Creating clusters for binary data I have a set of data with patients and their diseases. I would like to use hierarchical clustering or some kind of cluster analysis to make a dendrogram to see which diseases cluster together in this population. This is basically what it looks like, except with more diseases and more patients.
Moya    Hypothyroid Hyperthyroid    Celiac
   1       1           0             0
   1       1           0             0       
   0       0           1             1
   0       0           0             0
   1       1           0             0
   1       0           1             0
   1       1           0             0
   1       1           0             0
   0       0           1             1
   0       0           1             1

How would I go about making a dendrogram considering this is all binary data? Should I use Hierarchical clustering or UPGMA or something else?
 A: Many forms of clustering could work. Since you asked about constructing a dendrogram, it sounds like you want hierarchical clustering. Hierarchical agglomerative clustering is a popular class of methods. You'll have to choose the linkage function, which determines how clusters are merged. UPGMA (aka average linkage) is one example. A good discussion on this topic is available here.
For distance/dissimilarity-based clustering (including hierarchical clustering), you would need a distance measure that works for binary data. The Hamming distance is one example. The Hamming distance between two binary vectors is the number of elements that are not equal. In your example, the Hamming distance between two diseases would be the number of patients that are positive for one disease but not the other.
A: Latent class modeling would be one approach to finding underlying, "hidden" partitions or groupings of diseases. LC is a very flexible method with two broad approaches: replications based on repeated measures across subjects vs replications based on cross-classifying a set of categorical variables with no repeated measures. Your data would fit the second type.
All LC models have 2 stages: in stage 1, a dependent or target variable is identified and a regression model is built. In stage 2, the residual (a single "latent" vector) from the stage 1 model is analyzed and partitions are created capturing the variability (or heterogeneity) in that vector -- these are the "latent classes."
Freeware is out there for downloading that would probably work pretty well for you. One of these is an R module called polCA available here. Note that this approach is to be used only with binary data such as yours:
http://www.jstatsoft.org/article/view/v042i10
If you have about $1,000 to spend on a commercial product, Latent Gold is available from www.statisticalinnovations.com Having used on Latent Gold for years, I'm a big fan of that product for its analytic power and range of solutions. For instance, polCA is only useful for LC models with categorical information whereas LG works for true mixtures...plus, their developers are always adding new modules. The most recent addition builds LC models using hidden Markov chains. Bear in mind that LG is not an "end-to-end" data platform, i.e., it is not good for heavy data manipulation or lifting. 
Mplus is another commercially available product for this class of models with pricing similar to LG.
