Daniel Reda - Correlation


Published on

At a "Quantified Self and Science" meetup, Daniel Reda talks about the issue of correlation.

  • Notes from Eri Gentry watching Daniel's talk at QS:
    CureTogether. Started in mid-2008. A different kind of health website, one that gathered health data from people all over the web (anyone with any health condition): symptoms, treatments, efficacy. 4 million data points gathered. Results faced public.
    Ex: 93 depression treatments compared by 5,760 patients. As expected, less effective treatments were less popular, but surprisingly found that certain less popular treatments were very effective, such as music therapy and art therapy.
    Kept it simple. If you ask people 1. What do you feel 2. What did you try and 3. How did it work, you can get useful information. Noisy, yes, but not in a systematically wrong way. E.g, you’re probably better trying exercise than Paxil.
    Gary: concerned about being fooled by the data… but you would be fooled if the error were caused by something you didn’t’ expect. So – what about selection bias?
    Ex: Lot of migraine data came in, so they started examining it. A popular drug was Imitrex. Lots of people did well on it… but a lot of people reported worsening symptoms. Is there some way of distinguishing between these groups? Was there a difference in symptoms? Members who reported vertigo were 3x more likely to have a worsening reaction. Nothing was reported in the literature along these lines. Tested it. Correlation held. Confidence (p value) held.
    Hard to collect massive data on an individual basis (ie self tracking).
    Close: When you describe how to achieve p value, there are many elements not accessible to QSers (monte carlo simulations, large sets of data, etc). How to make things simpler… best way to test a theory is to gather more data. Look at it. Repeat experiment, look for same correlations. Stuff can still pop out with little data but we still trick ourselves. “It takes a special mind” a special sort of motivation
    Comment – we need to develop a common infrastructure
    Keep an eye on case crossover experimental design talk
    Need, as an individual, to look at your own data as well as the sea
    Comment: visualization affects our interpretation of data
    Median vs mean/exploratory data analysis not easily done as an individual

    TAKEAWAY: just gather data! You’ll find meaning or, if you don’t, you’ll know if you should run the experiment or a different one again
    Are you sure you want to  Yes  No
    Your message goes here
  • Be the first to like this