Cross Validated
2026-11-02 19:04 UTC
By clara_ra_lam
AI-113-20261102-social-media-4fa0b411
How can I calculate the association between compositional data (mixtures) and an external variable?
I have identified three archetypes in a dataset using Principal Convex Hull Analysis (PCHA). I would now like to investigate whether the composition (or mixture) of individuals across these three archetypes is associated with an external variable. To illustrate the problem, imagine a dataset of coffee drinkers in which PCHA identifies three archetypes: A: People who drink a lot of high-quality coffee. B: People who drink very little low-quality coffee. C: People who drink a lot of low-quality coffee. We have hundreds of individuals distributed across this archetypal space. Each individual is represented by a mixture of the three archetypes. For example, one individual might have a composition of A = 70%, B = 20%, and C = 10%. Importantly, these three proportions always sum to 1 (100%). I would like to investigate whether these individual compositions are associated with an external variable, for example, the number of hours each individual sleeps per week. My initial thought was to calculate Pearson or Spearman correlations between the external variable and each of the three archetype proportions. However, because the proportions are constrained to sum to 1, they are not independent, and this approach could lead to misleading conclusions. One difficulty I encounter is that an underlying association might be distributed across multiple archetypes. For example, suppose that people who sleep less tend to drink more coffee. However, in my archetypal decomposition, high coffee co…
I have identified three archetypes in a dataset using Principal Convex Hull Analysis (PCHA). I would now like to investigate whether the composition (or mixture) of individuals across these three archetypes is associated with an external variable. To illustrate the problem, imagine a dataset of coffee drinkers in which PCHA identifies three archetypes: A: People who drink a lot of high-quality coffee. B: People who drink very little low-quality coffee. C: People who drink a lot of low-quality coffee. We have hundreds of individuals distributed across this archetypal space. Each individual is represented by a mixture of the three archetypes. For example, one individual might have a composition of A = 70%, B = 20%, and C = 10%. Importantly, these three proportions always sum to 1 (100%). I would like to investigate whether these individual compositions are associated with an external variable, for example, the number of hours each individual sleeps per week. My initial thought was to calculate Pearson or Spearman correlations between the external variable and each of the three archetype proportions. However, because the proportions are constrained to sum to 1, they are not independent, and this approach could lead to misleading conclusions. One difficulty I encounter is that an underlying association might be distributed across multiple archetypes. For example, suppose that people who sleep less tend to drink more coffee. However, in my archetypal decomposition, high coffee co…
Full article content could not be extracted automatically. Read the original below.
Source:
Cross Validated
· stats.stackexchange.com