Introduction to openEBGM

2020-03-02

Background

William DuMouchel (1, 2) created an empirical Bayes (EB) data mining approach for finding “interestingly large” counts in contingency tables. DuMouchel’s approach works well even when most of the counts are zero or one (i.e., a sparse table). The benefit of DuMouchel’s model over simpler approaches such as the relative reporting ratio, \(RR\), is that Bayesian shrinkage corrects for the high variability in \(RR\) that results from small counts.

The rows and columns of the table represent levels of two different variables, such as food or drug products and adverse events:

Headache Nausea Vomiting
Product A 0 1 0
Product B 4 0 0
Product C 1 0 9

The relative reporting ratio is calculated as \(RR=\frac{N}{E}\), where \(N\) is the actual count for a cell in the table and \(E\) is the expected count under the assumption of independence between rows and columns. When \(RR = 1\), you observe the exact count you would expect to observe if no association exists between the two variables. When \(RR > 1\), you observe a larger count than expected. This approach works well for large counts; however, small counts cause \(RR\) to become quite unstable. For instance, an actual count of \(N = 1\) with an expected count of \(E = 0.01\) gives us \(RR = 100\) – which seems large – but a single event could easily occur simply by chance. The EB approach shrinks large \(RR\)s that result from small counts to a value much closer to the “null hypothesis” value of 1. The shrinkage is smaller for larger counts and negligible for very large counts. Shrinkage gives results that are more stable than the simple \(RR\) measurement.

DuMouchel’s model uses a Poisson(\(\mu_{ij}\)) likelihood (i.e. data distribution) for the actual cell counts, \(N_{ij}\), in row i and column j. The expected cell counts, \(E_{ij}\), are treated as constants. We are interested in the ratio \(\lambda_{ij}=\frac{\mu_{ij}}{E_{ij}}\), which is analagous to \(RR=\frac{N}{E}\). The prior on \(\lambda\) is a mixture of two gamma distributions, resulting in a posterior distribution for \(\lambda\) which is a mixture of two gamma distributions. Hence, the model is sometimes referred to as the Gamma-Poisson Shrinker (GPS) model. The posterior distribution of \(\lambda\) can be thought of as a Bayesian representation of \(RR\). Summary statsistics from the posterior distribution are used as attenuated versions of \(RR\).

Each cell in the contingency table will have its own posterior distribution determined both by that cell’s actual and expected counts (the data) and by the distribution of actual and expected counts in the entire table (the prior). Often, the Empirical Bayes Geometric Mean \((EBGM)\) of the posterior distribution is used in place of \(RR\). Alternatively, the more conservative percentiles (5th, 10th, etc.) can be used. The percentiles can also be used to construct Bayesian credible intervals. Similar to \(RR\), an \(EBGM\) (or lower percentile) much bigger than 1 represents an actual count much bigger than expected. Such cases might represent signals of interest, and the product/event pair can be further examined by subject matter experts to determine if the association might actually be causal in nature.

An extension of the GPS model, the Multi-Item Gamma-Poisson Shrinker (MGPS) model (2001), is currently being used by the U.S. Food and Drug Administration (FDA) to find higher-than-expected reporting of adverse events associated with food, drugs, etc. For instance, FDA’s Center for Food Safety and Applied Nutrition (CFSAN) uses the MGPS model to mine data from the CFSAN Adverse Events Reporting System (CAERS): https://www.fda.gov/Food/ComplianceEnforcement. (The variables forming the rows and columns of the contingency table are product and adverse event.) MGPS allows for product interactions, unlike the GPS model implemented in openEBGM (3), which can only use individual product-event pairs.

Purpose

The openEBGM package implements DuMouchel’s approach with some small differences. For example, the expected counts are calculated by counting unique “transactions” (2) in each row and column, not actual marginal totals. In the CAERS data, a unique report is a transaction. In some applications, a single transaction could occur in several cells. For instance, a single CAERS report might mention multiple products and/or adverse events. Using simple marginal totals would then count a single report multiple times.

This document teaches you how to prepare your data for use by openEBGM’s functions. Other vignettes give explanations and examples of more advanced topics:

References

  1. DuMouchel W (1999). “Bayesian Data Mining in Large Frequency Tables, With an Application to the FDA Spontaneous Reporting System.” The American Statistician, 53(3), 177-190. https://doi.org/10.1080/00031305.1999.10474456

  2. DuMouchel W, Pregibon D (2001). “Empirical Bayes Screening for Multi-item Associations.” In Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’01, pp. 67-76. ACM, New York, NY, USA. ISBN 1-58113-391-X. https://doi.org/10.1145/502512.502526

  3. Canida T, Ihrie J (2017). “openEBGM: An R Implementation of the Gamma-Poisson Shrinker Data Mining Model.” The R Journal, 9(2), 499-519. https://journal.r-project.org/archive/2017/RJ-2017-063/index.html


Preparing Your Data

openEBGM requires the input data to be a data frame of a particular form.

Data form

Data must be in tidy format (one column per variable and one observation per row). The columns can be of type factor, character, integer, or numeric. Missing values are not allowed - either replace NAs and empty strings with appropriate values or remove them from the data.

Column names

The input data frame must contain certain column names: var1, var2, and id. var1 and var2 are simply the row and column variables of the contingency table. The identifier (id) column allows openEBGM to properly handle marginal totals (for instance, this would be the report identifier variable in the aformentioned CAERS data). If the cells of the table actually represent mutually exclusive “events of interest”, the user can create a column of unique sequential identifiers with df$id <- 1:nrow(df).

Stratification can help reduce the effects of confounding variables. If stratification is used, any column whose name contains the substring ‘strat’ (case sensitive) will be treated as a stratification variable. If a continuous variable such as age is used for stratification, remember to categorize the variable.

Other columns are allowed, but will be ignored.

CAERS example

Here is a small subset of raw data from the publicly available CAERS data described above:

Only one product name is given per row, but we need to separate the adverse events into different rows:

Next we need to change column names:

Suppose we want to use gender and age as stratification variables:

Age is a continuous variable, so we need to create categories:

Now we have the data in the proper form:

Next, the Processing Raw Data with openEBGM vignette will demonstrate how to use data in this general form to find counts and simple disproportionality measures–\(RR\) and \(PRR\) (proportional reporting ratio).