 Proceedings
 Open access
 Published:
Using mixture models to characterize diseaserelated traits
BMC Genetics volumeÂ 6, ArticleÂ number:Â S99 (2005)
Abstract
We consider 12 eventrelated potentials and one electroencephalogram measure as diseaserelated traits to compare alcoholdependent individuals (cases) to unaffected individuals (controls). We use two approaches: 1) twoway analysis of variance (with sex and alcohol dependency as the factors), and 2) likelihood ratio tests comparing sex adjusted values of cases to controls assuming that within each group the trait has a 2 (or 3) component normal mixture distribution. In the second approach, we test the null hypothesis that the parameters of the mixtures are equal for the cases and controls. Based on the twoway analysis of variance, we find 1) males have significantly (p < 0.05) lower mean response values than females for 7 of these traits. 2) Alcoholdependent cases have significantly lower mean response than controls for 3 traits. The mixture analysis of sexadjusted values of 1 of these traits, the eventrelated potential obtained at the parietal midline channel (ttth4), found the appearance of a 3component normal mixture in cases and controls. The mixtures differed in that the cases had significantly lower mean values than controls and significantly different mixing proportions in 2 of the 3 components. Implications of this study are: 1) Sex needs to be taken into account when studying risk factors for alcohol dependency to prevent finding a spurious association between alcohol dependency and the risk factor. 2) Mixture analysis indicates that for the eventrelated potential "ttth4", the difference observed reflects strong evidence of heterogeneity of response in both the cases and controls.
Background
Diseaserelated traits (DRTs) may provide more powerful phenotypes than the disease itself for identifying alcohol dependency genes. For example, an alcohol DRT phenotype might be due to a single major gene with high penetrance, while alcohol dependency may result from the action of several genes and environmental factors. In characterizing a DRT, we first compare affected to unaffected individuals. If the DRT is due to 1 of many disease predisposing genes, then the responses in both affected and unaffected individuals may be a mixture with 2 or 3 components depending on the effect of genotype on the DRT in affected individuals and unaffected individuals. Lo et al. [1] successfully applied this idea in their study of working memory, a schizophreniarelated trait. Assuming a withingroup mixture of an exponential and a normal distribution, they found significant differences between normal controls and relatives of patients with schizophrenia. These results had not been noted when they compared these two groups using traditional 2sample tests comparing means or medians.
Methods
The sample
We considered Collaborative Study on the Genetics of Alcoholism (COGA) family data provided by Genetic Analysis Workshop 14 (GAW14) Problem 1 [2]. One affected individual was sampled at random from each of the 105 families providing data on the electrophysiological measures. We then randomly sampled one "purely unaffected" individual from those families, when such a person was available. The result was a sample of 105 cases, the alcoholdependent affected individuals, and 50 controls, the purely unaffected individuals. Seventythree percent of the affected individuals were male, and 22% of the unaffected individuals were male.
Variables
The 12 eventrelated potentials (ERPs), phenotypes ttth1, ttth2, ttth3, ttth4, ttdt1, ttdt2, ttdt3, ttdt4, ntth1, ntth2, ntth3, and ntth4, and one electroencephalogram (EEG) phenotype, ecb21, were considered, as well as the sex of the individual.
Statistical methods
The 2way analysis of variance used sex, disease status (affected vs. unaffected), and the sexdisease status interaction as factors.
A mixture model analysis incorporated the findings on sex obtained in the 2way analysis of variance, and was done on sexadjusted values for those traits in which we found a difference between males and females. The adjustment was Y_{Adj} = Y  d_{f}, where for females and d_{f} = 0 for males.
This analysis assumed that conditional on whether an individual is a case or control, the distribution of the trait has a 2component normal mixture distribution. If we let X = 1 when an individual is a control and X = 2 when an individual is a case, then the density of the trait, Y, is
f_{ x }(y) = Ï€_{1x}Ï†(y; Î¼_{1x}, Ïƒ) + Ï€_{2x}Ï†(y; Î¼_{2x}, Ïƒ) for x = 1, 2, Â Â Â (1)
where Ï†(y; Î¼, Ïƒ) denotes the normal density with mean, Î¼, and standard deviation, Ïƒ and Ï€_{1x}+ Ï€_{2x}= 1. Without loss of generality, Î¼_{1x}< Î¼_{2x}and 0 < Ï€_{ ix }< 1 for x = 1, 2.
The null hypothesis
H_{00}: Î¼_{i 2}= Î¼_{i 1}and Ï€_{i 2}= Ï€_{i 1}for i = 1, 2 Â Â Â (2)
can be tested against the alternative of equal component means and unequal mixing proportions
H_{01}: Î¼_{i 2}= Î¼_{i 1}and Ï€_{i 2}â‰ Ï€_{i 1}for i = 1, 2 Â Â Â (3)
using a likelihood ratio test (LRT) statistic. Under the null hypothesis, the LRT statistic has an asymptotic chisquare distribution with 1 df. We can also consider an alternative of unequal means and equal mixing proportion, i.e.,
H_{10}: Î¼_{i 2}â‰ Î¼_{i 1}and Ï€_{i 2}= Ï€_{i 1}for i = 1, 2. Â Â Â (4)
Finally we also consider an alternative of unequal means and unequal mixing proportions
H_{11}: Î¼_{i 2}â‰ Î¼_{i 1}and Ï€_{i 2}â‰ Ï€_{i 1}for i = 1, 2. Â Â Â (5)
If we reject H_{00} we might want to consider the alternative H_{11} given in (5) versus H_{01} using a 2 df chisquare test or H_{10} using a 1 df chi square test.
Following the same logic, we considered a set of 3component normal mixture models for cases and controls. Similar to model (1), we considered 3 component mixtures with equal within component variances. Thus the general equation for the mixture density is
f_{ x }(y) = Ï€_{1x}Ï†(y; Î¼_{1x}, Ïƒ) + Ï€_{2x}Ï†(y; Î¼_{2x}, Ïƒ) + Ï€_{3x}Ï†(y; Î¼_{3x}, Ïƒ) for x = 1, 2, Â Â Â (6)
where Ï€_{1x}+ Ï€_{2x}+ Ï€_{3x}= 1 and 0 < Ï€_{ ix }< 1 for x = 1, 2, i = 1, 2, 3. We again set Î¼_{1x}< Î¼_{2x}< Î¼_{3x}, and refer to the component having mean Î¼_{ ix }as the i^{th} component. As in comparing cases to controls assuming a 2component normal mixtures, we estimate the parameters and test hypotheses under various 3component normal mixture models. These include 1) equal parameter values for cases and controls (6 parameters); 2) unequal mixing proportions, but equal component means (8 parameters); 3) unequal component means but equal mixing proportions (9 parameters); 4) equal first component means and equal first component mixing proportions (9 parameters); 5) unequal mixing proportions and unequal within component means (11 parameters).
The expectationmaximization algorithm (EM) [3], a general approach to maximum likelihood estimation (MLE), is applied to estimate parameters Ï€_{ ix }, Î¼_{ ix }and Ïƒ for i = 1, 2 (or 3) and x = 1, 2.
A method of Maller and Zhou [4] allows us to test specific hypotheses using the LRT. However, when the mixing proportions are on the boundary of the parameter space and the parameters are not identifiable under the null model, the LRT does not follow the usual asymptotic chisquare distribution with degrees of freedom equal to the difference in the number of parameters between the 2 hypotheses. In this case, to select the model, we considered both the Akaike information criterion (AIC) [5] and the Bayesian information criterion (BIC) [6]. In using AIC and BIC, we selected the model with the smallest AIC or BIC value. We used both because in many model selection studies, it is found that AIC tends to select more complex models, while BIC tends to penalize complex models heavily, giving preference to simpler models. This appears to hold in selecting the number of components in the mixture analysis [7].
Results
2way analysis of variance
Table 1 shows the results of 2way analysis of variance. In each of the 13 DRTs, the sexdisease interaction was nonsignificant (all p > 0.17). For 7 traits, sex was a significant factor; disease was a significant factor for only 3 traits. This was unexpected, because based on the data description, we expected to observe differences between cases and controls on all measures. In Table 1 we report the confidence intervals for the means on comparing cases to controls and on comparing males to females.
Mixture analysis
The means observed for males were slightly lower than those observed for females wherever there was a significant sex difference. Thus the adjusted values for females were slightly smaller than the original values. The values of the adjustment used in the females for the electrophysiological measures, when there was a significant sex difference, range from 0.32 to 2.00, with the sex adjustment value for the ERP obtained at the parietal midline channel, ttth4, equal to 0.45.
We failed to reject H_{00} for the alternative H_{01} (equal means, unequal mixing proportions) for every DRT considered. We rejected H_{00} at the 0.05 level for alternative H_{10} (equal mixing proportions, unequal means) in the case of trait ttth2 (Ï‡^{2} = 6.3, df = 2, pvalue = 0.04). In the case of ttth4, the DRT obtained at the parietal midline channel, we reject H_{00} (Ï‡^{2} = 8.9, df = 3, pvalue = 0.03), H_{01} (Ï‡^{2} = 7.1, df = 2, pvalue = 0.03), and H_{10} (Ï‡^{2} = 4.4, df = 1, pvalue = 0.04) for H_{11}, indicating that we may have both unequal mixing proportions and unequal means for cases and controls. Thus, while the analysis of variance shows that the controls have a higher mean value of ttth4, the mixture analysis indicates more complex distribution differences. Both component means are higher in the controls than the cases (3.83 vs. 3.33 and 6.00 vs. 4.64), whereas the estimated proportion of controls in component with the higher mean is lower than that for the cases (0.04 vs. 0.31).
