There are often questions about nonlinear data here, but I find the tags available are confusing. To my best knowledge there are three tags for this: , , and . The latter two, which seem like good catch-all tags, are both deprecated. So the only option is nonlinear-regression, but there are plenty of questions here that start off having less to do with nonlinear regression (e.g. this post here about correlations for nonlinear data). What is the best approach here in these cases? Should the older tags be revised? Or is my interpretation wrong?

• Some comments: (1) Data aren't linear or nonlinear. Models of relationships between variables are. (2) 'Nonlinear' has multiple meanings in statistics realm; here are some: (i) the relationship between $y$ and $x$ is not well described by a straight line (a la $y = a + bx$), (ii) the model equation is not linear—not a sum of scaled values—in the parameters (a la $y = \beta_1^{x} + x^{\beta_2} + \varepsilon$ – fit that with OLS!), (iii) the variable(s) on the left-hand side of the model equation(s) are also on the right-hand side(s) (a la $y_t = \beta + \rho y_{t-1} + \varepsilon$). Commented May 6 at 18:16
• I have never heard the distinction for Point 1 before. Why is it wrong to say the data does not have a linear distribution? Commented May 6 at 23:41
• Shaw Hemelstrand, because the thing people are actually pointing at with a linear distribution is a model. Commented May 7 at 4:39

Not every aspect or element is a keyword. I don't think that it is sufficient that there are plenty of questions about non-linearity, it is also important that it is a well know and clearly defined topic.

It should be such that people are using it to search for questions. E.g. imagine a researcher specialised in a specific field of statistics relating to that keyword or the creator of a software package doing the thing that the keyword describes, being able to use that keyword.

Non-linear regression is an important topic (and reasonably defined). The are many books, articles, encyclopedia entrees, and software packages that have the keyword as title.

If there is some topic ABC, but nonlinear like "non-linear ABC", then ask the question: Is it helpful if it is tagged with a more specific "non-linear ABC" instead of just "ABC"? Are there people looking for the specific non-linear variation?

Relating to the question about "non-linear correlation" I don't think that the answer is positive. And actually, it is more a question about regression than correlation, but it is using the correlation as a name for 'effect size' or 'coefficient of determination'.

• I've noted my comments in the answer there. With respect to the questions you ask about the utility of the specificity here, I think both "nonlinear" and "nonlinear regression" are both helpful, in that nonlinearity seems to me a clear idea with a broad number of topics that are coverable (including nonlinear transformations, linearity of parameters, etc.), whereas nonlinear regression is usually more specific to applications of regression for fitting said data. I feel at minimum there should be some other category like "nonlinear data." Commented Apr 28 at 11:30
• @ShawnHemelstrand I think that a category 'nonlinear' would be too broad and vague. The power of tags and keywords is when you do not have too many of them. Otherwise it would be like IMDB having 'snow' as category and starts tagging all movies with at least some snowflakes in it with that tag. If people want to look for movies with snow in it then they will have to search for a 'snow' in the descriptions.. Commented Apr 28 at 13:20
• .... A more specific 'nonlinear' category like 'nonlinear data' might work, but I don't know what it means and if this is some actual topic or theme in statistics. Commented Apr 28 at 13:20
• A keyword should be such that a person reading a specific question and answer tagged with that keyword, if they think they want to see 'more like this' then searching on the keyword should be helpful. I don't think that a broad category like 'nonlinear' would do that. Commented Apr 28 at 13:24
• I'm thinking you are correct. My best guess is that "nonlinear data" would be more specific and have some utility, whereas "nonlinear" may be too vague. Commented Apr 28 at 13:28