--- title: "Analysis of bivariate binomial data: Twin analysis" author: Klaus Holst & Thomas Scheike date: "`r Sys.Date()`" output: rmarkdown::html_vignette: fig_caption: yes fig_width: 7.15 fig_height: 5.5 vignette: > %\VignetteIndexEntry{Analysis of bivariate binomial data: Twin analysis} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) library(mets) ``` Overview ========== When analysing bivariate binomial data with the aim of learning about the dependence present, possibly after adjusting for covariates, many models are available. * Random-effects logistic regression models covered elsewhere (`glmer` in `lme4`). In the mets package you can fit the * Pairwise odds ratio model * Bivariate Probit model + With random effects + Special functionality for polygenic random effects modelling such as ACE, ADE ,AE and so forth. * Additive gamma random effects model + Special functionality for polygenic random effects modelling such as ACE, ADE ,AE and so forth. Typically it can be hard or impossible to specify random effects models with special structure among the parameters of the random effects. This is possible with our models. To be concrete about the model structure, assume that we have paired binomial data $(Y_1, Y_2, X_1, X_2)$ where the responses are $Y_1, Y_2$ and the covariates are $X_1, X_2$. We start by giving a brief description of these different models. First we for bivariate data one can specify the marginal probability using logistic regression models \[ logit(P(Y_i=1|X_i)) = \alpha_i + X_i^T \beta i=1,2. \] These models can be estimated under working independence (Zeger and Liang, 1986). A typical twin analysis will consist of looking at both * Pairwise odds ratio model * Bivariate Probit model The additive gamma can be used for the same as the bivariate probit model but is more restrictive in terms of dependence structure, but is nevertheless still valuable to have also as a check of results of the bivariate probit model. Biprobit with random effects ============================= For these models we assume that given random effects $Z$ and a covariate vector $V_{12}$ we have independent probit regression models \[ \mbox{probit}(P(Y_i=1|X_i, Z)) = \alpha_i + X_i^T \beta + V_{12}^T Z, \quad i=1,2. \] where $Z$ follows a bivariate normal distribution with covariance $\Sigma$. The general covariance structure $\Sigma$ makes the model very flexible. We note that - Parameters $\beta$ are subject-specific - $\Sigma$ reflects the dependence structure The more standard $\mbox{logit}$ link rather than the $\mbox{probit}$ link is often used and implemented in, for example, Tutz and Hennevogl (1996). The advantage is that one obtains an odds-ratio interpretation of the subject-specific effects, but numerical integration is then needed to fit the model. Pairwise odds ratio model ------------------------ The pairwise odds ratio model specifies that given $X_1, X_2$ the marginal models are \[ logit(P(Y_i=1|X_i)) = \alpha_i + X_i^T \beta i=1,2 \] The primary object of interest are the odds ratio between $Y_{1}$ and $Y_{2}$ \[ \gamma_{12} = \frac{ P( Y_{ki} =1 , Y_{kj} =1) P( Y_{ki} =0 , Y_{kj} =0) }{ P( Y_{ki} =1 , Y_{kj} =0) P( Y_{ki} =0 , Y_{kj} =1) } \] given $X_{ki}$, $X_{kj}$, and $Z_{kji}$. We model the odds ratio with the regression \[ \gamma_{12} = \exp( Z_{12}^T \lambda) \] Where $Z_{12}$ are covariates that may influence the odds ratio between $Y_{1}$ and $Y_{2}$ and contains the marginal covariates \cite{carey-1993,dale1986global,palmgren1989,molenberghs1994marginal}. This odds-ratio is given covariates as well as marginal covariates. The odds-ratio and marginals specify the joint bivariate distribution via the so-called Placckett-distribution. One way of fitting this model is the ALR (alternating logistic regression) algorithm, described in several papers (Kuk, 2004; Kuk, 2007; Qaqish, 2012). We estimate the parameters in a two-stage procedure: * Estimating the marginal parameters via GEE * Using marginal estimates, estimate dependence parameters This gives efficient estimates of the dependence parameters due to orthogonality, though some efficiency may be gained for the marginal parameters by using the full likelihood or iterative fitting such as ALR. The pairwise odds-ratio model is very useful, but one does not have a random effects model. Additive gamma model ---------------------- Again we operate under marginal logistic regression models are \[ logit(P(Y_i=1|X_i)) = \alpha_i + X_i^T \beta i=1,2 \] First with just one random effect $Z$ we assume that conditional on $Z$ the responses are independent and follow the model \[ logit(P(Y_i=1|X_i,Z)) = exp( -Z \cdot \Psi^{-1}(\lambda_{\bullet},\lambda_{\bullet},P(Y_i=1|X_i)) ) \] where $\Psi$ is the laplace transform of $Z$ where we assume that $Z$ is gamma distributed with variance $\lambda_{\bullet}^{-1}$ and mean 1. In general $\Psi(\lambda_1,\lambda_2)$ is the laplace transform of a Gamma distributed random effect with $Z$ with mean $\lambda_1/\lambda_2$ and variance $\lambda_1/\lambda_2^2$. We fit this model by * Estimating the marginal parameters via GEE * Using marginal estimates, estimate dependence parameters To deal with multiple random effects we consider random effects $Z_i i=1,...,d$ such that $Z_i$ is gamma distributed with * mean: $\lambda_j/\lambda_{\bullet}$ * variance: $\lambda_j/\lambda_{\bullet}^2$, where we define the scalar $\lambda_{\bullet}$ below. Now given a cluster-specific design vector $V_{12}$ we assume that \[ V_{12}^T Z \] is gamma distributed with mean 1 and variance $\lambda_{\bullet}^{-1}$ such that critically the random effect variance is the same for all clusters. That is \[ \lambda_{\bullet} = V_{12}^T (\lambda_1,...,\lambda_d)^T \] We return to some specific models below, and show how to fit the ACE and AE model using this set-up. One last option in the model-specification is to specify how the parameters $\lambda_1,...,\lambda_d$ are related. We thus can specify a matrix $M$ of dimension $p \times d$ such that \[ (\lambda_1,...,\lambda_d)^T = M \theta \] where $\theta$ is d-dimensional. If $M$ is diagonal we have no restrictions on parameters. This parametrization is obtained with the var.par=0 option that thus estimates $\theta$. The DEFAULT parametrization instead estimates the variances of the random effects (var.par=1) via the parameters $\nu$ \[ M \nu = ( \lambda_1/\lambda_{\bullet}^2, ...,\lambda_d/\lambda_{\bullet}^2)^T \] The basic modelling assumption is now that given random effects $Z=(Z_1,...,Z_d)$ we have independent probabilites \[ logit(P(Y_i=1|X_i,Z)) = exp( -V_{12,i}^T Z \cdot \Psi^{-1}(\lambda_{\bullet},\lambda_{\bullet},P(Y_i=1|X_i)) ) i=1,2 \] We fit this model by - Estimating the marginal parameters via GEE - Using marginal estimates, estimate dependence parameters Even though the model not formally in this formulation allows negative correlation in practice the parameters can be negative and this reflects negative correlation. An advantage is that no numerical integration is needed. The twin-stutter data ------------------------ We consider the twin-stutter where for pairs of twins that are either dizygotic or monozygotic we have recorded whether the twins are stuttering \cite{twinstut-ref} We here consider MZ and same sex DZ twins. Looking at the data ```{r} library(mets) data(twinstut) twinstut$binstut <- 1*(twinstut$stutter=="yes") twinsall <- twinstut twinstut <- subset(twinstut,zyg%in%c("mz","dz")) head(twinstut) twinstut <- subset(twinstut,tvparnr < 3000) ``` Pairwise odds ratio model ------------------------- We start by fitting an overall dependence OR for both MZ and DZ even though the dependence is expected to be different across zygosity. The first step is to fit the marginal model adjusting for marginal covariates. We here note that there is a rather strong gender effect in the risk of stuttering. ```{r} margbin <- glm(binstut~factor(sex)+age,data=twinstut,family=binomial()) summary(margbin) ``` Now estimating the OR parameter. We see a strong dependence with an OR at around 8 that is clearly significant. ```{r twostage1} bina <- binomial_twostage(margbin,data=twinstut,var.link=1, clusters=twinstut$tvparnr,detail=0) summary(bina) ``` Now, and more interestingly, we consider an OR that depends on zygosity and note that MZ twins have a much larger OR than DZ twins. This type of result is somewhat complex to interpret: one possibility is a genetic effect, another is a stronger shared environmental effect for MZ twins. ```{r twostage2} # design for OR dependence theta.des <- model.matrix( ~-1+factor(zyg),data=twinstut) bin <- binomial_twostage(margbin,data=twinstut,var.link=1, clusters=twinstut$tvparnr,theta.des=theta.des) summary(bin) ``` We now consider further regression modelling of the OR structure by considering possible interactions between sex and zygosity. We see that MZ has a much higher dependence and that males have a much lower dependence. We tested for