Why not just use age as a continuous numerical variable? Why do you need to "bin"?

What is your approach to do that? Remember, I don't want to partial out age in a regression-type of setting. This question is concerned with the creation of sub sets of a data set that match the original data set's distribution of age.

Take every n-th quantile. This assures that the age distribution will be similar. If you have a large dat set (and sub set), random sampling will also do the job.

It is often hard for a person to give an appropriate answer when the question does not include details of the idea being tested. Often the questions seem to assume that problems do not have contexts, as though solutions were general to all similar questions. But, indeed, as I tell my students in statistics, a question without a context has an infinite number of bad answers, and probably only one or two good ones. But, a question in context can usually get at those few good answers very easily. And, by context, I mean hypotheses to be tested, data structure and so on.

James, I fully agree. 99% of questions on RG should really only be answered with your comment. My question was meant to be unspecific. I don't have a particular problem but noticed that people do bins of age in various types of analyses and I have been wondering a lot on how people choose the bins. Really, it seems that they often just guess a number of bins and stick to that (when in many types including age as a continuous term in the model would have been easy). This is why I asked for "your experience" for your particular problems. This question is to be seen rather as a survey than as a discussion on a particular problem.

Jochen, this is what I often heard. The question is now, with what rule do you determine n?

You mean how large your training set should be compared to your test set? I would go for a 1:1 ratio, so n=2.

Jochen, you suggested to "Take every n-th quantile.". I am referring to how many quantiles or bins one should select. Shouldn't we decide on this number on some criterion that is formalizing the mismatch between the original distribution and the partitioned distributions?

I have seen a lot of studies in which people measure things, such as age, length, etc. after which, they divide those more precise measurements into bins. I always thought that was a carry over from pre-computer days, when it simplified analysis. Today, with computers, I just can't figure out why, in the vast majority of cases, people want to REDUCE their information by binning, when analyses are available to use the original value of the variable. Hence my answer - without a context, how can one decide how to answer? In fact, binning always seemed to be arbitrary in some way. After all, why one instead of two years, or why a month instead of a week? I have never seen binning in which the logic behind the size of the bins was justified with scientific or logical reasoning.

Andreas, is it the question how many subsets you should create? I thought you want to devide your data into two sets: one training set and one test set. If both these two sets should be similar in the distribution of "age", then one simple way to create these two groups is to sort all your data and take the 1st value into set A, the 2nd into set B, the 3rd into A, B, A, ... this is what I suggested before.

I am not sure if I got you.

If I understand your question, it depends on how you want to allocate to training and test samples.

Here's some examples. I'm not going to do a lot of training, so that's easy. I want pretty accurate estimation of my prediction accuracy from the test set. OK - 50% training, 50% test. Age is the only variable I consider important - either because it is the only predictor, or I think my (complicated) predictors are pretty independent and it is the only confounder I can think of. For a 50/50 split, I order them by age, select the age ordered observations 2 at a time, randomly assign one of the pair to training, the other to test. There may be 1 left over, and ... whatever.

If the training is more complicated, then I may use 5-fold cross validation. Then I will select contiguous groups of 5 from my age-ordered data, randomly assign to one of 5 groups, and repeat the training and prediction with each of the 20% left out.

Hope that answers your question.

I agree with James. Binning, from statistical point of view, only causes loss of information. You want to avoid it at all cost. There are some exceptions where people do not want to give out their accurate age. Binning could be used to reduce respondents' reluctance. But, in this case, binning is better than the worst, still it is not the best. It seems that many users misunderstood the purpose of binning and use it without much discretion. Often students use 'standardized' demographic questionnaires for their research. The results? Myriads of wrong information about the influence of age. This issue is related to the practice of "dichotomization" If you search the term in the database, you can easily find some articles.