The grind samples are done, and Kendall will be running the analyses towards the end of next week. I thought in the meantime we could discuss what sort of things we should look for in the data (settling this before the data is produced allows for some types of testing that are precluded when one peeks at the data first)

Here's the breakdown of the 20 pairs of samples we received

----------------------------
GRINDER        HOPPER SINGLE
----------------------------
Ascaso I2..........1........           
Baratza Vario......2......1.
Cimbali Junior.....1......1.
Compak K10.........2......3.
Lelit PL53.........1........
Macap MXK.................1.
Macap M7K..........1......1.
Mazzer Kony........1......1.
Mazzer Robur.......2*.......
Versalab M3...............1.
----------------------------
* Both pairs from the same grinder


The bad news is that we are under-represented on conventional flat burr grinders; the good news is we have lots of duplication on some grinders and lots of hopper/no hopper sets.

Here's my rough take, please contradict or add -- the better we understand the issues the better we can analyze the data:

We will be able to nail the question whether hopper grinds differ from single dose grinds at the same dose and flow and how they differ.

We will be able to nail the question whether grind samples from the same grinder prepared at different homes can be adequately standardized by standardizing coffee, dose, basket and flow.

We may not be able to learn much about the greater consistency of large conicals; that is, whether there is a difference in how conicals change grind with how flats do it.

You can grind "finer" in two ways:

• by producing more fines to restrict the flow without changing the size or way the coarse particles extract;
• or by making the coarse particles smaller, and so both restricting the flow and increasing the extraction rate.
• the forgiveness factor of conicals may be because they work by mostly increasing fines, leaving extraction rates more constant, so that this can be controlled more simply by the length of the shot

We'll have lucked out if the Vario, Cimbali and Versalab samples fall onto one group, and the conicals into another; otherwise, we may just end up with hints about the consistency factor.

That would be cool very ; but there's a but.

To compare the overall shape, you have to compare single samples. We might be able to do that; if all the Vario and K10 samples submitted by different people are really close, the differences between different brands of grinders will be meaningful. If the output of the same grinder from different sites is as different as the outputs of different brands of grinders, only the paired sample data will be meaningful; and that just tells us how the grind changes.

But we should do the quality control test. If the raw distributions from the same brand of grinder are really close, we have much higher quality data than we expected, and can look at the different grinders directly.

Aleksi is right. If you test for patterns in the same data in which you've discovered them, success is both guaranteed and meaningless. It's equally meaningless to test for thousands of patterns and then declare ten or so significant at the 1% level. "Reanalyses" or "meta-analysis" of social survey or epidemiological data frequently produce pseudo-discoveries of this sort, since people forget to modify their significance levels to account for all the testing done by previous researchers.

The proper practice is to take a subset of the data, look for prominent patterns in it, then test the rest of the data to see if those patterns indeed hold. We have too few samples to do that; so we need to decide ahead of time what to test.

I'm looking closely at the samples previously produced by Kendall and John to get ideas.