If I flip one coin 100 times the outcome is likely to be close to the perception you would expect.
If I flip 8 million coins simultaneously for 100 iterations it is possible that a number of the coins will display massive outlier behaviour (98 Heads / 2 tails etc).
Now I'm confused what is meant by spurious correlation. You're talking about an event that is known to happen, and happens infrequently, and if you have enough samples it's guaranteed to happen at least once with high probability.
Spurious correlation seems to be related to the use of ratios of independent variables making correlation "out of thin air." The word "spurious" also seems to be used in relation to hidden causation.
Neither of these two uses are related to this example, nor does the problem (in the sense in the previous paragraph) "increase" with more data.
OK, my example was not perfect, I was just trying to make it accessible to show why spurious correlations in data have a greater likelihood in larger datasets, especially super massive. The sequence was analogous to the correlation.
Given enough data it will appear but it cannot be trusted as any sort of indicator.
Yeah I read that as well as the wikipedia article on "Spurious correlation" proper (which appears to be a different concept) [1]. Neither one seems to be related to the amount of data being processed, nor to the description the author gives of unlikely but statistically robust events.
Without an explanation I can only conclude that the author misused the term. It's a shame when he's (she's?) writing to inform.
That is your choice of course but it is classic Dunning Kruger. You are not required to form an opinion about concepts you are not au fait with and to do so reveals more about you than it does about the author.
I would read the Tom Nichols paper above; it changed the way I approach certain topics.
In all honesty though; I am struggling to see what further explanation you require?
In larger datasets the likelihood of spurious correlation increases. This is because in large data sets, large deviations are vastly more attributable to variance (or noise) than to information (or signal).
Figure 18 shows the swelling number of potential spurious relationships. The idea is as follows. If I have a set of 200 random variables, completely unrelated to each other, then it would be near impossible not to find in it a high correlation of sorts, say 30 percent, but that is entirely spurious. There are techniques to control the cherry-picking (one of which is known as the Bonferoni adjustment), but even then they don’t catch the culprits—much as regulation doesn’t stop insiders from gaming the system.
Thanks, the linked article gives a much better explanation.
It's not about my personal understanding of something, but about the apparent contradiction with other sources I trust and the unnecessary vagueness about a subject that requires precision in terminology. "Statistically robust but unlikely" has many meanings, but "spurious correlation" has only one. I am training to be an expert in mathematics, so I do feel somewhat entitled to discuss the finer (though perhaps overly pedantic) points. It's literally my job to rigorously reason about these things. So I notice that the word "large" means something very specific here (many independent variables), while you appear to still mean "lots of data," which can happen with few variables just as well. "Big data" encompasses both (volume and variety), which is why it's confusing. So my need for a further explanation is because "in large data sets large deviations happen more often" does not appear to address my question if I'm specifically asking why volume would cause more spurious correlations (in fact, it does not). My questions are admittedly vague as well, but this should clear up my confusion.
I'm not saying the author doesn't know what spurious correlation is, just that the author used the term incorrectly, or too vaguely to be called correct, in that instance. There are many reasons to do this intentionally, I'm sure, but as a consequence I will ask questions with trivial but technical answers. I ask a hundred such questions every day, most of which I can answer myself, but I'll continue to ask them even when (if ever) I'm considered an expert in any topic.
8 comments
[ 4.6 ms ] story [ 13.8 ms ] threadCan someone explain why this is the case? I don't see how spurious correlations wouldn't appear in small data just as well as big data.
If I flip one coin 100 times the outcome is likely to be close to the perception you would expect.
If I flip 8 million coins simultaneously for 100 iterations it is possible that a number of the coins will display massive outlier behaviour (98 Heads / 2 tails etc).
Spurious correlation seems to be related to the use of ratios of independent variables making correlation "out of thin air." The word "spurious" also seems to be used in relation to hidden causation.
Neither of these two uses are related to this example, nor does the problem (in the sense in the previous paragraph) "increase" with more data.
Given enough data it will appear but it cannot be trusted as any sort of indicator.
Prob just best to read http://en.wikipedia.org/wiki/Spurious_relationship
Without an explanation I can only conclude that the author misused the term. It's a shame when he's (she's?) writing to inform.
[1]: http://en.wikipedia.org/wiki/Spurious_correlation
http://thefederalist.com/2014/01/17/the-death-of-expertise/
That is your choice of course but it is classic Dunning Kruger. You are not required to form an opinion about concepts you are not au fait with and to do so reveals more about you than it does about the author.
I would read the Tom Nichols paper above; it changed the way I approach certain topics.
In all honesty though; I am struggling to see what further explanation you require?
In larger datasets the likelihood of spurious correlation increases. This is because in large data sets, large deviations are vastly more attributable to variance (or noise) than to information (or signal).
http://valbonneconsulting.files.wordpress.com/2013/10/traged...
Figure 18 shows the swelling number of potential spurious relationships. The idea is as follows. If I have a set of 200 random variables, completely unrelated to each other, then it would be near impossible not to find in it a high correlation of sorts, say 30 percent, but that is entirely spurious. There are techniques to control the cherry-picking (one of which is known as the Bonferoni adjustment), but even then they don’t catch the culprits—much as regulation doesn’t stop insiders from gaming the system.
http://www.wired.com/2013/02/big-data-means-big-errors-peopl...
It's not about my personal understanding of something, but about the apparent contradiction with other sources I trust and the unnecessary vagueness about a subject that requires precision in terminology. "Statistically robust but unlikely" has many meanings, but "spurious correlation" has only one. I am training to be an expert in mathematics, so I do feel somewhat entitled to discuss the finer (though perhaps overly pedantic) points. It's literally my job to rigorously reason about these things. So I notice that the word "large" means something very specific here (many independent variables), while you appear to still mean "lots of data," which can happen with few variables just as well. "Big data" encompasses both (volume and variety), which is why it's confusing. So my need for a further explanation is because "in large data sets large deviations happen more often" does not appear to address my question if I'm specifically asking why volume would cause more spurious correlations (in fact, it does not). My questions are admittedly vague as well, but this should clear up my confusion.
I'm not saying the author doesn't know what spurious correlation is, just that the author used the term incorrectly, or too vaguely to be called correct, in that instance. There are many reasons to do this intentionally, I'm sure, but as a consequence I will ask questions with trivial but technical answers. I ask a hundred such questions every day, most of which I can answer myself, but I'll continue to ask them even when (if ever) I'm considered an expert in any topic.