22
$\begingroup$

Over the course of almost nine years, we have amassed a body of good answers to many clearly stated questions: one of our site's principal goals. To be useful, though, these answers have to be accessible--and that means searchable.

"Searchability" is the capability of offering links to likely answers based on the initial formulation of a question. Two search mechanisms are currently available to everyone:

  1. The site search mechanism at https://stats.stackexchange.com/search.

  2. External search engines such as Google, etc.

More often than not, new questions posted here have been asked and answered. Our initial review activities ought to focus on linking them to answers, leaving us to address the genuinely new questions that remain. However, more and more, newcomers plead ignorance: they are unaware of the search tools available or uncertain how to use them.

The purpose of this Meta thread is to collect your advice about searching our site so we can refer users here for quick help and advice about finding answers to their questions. What kinds of searches work? How can a user identify appropriate keywords or tags? What special options in the site search mechanism can be helpful? How can external search engines help? What tips and tricks can you share?

Please consider illustrating your replies with sample questions from our site, showing the search or sequence of searches you constructed to find the answers.

$\endgroup$
15
$\begingroup$

This reply illustrates a search strategy by means of an example. The strategy consists of an iterative procedure whose objectives are, in order,

  1. Discover appropriate key words and search terms to include.
  2. Find questions and answers related to these terms. Mine them for related terms (which might be new or unfamiliar: that's ok).
  3. Refine the search by including and excluding related terms, as appropriate.
    • Examine top hits by relevance and, separately, by voting.
    • When there are still too many hits, limit them by asking for posts with positive upvotes (using the keyword score:1).
  4. Repeat as necessary.

If you are the original proposer of the question, this activity--which usually takes less than a minute--often will answer your question. But if it does not, it likely will help you ask your question in a clearer and more focused way, which will help us answer the question you came to the site with.

If you are a reviewer, this activity often identifies interpretations you might not have considered upon a first reading, enabling you to guide the OP to clarify their question. Sometimes it will turn up existing answers and you can vote to close the current post as a duplicate.


Searching comments (optional)

To construct this example, I needed to find threads on our site where I, user "whuber," had recommended or mentioned searches. Because such recommendations would occur primarily in comments, and comments are not targeted by the site search engine, I started with Google. As you likely know, you can focus any Google search on a specified site using site: syntax, as in

site:stats.stackexchange.com whuber search

This input requested a search of all pages served at our site that include both "whuber" and "search."

Carrying out the site search strategy

The first hit in this Google search is a closed thread at Understanding MLR about "MLR." The sequence of interactions with the original poster (OP) went like this:

  1. Discover key words. I posted a comment asking for an explanation of what "MLR" was intended to mean, along with a suggestion about one possible interpretation.

    • It's often a good idea to ask posters for clarification rather than assuming you know what they intended (unless you are very good at long-distance mind reading).
    • However, sometimes unclear acronyms can be resolved by searching the web for the acronym along with the term "statistics" or "mathematics." In this case, the hits make it clear that MLR really does mean multiple regression. It looks like I didn't take the time to do this search before I commented, mea culpa.
    • The OP patiently confirmed this interpretation in a comment.
  2. Conduct a broad site search. I recommended searching on themultiple-regression tag along with keywords like "rank," which was suggested by terms used by the OP in their original question formulation. I couldn't do much more than this because grammatical distortions in the question statement made it unclear what the specific question might be. The search returned hundreds of results.

  3. Use initial results to refine the search. This involves several tactics:

    • Sort the results by relevance and then by votes: both have their advantages.
    • Skim through the first few hits to identify related keywords or other technical terms that are relevant to the question.
    • When there are many results, skim through the first page or two of hits to identify unrelated keywords that can be used to exclude irrelevant hits. A common example is to exclude "logistic" when the focus is on ordinary regression, by including -logistic in the search term.
    • In this case, one highly voted thread refers to "rank deficiency" and several highly-voted threads refer to "multicollinearity."

This already is progress: a tiny bit of searching will have revealed useful keywords that even a novice can use to find an answer, whether or not they are familiar with those terms.


Today these initial results, along with a review of the original question, suggest simply adding the keyword "unique" to the search and focusing on upvoted posts. The search expression becomes multiple-regression rank unique score:1.

  • Top hits (by relevance) indicate that "regularization" and "identifiability" might be useful keywords for subsequent searches. An identifiability post might contain an answer, or at least provide useful information to help the OP formulate and refine their question.

  • A top hit by votes turned up a nice mathematical analysis of least squares solutions which, if the OP is of a mathematical bent, might also answer their question.

$\endgroup$
  • $\begingroup$ I usually find sorting by votes to give more relevant results, unless the query is very broad, then it gives some irrelevant, very popular questions on top. $\endgroup$ – Tim Jan 23 at 16:27
  • $\begingroup$ @Tim The relevance algorithm is pretty good, so I pay attention to it. Sorting by votes will turn up good hits, too, but sometimes you have to look pretty far down the list to find them. You can (sort of) combine the approaches by including something like score:5 (with a relatively high vote threshold) in the search and focusing on the relevant hits. $\endgroup$ – whuber Jan 23 at 17:47
  • 4
    $\begingroup$ One thing you could mention is use of wildcard: valida* to search for validation or validated or ... valid* woud also include validity $\endgroup$ – kjetil b halvorsen Jan 24 at 10:09

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .