Calling all statisticians

[shermarama]

I'M ON THE TRAIN. And doing some work. I'm writing a paper where I find myself saying 'this correlation coefficient is fairly high (0.82) but one of the sets of numbers being correlated is questionable, and so this apparent correlation may be not be valid.' So what I need is a way to say whether these results are so dodgy that making comparisons with them is stupid, or if they're just a bit dodgy and the comparison is therefore interesting, if far from bulletproof.

The dodgy results themselves are the gradients of some lines of best fit. The lines of best fit often have a poor correlation with the results that generate them - more than half of them have coefficients less than 0.7, and two are so poor that the correlations (in a known zero-order reaction, measuring the amount of the breakdown compound being generated against time) are negative. Also, the lines are being calculated from only four or five points of data, (five or six with a forced zero intercept, which should be true) and in one case only three, or four with the intercept. That one has a correlation of 0.95, mind.

I have tried calculating the standard error but with so few results, I think the standard error formula is meaningless. The largest calculated error is about 25%, which looking at the graph it's being generated from is surely bollocks. And this is why I need a statistician, to tell me whether this really is bollocks, or just plausible enough to have to be considered.

My only interest in this is to compare some test methods and show which test methods give bobbins for results, by the way. This one is already going to lose on the basis of taking lots of intervention and high-tech equipment to give you those meagre four or five poorly-correlating results. It's just a question of whether I should dismiss the actual results, too. Anyone got any clues?

Comments

[friend-of-tofu.livejournal.com]

http://www.flickr.com/photos/sluggerotoole/153603564/

Sorry!

[deathboy.livejournal.com]

is any graph with only 4 or 5 datapoints useful for much?

[shermarama]

Well, this is it. The method only produces that few results because it's a bit rubbish. But are the results totally ignorable rubbish as a result, or do they have enough credibility to justify observations about how similar they are to some other results?

[shermarama]

And also, I haven't made this method up, other people have used it and published papers with observations about the things being tested based entirely on results like these, so someone must think they're valid.