10 things statistics taught us about big data analysis

Tweet about this on TwitterShare on Facebook495Share on Google+165Share on LinkedIn708Email this to someone

In my previous post I pointed out a major problem with big data is that applied statistics have been left out. But many cool ideas in applied statistics are really relevant for big data analysis. So I thought I'd try to answer the second question in my previous post: "When thinking about the big data era, what are some statistical ideas we've already figured out?" Because the internet loves top 10 lists I came up with 10, but there are more if people find this interesting. Obviously mileage may vary with these recommendations, but I think they are generally not a bad idea.

  1. If the goal is prediction accuracy, average many prediction models together. In general, the prediction algorithms that most frequently win Kaggle competitions or the Netflix prize blend multiple models together. The idea is that by averaging (or majority voting) multiple good prediction algorithms you can reduce variability without giving up bias. One of the earliest descriptions of this idea was of a much simplified version based on bootstrapping samples and building multiple prediction functions - a process called bagging (short for bootstrap aggregating). Random forests, another incredibly successful prediction algorithm, is based on a similar idea with classification trees.
  2. When testing many hypotheses, correct for multiple testing This comic points out the problem with standard hypothesis testing when many tests are performed. Classic hypothesis tests are designed to call a set of data significant 5% of the time, even when the null is true (e.g. nothing is going on). One really common choice for correcting for multiple testing is to use the false discovery rate to control the rate at which things you call significant are false discoveries. People like this measure because you can think of it as the rate of noise among the signals you have discovered. Benjamini and Hochber gave the first definition of the false discovery rate and provided a procedure to control the FDR. There is also a really readable introduction to FDR by Storey and Tibshirani.
  3. When you have data measured over space, distance, or time, you should smooth This is one of the oldest ideas in statistics (regression is a form of smoothing and Galton popularized that a while ago). I personally like locally weighted scatterplot smoothing a lot.  This paperis a good one by Cleveland about loess. Here it is in a gif. loessBut people also like smoothing splines, Hidden Markov Models, moving averages and many other smoothing choices.
  4. Before you analyze your data with computers, be sure to plot it A common mistake made by amateur analysts is to immediately jump to fitting models to big data sets with the fanciest computational tool. But you can miss pretty obvious things like this if you don't plot your data. baThere are too many plots to talk about individually, but one example of an incredibly important plot is the Bland-Altman plot, (called an MA-plot in genomics) when comparing measurements from multiple technologies. R provides tons of graphics for a reason and ggplot2 makes them pretty.
  5. Interactive analysis is the best way to really figure out what is going on in a data set This is related to the previous point; if you want to understand a data set you have to be able to play around with it and explore it. You need to make tables, make plots, identify quirks, outliers, missing data patterns and problems with the data. To do this you need to interact with the data quickly. One way to do this is to analyze the whole data set at once using tools like Hive, Hadoop, or Pig. But an often easier, better, and more cost effective approach is to use random sampling . As Robert Gentleman put it "make big data as small as possible as quick as possible".
  6. Know what your real sample size is.  It can be easy to be tricked by the size of a data set. Imagine you have an image of a simple black circle on a white background stored as pixels. As the resolution increases the size of the data increases, but the amount of information may not (hence vector graphics). Similarly in genomics, the number of reads you measure (which is a main determinant of data size) is not the sample size, it is the number of individuals. In social networks, the number of people in the network may not be the sample size. If the network is very dense, the sample size might be much less. In general the bigger the sample size the better and sample size and data size aren't always tightly correlated.
  7. Unless you ran a randomized trial, potential confounders should keep you up at night Confounding is maybe the most fundamental idea in statistical analysis. It is behind the spurious correlations like these and the reason why nutrition studies are so hard. It is very hard to hold people to a randomized diet and people who eat healthy diets might be different than people who don't in other important ways. In big data sets confounders might be technical variables about how the data were measured or they could be differences over time in Google search terms. Any time you discover a cool new result, your first thought should be, "what are the potential confounders?"correlation
  8. Define a metric for success up front Maybe the simplest idea, but one that is critical in statistics and decision theory. Sometimes your goal is to discover new relationships and that is great if you define that up front. One thing that applied statistics has taught us is that changing the criteria you are going for after the fact is really dangerous. So when you find a correlation, don't assume you can predict a new result or that you have discovered which way a causal arrow goes.
  9. Make your code and data available and have smart people check it As several people pointed out about my last post, the Reinhart and Rogoff problem did not involve big data. But even in this small data example, there was a bug in the code used to analyze them. With big data and complex models this is even more important. Mozilla Science is doing interesting work on code review for data analysis in science. But in general if you just get a friend to look over your code it will catch a huge fraction of the problems you might have.
  10. Problem first not solution backward One temptation in applied statistics is to take a tool you know well (regression) and use it to hit all the nails (epidemiology problems). hitnailsThere is a similar temptation in big data to get fixated on a tool (hadoop, pig, hive, nosql databases, distributed computing, gpgpu, etc.) and ignore the problem of can we infer x relates to y or that x predicts y.
  • Brian

