Tuesday, 21 April 2015

By all means experiment, but please do so intelligently


Experimentation and Design of Experiments

  

Experimentation is a ‘new’ buzzword, some even claiming “Big Data” demands experimentation.
I agree, and disagree. 

The need for experimentation has always been here – why would you launch a full scale campaign costing hundreds of thousands when you can spend some time refining and enhancing your offer and targeting?

It’s just that, with “Big Data”, data is available at a more granular level and hence we are able to detect more subtle events and sequences of events that can be seen as precursors to various courses of action. Today, organisations can quite easily know where their customers are, and allied with their recent behaviour, can create tailor-made offers. If I just had lunch at a fried chicken restaurant, and my bank has an offer at a yogurt bar nearby, it might be the exact thing to balance my tummy. Hence the bank could send me an offer for the yogurt bar.

Technology has helped in 2 further ways. Firstly, not only is “Big Data” available for analysis at a very granular level, but we are able to do so without having a Computer Science Degree. Hence the number of potential experiments that can be attempted has increased. Secondly, technology has also increased the ability to execute these experiments; apps, text messages are all avenues where we can be reached.

But what, to me, is critical is the design of the experiments. Just like with “Big Data” analytics, while technology has put the ability to analyse data in the hands of many, there are fundamentals that put probability on your side. Proper experimental design is critical and the costs of not doing so can be catastrophic.

The articles linked below highlight some failures in the medical field that have huge implications and that could be avoided simply with proper experimental design. The sad part is that the failures are so basic that a little thought would have avoided them:
                Ensure large enough a sample to conduct experiments (power tests)
                Do not bias the experiment by selecting groups to affect success probabilities 
(random selection)
                Do not bias the experiment by your own behaviour (double blind tests)

By all means experiment, but please do so intelligently.


No comments:

Post a Comment