Can we do better than R-squared?(tomhopper.me)
tomhopper.me
Can we do better than R-squared?
http://tomhopper.me/2014/05/16/can-we-do-better-than-r-squared/
3 comments
The biggest problem: People hide between such quantities as R-squared and p-Value. While they are just indicators on how the model fits the data, people turn them into a monolithical, undoubtable truth. That is why studies are being published with 7 participants,
I would even go further. The biggest problem isn't just using p values and R^2. I think the biggest problem is that a lot of people didn't learn statistics properly.
Properly is a vague term. So what do I mean? Instead of obsessing with tons of techniques going back to the basic and actually learn how to do design studies, work with data, learn statistical reasoning and critical thinking.
I took quite a few courses in statistics because I liked it. But a lot other people – especially those that apply statistics – maybe take one or two courses in stats and then do research / studies. The results can be pretty terrible. In conclusion, more fundamentals and less icing on the cake.
Properly is a vague term. So what do I mean? Instead of obsessing with tons of techniques going back to the basic and actually learn how to do design studies, work with data, learn statistical reasoning and critical thinking.
I took quite a few courses in statistics because I liked it. But a lot other people – especially those that apply statistics – maybe take one or two courses in stats and then do research / studies. The results can be pretty terrible. In conclusion, more fundamentals and less icing on the cake.
Statistics are a tool that is quite easily abused. Look at the damage a regression can do in the hands of someone who only had a couple weeks to learn it at the tail end of a stats or econometrics class.
Stats are a field where it does pay to be a degree snob - it's best to have folks who studied it extensively, and then actually used it in the real world.
Stats are a field where it does pay to be a degree snob - it's best to have folks who studied it extensively, and then actually used it in the real world.
While I agree with the premise that most people do not know enough statistics, there are several underlying problems to that.
When I was first encountered to statistical topics as an undergrad, the lecturer himself was not confident of the matter and taught a very bare subset that consisted of linear fitting, error propagation, means and standard deviation. Even back then I had the feeling that the equations provided were insufficient and contained a lot of things that were not motivated or explained.
Nowadays I can see why I was not introduced to statistics and probabilities back then as I was introduced to algebra, analysis and quantum mechanics. The field of statistics is complex, full of contradicting best-practices and analytically challenging. Let alone Bayes-vs.-Frequentist, etc. In a way I doubt that all researchers working with Poisson-distributed data once in their life as scientists can work through the details Poissonian statistics analytically.
Maybe an illustrative introduction would be more beneficial. I imagine people could perform statistical experiments before working with real data and experience first hand how misleading a small data subset can be for example, or how fundamentally data plots can change their face. Maybe then people would stop being overenthusiastic about their N=20 experiments.
When I was first encountered to statistical topics as an undergrad, the lecturer himself was not confident of the matter and taught a very bare subset that consisted of linear fitting, error propagation, means and standard deviation. Even back then I had the feeling that the equations provided were insufficient and contained a lot of things that were not motivated or explained.
Nowadays I can see why I was not introduced to statistics and probabilities back then as I was introduced to algebra, analysis and quantum mechanics. The field of statistics is complex, full of contradicting best-practices and analytically challenging. Let alone Bayes-vs.-Frequentist, etc. In a way I doubt that all researchers working with Poisson-distributed data once in their life as scientists can work through the details Poissonian statistics analytically.
Maybe an illustrative introduction would be more beneficial. I imagine people could perform statistical experiments before working with real data and experience first hand how misleading a small data subset can be for example, or how fundamentally data plots can change their face. Maybe then people would stop being overenthusiastic about their N=20 experiments.
AIC? BIC? The article mentions adjusted R^2, which would still be an improvement over plain R^2 if you have a number of predictors in your model.
And the mistake is here :
> Minitab calculates predicted R-squared by systematically removing each observation from the data set, estimating the regression equation, and determining how well the model predicts the removed observation. Like adjusted R-squared, predicted R-squared can be negative and it is always lower than R-squared.
