Experience tells me that small changes in the data can lead to substantial changes in the solution of a cluster analysis. This is especially true when the space is sparsely populated, as is the case with sequence analysis of lifecourses. Small changes in parameterisation (e.g., substitution costs) can lead to substantial differences in the cluster solution.
However, recently I came across an extreme case of sensitivity. Continue reading Cluster analysis is unstable, we knew that!
There are really three main ways of interacting with Stata:
Batch mode is critical to reproducible, self-documenting code. Even if you don’t use it, its existence is a reflection of the idea of a single do-file that takes you from data to results in a single movement. Most people use Stata through the GUI, most of the time, though, even when their goal is a pristine goal-oriented do-file. Stata’s GUI is clean, efficient and pleasant to use.
Console mode, on the other hand, is like the dark ages: a text mode interface (reminiscent of the bad old days before Windows 3.1, of DOS and mainframes). Why would anyone use it? Continue reading Tips for Stata console mode
One of the key anxieties about sequence analysis is how to set substitution costs. Many criticisms of optimal matching focus on the fact that we have no theory or method for assigning substitution costs (Larry Wu’s 2000 SMR paper is a case in point). Sometimes analysts opt for using transition-probability-derived costs to avoid the issue.
My stock line is that substitution costs should describe the relationships between states in the basic state space (e.g., employment status), and that the distance-measure algorithm simply maps an understanding of the basic state-space onto the state-space of the trajectories (e.g., work-life careers). Not everyone finds this very convincing, however.
I’ve been playing recently with a set of tools for evaluating different distance measures, and it struck me that I could use them to address this issue. Rather than compare hand-composed substitution matrices, however, I felt an automated approach was needed: randomise the matrices, and see what the results looked like. Can we improve on the analysts’ judgement? What are the characteristics of “good” substitution cost matrices?
Continue reading Substitution costs in sequence analysis — randomise!