Open stephpenn1 opened 3 years ago
6/16 meeting (SP/ BBL):
From the methods of Terrer et al. 2021:
We coded 19 potential moderators (Extended Data Table 1). Including all 19 moderators in a meta-regression risks overfitting the model, so we applied the R package metaforest50 to identify potentially relevant moderators in predicting βsoil across the complete dataset of 108 studies. The approach is based on the machine-learning ‘random forest’ algorithm, which is robust to overfitting, and is integrated in a meta-analytic context by incorporating the variance and weight of each experiment as in classic meta-analysis (see above). ...Unlike maximum likelihood model-selection approaches, this method can handle many potential predictors and their interactions and considers nonlinear relationships. Partial dependence plots were produced that visualize the association of each moderator with the effect
This is basically what we're going to want to do.
Their code is at https://github.com/cesarterrer/SoilC_CO2
An example looks like this:
# Soil C stocks
library(funModeling)
dat$stocks_categorical <- funModeling::equal_freq(dat$amb, 3)
stocks <- rma.mv(yi, vi, data=dat, mods=~stocks_categorical -1, random = ~ 1 | Site / obs)
stocks
make_pct(coef(summary(stocks)))
stocks.n <- dat %>% group_by(stocks_categorical) %>% summarise(n = n())
stocks.df <- coef(summary(stocks)) %>% mutate(type="Soil C stocks",
factor=levels(dat$stocks_categorical),
size=stocks.n$n)
From 5/17 meeting (SP / BBL):