I am a graduate student in statistics and as such involved in a couple of collaborations with applied scientists (economists, foresters, …). These collaborations are fun (most of the time) and I do learn a lot, but there are also some complications, for example:

  • Sometimes my view of what a good statistical model is differs from the background of my collaborators and the common practices in their field. It is then difficult to convince them of trying out something new, either because they struggle to understand the model or because they are reluctant to change their habits
  • When proposing to use different statistical methods, I often have the impression that my collaborators consider this a criticism of their “standard” methods. However, it is by no means my intention to criticize anybody for their statistical knowledge or habits
  • And finally there is the other extreme: Some people expect too much. They think that I can miraculously extract interesting information from their data without their assistance. Of course, this is not true, especially if I miss the subject-specific background

I could probably think of more points but these are the first that came to my mind.

The questions I am asking you are:

  1. Do you experience the same or similar difficulties in your collaborations? How do you confront them? Generally, what do you do to be a good statistical collaborator?
  2. Are there any third-party resources on this topic, i.e., the soft skills needed in collaborations between statisticians and applied scientists?

Note: This question is more or less the converse of this one.


3 Answers 3


You are getting good advice, but as your experience widens, it will diversify.

Other possibilities include:

  1. Scientists should have considerable subject-matter expertise, for example on measurement and what kind of relationships make physical (biological, whatever) sense. Showing that you respect their expertise is a natural and congenial way to establish a good relationship.

  2. Scientists may know statistical stuff you don't. For example, most astronomers know more about irregular time series and non-detection problems than many statisticians do. Many fields use circular statistics, which even a full statistical education rarely includes.

  3. Graphs are often a lingua franca. Curiously or not, economists often distrust graphs as they are schooled to treat statistics in a highly formal fashion (your mileage may vary) and to avoid subjectivity (meaning, judgement).

  4. Sometimes you need to back off. If scientists don't know what they expect, but merely ask for the analysis or something that's publishable, they're wasting your time and you've better things to do. If the data are a haphazard mess, then they can't be rescued by any smart analysis.

Always establish an escape route. Your conditions could include (a) agreeing only to preliminary discussion (b) a limit on your time or other commitment (c) the right to back off if they won't follow your advice (d) some kind of idea on conditions for co-authorship. Beware the situation when a scientist just keeps coming back for a little more. Also, beware the situation in which you're treated like a person from the gas company or a plumber: you are called in to clear up a mess but they feel no obligation to maintain a relationship once that's done.

I am not a statistician but write from experience in so far as I know more statistics than most of my scientist colleagues. If each party respects the other, the relationship can be highly fruitful.

  • $\begingroup$ Great advice. I will echo #4...the moment I feel that I am being treated as the p-value dancing monkey (aka, the client just wants p-hacking)...I end the collaboration. The key is to doing so respectfully, and not burning any bridges (as they may recommend you to others, and those could be fruitful collaborations). To that end, the comments in the penultimate ¶ above would be key. $\endgroup$
    – Gregg H
    Apr 18, 2018 at 17:56
  • $\begingroup$ Common expectations are (a) there's one test [sic] or one method that is the solution (b) explaining what's wanted and what the answer will be will take just a few minutes. Counter-example: one scientist colleague asked the kurtosis of a uniform distribution and 1.8 emerged from my memory as the answer. Total conversation time: about 10 seconds (although I did look it up afterwards to check). $\endgroup$
    – Nick Cox
    Apr 18, 2018 at 18:01

Of course, your attitude is everything. If your clients/collaborators feel that you are there to support—as opposed to judge—that will go a long way. But, even then, there are issues that pop up. The two bullets you mention are key.

First, always stress that you want them to produce the very best science, and while you recognize that there may be discipline specific conventions, that doesn't mean there may not be better ways to accomplish the task. To that end, your two best friends would be: (1) the research question, and (2) any and all of the model assumptions. If the answer to the RQs can be obtained (even imperfectly) from the "conventional" approach, it probably will be reasonable. If the violations of the assumptions become too egregious...then you can reference back to wanting to produce the best science.

Hope my reflections are useful to you.


Hard skills are your foot in the door, and soft skills are the key to actually implementing a solution. Being the smartest person in the room doesn't earn you points.

That being said, you don't have to learn on your own. As cliche as it is, Dale Carnegie's How to Win Friends and Influence People actually can make you a better person. In the same vein, behavioral economics-type podcasts are good at surfacing research, making you think critically, and keeping it lively. See Freakonomics, for example.

Reading and listening are great, but you actually have to change how you act in order to affect good results.

Specific to your case, I've had success by trying all methods and comparing to an agreed-upon metric of "goodness". There's no need to argue if you can objectively test which model is best. This can be in minimizing error, having the best explanatory value, yielding the best "story", etc.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.