Tell me more ×
Cross Validated is a question and answer site for statisticians, data analysts, data miners and data visualization experts. It's 100% free, no registration required.
up vote 2 down vote favorite

I am trying to quantify my concerns regarding a proposed incinerator in our community. The company is basing its potential to emit dioxins (a class of chlorinated organic compounds with a reference dose of 1.7E-8 grams per 150 lb person per year) on three emission measurements (mg/m^3 exhaust):

0.0002139; 0.0000014; 0.00000186

The company has submitted their potential to emit as the mean of these three values, which is near legal limits. If the plant is built, they may be required to measure dioxins only once on completion, or possibly every year thereafter.

My initial thoughts have been to find the probability that these samples were drawn from a distribution with a mean above a certain threshold (e.g. the legal limit). As sample measurements, I would assume a normal distribution, though I guess with some non-zero skewness. I don't have any prior knowledge about the variance other than the samples.

But, I am now confronted with an infinite number of possible distributions, and am unsure how to proceed. I suspect this is a well studied area of statistics (risk analysis?), and any pointers in common methods would be much appreciated. I am a biochemist, and statistics is not my first language.

share|improve this question
As @Greg Snow points out (+1), there aren't enough data here to do anything definitive. You need a different strategy. There's a huge one out there: focus on the burden of proof. If the incinerator operator has to show that they meet the legal limit, then they will have to collect enough data to demonstrate they are highly likely to continue meeting the limit. This means that an upper bound of their values must lie beneath that limit. If you win this argument, you give the operator a strong incentive to collect lots of data, which will help both sides understand what the risks really are. – whuber♦ Oct 26 '12 at 19:15
To put it more technically: normally, it is not sufficient to have an average value lie below a regulatory limit. The operator must show that an upper confidence limit of the mean is below that limit. In the US, you should be able to find requirements to use UCLs written into the state and federal guidance documents, the regulations, and in some cases the very laws that govern this situation. (In some documents such limits are vaguely referred to as "UPLs" and even "ULs" ("upper limits").) – whuber♦ Oct 26 '12 at 19:18
Thanks! What you say makes much sense, but in our state much seems to be left to the DEQ permit writer, and it has been his position that it is not his responsibility to vet the applicant's submitted values but just to require standard testing (though the testing protocol depends on whether criteria threshold are exceeded in the application!). But we are making headway -- permit writer booted for conflict of interest. If application is granted this will be challenged. – roverred Oct 26 '12 at 21:52

1 Answer

up vote 1 down vote accepted

This is a case where a confidence interval is probably better than the significance test. You can construct a confidence interval using t procedures if you are willing to assume normality. With a positive skew (assuming you think that because the values cannot go below 0) you might take the log of the data, calculate the confidence interval using the t procedure, then transform the limits back to the original scale. The confidence interval gives a plausible region for where the true mean emmissions could be given the data. If there are values within the interval that are above the legal limit then that would argue at least for more data.

Possibly even better would be a prediction interval which says what interval you expect most of the individual measurements to be in, not just the mean. I remember seeing a comic that had 2 people standing on the side of a crater that is clearly the site of a recent explosion and one of them is saying something like "this factory is still the safest on average". Health concerns should look at how high the emmissions could get, not just what their average is.

With only 3 datapoints it would be best to see if there is other data on other similar plants or other information that could be used to justify assumptions of normality or other possible distributions used. You may also want to consult with a local statistician, the other side certainly will if they don't like your results and "some guy on the internet said ..." will probably not carry much weight.

share|improve this answer
Thank you! I only vaguely remembered the t-test for comparing whether two distributions were different way back in stats class. Wikipedia has a good write up, and indeed it seems exactly what I need. Unfortunately no more data available, as this is 'experimental' technology; though the company claims to have data showing cleaner emissions, they will not release it due to an odd argument about it being proprietary. Anyway, I think the burden should be on them to provide more data, and this will help. – roverred Oct 26 '12 at 20:21

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.