Statistical/methodological prep for a career in neuroscience research?

Shea Levy writes:

I’m currently a software developer, but I’m trying to transition to the neuroscience research world. Do you have any general advice or recommended resources to prepare me to perform sound and useful experimental design and analyses? I have a very basic stats background from undergrad plus eclectic bits and pieces I’ve picked up since, and have a fairly strong mathematical background.

I’m not sure! I think my book with Jennifer is a good place to start on statistical analysis, and for design I recommend the classic book by Box, Hunter, and Hunter.

I also recommend this paper by Microsoft engineers Ron Kohavi, Alex Deng, Brian Frasca, Roger Longbotham, Toby Walker, Ya Xu, a group of software engineers (I presume) at Microsoft. These authors don’t seem so connected to the statistics literature—in a follow-up paper they rediscover a whole bunch of already well-known stuff and present it as new—but this particular article is crisp and applied, and I like it.

Maybe readers have other suggestions?

12 thoughts on “Statistical/methodological prep for a career in neuroscience research?

  1. Stats is a crucial but limited part of the skillet of a neurostats person. You need to know the experiments. From the questions. Be willing to deal with a world of rather fuzzy ideas.

    So I would recommend reading dayan and Abbot, arbibs fat handbook, spikes, etc.

    And yes, Andrew’s book but also Bishop ( for ml), Mackay (Bayes), and some intro to ai book. And definitely the kass book.

    And then one of the fat neuroscience tomes for the experimental background.

    I would also do a summer school at cshl, MBL, or if you like movement our Cosmo school.
    And the sand meeting.

    Also, meet neuroscientists when you can.

    It is great fun. Say hi if you ever come to Chicago.

  2. The SMART Group at Hopkins, PENNsive at Penn, and Rebecca Betensky’s group at Harvard are some great hubs of varied neurostats research. Searching out their papers relevant to your interests may be useful.

    http://www.smart-stats.org
    http://www.med.upenn.edu/pennsive/
    http://www.hsph.harvard.edu/biostats/research/divisions/neurostat/seminars_workinggroups.html

    It’s not showing up on the CRC website yet (https://www.crcpress.com/Chapman–HallCRC-Handbooks-of-Modern-Statistical-Methods/book-series/CHHANMODSTA), but there is also a handbook on statistical methods for neuroimaging coming out.

  3. I’m no statistician, but I am a Neuroscientist. Most neuroscience research is still done using t-tests and plunger plots ±1SEM so the bar to entry is spectacularly low[1]. If you want to enter from the deep end, then read around Paninski, he has a nice reading list here: http://www.stat.columbia.edu/~liam/teaching/neurostat-fall13/

    And general ideas of some of the software from Emory Brown: http://www.neurostat.mit.edu/software

    And for a statistical model of how brains may work, Karl Friston has been developing a bayesian predictive framework, a good introduction paper is from Bastos et al., http://www.ncbi.nlm.nih.gov/pubmed/23177956 & general background: http://www.ncbi.nlm.nih.gov/pubmed/23663408 — there are many more bayesian models applied to various sensory and motor systems across neuroscience…

    [1] http://www.nature.com/neuro/journal/v14/n9/full/nn.2886.html

  4. Thanks Andrew and commenters! Looks like this will be a really good start.

    Konrad: Yeah, been working my way through Kandel as my “fat neuroscience tome” of choice. What do you think the summer school brings specifically in this context?

    Vanessa: Already have a few lab job inquiries in flight, I definitely learn best by completely engaging with a topic on as many fronts as possible.

  5. I’ve gone the other way, a neuroscience PhD who is now a data person full-time. The most important things you can learn are the techniques specific to the research methodologies you are interested in. For example, there are specific timing sequences used to design fMRI experiments, to detect the blood-flow signal at the correct times, or to analyze event-related potentials or neural spikes. Programming in Matlab or R will also give you a leg up – that’s a skill that a lot of students lack going in to the lab setting, and labs are anxious to find people with those skills. If you’re thinking about computational neuroscience, then data modeling and deep learning are critical. It’s true that a lot of neuroscience experiments are analyzed in basic ways, as one of the commentators above suggested, but I think that is becoming less true over time, and as I’ve implied, sometimes the creation of the experiment requires more statistical sophistication than the analysis of the results! Good luck to you. Neuroscience is fun.

  6. “Do you have any general advice or recommended resources to prepare me to perform sound and useful experimental design and analyses? I have a very basic stats background”

    Don’t even worry about stats (except, obviously, parameter estimates), and pay no attention to it in the papers. The vast majority of the use cases will be “one group is different than the other”, etc. Experimental design is all about distinguishing between different possibilities (which is rather unique for every experiment), while the usual use of stats only distinguishes between two possibilities (effect vs no effect). Any design that can let us distinguish between two real explanations will also rule out a “nil-null” hypothesis, so the focus on that is extremely misguided.

    When you read something in the textbook follow the sources. Read the limitations mentioned by the original authors. See if you can find any skeptical publications on the topic. Have these limitations or alternative explanations for the results ever been addressed? Think up your own alternative explanations. Finally, do not be intimidated by thousands of publications on a topic.

  7. I found the following paper to have several useful concepts that speak well to the large group of neuroscientists that I work with.
    Concepts such as measurement error, reliability, experimental design, etc.

    ‘Elementary statistical methods and measurement error’ Vardeman et al. (2010). The American Statistician. Vol 54 (1), p 46-51.

  8. +1 on all of Konrad Kording’s suggestions. Rob Kass’s book (as linked in the comment above) and his lecture notes (http://www.stat.cmu.edu/~kass/) are good. While good fundamentals are essential in all areas of neuroscience, the specifics will vary (as usual). Neural coding people are often focused on point process models, generalized linear models, etc., whereas fMRI people may focus on something entirely different.

    I would start with some neural data you might be interested in and just play with it in e.g. an IPython Notebook or some other friendly workflow. Extract some statistics of interest, plot some things, fit a model, and become familiar with what is in the data.

Leave a Reply

Your email address will not be published. Required fields are marked *