Genetic and Neurophysiological Correlates of the Age of Onset of Alcohol Use Disorders in Adolescents and Young Adults

Abstract

Discrete time survival analysis was used to assess the age-specific association of event-related oscillations (EROs) and CHRM2 gene variants on the onset of regular alcohol use and alcohol dependence. The subjects were 2,938 adolescents and young adults ages 12–25. Results showed that the CHRM2 gene variants and ERO risk factors had hazards which varied considerably with age. The bulk of the significant age-specific associations occurred in those whose age of onset was under 16. These associations were concentrated in those subjects who at some time took an illicit drug. These results are consistent with studies which associate greater rates of alcohol dependence among those who begin drinking at an early age. The age specificity of the genetic and neurophysiological factors is consistent with recent studies of adolescent brain development, which locate an interval of heightened vulnerability to substance use disorders in the early to mid teens.

Appendices

The members of Collaborative Study on the Genetics of Alcoholism (COGA)

Principal Investigators: B. Porjesz, V. Hesselbrock, H. Edenberg, L. Bierut, includes 10 different centers: University of Connecticut (V. Hesselbrock); Indiana University (H.J. Edenberg, J. Nurnberger Jr., T. Foroud); University of Iowa (S. Kuperman, J. Kramer); SUNY Downstate (B. Porjesz); Washington University in St. Louis (L. Bierut, A. Goate, J. Rice, K. Bucholz); University of California at San Diego (M. Schuckit); Rutgers University (J. Tischfield); Texas Biomedical Research Institute (L. Almasy), Howard University (R. Taylor) and Virginia Commonwealth University (D. Dick). Other COGA collaborators include: L. Bauer (University of Connecticut); D. Koller, S. O’Connor, L. Wetherill, X. Xuei (Indiana University); Grace Chan (University of Iowa); S. Kang, N. Manz, M. Rangaswamy (SUNY Downstate); J. Rohrbaugh, J.-C. Wang (Washington University in St. Louis); A. Brooks (Rutgers University); and F. Aliev (Virginia Commonwealth University)

A. Parsian and M. Reilly are the NIAAA Staff Collaborators. We continue to be inspired by our memories of Henri Begleiter and Theodore Reich, founding PI and Co-PI of COGA, and also owe a debt of gratitude to other past organizers of COGA, including Ting-Kai Li, currently a consultant with COGA, P. Michael Conneally, Raymond Crowe, and Wendy Reich, for their critical contributions. This national collaborative study is supported by NIH Grant U10AA008401 from the National Institute on Alcohol Abuse and Alcoholism (NIAAA) and the National Institute on Drug Abuse (NIDA).

Appendix: Methodological details

Survival analysis models

Survival analysis models may be distinguished by assumptions made about the effects of the covariates on the hazard. Some models assume that these effects are time-invariant while others enable the estimation of time-varying effects, as well as the use of time-varying covariates.

The hazard function λ(t) is defined as the instantaneous rate of the occurrence of the event.

A commonly used survival model is the Cox proportional hazards model,

$$ \log(\lambda_i(t \vert \hbox{x}_{\rm i})) = \alpha_0(t) + \hbox{x}_{\rm i}^{\prime} \beta $$

where x _i is the vector of covariates for individual i ($\hbox{x}_{i}^{\prime}$ is the transpose of x _i ), β is the vector of coefficients to be estimated and α ₀(t) represents a time-varying baseline hazard.

The assumption in this model is that the hazard due to the covariates is constant over time, in other words, that the effect of a covariate does not change over the interval studied. It is possible to extend this model to enable time-varying effects by substituting β(t) for β in the original model. To model time-varying effects, it may be easier for computational purposes to use a discretized model for log (λ(t)), which enables piecewise estimation of the effects of covariates. Estimates of the parameters could be made using a Poisson log-linear model (Rodriguez 2007). An alternative strategy, DTSA, is to use a discretized model for logit(λ(t)).

The discrete time survival model is

$$ \hbox{logit}(\lambda_i(t_{j} \vert \hbox{x}_{\rm i})) = \alpha_j + \hbox{x}_{\rm i}^{\prime} \beta_j $$

with j ranging over the time intervals. We use

$$ \hbox{logit}(\lambda_i(t_{j} \vert \hbox{x}_{\rm ij})) = \alpha_j + \hbox{x}_{\rm ij}^{\prime} \beta_j $$

to account for the possibility of time-varying covariates and time-varying effects.

The DTSA model parameters can be calculated by creating pseudo-observations, as many for each individual as there are time ranges starting from the first range to the one in which the outcome or censoring occurs. Each pseudo-observation contains covariate information corresponding to the form of the model, in terms of time-invariant and time-varying parameters used. Parameters are estimated by standard logistic regression algorithms (Singer and Willett 1993, 2003a, b; Willett and Singer 1993; Rodriguez 2007).

Treatment of familial data and population structure

Since most of the subjects in the study are from multi-member families it is necessary to account for correlations in the phenotypic data which arise from common genetic and environmental factors within families, and also to account for population stratification. As in a number of other recent papers (Kang et al. 2010), we use genetic relatedness information to model the covariance structure of the phenotypic data. We base our treatment of this problem on the exposition of the generalized estimating equations (GEE) method found in Liang and Zeger (1993) and the more detailed explanation of Hanley et al. (2003) of GEE model construction, and a similar approach based on pedigree information (Yang et al. 2011). The methodology of GEE is to form a weighted regression model in which the weights are a function of the covariance structure of the phenotypic data estimated from the data itself. In the method proposed here, the weights are instead estimated from the genetic relatedness structure of the subjects.

The method is as follows: Given a large enough set of SNPs from the sample, no pair of which is in linkage disequilibrium (LD), the allelic frequency for each SNP is determined. Then the pairwise relationship between all members of each multi-member family is calculated using the algorithm of Choi et al. (2009). This is equivalent to constructing a block-diagonal version of the kinship matrix $\Upphi$ (with elements ϕ _ij ) (Choi et al. 2009, Eq. 3), with the inbreeding coefficients assumed to be zero. This matrix corresponds to the variance-covariance matrix of the phenotypic data as used in the GEE method. The weights assigned to each individual in the regression model in the following manner: Each individual who is not a member of a multi-member family is assigned weight 1. Suppose individual is member i of family with n members $1, \ldots, n$. Then the weight assigned to that person is 1/(1 + 2 ∑ ^j=n _{j =1} ϕ _{ij, i ≠ j} ). This corresponds to the determination of weights in the GEE model (Hanley et al. 2003).

Population stratification was dealt with by using the principal component scores derived from the complete kinship matrix $\Upphi$ as additional independent variables in the regression analysis. This was found to be a satisfactory method in Astle and Balding (2009).

Complete DTSA results for delta ERO and CHRM2 SNPs for regular alcohol use and alcohol dependence in both the entire sample and illicit drug use subsample are found in Tables 5 and 6.

Table 5 DTSA: delta ERO p values for regular alcohol use and alcohol dependence in the entire sample and illicit drug use subsample

Table 6 DTSA: CHRM2 SNP p values for alcohol dependence in the entire sample and illicit drug use subsample