# Compositional data analysis - what's the "method"?

Let $\textbf{X} = (X_1, \ldots, X_n)$ be a vector of responses, where $X_i = (p_1, \ldots, p_k)$ is itself a vector of probabilities.

What method does one use to analyze such data? I want the logic/steps/ideas/concepts behind the method outlined in a short summary, so that I know what's going on before I study it in depth.

To be more specific: I am particularly interested in how one would go about developing the analogue of a linear model: we have some covariates for each response, and we may be interested in whether the mean of the probabilities is influenced by some linear function of the covariates, and then we want to find these parameters, test them, do confidence intervals, and predictions.

What's the "general" idea here? One way could be to just straight-up assume that $X_i$ follows a Dirichlet distribution, but I am more interested in the standard "log-ratio analysis" approach, and what the underlying plan is.

• Hi Margin. Welcome to CV! Can you say whether your outcome variables (your $X_{i}$s, right?) are probabilities based on count data? Or are they continuous between 0 and 1 probability measures? You can edit your question using the "edit" at its lower left to clarify and improve. Mar 12, 2017 at 21:39
• Does $\sum_{j=1}^k p_j =1$ for given $X_i$? Also can we assume that you are having "standard" univariate predictor variables? Or these are compositional data themselves? (By the way, usually the response variables are denoted as $Y$ and $X$ is reserved for the explanatory variables.) Mar 13, 2017 at 0:21

The Logistic-normal model has a number of advantages over the Dirichlet models. The LN models allow for complex covariance of the variables $(X_1,\dots,X_n)$ beyond the covariance introduced by the sum constraint.