interaction in this model and these were not significant. ```{r twostage3} twinstut$cage <- scale(twinstut$age) theta.des <- model.matrix( ~-1+factor(zyg)+factor(sex),data=twinstut) bina <- binomial_twostage(margbin,data=twinstut,var.link=1, clusters=twinstut$tvparnr,theta.des=theta.des) summary(bina) ``` Bivariate Probit model ------------------------ ```{r} library(mets) data(twinstut) twinstut <- subset(twinstut,zyg%in%c("mz","dz")) twinstut$binstut <- 1*(twinstut$stutter=="yes") head(twinstut) twinstut <- subset(twinstut,tvparnr < 2000) ``` First testing for same dependence in MZ and DZ that we recommend doing by comparing the correlations of MZ and DZ twins. Apart from regression correction in the mean this is an un-structured model, and the useful concordance and casewise concordance estimates can be reported from this analysis. ```{r biprobit1} b1 <- bptwin(binstut~sex,data=twinstut,id="tvparnr",zyg="zyg",DZ="dz",type="un") summary(b1) ``` Polygenic modelling ------------------- We now turn attention to specific polygenic modelling where special random effects are used to specify ACE, AE, ADE models and so forth. This is very easy with the bptwin function. The key parts of the output are the sizes of the genetic component A and the environmental component, and we can compare with the results of the unstructed model above. Also formally we can test if this submodel is acceptable by a likelihood ratio test. ```{r bptwin1} b1 <- bptwin(binstut~sex,data=twinstut,id="tvparnr",zyg="zyg",DZ="dz",type="ace") summary(b1) ``` ```{r bptwin2} b0 <- bptwin(binstut~sex,data=twinstut,id="tvparnr",zyg="zyg",DZ="dz",type="ae") summary(b0) ``` Additive gamma random effects ----------------------------- Fitting first a model with different size random effects for MZ and DZ. We note that as before in the OR and biprobit model the dependence is much stronger for MZ twins. We also test if these are the same by parametrizing the OR model with an intercept. This clearly shows a significant difference. ```{r addgamma1} theta.des <- model.matrix( ~-1+factor(zyg),data=twinstut) margbin <- glm(binstut~sex,data=twinstut,family=binomial()) bintwin <- binomial_twostage(margbin,data=twinstut,model="gamma", clusters=twinstut$tvparnr,detail=0,theta=c(0.1)/1,var.link=1, theta.des=theta.des) summary(bintwin) # test for same dependence in MZ and DZ theta.des <- model.matrix( ~factor(zyg),data=twinstut) margbin <- glm(binstut~sex,data=twinstut,family=binomial()) bintwin <- binomial_twostage(margbin,data=twinstut,model="gamma", clusters=twinstut$tvparnr,detail=0,theta=c(0.1)/1,var.link=1, theta.des=theta.des) summary(bintwin) ``` Polygenic modelling ----------------------- First setting up the random effects design and the relationship between variance parameters. We see that the genetic random effect has size 1 for MZ and 0.5 for DZ subjects, who have shared and non-shared genetic components with variance 0.5 such that the total genetic variance is the same for all subjects. The shared environmental effect is the same for all. Thus two parameters with these constraints. ```{r polygenic1} out <- twin_polygen_design(twinstut,id="tvparnr",zygname="zyg",zyg="dz",type="ace") head(cbind(out$des.rv,twinstut$tvparnr),10) out$pardes ``` Now, fitting the ACE model, we see that the variance of the genetic, component, is 1.5 and the environmental variance is -0.5. Thus suggesting that the ACE model does not fit the data. When the random design is given we automatically use the gamma fralty model. ```{r polygenic2} margbin <- glm(binstut~sex,data=twinstut,family=binomial()) bintwin1 <- binomial_twostage(margbin,data=twinstut, clusters=twinstut$tvparnr,detail=0,theta=c(0.1)/1,var.link=0, random.design=out$des.rv,theta.des=out$pardes) summary(bintwin1) ``` For this model we estimate the concordance and casewise concordance as well as the marginal rates of stuttering for females. ```{r} concordanceTwinACE(bintwin1,type="ace") ``` The E component was not consistent with the fit of the data and we now consider instead the AE model. ```{r polygenic_ae} out <- twin.polygen.design(twinstut,id="tvparnr",zygname="zyg",zyg="dz",type="ae") bintwin <- binomial_twostage(margbin,data=twinstut, clusters=twinstut$tvparnr,detail=0,theta=c(0.1)/1,var.link=0, random.design=out$des.rv,theta.des=out$pardes) summary(bintwin) ``` Again, the concordance can be computed: ```{r} concordanceTwinACE(bintwin,type="ae") ``` SessionInfo ============ ```{r} sessionInfo() ```