Outlier removal in Glycaemic Index testing

#1
Hi all,

Hoping you can clarify for me on the appropriateness of method of outlier removal in GI testing.

It is part of the ISO standard that outliers outside 2 standard deviations can be removed to calculate GI.

However, some research centres stop after one round of outlier removals to get a final GI value while others recalculate the mean and standard deviation and do a 2nd (and 3rd and 4th if necessary) round of outlier removal until there are no more outliers.

I guess the question is, which is more appropriate?

Thanks for any input.
 

rogojel

TS Contributor
#2
At the risk of contradicting an ISO standard, I would say that even BLINDLY removing outliers in the first round is questionable. Generally one should only remove outliers where their cause is known - like a faulty measurement.

Repeating this procedure seems to me to be completely wrong - I would guess that the distribution would probably be quite skewed, which could be perfectly natural, and the method is doing nothing else except arbitrarily chopping off the tail of the distribution until the skewness disappears.


regards
rogojel
 
#3
I agree with what Rogojel said.

It is part of the ISO standard that outliers outside 2 standard deviations can be removed to calculate GI.
If that is a part of an ISO standard, then I wonder which ISO number? (There are many ISO standards like ISO 9000 and ISO 14000.) Do you have a link to that standard?
 
#4
I agree with both yours and rogojels opinions on the (lack of) scientific merit in blind removal of outliers but we need to work within the framework of the ISO to get products certified as low GI by these research centres. The ISO number is 26642:2010, but I don't know where you can get a free, unrestricted copy.

Thanks for your input.
 
#5
I assume that the products that are relevant for GI testing are products with quite low GI values, thus values that are close to zero. Therefore I assume that they have a (positive) skewed distribution. So that only the large values will be classified as "outliers", by being larger than "mean+2*standard deviations".

Suppose the GI is exponentially distributed with a parameter lambda=1 so that the mean and standard deviation both are 1. From the distribution function 1- exp(-(1+1*2) = 0.9502. So the largest 5% of the observations will be thrown away. After another calculation of mean and standard deviation more data will be thrown away. (No low values will be taken away, since there are no values below zero.)

The procedure is not only inefficient. It also biases the result. But I guess that getting an even lower GI value is only welcome from the managements point of view. Usually statistics tries to be a guide to get to get a correct opinion about reality. But this procedure is deceiving.

I am not blaming Tempellis, but maybe the company management should be reminded that there are not only ISO standards to obey. There are also laws that forbid cheating the customer. And if you say that the product has a lower GI than it really has, then it is cheating!
 

Lazar

Phineas Packard
#6
I think it can be reasonable to remove outliers as part of sensitivity analysis but as sensitivity/supplementary analysis.
 

rogojel

TS Contributor
#7
Just to add to Greta's post the procedure to repeatedly remove the outliers is not part of the ISO , and even the first step is not mandatory only allowed. So, there is no need to do this in a questionable way, it is only a loophole in the standard.

regards
rogojel
 
#8
I think of the idea of classifying values outside of "mean +/- 2 standard deviations" as a symptom of the disease of believing that everything is normally distributed.

But I believe that many things, like pollutants and substances that can be a health hazard, have a skewed distribution with values close to zero. Sometimes large values will occur and they are as real as any other values. And it is necessary for the food consumer (in this case) to be informed about that.

I think of an outlier as a value that "does belong there there", like because of contamination or crude mistakes and so on. High values are not necessarily outliers.

(But maybe I used to strong words above, when I was a little bit upset by the potential bad statistics by the ISO standard. :) )

It is good to hear that the step is not mandatory in this ISO standard.
 

Miner

TS Contributor
#9
I assume that the products that are relevant for GI testing are products with quite low GI values, thus values that are close to zero. Therefore I assume that they have a (positive) skewed distribution. So that only the large values will be classified as "outliers", by being larger than "mean+2*standard deviations".
GI is a 0 - 100 scale. Some foods such as peanuts have a GI of 7, while glucose has a GI of 95. Many foods fall into the 30 - 70 range.

GI is partially based on blood glucose levels tested by a finger prick test at a specific time after eating a specific quantity of the food. It is not unusual to obtain occasional bad readings from the test. Reasons vary from not hitting a capillary when pricking the finger and squeezing too much to get blood, not cleaning the skin first, etc.

ASTM had similar test procedures such as take X readings, throw out the high and low and use the remaining data. It always focused more on the vagaries of the test itself, rather than the distribution. For example, you might get a low tensile strength because the sample slipped in the jaws of the tester.
 

hlsmith

Less is more. Stay pure. Stay poor.
#10
White boiled potato is the god of GI.

Good point about the accuracy and reliability of the test used to acquire the GI value.
 
#11
Just to be clear, I work for clients who get GI testing done at centres which work to the ISO standard and then certify our products. We don't have input into their analysis.

Yes the ISO says it is not mandatory to remove outliers but I have yet to work with a centre that does not do it. As per my original question, we have worked with 3 centres (in three different countries) and 2 of them remove outliers once, the third does repeated removals.

Removal of outliers isn't a real issue for products with very low GIs (like peanuts) because they would still be low GI with or without outliers. Where it becomes an issue is for products at the boundary of low and medium GI (GI=55) or medium to high GI (GI=70).

Concerning the distribution, in the products the centres have tested for us with a GI around 50 the removal of high outliers outnumbers removal of low outliers by about 10 to 1 (estimate). Also, about 30-50% of the tests produce an outlier.

This has been very useful to discuss one of the queries I have with the scientific validity of GI testing.
 
Last edited:

rogojel

TS Contributor
#12
hi,
you could consider sending a range of products with varying GIs to all the labs and see how consistent their results are. That would give you an objective basis to rause the doubts - something like a measurement system analysis we do in six sigma.

I seem to remember that the FBI did something similar with DNA testing and got completely inconsistent results .

Regards
rogojel
 

Miner

TS Contributor
#13
Removal of outliers isn't a real issue for products with very low GIs (like peanuts) because they would still be low GI with or without outliers. Where it becomes an issue is for products at the boundary of low and medium GI (GI=55) or medium to high GI (GI=70).
For those unfamiliar with GI, high is bad, low is good. There is commercial motivation for the numbers to come in lower than the dividing line between categories.
 

hlsmith

Less is more. Stay pure. Stay poor.
#14
High is cpnsidered less than ideal for regular public consumption in every day foods. The corn syrup people may argue. There are medical uses for manufactured sugars (e.g., dextrose) used in these product to address hypoglycemia, etc.

Interesting thread, just wanted to point out that if you truncate off the outliers defined by standard deviation, then you just get new ones. Many good points were made about testing if they really are outliers due to systematic error.
 

Miner

TS Contributor
#15
High is cpnsidered less than ideal for regular public consumption in every day foods. The corn syrup people may argue. There are medical uses for manufactured sugars (e.g., dextrose) used in these product to address hypoglycemia, etc.
True. I was looking from the normal dietary perspective where it is undesirable to have glucose levels spike quickly then plummet. Obviously, glucose tablets and similar items have the opposite need (i.e., get glucose levels up quickly).