RPC

The Robust Profile Clustering (RPC) model is a population-based clustering technique that adjusts for differences that may exist within different subpopulations. Here, participants are clustered at two levels: (1) globally, where subjects are assignd to an overall population-level cluster using an overfitted mixture model, and (2) locally, where variations to global patterns are accommodated via a Beta-Bernoulli process dependent on subpopulation differences.

Getting Started

The code and supporting materials are run using MATLAB software. To run the example data, you will need:

• SourceCode/RPC_example.m - run RPC and create output data
• SourceCode/drchrnd.m - Dirichlet random generator function
• SourceCode/heatmap.m - function file to generate desired heatmap figures (this file is only needed if running MATLAB 2016a or earlier)
• ExampleData.mat - input data source

Example Data

The example dataset is a MAT-file that contains the following variables:

• subdata: 1800x50 matrix. This matrix is the input dataset containing subject level data for 50 variables. Each variable assigned a single categorical value (1,2,3,4).
• sub_nu: binary 50x3 matrix. This matrix is used as a reference to illustrate the true probability of allocation for each variable within each subpopulation to global (ν= 1) or local (ν= 0).
• subpop_i: 1800x1 vector. This vector contains subpopulation ID for the 1800 subjects included in the dataset.
• Subpop1_true: 50x3 matrix. This matrix contains the 3 modal cluster patterns modally expected from subpopulation 1. Each column represents a different global pattern with a deviation reflected in 13 of the 50 variables.
• Subpop2_true: 50x3 matrix. This matrix contains the 3 modal cluster patterns modally expected from subpopulation 2. Each column represents a different global pattern with a deviation reflected in 24 of the 50 variables.
• Subpop3_true: 50x3 matrix. This matrix contains the 3 modal cluster patterns modally expected from subpopulation 3. Each column represents a different global pattern with no deviation in any of the 50 variables.

Example Output

Execution of the RPC_example.m file with Example dataset should generate the following results:

• Posterior median estimates of all RPC model parameters (π,λ,θ_0,θ_1,ν,β) saved as RPCparm_ex.mat
• Text file outputs of π,ν,β (pis.txt, nus.txt, betas.txt).
• Text file output of predicted modal pattern of global clusters (GlobalPattern_predicted.txt)
• Trace plot figure of π (pis.png)
• Trace plot figure of β (betas.png)
• Dendrogram plot from Posterior pairwise label switching step (dendrogram.png)
• Heat map comparison side-by-side plots of predicted deviations by subpopulation and actual deviations by subpopulation. (G_deviations.png) These results are stored in /Output for reference.

Application to Nutritional Cohort Studies

Supporting files and data are provided to apply the RPC method to two national cohorts.

1. National Health and Nutrition Examination Survey - RPC was applied here to examine racial/ethnic heterogeneity amongst low-income adult women
2. Hispanic Community Health Study/Study of Latinos - RPC was applied here to examine geographic and ethnic heterogeneity amongst US Hispanic/Latino residents

Authors

Briana Stephenson, Amy Herring, Andrew Olshan