stats-208-jackknife

Analysis of Predictive inference with jackknife+

Extensions

Simulations

Real-World Data and Modeling

We gathered data from California Cooperative Oceanic Fisheries Invesitgations (CalCOFI) from the CalCOFI website. The downloaded CSV files (for the bottle and cast data) should go in data/raw/CalCOFI_Database_194903-202001_csv_22Sep2021/ to work with our written scripts. From here, we joined and processed the data to include fully-present observations across our response variable (salinity, denoted as Salnty) and 20 predictor variables: Distance, Bottom_D, Wind_Spd, Depthm, T_degC, O2ml_L, STheta, O2Sat, Oxy_µmol/Kg, ChlorA, Phaeop, PO4uM, SiO3uM, NO2uM, NO3uM, NH3uM, DarkAs, MeanAs, R_DYNHT, and R_Nuts. Please see the CalCOFI website for a codebook explaining each feature.

After processing, we were left with 6,102 complete observations. Similar to Barber, et al., we had a training set of 200 observations with the rest as our test set. We wanted to see how the jackknife+ would perform with a smaller number of predictors. Beyond this, though, we aimed to further test the generalizability of this method by constructing two different models: LASSO (with a hyperparameter value identical to the one proposed for the ridge regression simulations, and a boosting regressor (both used the default arguments in the model object from scikit-learn). Upon running these trials, we noticed that the performance remained similar across the models and interval types: the jackknife+ slightly outperformed the jackknife and met the coverage rate $1 - \alpha$ as proposed in the paper.