# What are some good interview questions for statistical algorithm developer candidates?

I'm interviewing people for a position of algorithm developer/researcher in a statistics/machine learning/data mining context.

I'm looking for questions to ask to determine, specifically, a candidate's familiarity, understanding and fluidity with the underlying theory, e.g. basic properties of expectation and variance, some common distributions, etc.

My current go-to question is: "There is an unknown quantity $X$ which we would like to estimate. To this end we have estimators $Y_1, Y_2, \ldots, Y_n$ which, given $X$, are all unbiased and independent, and each has a known variance $\sigma_i^2$, different for each one. Find the optimal estimator $Y=f(Y_1,\ldots, Y_n)$ which is unbiased and has minimal variance."

I'd expect any serious candidate to handle it with ease (given some time to work out the calculations), and yet I'm surprised at how many candidates which are supposedly from relevant fields fail to make even the smallest bit of progress. I thus consider it a good, discriminative question. The only problem with this question is that it is only one.

What other questions can be used for this? Alternatively, where can I find a collection of such questions?

• For many machine learning people (including good ones), that question is way out of their comfort zone. This is an obvious statistician question. Commented May 3, 2016 at 14:44
• This question is legitimately borderline on / off topic. However, it has many views, several upvotes, an answer w/ several upvotes, &, moreover, is CW. It could stay open, IMO. Commented May 3, 2016 at 16:41
• The go-to question might be worded in a confusing manner. For example, the use of $X$ with a capital would make $X$ seem random. But since you're mentioning minimum variance, it would seem like you want $X$ to be non-random (in which case, why does the variance of the estimators not have a written dependence on $X$?) Commented May 3, 2016 at 19:04
• A point of caution, Google did a big study of their internal HR process and found that interviewer scores didn't correlate at all with subsequent job performance!! My impression of the literature here is that (1) puzzle type questions are the absolute worst, serving only to make the interviewer feel smart (i.e. 0 forecasting power) and (2) resume, experience based questions may have predictive value. Past performance forecasts future performance & you may want to focus questions to ascertain what their past performance was, but the interview is far less informative than interviewers think. Commented May 3, 2016 at 19:44
• Unbiasedness is guaranteed by having the weights sum to unity. However, even limiting your solution to linear combinations of the estimators, it's almost always going to be the case that multiple estimators based on the same data will be highly correlated. (If they are truly independent, then they would be applied to disjoint, independent subsets of the data.) It's not at all evident that a linear combination of estimators will be optimal, though.
– whuber
Commented May 5, 2016 at 21:14

What do you want your statistical developer to do?

The US Army says "train you will fight, because you will fight like you were trained". Test them on what you want them to do all day long. Really, you want them to "create value" or "make money" for the company.

Boss 101

Think "show me the money."

• Money grows on trees called employees. You put in a "dime" (their wages) and they pay you a "quarter" (their value).
• If you can't relate their job to how they make money for the company then neither you nor they are doing their job correctly.

Note: If your symbolic manipulation question doesn't cleanly connect to the "money" then you might be asking the wrong question.

There are 3 things every employee has to do to be an employee:

• Be actually able to do the job
• Work well with the team
• Be willing/motivated to actually do the job

If you don't get these down rock solid, no other answer is going to do you any good.

If you can replace them with a good piece of software or a well-trained teenager, then you will eventually have to do it, and it will cost you.

Data 101

What they should be able to do:

• use your internal flavors of software (network, os, office, presentation, and analysis)
• use some industry standard flavors of software (Excel, R, JMP, MatLab, pick_three)
• get the data themselves. They should know basic data sets for basic tasks. They should know repositories. They should know which famous data is used for which task. Fisher Iris. Pearson Crab. ... there are perhaps 20 elements that should go here. UCI, NIST, NOAA.
• They should know rules of handling data. binary data (T/F) has very different information content than categorical (A,B,C,D) or continuous. Proper handling of the data by data-type is important.
• A few Basic statistical tasks include: are these two the same or different (aka cluster/classify), how does this relate to that
(regression/fitting including linear models, glm, radial basis,
difference equations), is it true that "x" (hypothesis testing), how many samples do I need (acceptance sampling), how do I get the most
data from few/cheap/efficient experiments (statistical Design of
experiment) - disclaimer, I'm engineer not statistician You might ask them the question "what are the different fundamental tasks, and how do you test that the statistician can do them efficiently and correctly?
• access/use the data themselves. This is about formats and tools.
They should be able to read from csv, xlsx (excel), SQL, and
pictures. (HDF5, Rdata) If you have a custom format, they should
be able to read through it and work with the tools quickly and
efficiently. They should know strength/weakness of the format. CSV is quick use, been around forever, fast prototype, but bloated, inefficient and slow to run.
• process the data properly, using best practices, and not committing sins. Don't throw away data, ever. Don't fit binomial data with a continuous line. Don't defy physics.
• come up with results that are repeatable and reproducible. Some
folks say "there are lies, damn lies, and statistics" but not at my
company. The same good input gives the same good output. The output isn't a number, it is always a business decision that informs a
technical action and results in a business result. Different tests may set the dial at 5.5, or 6.5, but the capability is always above 1.33.
• present findings in the language and at the level that the decision
makers, and/or minion-developers, and/or themselves in a year, can
understand with the least errors. A beautiful thing is being able to explain it so your grandma gets it. This (link) is my answer, but I like it.

Analytic zingers:

I think impossible questions are great. They are impossible for a reason. Being able to know whether something is impossible out the gate is a good thing. Knowing why, having some ways of engaging it, or being able to ask a different question can be better.