Fitting Incident Time to a Distribution

Incident times follow a von Mises distribution centered near 5:30pm.

Introduction

One goal of the SARBayes project is to forecast the probability of survival for lost persons. Such models could be useful in deciding to continue searching, and researchers making motion models can use survival predictions when generating probability maps of the lost person's location. We are analyzing data from the International Search & Rescue Incident Database (ISRID) to describe the probability of survival as a function of various features, such as age or temperature.

Intuitively, a lost person should stand a better chance of survival when his or her absence is noticed and reported early. Therefore, one feature we have chosen to examine is the time at which an incident is reported—the incident time. Here, we present how we modeled the frequency of incidents as a function of the time of day.

Methods

The Data

In ISRID, the date and time of the incident are recorded together in the incident date field. When the time is missing, the spreadsheet library reading the field provides a default time of midnight, 12:00 AM. Therefore, we confined our analysis to the \(\) instances of incident time that were not at midnight. When these times are organized into half-hour bins, we can calculate the frequency of each bin and display the data in a histogram.

time-distribution

The incident frequency, as shown above, roughly follows a bell-shaped distribution. As expected, fewer incidents are reported in the early hours of the morning, and a surge of incidents occurs late in the afternoon. The surge appears to be centered around hour 17 or 18—roughly 5:30 PM.

Parameter Estimation

Because incident time wraps around (e.g. hour 0 is the same as hour 24), we chose to fit incident time to the circular von Mises distribution, which is on \(\). This section summarizes how to fit the parameters of our von Mises distribution. Readers willing to take that step on faith may safely skip to the Results section.

The two parameters of a von Mises distribution are \(\), a real number describing the mean, and \(\), a positive real number describing the distribution's concentration. As \(\) increases, the values cluster more tightly around \(\). We converted each time on the \(\) hour range to radians on \(\), and stored the results in a list \(\). If we think of each time as a complex number, we can compute an average, \(\), of the times.

\[\]

According to the von Mises Wikipedia article, the argument of \(\) is a biased estimator of \(\).

\[\]

Using another equation from the article, the expectation value of the square of the modulus of \(\) is

\[\]

where \(\) is the modified Bessel function of order \(\). To estimate \(\), we simply iterated in increments of \(\) to find a value that best satisfied

\[\]

Results

We found \(\), \(\), and \(\). When converted back to the time of day, \(\) is equivalent to 5:20 PM. Using these parameters, we also sampled \(\) values from the distribution. The frequency of the these values are shown below in another histogram. Although the curve is smoother, this fitted distribution approximates the actual data fairly well.

time-distribution_fitted

Conclusion

The vast majority of the 75,000-some cases in our copy of ISRID do not have an incident time. Sampling a distribution with the parameters found here could be used to fill in missing values for incident time, if doing so was found to improve a survival model (another article).  Probability maps can also take the incident time distribution into account, as lost person behavior may depend on the time of day and how much sunlight is available.

Author: Jonathan Lee

Jonathan is a senior at Thomas Jefferson High School for Science and Technology in Northern Virginia. He joined the SARBayes project in the summer of 2014 and has worked on the MapScore website and survival modeling.

One thought on “Fitting Incident Time to a Distribution”

Comments are closed.