    Great post! I am curious about #3 - are you thinking of this in terms of preparing predictor variables for modeling or generally in terms of visualizing or expressing a relationship descriptively?

  • Seref Arikan

    Brilliant stuff. I have been looking for this type of material for quite some time now, but statistics books rarely focus on these type of gems. As a PhD student who had to do lots of self learning, I wish had a lot more of this type of wisdom around, especially in the form of books. Any suggestions you may have would be great 😉

  • Trust me, I’m a statistician

    In #2, Pro. Hochberg's name is misspelled

  • Chandrika B-Rao

    Excellent summarization. However, half knowledge is dangerous and suggestion 1 can be misused. It is taken as license to combine results of multiple heterogeneous models without any understanding of the bases of the models and the right way to combine them. Exactly how the averaging over multiple models is done is crucial to the reliability of the combined result. I have seen misuse of suggestion 3 as well - 50% inhibitory constants that are lower than the lowest tested concentration or higher than the highest tested concentration of a compound in a bioassay. This comes from blind use of software prediction involving smoothing and extrapolation that the user is often unaware of. Combined with suggestion 4, this problem could be overcome.

  • SandraPickering

    Thank you - some of my 'favourites' and a few new ones.
    I appreciate having all of them gathered in one place.

  • http://choicewords.tumblr.com/ A Corresponder

    Please credit the art! The bottom one is clearly Allie Bosch's from hyperbole and a half.

    • pls

      second this. Good art is good.
      (Btw. surprisingly wasn't sure about the association, bcs unlikely. Thx)

  • http://www.matrixdatalabs.com/ Matrix DataLabs

    Thanks for this! The basic ideas in statistics relating data to information are always at the heart of predictive analytics, knowledge discovery, data mining, etc whether with big data or not. Just one comment: in aggregation models it tends to work better to consider weighted combinations (this is actuallly what bagging does, searching iteratively a solution over the set of convex combinations of very simple learners using a gradient algorithm)

  • Nick Jachowski

    Great article Jeff! I like that you differentiate between predictive and descriptive modeling. Too often practitioners conflate the two which leads to sub-optimal results! If you're trying to make a prediction - great. If you're trying to describe what you think is going on too, that's also great. But rarely are the two one and the same...

  • http://www.twentylys.com/ TwentyLYS

    Great stuffs. Although not sure if I get number 1 quite correctly, average out prediction models results to get better accuracy. How about the assumptions each prediction model makes? I think if we use bootstrapping then it re-samples of sample and repeats the process large number of times. i.e runs the same model but large number of times to output the shape of the distribution. As far as I know Random Forest applies the same technique of bootstrap aggregating algorithm and uses some sort of modified decision trees method to reduce correlation among spitted trees.

  • Gianluca Rossi

    Great article! Unfortunately the link to the article by Cleveland about loess is broken