In the main vignette, I discussed variable importance defined using R-squared. I also mentioned that all of the analyses were carried out using a condititonal variable importance measure. In this document, I will discuss all three types of variable importance that may be computed using vimp.
In general, I define variable importance as a function of the true population distribution (denoted by \(P_0\)) and a predictiveness measure \(V\) – large values of \(V\) are assumed to be better. Currently, the measures \(V\) implemented in vimp are \(R^2\), classification accuracy, area under the receiver operating characteristic curve (AUC), and deviance. For a fixed function \(f\), the predictiveness is given by \(V(f, P)\), where large values imply that \(f\) is a good predictor of the outcome. The best possible prediction function, \(f_0\), is the oracle model – i.e., the prediction function that I would use if I had access to the distribution \(P_0\). Often, \(f_0\) is the true conditional mean (e.g., for \(R^2\)). Then the total oracle predictiveness can be defined as \(V(f_0, P_0)\). This is the best possible value of predictiveness.
I define variable importance measures (VIMs) as contrasts in oracle predictivness. The oracle models that I plug in determine what type of variable importance is being considered, as I outline below. For the remainder of this document, suppose that I have \(p\) variables, and an index set \(s\) of interest (containing some subset of the \(p\) variables). Throughout this document, I will use the VRC01 data (Magaret et al. 2019), a subset of the data freely available from the Los Alamos National Laboratory’s Compile, Neutralize, and Tally Neutralizing Antibody Panels database. Information about these data is available here. Throughout, I will also use a simple library of learners for the Super Learner (this is for illustration only; in practice, I suggest using a large library of learners, as outlined in the main vignette). Finally, I will use the area under the receiver operating characteristic curve (AUC) to measure importance.
# read in the data
data("vrc01")
# subset to the columns of interest for this analysis
library("dplyr")
library("tidyselect")
# retain only the columns of interest for this analysis
y <- vrc01$ic50.censored
X <- vrc01 %>%
select(starts_with("geog"), starts_with("subtype"), starts_with("length"))
learners <- "SL.glm"The reduced oracle predictiveness is defined as \(V(f_{0,-s}, P_0)\), where \(f_{0,-s}\) is the best possible prediction function that does not use the covariates with index in \(s\). Then the conditional VIM is defined as \[V(f_0, P_0) - V(f_{0,-s}, P_0).\] This is the measure of importance that I estimated in the main vignette. To estimate the conditional VIM for family history of heart disease, I can use the following code:
The marginal oracle predictiveness is defined as \(V(f_{0,s}, P_0)\), where \(f_{0,s}\) is the best possible prediction function that only uses the covariates with index in \(s\). The null oracle predictiveness is defined as \(V(f_{0, \emptyset}, P_0)\), where \(f_{0,\emptyset}\) is the best possible prediction function that uses no covariates (i.e., is fitting the mean). Then the marginal VIM is defined as \[V(f_{0,s}, P_0) - V(f_{0,\emptyset}, P_0).\] To estimate the marginal VIM for family history of heart disease, I can use the following code:
The Shapley population VIM (SPVIM) generalizes the marginal and conditional VIMs by averaging over all possible subsets. More specifically, the SPVIM for feature \(j\) is given by \[\sum_{s \subseteq \{1,\ldots,p\} \setminus \{j\}} \binom{p-1}{\lvert s \rvert}^{-1}\{V(f_{0, s \cup \{j\}}, P_0)) - V(f_{0,s}, P_0)\};\] this is the average gain in predictiveness from adding feature \(j\) to each possible grouping of the other features. To estimate the SPVIM for family history of heart disease, I can use the following code (note that sp_vim returns VIM estimates for all features):
In some cases, there may be confounding factors that you want to adjust for in all cases. For example, in HIV vaccine studies, we often adjust for baseline demographic variables, including age and behavioral factors. If this is the case, then the null predictiveness above can be modified to be \(V(f_{0,c}, P_0)\), where \(c\) is the index set of all confounders.
The three VIMs defined here may be different for a given feature of interest. Indeed, we can see this for whether or not subtype is 01_AE in the VRC01 data:
subtype_01_AE_cond
#> Variable importance estimates:
#> Estimate SE 95% CI VIMP > 0 p-value
#> s = 5 0.001778934 0.05614002 [0, 0.1118113] FALSE 0.4873607
subtype_01_AE_marg
#> Variable importance estimates:
#> Estimate SE 95% CI VIMP > 0 p-value
#> s = 1 0.01302083 0.1073992 [0, 0.2235194] FALSE 0.4517514
# note: need to look at row for s = 5
all_vim_spvim
#> Variable importance estimates:
#> Estimate SE 95% CI VIMP > 0 p-value
#> s = 1 0.000000000 0.021791723 [0, 0.04271099] FALSE 0.5000000
#> s = 2 0.011043008 0.017183290 [0, 0.04472164] FALSE 0.4379094
#> s = 3 0.010119884 0.028051877 [0, 0.06510055] FALSE 0.4445270
#> s = 4 0.012518761 0.034394897 [0, 0.07993152] FALSE 0.4325797
#> s = 5 0.000000000 0.013993958 [0, 0.02742765] FALSE 0.5000000
#> s = 6 0.000000000 0.006363417 [0, 0.01247207] FALSE 0.5000000
#> s = 7 0.000000000 0.009907345 [0, 0.01941804] FALSE 0.5000000
#> s = 8 0.015550728 0.019018377 [0, 0.05282606] FALSE 0.4132343
#> s = 9 0.000000000 0.023765091 [0, 0.04657872] FALSE 0.5000000
#> s = 10 0.000000000 0.018762896 [0, 0.03677460] FALSE 0.5000000
#> s = 11 0.014845056 0.018106896 [0, 0.05033392] FALSE 0.4169983
#> s = 12 0.034118865 0.033878081 [0, 0.10051868] FALSE 0.3212957
#> s = 13 0.000000000 0.013849446 [0, 0.02714442] FALSE 0.5000000
#> s = 14 0.007543575 0.023941056 [0, 0.05446718] FALSE 0.4580642
#> s = 15 0.000000000 0.015921007 [0, 0.03120460] FALSE 0.5000000
#> s = 16 0.043251122 0.065473115 [0, 0.17157607] FALSE 0.3019106
#> s = 17 0.035069436 0.064557522 [0, 0.16159985] FALSE 0.3367702
#> s = 18 0.044388824 0.051888790 [0, 0.14608898] FALSE 0.2863218
#> s = 19 0.001066722 0.020758848 [0, 0.04175332] FALSE 0.4940285
#> s = 20 0.028659642 0.054194553 [0, 0.13487901] FALSE 0.3592621
#> s = 21 0.017929705 0.040749285 [0, 0.09779684] FALSE 0.4060425This is simply a function of the fact that the VIMs are different population parameters. All three likely provide useful information in practice:
To choose a VIM, identify which of these three (there may be more than one) that best addresses your scientific question.
Magaret, CA, DC Benkeser, BD Williamson, BR Borate, LN Carpp, IS Georgiev, I Setliff, et al. 2019. “Prediction of VRC01 Neutralization Sensitivity by HIV-1 Gp160 Sequence Features.” PLOS Computational Biology. https://doi.org/10.1371/journal.pcbi.1006952.