(linked from the article, and better explained than anywhere in the article).
The mistake is in what is compared against what. Ideally a model would be independent from the data entirely. This is what you get if you correctly derive a model from first principles, and then verify that observed data points agree. The independence is perfect and it is valid to compare the two data sets. Any agreement builds confidence that the model is correct.
However that's not what's being done here. The "model" is a linear combination of the data set. So you're calculating the model generated from n data points with the model generated from n-1 data points, for each point. So you're effectively comparing the data against itself (which is also what R-squared is doing of course).
The correct way to go about this is to randomly divide your dataset in 2 pieces, ideally both should have a decent size. The second part you give to your colleague, with strict instructions to shoot you if you even ask to see that data. You build your model with your data points, and then ask your colleague to compare it with his data, and report only if they're close or not, NOT the values. If your fit is not close to his fit ("too much" - some fuziness here), you're overfitting.
The key difference here is the random part. The "adjusted predictive R squared" does not create independence between the compared data sets. This means it can't work correctly.
I realize this is the typical mathematician argument. I'm claiming here that there exists a way to make R squared and predictive R squared both be way too high, without giving any method to do so. But this is obviously wrong nonetheless.
> Minitab calculates predicted R-squared by systematically removing each observation from the data set, estimating the regression equation, and determining how well the model predicts the removed observation. Like adjusted R-squared, predicted R-squared can be negative and it is always lower than R-squared.
(linked from the article, and better explained than anywhere in the article).
The mistake is in what is compared against what. Ideally a model would be independent from the data entirely. This is what you get if you correctly derive a model from first principles, and then verify that observed data points agree. The independence is perfect and it is valid to compare the two data sets. Any agreement builds confidence that the model is correct.
However that's not what's being done here. The "model" is a linear combination of the data set. So you're calculating the model generated from n data points with the model generated from n-1 data points, for each point. So you're effectively comparing the data against itself (which is also what R-squared is doing of course).
The correct way to go about this is to randomly divide your dataset in 2 pieces, ideally both should have a decent size. The second part you give to your colleague, with strict instructions to shoot you if you even ask to see that data. You build your model with your data points, and then ask your colleague to compare it with his data, and report only if they're close or not, NOT the values. If your fit is not close to his fit ("too much" - some fuziness here), you're overfitting.
The key difference here is the random part. The "adjusted predictive R squared" does not create independence between the compared data sets. This means it can't work correctly.
I realize this is the typical mathematician argument. I'm claiming here that there exists a way to make R squared and predictive R squared both be way too high, without giving any method to do so. But this is obviously wrong nonetheless.
> I realize this is the typical mathematician argument.
Interestingly, a mathematician's argument on this exact method was actually given in 1977, in "An Asymptotic Equivalence of Choice of Model by Cross-Validation and Akaike's Criterion". [1]
What Stone showed was that (modulo technicalities) performing this procedure - leave one out cross validation, the procedure you claim "cannot work correctly" - was asymptotically equivalent to minimising the AIC [2] of the model. So we have an asymptotic guarantee (for whatever that's worth) on the behaviour of this procedure.
[1]: http://www.jstor.org/stable/2984877
[2]: http://en.wikipedia.org/wiki/Akaike_information_criterion
Interestingly, a mathematician's argument on this exact method was actually given in 1977, in "An Asymptotic Equivalence of Choice of Model by Cross-Validation and Akaike's Criterion". [1]
What Stone showed was that (modulo technicalities) performing this procedure - leave one out cross validation, the procedure you claim "cannot work correctly" - was asymptotically equivalent to minimising the AIC [2] of the model. So we have an asymptotic guarantee (for whatever that's worth) on the behaviour of this procedure.
[1]: http://www.jstor.org/stable/2984877
[2]: http://en.wikipedia.org/wiki/Akaike_information_criterion
Do you really need a colleague here or just self-honesty once you have split the dataset?