,
13 tweets,
6 min read
Read on Twitter
Yesterday I promised to write some tweets about our recent article in @PNASNews. A nerdy thread on technical details first! We used a cool model to analyze farmers' crop variety evaluation and test if seasonal climate could explain differences. doi.org/10.1073/pnas.1… 
The model is called "Plackett-Luce trees". This model produced the nice graph above. The model splits the dataset in different segments of farmers that each have very different observations on the crop varieties. The example shows bean varieties in Nicaragua.
Our input data consisted of farmers' rankings. We already knew that farmers can easily and very reliably tell which variety does better than another variety in a plot next to it. This is the core principle of our #citizenscience approach. (See doi.org/10.1007/s13593…)
The values on the x axis of that graph above are the probabilties that a variety would "beat" all other varieties, according to what farmers' rank in their own fields. So higher values indicate varieties that did better.
We used this model to see if seasonal climate had any influence on farmers' observations. We fed many climate covariates to the model, obtained from satellite data by @PaliwalAmbica.
Lo and behold, the model told us that the bean farmers had very different observations when the night temperature was higher than 18.7 degrees Celsius. Which happens to be very close to what the textbooks say about heat stress tolerance in beans!
In other words, farmers can accurately observe heat stress tolerance under farm conditions. That is a big deal, because it means that the huge breeding effort to develop this heat stress tolerance in beans has paid off and is visible where it really matters: in farmers' fields.
The problem was that we only realized that we needed the blessed Plackett-Luce model after we had already started analyzing the data... This was the perfect model for our data, but it was not available in any software! So we freaked out a little bit...
Luckily we found @HeathrTurnr who did an amazing job to implement Plackett-Luce trees in #rstats (with @firthcity and @IKosmidis_). This was our first time to have such professional support in software implementation, a great experience!
So that led to another paper, describing a new implementation of Plackett-Luce model and trees in #rstats. The cool thing about open-source software development is that the code can also help others to apply the code to completely different problems. See arxiv.org/abs/1810.12068
We tweaked how the results are displayed a bit. The partitioning part that makes those decision trees is based on software developed by @AchimZeileis and others. Normally is shows the results as log-odds, like here. hturner.github.io/PlackettLuce/a…
@desousakaue converted the log-odds to probabilities that one variety beats all other varieties in the trial. He added error bars and a vertical gray line and flipped the axes. Perhaps @AchimZeileis likes it so much that this can be the default way of showing this type of graph!
I will write some other tweets and threads about this work in the coming days. Stay tuned!
