Share
Embed
Transcript
40 slides
-
Estimating risk of seizure recurrence for people with epilepsy Laura
Bonnett, NIHR Post-Doctoral Fellow -
Statistical Methods log = + log( ) Poisson
Negative Binomial Zero-inflated Poisson Zero-inflated Negative Binomial Expected number of events for subject i Length of time subject i is at risk Covariate design matrix Parameter estimates -
Statistical Methods log = + log( ) Poisson
Negative Binomial Zero-inflated Poisson Zero-inflated Negative Binomial Expected number of events for subject i Length of time subject i is at risk Covariate design matrix Parameter estimates -
Statistical Methods Extended Cox models for multiple events of
the same type: Independent increment – Anderson-Gill (AG) Marginal – Wei, Lin and Wiessfeld (WLW) Conditional – Prentice, Williams and Peterson (PWP) -
Statistical Methods , = Ι ()0 ()() Anderson-Gill Hazard function
for ith subject at time t Common baseline hazard for all events over time Vector of covariate processes for ith individual Fixed vector of coefficients Predictable process taking values in {1,0} indicating when the ith individual is under observation Anderson & Gill (1982). Cox’s regression model for counting processes: a large same study. Annals of Statistics 10(4):1100-20 -
Anderson-Gill Statistical Models Efficient Accommodates heterogeneity via robust
standard errors Simple to visualise & set up Strong distributional assumption – event does not change the subject Lacks the detail & versatility of event-specific models Pros Cons -
Statistical Methods , = ()0 ()() WLW Hazard function for
kth event of the ith subject at time t Common baseline hazard for all events over time Vector of covariate processes for ith individual Fixed vector of coefficients At-risk process is 1 until the kth event, unless censored Wei, Lin & Weissfeld(1989). Regression analysis of multivariate incomplete failure time data by modelling marginal distributions. J. Amer. Statist. Assoc. 84:1065-73 -
WLW Statistical Models Allows for changes in the model
effects over time Datasets get quite large Semi-restricted risk set – subjects to be at risk of the event (where = , , …) even if they have only had one event. Pros Cons -
Statistical Methods , = ()0 ()() PWP Hazard function for
kth event of the ith subject at time t Common baseline hazard for all events over time Vector of covariate processes for ith individual Fixed vector of coefficients At-risk process on the interval from the k-1th event to the kth event unless censored Prentice, Williams, Peterson (1981). On the regression analysis of multivariate failure time data. Biometrika 68(2):373-79 -
PWP Statistical Models Underlying intensity function can vary from
event to event Natural interpretation Easy to set up Biased when an important covariate is omitted due to loss of balance in later strata Risk sets for later event numbers will get small, making estimates of per-stratum risk unstable Pros Cons -
Statistical Methods Extended Cox models for multiple events of
the same type: Counting process - Anderson-Gill (AG) Marginal – Wei, Lin and Wiessfeld (WLW) Conditional – Prentice, Williams and Peterson (PWP) -
Statistical Methods PWP-CP for risk of seizure recurrence
Comparison with published Cox models Risk of 1st seizure after randomisation Chance of 12-month remission Multivariable model development via forwards and backwards selection according to AIC Treatment forced into each model -
Data Standard versus New Antiepileptic Drug (SANAD) Study (n=2437)
Un-blinded RCT Hospital based outpatient clinics in U.K. December 1999 → August 2004 → January 2006 Outcomes Time to treatment failure Time to 12 month remission Arm Patients Standard Drug Typical Seizure Type Randomised Drugs A 1721 CBZ Focal Onset CBZ, GBP, LTG, TPM, OXC* B 716 VPS Generalised Onset VPS, LTG, TPM Marson, A.G., et al., The SANAD study of effectiveness of carbamazepine, gabapentin, lamotrigine, oxcarbazepine, or topiramate for treatment of partial epilepsy: an unblinded randomised controlled trial. Lancet, 2007. 369(9566): p. 1000-15. Marson, A.G., et al., The SANAD study of effectiveness of valproate, lamotrigine, or topiramate for generalised and unclassifiable epilepsy: an unblinded randomised controlled trial. Lancet, 2007. 369(9566): p. 1016-26. -
Discussion Model validation required Investigate in alternative data Compare to
count-based models Better understanding of treatment effects Improved statistical power -
Count & Rate Based Models: Poisson Statistical Models Very
simple Commonly used to compare event rates across different groups Ignores heterogeneity amongst patients within different groups Assumes independence of events within individuals Tends to underestimate the underlying true rate Pros Cons -
Count & Rate Based Models: Negative Binomial Statistical Models
Simple & straightforward to use Does not require complicated data files Offers more valid mean rates than Poisson Allows different patient risks Pros Cons -
Count & Rate Based Models: ZIP & ZINB Statistical Models
Handle data with excess 0s, and potential over-dispersion Simpler models do not allow mixing with respect to the 0s Mixture effect is require as variance is not just in the number of 0s Pros Cons