PhViD (version 1.0.3)

BCPNN: Bayesian confidence propagation neural network

Description

Bayesian confidence propagation neural network (Bate et al. 1998, Noren et al. 2006) extended to the multiple comparison framework.

Usage

BCPNN(DATABASE, RR0 = 1, MIN.n11 = 1, DECISION = 1, DECISION.THRES = 0.05, 
RANKSTAT = 1, MC = FALSE, NB.MC = 10000)

Arguments

DATABASE

Object returned by the function as.PhViD.

RR0

Value of the tested risk. By default, RR0=1.

MIN.n11

Minimum number of notifications for a couple to be potentially considered as a signal. By default, MIN.n11 = 1.

DECISION

Decision rule for the signal generation based on 1 = FDR (Default value)

2 = Number of signals

3 = Ranking statistic. See RANKSTAT

DECISION.THRES

Threshold for DECISION. Ex 0.05 for FDR (DECISION=1).

RANKSTAT

Statistic used for ranking the couples:

1 = Posterior probability of the null hypothesis

2 = 2.5% quantile of the posterior distribution of IC.

MC

If MC=TRUE, the statistic of interest (see RANKSTAT) is calculated by Monte Carlo simulations which can be very long. If MC=FALSE, IC is approximated by a normal distribution (which can be very crude for small counts

NB.MC

If MC=TRUE, NB.MC indicates the number of Monte Carlo simulations to be done

Value

• ALLSIGNALS
  Data.frame summarizing the results of all couples with at least MIN.n11 notifications ordered by RANKSTAT. It contains notably the labels, the cell counts, the expected counts ($n1. \times n.1 / N$, see as.PhViD), RANKSTAT, the ratios(count/expected count), the marginal counts and the estimations of FDR, FNR, Se et Sp. If RANKSTAT!=1, the last column is the posterior probability of the null hypothesis.
• SIGNALS
  Same Data.frame as ALLSIGNALS but restricted to the list of generated signals.
• NB.SIGNALS
  Number of generated signals.
• INPUT.PARAM
  Parameters entered in the function.

encoding

UTF-8

Details

The BCPNN method is based on the calculation of the Information Component IC. If MC = FALSE, the bayesian model used is the beta-binomial proposed by Bate et al. (1998). The statistic of interest (see RANKSTAT) is calculated by the normal approximation made in Bate et al. (1998) with the use of the exact expectation and variance proposed by Gould (2003). If MC = TRUE, the model is based on the Dirichlet-multinomial model proposed more recently in Noren et al. (2006). In this case, the statistic of interest is calculated by Monte Carlo simulations.

References

Ahmed I, Haramburu F, Fourrier-Réglat A, Thiessard F, Kreft-Jais C, Miremont-Salamé G, Bégaud B, Tubert-Bitter P. Bayesian pharmacovigilance signal detection methods revisited in a multiple comparison setting. Stat Med. 2009 Jun 15;28(13):1774-1792.

Bate A, Lindquist M, Edwards IR, Olsson S, Orre R, Lansner A, De Freitas RM, A Bayesian Neural Network Method for Adverse Drug Reaction Signal Generation European Journal of Clinical Pharmacology, 1998, 54, 315-321.

Gould AL, Practical Pharmacovigilance Analysis Strategies Pharmacoepidemiology and Drug Safety, 2003, 12, 559-574

Noren, GN, Bate A, Orre R, Edwards IR, Extending the methods used to screen the WHO drug safety database towards analysis of complex associations and improved accuracy for rare events Statistics in Medicine, 2006, 25, 3740-3757.

Examples

## start
data(PhViDdata.frame)
PhViDdata <- as.PhViD(PhViDdata.frame)
res <- BCPNN(PhViDdata)
## end