tIFA

The tIFA R package presents a set of functions that perform missing data imputation using a truncated infinite factor analysis model designed for high-dimensional metabolomics data.

Installation

You can install the development version of tIFA like so:

# install remotes package if not already installed
# install.packages("remotes")
remotes::install_github("kfinucane/tIFA")

# load the library
library(tIFA)

Example imputation

To illustrate the tIFA functionality, here a test dataset is generated with two missing values.

# load the library
library(tIFA)

# generate example data
example_data <- matrix(abs(rnorm(100)), nrow = 5)

# add missingness to example data coded as 0.001
example_data[4, 2] <- 0.001
example_data[2, 18] <- 0.001

This example data can then be passed into the tIFA_model() function which contains all package functionality. Here, input_data is the dataset with missing values, in matrix format. The coding argument refers to how the missing values are coded. If your missing values all have a value of NA, for example, you should input coding = NA. Here, for the example, missing entries are coded with the value 0.001. Here, the value k.star = 3 refers to the practical non-infinite number of latent factors used by the tIFA model; usually this defaults to k.star = 5 but for this small example dataset we will use a smaller number.

The remaining parameters refer to the MCMC procedure; here n.iters = 100 runs the MCMC chain for 100 iterations, with burn = 50 controlling the number of draws discarded in a burn and thin = 5 stating that every fifth post-burn draw from the MCMC chain should be retained. Please note that burn should have a value less than n.iters, and thin should divide into both burn, and n.iters - burn with no remainder.

# run tIFA model
# short chain for example
res <- tIFA_model(input_data = example_data, coding = 0.001, n.iters = 100, k.star = 3,
                  burn = 50, thin = 5)

The results of the tIFA imputation can be accessed as follows.

# checkout imputed dataset
res$imputed_dataset

# checkout further information on imputed entries
res$imputation_info

The input dataset, with missing values imputed according to the tIFA imputation method is contained in res$imputed_dataset in matrix format. Details on the imputation can be accessed from res$imputation_info. This is a dataframe with number of rows equal to the number of missing values in the dataset, and seven columns. The entry_row and entry_col columns contain the row and column index of each missing value in the input dataset. The imputed_val column contains the imputed value, and cred_int_upper and cred_int_lower contain the upper and lower limits of hte 95% credible interval for the imputed point. Finally, the miss_mech column gives the final inferred missingness type of each point, with miss_mech_unc providing the corresponding uncertainty.

If one wishes to change the hyperparameters of the tIFA model from their defaults, this can be done as follows. All parameters are as given in the tIFA paper.

res <- tIFA_model(input_data = example_data, coding = 0.001, n.iters = 100, k.star = 3,
                  burn = 50, thin = 5, mu_varphi = 0.1, kappa_1 = 3L, kappa_2 = 2L,
                  a_sigma = 1L, b_sigma = 0.3, a_1 = 2.1, a_2 = 3.1)

References

Finucane, K., Brennan, L., and Gormley, I. C. (2024). “Missing data imputation using a truncated infinite factor model with application to metabolomics data.” arXiv preprint: https://doi.org/10.48550/arXiv.2410.10633