3 ms·
You don't need thousands of cases, but you probably do need on the order of around a thousand samples. A hundred individuals with ten samples each over time is
by CuriouslyC 9y ago
You don't need thousands of cases, but you probably do need on the order of around a thousand samples. A hundred individuals with ten samples each over time isn't that challenging. I've done it successfully with mouse work at that scale.
I've found that in biological systems, individual variation can be as large as population variation, particularly for something "reactive" like the immune system, which is tied into stress response.
- mrow84 9y agoDon't you have to be concerned about differences between the structure of individual variability as opposed to population variability? Perhaps for the distinctions you are trying to draw it is not an issue?
- CuriouslyC 9y agoI haven't found that to be a big issue. I choose clustering and dimensionality reduction hyperparameters to penalize clusterings with few repeated individuals (high entropy), while trying to maximize the entropy of the distribution of cluster labels. To prevent cluster count explosion, you can put a regularization term/prior on the total number of clusters. An exponential or gamma with the expectation set to the square root of the number of distinct individuals is a good starting point. Of course a big part of any analysis is responding to the data, no single approach will work every time.
- mrow84 9y agoAll sounds very reasonable - thanks for the insight!