G, Tian X: The impact of diagnostic error on testing genetic association in case-control studies. Stat Med. 2005, 24: 869-882. 10.1002/sim.1976.View ArticlePubMedGoogle Scholar
- Ioannidis JP, Trikalinos TA, Ntzani EE, Contopoulos-Ioannidis DG: Genetic associations in large versus small studies: an empirical assessment. Lancet. 2003, 361: 567-571. 10.1016/S0140-6736(03)12516-0.View ArticlePubMedGoogle Scholar
- Lohmueller KE, Pearce CL, Pike M, Lander ES, Hirschhorn JN: Meta-analysis of genetic association studies supports a contribution of common variants to susceptibility to common disease. Nat Genet. 2003, 33: 177-182. 10.1038/ng1071.View ArticlePubMedGoogle Scholar
- Hirschhorn JN, Altshuler D: Once and again-issues surrounding replication in genetic association studies. J Clin Endocrinol Metab. 2002, 87: 4438-4441. 10.1210/jc.2002-021329.View ArticlePubMedGoogle Scholar
- Ioannidis JP: Genetic associations: false or true?. Trends Mol Med. 2003, 9: 135-138. 10.1016/S1471-4914(03)00030-3.View ArticlePubMedGoogle Scholar
- Lansbury PTJ: Back to the future: the 'old-fashioned' way to new medications for neurodegeneration. Nat Med. 2004, 10 Suppl: S51-7. 10.1038/nrn1435.View ArticlePubMedGoogle Scholar
- Press MF, Hung G, Godolphin W, Slamon DJ: Sensitivity of HER-2/neu antibodies in archival tissue samples: potential source of error in immunohistochemical studies of oncogene expression. Cancer Res. 1994, 54: 2771-2777.PubMedGoogle Scholar
- Burd L, Kerbeshian J, Klug MG: Neuropsychiatric genetics: misclassification in linkage studies of phenotype-genotype research. J Child Neurol. 2001, 16: 499-504.View ArticlePubMedGoogle Scholar
- Mote VL, Anderson RL: An investigation of the effect of misclassification on the properties of chisquare-tests in the analysis of categorical data. Biometrika. 1965, 52: 95-109.PubMedGoogle Scholar
- Gordon D, Ott J: Assessment and management of single nucleotide polymorphism genotype errors in genetic association analysis. Pac Symp Biocomput. 2001, 18-29.Google Scholar
- Carroll RJ, Gail MH, Lubin JH: Case-control studies with errors in covariates. J Am Stat Assoc. 1993, 88: 185-199.Google Scholar
- Mitra SK: On the limiting power function of the frequency chi-square test. Ann Math Stat. 1958, 29: 1221-1233.View ArticleGoogle Scholar
- Cochran WG: The chi-square test of goodness of fit. Ann Math Stat. 1952, 23: 315-345.View ArticleGoogle Scholar
- Tosteson TD, Buzas JS, Demidenko E, Karagas M: Power and sample size calculations for generalized regression models with covariate measurement error. Stat Med. 2003, 22: 1069-1082. 10.1002/sim.1388.View ArticlePubMedGoogle Scholar
- Sasieni PD: From genotypes to genes: doubling the sample size. Biometrics. 1997, 53: 1253-1261.View ArticlePubMedGoogle Scholar
- Czika W, Weir BS: Properties of the multiallelic trend test. Biometrics. 2004, 60: 69-74. 10.1111/j.0006-341X.2004.00166.x.View ArticlePubMedGoogle Scholar
- Box GEP, Hunter WG, Hunter JS: Statistics for Experimenters. Wiley series in probability and mathematical statistics. 1978, New York, John Wiley and SonsGoogle Scholar
- Corder EH, Saunders AM, Strittmatter WJ, Schmechel DE, Gaskell PC, Small GW, Roses AD, Haines JL, Pericak-Vance MA: Gene dose of apolipoprotein E type 4 allele and the risk of Alzheimer's disease in late onset families. Science. 1993, 261: 921-923.View ArticlePubMedGoogle Scholar
- Sloane PD, Zimmerman S, Suchindran C, Reed P, Wang L, Boustani M, Sudha S: The public health impact of Alzheimer's Disease, 2000-2050: potential implication of treatment advances. Annu Rev Public Health. 2002, 23: 213-231. 10.1146/annurev.publhealth.23.100901.140525.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.