As noted previously, timeliness and accuracy are the two of most important characteristics when we evaluate an infectious disease surveillance system. Our results show that when an infectious disease required a time-consuming process for diagnosis, such as the dengue fever using the previously mentioned protocol, the actual daily number of infected cases and cumulative positive cases are potentially underestimated. The proposed method dynamically updates the parameters daily by making use of the most recently available information on suspect cases, and then performed estimates with a lower absolute relative bias than when using observed daily lab-confirmed cases only. As shown in Table 1, the proposed method performed a lower median absolute relative bias (ABS range 1.7% ~ 8.5%) than those solely based on daily confirmed cases (ABS range 4.7% ~ 67.7%) between July 6 and December 31. These dates covered the rising stage and around the peak stage which were of public health interest. The proposed method provides a more accurate estimate of the epidemic curves when applied to the dengue fever dataset for Taiwan during the 2006-2007 season. Based on these results, this approach can be used for the real-time evaluation of the severity of a disease outbreak when case classification requires that a confirmed case involves a time-consuming process.

In this study, we first established the different distributions for the onset-to-confirmation time of the positive cases and negative cases. Next, either a gamma distribution was assumed in order to estimate the probability of being a confirmed case given cases status in equation (1), or, alternatively, a nonparametric approach was used. We actually experimented with several types of distribution. The estimates using a log-normal distribution were numerically very similar to the results for the gamma distribution. The estimates using a Weibull distribution did not perform as well as the gamma distribution applied in our dengue fever data. From Figure 4, we learn that the shape parameter *α* changed from 0.5 to 2 and therefore an exponential distribution may not be appropriate. For simplification, we have chosen to present only the results from the gamma distribution as one example of a parametric approach and compare this with a nonparametric approach. As shown in Figure 2 for daily new cases, the differences in the estimates based on parametric approach with Gamma distribution and those with nonparametric approaches were minor. The Figure 3 and Table 1 for cumulative cases showed that a gamma distribution is a more appropriate assumption for the onset-to-diagnosis time when estimating the probability of being a positive case using the dengue fever example; nonetheless, the difference between the gamma and the nonparametric method is again only slight except towards the end stage of the epidemic after January 1. The reason that the nonparametric method did not work well after January 1 is because *P* (*Y*
_{
i
}= 1), *P* (*Y*
_{
i
}= 0), and *P* (*T* | *y*
_{
i
}) had not changed substantially, resulting in a near constant estimate of the daily positive cases.

In practice, any form of the probability of being a positive case can be assumed. It is also not restricted to certain distributions when the models are adapted to different types of infectious disease. When applying this approach to other diseases, researchers should investigate several distributions according to the shape of their data and choose an appropriate one based on some appropriate measures, for instance, those shown in Table 1.

Our method estimated the probability of being a positive case based on the data within a 1-year "moving window" before date *c* and updated *P*
_{
i
}
*(c)* and *E*
_{
i
}
*(c)* everyday. The epidemic profiles of dengue fever are different from one year to another in Taiwan. Choosing the data from most recent one year was done in order to insure that there was enough information to cover a whole epidemic season. In the early stage of the 2006-2007 season, the data from the 2005-2006 season actually contributed more to estimating the daily cases counts. In this study, even the epidemic profiles were not necessarily the same between the 2005-2006 and 2006-2007 seasons, the proposed methods performed well.

The study shows that before the first positive case appeared on July 6, the proposed method did not work well and are not that useful (Table 1). Our method worked well after the first positive case appeared during the 2006-2007 season. Indeed, it needed only four days to be able to consistently estimate the final status curve. In the 2006-2007 season, Taiwan CDC activated a central command center for intensively dengue epidemic control on October 2. The task of this command center included expanded blood sample collection and it is likely that this resulted in more suspect cases for laboratory confirmation, which might have led to a lower proportion of positive cases. This would influence the estimation of probability of being a positive case over the following few days. As we can see on Figure 4, it also influenced the estimation of the parameters for negative cases. While our manuscript was being prepared, the Taiwan CDC changed their laboratory protocol for dengue fever to one that requires only a single laboratory test for dengue surveillance and control. The result is a substantial reduction in the waiting time for laboratory confirmation. However, confirmation time can never be completely avoided with dengue fever. A situation where a large number of serum specimens are sent for diagnosis at the same time will result in overloading at the laboratory, which might increase the confirmation waiting time. As described previously, the estimation used information based on a "moving window" time period before the estimated date and the parameters of the model are updated everyday. Since the observed confirmed cases counts on date *c* are always underestimated as long as there is a time lag, our method potentially can be applied while waiting for further investigation of the status of cases.

There are some limitations to our method. Firstly, the approach needs sufficient historical data to be available in order estimate the parameters of the model; therefore our model cannot be applied effectively to an emerging disease, such as SARS or avian flu. Secondly, we used confirmed cases, the dates of onset of which were within 1-year before the date estimated and if a case needs more than 1-year for diagnosis such a case might never provide any information to the parameter estimation; in such a circumstance a different "moving window" needs to be chosen. Thirdly, when missing diagnosis dates exist, the estimated curve using the nonparametric method cannot converge with the final status curve. There were 134 and 289 cases missing confirmation results for the 2005-2006 season and the 2006-2007 season, respectively, at the time that the manuscript was prepared. The nonparametric method estimates by plugging in the cumulative proportion of confirmed data given the final status. As we mention before, at the end stage of epidemic, the probabilities in equation(2) almost remained unchanged. In this study, we could only assume that the proportion of positive cases out of all suspect cases among the missing observations were similar to those having results, which basically assumes that the missing data were missing at random thus ignorable.