BACKGROUND: Current quantitative signal detection methods have been primarily developed for the purpose of detecting signals from spontaneous reports. These methods are not always appropriate for cohort data. More recently, parametric time-to-event models have been proposed to model hazard functions with the ultimate aim of detecting adverse drug reactions (ADRs). The rate of occurrence of ADRs after starting a drug will depend upon the causal mechanism and therefore will often vary with time, in contrast to events not associated with the drug, which will tend to occur at a constant background rate. After starting treatment, the onset of ADRs will be rapid for some but delayed for others. A non-constant rate over time may indicate a drug-event relationship. OBJECTIVE: The aim of this study was to propose a simple test to detect signals of ADRs in cohort data and to investigate the power of this test using simulated data. A signal detection tool using the proposed test to improve the power of detection is also described. METHOD: In order to test for a non-constant hazard (rate of occurrence), the hazard function was estimated using the model shape parameter for the Weibull function. If the shape parameter was found to be significantly different (p < 0.05) from the value one (the value for a constant hazard) a signal was raised. Simulation of background event rates used were 1%, 5% and 10% of the cohort size. The ADR rate was varied in proportion to the background rate; a 10%, 20% and 50% increase in the background rate was explored. The time of occurrence of the ADR will dictate the shape of the hazard function, therefore the ability of the model to detect a signal depending when the highest risk for ADR was also explored. The power of the test was investigated by simulation. RESULTS: The Weibull Shape Parameter (WSP) test was most powerful at detecting signals that occur shortly after starting treatment. These preliminary simulations had low power when the underlying hazard function was symmetrical (e.g. when ADRs occurred in the middle of the study period). The power of the test was improved by censoring the data as this broke the symmetry of the hazard function. A tool that censored the data at regular intervals and repeated the WSP test was found to correctly detect ADR or no ADR around 90% of the time when the sample size was at least 5000. CONCLUSION: The WSP test is simple to implement using standard statistical software, and can be used to detect non-constant hazards over time in order to raise signals of time-dependent ADRs. When there is no pre-specified event of interest or the time of the ADR is uncertain, the WSP tool should be used instead of the WSP test. These methods do not require any external data for comparative purposes and thus can be implemented in a single cohort of participants exposed to a drug.
BACKGROUND: Current quantitative signal detection methods have been primarily developed for the purpose of detecting signals from spontaneous reports. These methods are not always appropriate for cohort data. More recently, parametric time-to-event models have been proposed to model hazard functions with the ultimate aim of detecting adverse drug reactions (ADRs). The rate of occurrence of ADRs after starting a drug will depend upon the causal mechanism and therefore will often vary with time, in contrast to events not associated with the drug, which will tend to occur at a constant background rate. After starting treatment, the onset of ADRs will be rapid for some but delayed for others. A non-constant rate over time may indicate a drug-event relationship. OBJECTIVE: The aim of this study was to propose a simple test to detect signals of ADRs in cohort data and to investigate the power of this test using simulated data. A signal detection tool using the proposed test to improve the power of detection is also described. METHOD: In order to test for a non-constant hazard (rate of occurrence), the hazard function was estimated using the model shape parameter for the Weibull function. If the shape parameter was found to be significantly different (p < 0.05) from the value one (the value for a constant hazard) a signal was raised. Simulation of background event rates used were 1%, 5% and 10% of the cohort size. The ADR rate was varied in proportion to the background rate; a 10%, 20% and 50% increase in the background rate was explored. The time of occurrence of the ADR will dictate the shape of the hazard function, therefore the ability of the model to detect a signal depending when the highest risk for ADR was also explored. The power of the test was investigated by simulation. RESULTS: The Weibull Shape Parameter (WSP) test was most powerful at detecting signals that occur shortly after starting treatment. These preliminary simulations had low power when the underlying hazard function was symmetrical (e.g. when ADRs occurred in the middle of the study period). The power of the test was improved by censoring the data as this broke the symmetry of the hazard function. A tool that censored the data at regular intervals and repeated the WSP test was found to correctly detect ADR or no ADR around 90% of the time when the sample size was at least 5000. CONCLUSION: The WSP test is simple to implement using standard statistical software, and can be used to detect non-constant hazards over time in order to raise signals of time-dependent ADRs. When there is no pre-specified event of interest or the time of the ADR is uncertain, the WSP tool should be used instead of the WSP test. These methods do not require any external data for comparative purposes and thus can be implemented in a single cohort of participants exposed to a drug.
Authors: Preciosa M Coloma; Martijn J Schuemie; Gianluca Trifirò; Rosa Gini; Ron Herings; Julia Hippisley-Cox; Giampiero Mazzaglia; Carlo Giaquinto; Giovanni Corrao; Lars Pedersen; Johan van der Lei; Miriam Sturkenboom Journal: Pharmacoepidemiol Drug Saf Date: 2010-11-08 Impact factor: 2.890
Authors: Yvonne Buggy; Victoria Cornelius; Carole Fogg; Rachna Kasliwal; Deborah Layton; Saad A W Shakir Journal: Drug Saf Date: 2013-07 Impact factor: 5.606
Authors: Joep H G Scholl; Florence P A M van Hunsel; Eelko Hak; Eugène P van Puijenbroek Journal: Pharmacoepidemiol Drug Saf Date: 2019-06-12 Impact factor: 2.890