Applying similar methods we used LRTs to find the most parsimonious 3component normal mixture distribution for ttth4. Upon doing this we rejected hypotheses of equal mixing proportions for cases and controls (Ï‡^{2} = 12.6, df = 2, pvalue = 0.002) and of equal component means for cases and controls (Ï‡^{2} = 13.4, df = 3, pvalue = 0.004). Upon exploring further, we could not reject a hypothesis that cases and controls had equal means and proportions in the first component, i.e., the component with the lowest mean (Ï‡^{2} = 0.2, df = 2, pvalue > 0.9).
Using AIC and BIC, we compared the likelihoods of our most parsimonious models accounting for the differences in cases and controls. That is, we compared the likelihoods of a 1component normal density model (with cases and controls having unequal means and equal variances), to a 2component normal mixture model (with cases and controls having unequal mixing proportions and unequal component means), and to a 3component normal mixture model (with cases and controls having unequal mixing proportions and unequal means for 2 out of 3 of the components). The AIC values of the single normal density, 2component mixture model and 3component mixture model are 381.2, 380.8, and 371.5, respectively. The BIC values for the above three models are 390.3, 402.1, and 398.8, respectively. AIC leads to a 3component mixture model, while a single density model is indicated by BIC. When there are inconsistencies in model selection based on AIC and BIC, Leroux [8] recommends the choice of the number of components might be based on a direct comparison of the fitted frequency distributions. Figure 1(A, B) contains the density histograms in cases and controls for this trait, ttth4. It shows that a single normal density does not appear to be sufficient. Based on this, we have selected the 3component mixture model as most appropriate. Thus our selected model for the distribution of ttth4 is
f_{1}(y) = 0.19 Ï†(y; 2.76, 0.41) + 0.77 Ï†(y; 4.09, 0.41) + 0.04 Ï†(y; 6.05, 0.41) for controls
and
f_{2}(y) = 0.19 Ï†(y; 2.76, 0.41) + 0.52 Ï†(y; 3.52, 0.41) + 0.29 Ï†(y; 4.78, 0.41) for cases.
Figure 1(C, D) plots these mixtures.
Discussion
In the case of ttth4, the first component mean and corresponding mixing proportion are the same for cases and controls, and there is a general shift, in the direction that the mean ttth4 is lower for alcoholdependent individuals than their unaffected relatives in the other 2 component means. From the final model, we can see that the explanation for a lower mean in the cases is the lower mean and a lower estimated proportion in the second component compared to the control group.
An interesting result is that, with sex controlled, there are few significant differences between cases and controls, namely ttth4, ttdt3, and ttdt4. For moderate estimated effect size
, with a sample size n = 50 in each group, power is equal to or larger than 0.50. Thus we have reasonable power to detect differences between cases and controls. Whenever our alcoholdependent sample has a larger percentage of males than our control sample, any differences observed between cases and controls may reflect these sex differences rather than differences in the disease groups. Regardless of the sample makeup, taking sex into account should always be done when studying factors related to alcoholism. Another reason we do not see large differences between our controls and the cases may be that these controls all have a family history of alcoholism.
In this study we report significant findings observed on investigating 13 correlated measures. As in any study in which a large number of tests have been done, we would expect some significant findings due to chance. Thus the results here must be considered as preliminary. On the other hand, given that these measures were included in the COGA dataset [2] as being potential alcohol risk factors, it is rather surprising that so few significant findings are observed on comparing cases to controls.
Conclusion
Twoway analysis of variance (sex and disease) indicates that controlling for sex there is a significant difference between alcoholdependent cases and controls for only 3 ERPs, namely ttth4, ttdt3, and ttdt4. Comparison of both the 2component and 3component normal mixture parameters for ttth4, the ERP obtained at the parietal midline channel, indicate these differences may reflect the same mixing proportion and mean in the component having the lowest mean, but unequal mixing proportions and unequal component means in the other 2 components.
Abbreviations
 AIC:

Akaike information criterion
 BIC:

Bayesian information criterion
 COGA:

Collaborative Study on the Genetics of Alcoholism
 DRT:

Diseaserelated trait
 EEG:

Electroencephalogram
 EM:

Expectationmaximization algorithm
 ERPs:

Eventrelated potentials
 GAW14:

Genetic Analysis Workshop 14
 LRT:

Likelihood ratio test
 MLE:

Maximum likelihood estimation
References
Lo Y, Matthysse S, Rubin DB, Holzman PS: Permutation tests for detecting and estimating mixtures in task performance within groups. Stat Med. 2002, 21: 19371953. 10.1002/sim.1140.
Edenberg HJ, Bierut LJ, Boyce P, Cao M, Cawley S, Chiles R, Doheny KF, Hansen M, Hinrichs T, Jones K, Kelleher M, Kennedy GC, Liu G, Marcus G, McBride C, Murray SS, Oliphant A, Pettengill J, Porjesz B, Pugh EW, Rice JP, Rubano T, Shannon S, Steeke R, Tischfield JA, Tsai YY, Zhang C, Begleiter H: Description of the data from the Collaborative Study on the Genetics of Alcoholism (COGA) and singlenucleotide polymorphism genotyping for Genetic Analysis Workshop 14. BMC Genet. 2005, 6 (Suppl 1): S210.1186/147121566S1S2.
Dempster AP, Laird NM, Rubin DB: Maximum likelihood from incomplete data via the EM algorithm. J Roy Stat Soc B Met. 1977, 39: 138.
Maller RA, Zhou S: Survival Analysis with LongTerm Survivors. 1996, New York: Wiley
Akaike H: Information theory and an extension of the maximum likelihood principle. 2nd International Symposium Information Theory. Edited by: Petrov BN, Csaki F. 1973, Budapest: Akademiai Kiado, 267281.
Schwartz G: Estimating the dimensions of a model. Ann Stat. 1978, 6: 461464.
Biernacki C, Govaert G: Choosing models in modelbased clustering and discriminant analysis. J Stat Comput Sim. 1999, 64: 4971.
Leroux BG: Consistent estimation of a mixing distribution. Ann Stat. 1992, 20: 13501360.
Acknowledgements
The authors thank the members of the Stony Brook University Applied Mathematics and Statistics Department's Statistical Genetics Research Group which has met with them weekly throughout this past year and has given constructive criticism and ideas for efficiently implementing the proposed research.
Author information
Authors and Affiliations
Corresponding author
Additional information
Authors' contributions
NRM, SJF, KQY, and GAC conceived of the study, participated in its design and coordination, and helped to draft the manuscript. NRM presented this work. TD carried out all of the analyses including the genetic analyses, data reduction, and statistical analyses.
Rights and permissions
Open Access This article is published under license to BioMed Central Ltd. This is an Open Access article is distributed under the terms of the Creative Commons Attribution License ( https://creativecommons.org/licenses/by/2.0 ), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
About this article
Cite this article
Duan, T., Finch, S.J., Ye, K.Q. et al. Using mixture models to characterize diseaserelated traits. BMC Genet 6 (Suppl 1), S99 (2005). https://doi.org/10.1186/147121566S1S99
Published:
DOI: https://doi.org/10.1186/147121566S1S99