Coffee Data Science
Data on Waved Coffee
Analyzing other people’s taste data
Waved Coffee came out with a device that claims to use electromagnetic waves to changed the taste of coffee. Their machine is not yet widely available, and they have made claims backed with a panel of tasters. They kindly shared the data with me, and I’m taking a look at reviewing it.
I have no association with them other than asking if I could have the raw data.
For this data, they made three brew waters for N = 74 brews:
- Tonik (N = 32)
- Jamm (N = 20)
- Lylac (N = 22)
For each brew water, six variants were explored for each brew, so there was 74 x 6 total data points:
- Untreated Water (Baseline)
- Untreated Water (Baseline #2) with minerals added post-brew
- Waved 1: Brewed with waved water
- Waved 3: Brewed with waved water
- Waved 4: Prototype #1
- Waved 6: Prototype #2
Taste was evaluated based on a scale of 1 to 10:
- Sweetness
- Acidity
- Bitterness
- Dryness
- Body
- Preference
- Average across metrics 1 through 6
Each data point is assumed to be one taster across all the brew water variants.
Robustness of Flavor Metrics
Let’s first understand the correlation between each metric to better understand how independent each taste metric was. For these tests, I used all the 74 x 6 (N = 444) in the calculations.
A positive correlation means the two variables are proportional to one another, and a negative correlation means the two variables are inversely proportional to each other. I used the absolute because only a few lower correlations had negative numbers, so I thought this would clarify their connection to each other.
Correlation is between -1 (inversely correlated) to 0 (uncorrelated) to 1 (correlated), and I represent that correlation coefficient as a percentage for visual ease.
Sweetness has a mild correlation at 54% to Preference and Average. Dryness and Bitterness had a weaker correlation. Otherwise, all the components don’t have a strong correlation to each other.
Another way to look at this data is with a two tailed t-test to determine if the difference in the two distributions are statistically significant. A value less than 0.05 is considered statistically significant.
Get Robert McKeon Aloe’s stories in your inbox
Join Medium for free to get updates from this writer.
All the differences between variables statistically significant except for (Sweetness and Body) and (Acidity and Preference).
I looked at score distributions in a stacked bar chart where each column is the taste metric and then each blow is 1 through 10 in score.
I separated out Sweetness and Body as well as Acidity and Preference. They seem to have similar overall distributions.
Brew Water Data
Now, let’s look at the brew water, starting with correlation. Most of the correlations are weak, which could suggest any differences between them are more probable.
Moving on to a t-test using the average score, Waved 1, 4, and 6 are statistically significant in their taste difference to the Baseline, but only Wave 4 is statistically significant to Baseline #2 (minerals added post-brew).
The other key insight is that all the Waved tastes don’t have statistically significant differences to each other. Waved 3 and 4 are the closest, but most are no where near the 0.05 threshold.
Looking at the actual taste differences, this chart uses the row variant and subtracts the column variant. So Waved 1 is better than the baseline by 0.19 on average, and it is highlighted in green because it is statistically significant.
While these values are statistically meaningful, they are not very large. This can be explained in some scatter plots so one can see the noise.
Starting with Baseline #2 compared to Baseline, the values are all over the place.
Looking at Waved 1 and Waved 4, it is difficult to see improvements. Waved 4 is slightly clearer without squinting too hard.
Looking at Waved #4 across the water recipes, it is also not clear that any one of those is better.
I didn’t compare the water recipes otherwise because the datasets were uneven, and it was unclear if there were fair comparison data points.
Overall, I can say that their data is reasonable sound, and one could conclude from their data that using the Waved device improves taste. However, the difference they are seeing is noisy and overall small. Noisy means that the perceived differences varies greatly taster to taster.
I would love to try out the device and run my own tests. I’m curious how it performs relative to using ultrasonic mixing.
If you like, follow me on YouTube and Instagram where I post videos of espresso shots on different machines and espresso related stuff. You can also find me on LinkedIn. You can also follow me on Medium and Subscribe.
Further readings of mine:
My Second Book: Advanced Espresso
My First Book: Engineering Better Espresso
