New adaptive lasso approaches for variable selection in automated pharmacovigilance signal detection

Abstract

Background: Adverse effects of drugs are often identified after market introduction. Post-marketing pharmacovigilance aims to detect them as early as possible and relies on spontaneous reporting systems collecting suspicious cases. Signal detection tools have been developed to mine these large databases and counts of reports are analysed with disproportionality methods. To address disproportionality method biases, recent methods apply to individual observations taking into account all exposures for the same patient. In particular, the logistic lasso provides an efficient variable selection framework, yet the choice of the regularization parameter is a challenging issue and the lasso variable selection may be inconsistent. Methods: We propose a new signal detection methodology based on the adaptive lasso in the context of high dimension. We derived two new adaptive weights from (i) a lasso regression using the Bayesian Information Criterion (BIC), and (ii) the class-imbalanced subsampling lasso (CISL), an extension of stability selection. The BIC is used in the adaptive lasso stage for variable selection. We performed an extensive simulation study and an application to real data, where we compared our methods to the existing adaptive lasso, and recent detection approaches based on lasso regression or propensity scores in high dimension. For both studies, we evaluate the methods in terms of false discoveries and sensitivity. Results: In the simulations and the application, both proposed adaptive weights show equivalent or better performances than the other competitors, with an advantage for the CISL-based adaptive weights. CISL and lasso regression using BIC are solid alternatives. Conclusion: Our proposed adaptive lasso is an appealing methodology for signal detection in pharmacovigilance. Although we cannot rely on test theory, our approaches show a low and stable False Discovery Rate in all simulation settings. All methods evaluated in this work are implemented in the adapt4pv R package.

On the use of cross‐validation for the calibration of the adaptive lasso

“…The cross‐validation scheme described in Algorithm 1 is standard for the calibration of the lasso and the weighted lasso (i.e., when weights are given a priori). But, it is also commonly used for the calibration of the adaptive lasso (Chang et al., 2020; Courtois et al., 2021; Dessie et al., 2021; Pollard et al., 2021). Indeed, for the calibration of the one‐step lasso, for example, many statistical analysts implement Algorithm 2: They first compute initial estimates on the total sample, from which they derive the weight vector $${\bf w}\in \mathbb {R}^p_{\ge 0}$$ (Step 1).…”

“…In particular, when using the popular glmnet R package (Friedman et al., 2010), it is tempting to implement the naive scheme by simply applying the cv.glmnet function to select the hyperparameter of the adaptive lasso after computing the initial estimates on the total sample and considering the resulting weight vector as given a priori; see Algorithms 2, A1, and A3 for the complete description of this strategy in the case of the one‐step lasso, ols‐adaptive lasso, and ridge‐adaptive lasso, respectively. This improper cross‐validation scheme was used in the illustrative example of section 2.8.1 of Bühlmann and van De Geer (2011), and can also be found in the original version of the adapt4pv R package (Courtois et al., 2021). Consequently, many statistical analysts still seem to apply this naive and improper cross‐validation scheme for calibrating the adaptive lasso (Chang et al., 2020; Dessie et al., 2021; Pollard et al., 2021).…”