This repository aims to be a pipeline for processing the Sleep Heart Health Study (SHHS) dataset for cluster analysis. SHHS is available for download through the National Sleep Research Repository (NSRR). The R scripts and library functions here work with the CSV files downloaded from the NSRR listed in shhs-process.R. (Users will have to apply to download those files through the NSRR).
Currently, these scripts are intended to be run in a certain order. We are working on combining everything into one package, but to replicate the analysis done in my thesis, it is best to run the scripts.
Loads data into R workspace. The key files are a data csv and a datadict csv (the dictionary of all the variables). Note: You need to adjust file pointers to point to the directories where the files are located.
- Subsets the data from shhs1-data.csv, cleans the data (recoding binary vars as factors, fixing data entry errors, etc.) and builds a data frame of relevant columns.
- workingcvd, workinginterim, workingSHHS2 all do the same thing for those respective datasets.
- Input: data frame from workingSHHS1 (df)
- Calculates a heterochoric correlation (hc) for df, plots correlation heatmap
- Finds ideal number of factors through parallel analysis and VSS.
- Find factor loadings for n factors.
- Plots factor plot and factor diagram.
At this point, one would go back and select only the columns they want in df (that is, the important factors).
Function for plotting scree plot/parallel analysis in a nice, publication-worthy way.
Code taken from https://sakaluk.wordpress.com/2016/05/26/11-make-it-pretty-scree-plots-and-parallel-analysis-using-psych-and-ggplot2/
Sakaluk, J. K., & Short, S. D. (2016). A Methodological Review of Exploratory Factor Analysis in Sexuality Research: Used Practices, Best Practices, and Data Analysis Resources. Journal of Sex Research.
Calculates Gower distance matrix for hc.
- Calculates clusterings for hclust (many linkage method options), diana, and pam.
- Calculates connectivity, Dunn index, silhouette width for each of those clusterings, with cuts from 2-10.
- Plots a line graph for connectivity, Dunn, silhouette vs. number of clusters (for different clusterings)
Ideas and concepts of comparing clustering solutions came from: A consensus framework for clustering microarray data. Ted Laderas and Shannon McWeeney. OMICS: A Journal of Integrative Biology. 2007. 116-128.
At this point, pick a particular clustering method and number of clusters