[Mitch]

The dirty truth in the synthetic literature is that reaction yields are inflated. After many long discussions with my synthetic friends, they all admit that the yields reported in their papers or the yields they report to the boss are often the maximum yield they ever obtained. I will admit that when I was a stereotypical synthetic, I too would always report the highest yield during group meeting or to fellow colleagues. Regardless of those acts, it is still wrong and highly unscientific.

Average

It would be far more appropriate to report the average yield obtained for the reaction along with the standard deviation. When would we ever allow our g-chem students to only report their "best" titration for one of their labs? When did reporting the most statistically aberrational data-point become the norm? This behavior has become far too common for synthetics. Reaction yields should not be the yardstick used to measure one's scientific manliness or ego.

I'm often comforted by my friendly synthetics that if a yield was abnormally high, they would investigate further and probably not trust it. I usually follow up those types of statements with, "do you check to see if that abnormal yield could be q-tested out?" The usual answer I get back, "q-test?"

This rant is meant to respectfully ask the synthetics in the audience to boldly begin reporting the average yield obtained for a reaction along with the standard deviation. By doing this, the stellar yields would still be deducible from this combination of yield + standard deviation, and the new format would be more scientifically meaningful.

Standard Deviation


- Mitch

P.S. Another Chem 2.0 post is in the works. I've recently extracted 10,000 Chemists' names from ACS's rss feeds and dumped them into my MySQL database. What will Mitch do with the names of 10,000 Chemists? Stay tuned!

[mir]

I suppose its because there is a lot of money and prestige in finding a synthesis that works. I see with my tutor and his group, they also reports the highest yields. Luckily for me, I have a bachelor in Analytical Chemistry. So I see the trouble in it. Its much better for everyone to describe nature as it is - to be completely honest about everything you do in the lab. And that means: Write good lab-notes as you do it!

Btw - I have never thought of yields in organic chemistry then more than a hint. Yields described in literature is never meant to be used in model making or statistics. 

[Bronwen Dekker]

How many times do you repeat a reaction using more-or-less-exactly the same conditions?

Would you be satisfied with a median and range?

Thinking back to reactions that I did do repeatedly, I do not think that the yields would have fallen into a nice bell-shaped distribution (what you need to for a meaningful SD)? Admittedly, I was rather a vague regarding syntheses - as long as the product was pure and there was enough to be getting on with...

[mir]

I do not think that the yields would have fallen into a nice bell-shaped distribution

I agree. If your reaction is very fast, small changes in the environment will change the path completely. And if the substrate to your product have to go through a lot of intermediates, there's lots of things that might go wrong. And if your product is degrading at the same moment, the distrubution would probably not look like mr. Gauss, maybe more like the Himalayas.

[Mitch]

How many times do you repeat a reaction using more-or-less-exactly the same conditions?

I would hope that a reaction reported in the literature would of been repeated at least 3 times to test reproducibility. This could be forgiven if it was a rather expensive or long synthesis, but the beginning reactions should of been done in triplicate.

If you had a non-gaussian distribution of yields, that would also be something very relevant to put in your experimental section too.

[Yggdrasil]

Yes, and I'm sure that you can assess whether your distribution is non-Gaussian using only three data points.     You'd probably need at least ten times that amount of data points to tell whether the distribution is Gaussian or not, at which point your adviser will start wringing your neck for wasting time and money.

[Mitch]

Yes, and I'm sure that you can assess whether your distribution is non-Gaussian using only three data points.   

You could still report the mean and std.dev. with 3 points. I never said ascertaining whether the reaction produced a Gaussian distribution was the goal.

[Bronwen Dekker]

I am pretty sure that the standard deviation equation 'assumes' that the data is Gaussian (i.e the distrubution is 'normal').

Perhaps have a look at these:
http://www.robertniles.com/stats/stdev.shtml
http://www.shodor.org/UNChem/math/stats/index.html
http://www.physics.hmc.edu/analysis/statistics.php

But it is probably a good idea to talk to a statistics guru or find a good book.

I did not really mean to sidetrack from your point, though. I definitely agree that there should be a greater level of honesty in reporting results. I was just suggesting that we consider 'median and range' as an alternative for situations where we get the feeling that the results are a bit 'all over the place'. I recently published a protocol that had a little trouble getting through review, because the percentage expected yields of a precursor were lower than those described in the literature. The author answered by saying that the literature values were basically premature and that after years of many different people doing the reaction the more honest expectation was the lower yield.

Perhaps Yggdrasil and mir have different sugguestions for how we would like people to report the data? If the experiment has only been done three times (or less...), one could just write all three yields down?

It would be useful if people admitted how many times the reaction was done?

[billnotgatez]

When I did my study I did the experiment 6 times, but I wanted to do 7. My statistics buddy said that sometimes people publish with only 2 and are excepted, but he knew that 7 times starts putting things into statistical significance and the degrees or freedom do not make things crazy. Most likely 30 trials would reduce all doubt. When I published I used mean, median, number of cases, standard deviation. I do regret that I did not publish minimum and maximum. Having worked in the publishing of data field for a while I do know that most people only want to hear one statistic and mean is the favorite. I wish I had a nickel for all the times people eyes glazed over or went to sleep as soon as they heard the words standard deviation. The reason to publish the min, max, mean, median, n of cases, and significant modes (if any) is to prove authenticity and it should be done in a simple quick table. Rambling prose sprinkled with statistical terms just does not keep the attention span.

Having said all of the above I see no problem in a person saying that they did this synthesis several times and the BEST results was the maximum. If they leave out the word BEST they are being dishonest. If the best was once out of a hundred and they rest were very poor then they should add that in. On the other hand, if you do not do a table of results in a publication you are not very professional.

What I said above is the staunchest thing I have said in a long time. Usually I am more congenial (read wishy-washy).

[Maz]

Just to clarify:

The standard deviation of a probability distribution is defined as the square root of the variance.  It does not inherently assume a normal distribution.  If your data happens to fit a Gaussian, then the std. dev. formula simplifies to the familiar form we all know.  However, you can take std. dev's on a poisson distribution, for example, where it simplifies to the square root of the average.  You can take a std. dev. on any probability distribution. 

see:  http://mathworld.wolfram.com/StandardDeviation.html  for more details. 

[Mitch]

Bronwen Dekker: I would take the median + range. Anything but just the maximum yield ever recorded. How many times that reaction was ran at a similar volume and concentration would also be useful

billnotgatez: You would be very hard pressed to find any paper in the current literature that specifically says "the best yield obtained was x%".