OBJECTIVES:  When number of events are reported over varying duration of times, meta-analyzing these data using a compound Poisson process can help generate a survival curve for these events with time-varying hazard. We identify the structural assumptions underlying these methods, determine the number of knots possible given the data, where hazard changes, and apply these methods to group-level data from the literature on incident seizures following stroke. METHODS:  We searched PubMed and EMBASE for observational studies of post-stroke seizures from 1997 to 2016. We built a likelihood function based on a compound Poisson process that used different Poisson distribution across multiple overlapping intervals to model the counts of events. We divided 5 years into 3 intervals: within 7 days, from 8 days to 1 year, and from 1 year to 5 years after stroke. We compare this time classification to alternatives, which also include a constant hazard model over time using a single Poisson process or an exponential parametric model. Goodness-of-fit was demonstrated using AIC and BIC criteria. RESULTS:  Our preliminary results, based on 12 studies published in the past 20 years suggest that the compound Poisson process with time varying hazard is the best model based on the lowest AIC and BIC, and the Poisson process with constant hazard is also better fitting than using a parametric survival approach. We found that hazard rate of seizures is 0.014, 0.023, and 0.173 for 3 intervals respectively: first 7 days, from 8 days to 1 year, and from 1 year to 5 years after stroke, indicating that the risk of seizure after stroke changes over time. CONCLUSIONS:  The compound Poisson process could be a better way to meta-analyze count-data over different exposure periods, present the summary as survival curves that are readily interpretable by applied researcher and clinicians, and can spur meaningful clinical action.