Cell-Site Analysis - Forensic Data Science Laboratory

Research Grant Application (under development)

– Objectives

To develop relevant datasets and to develop and validate quantitative measurement and statistical modelling methods that will enable the adoption of the new paradigm for evaluation of evidence in a high-volume branch of forensic science in which so far it has made little progress: Cell-site analysis.

– Background

Current widespread practice in cell-site analysis is to go to the crime scene and survey which cells are connected to from that location, go to the alibi location and survey which cells are connected to from that location, then present as factual evidence the survey results and the call-data records (CDRs) detailing the IDs of the cells the telephone of interest connected to during the time period when the crime was committed. Interpretation is left to the legal-decision maker, or a technical expert expresses a posterior-probability-type opinion based on their subjective interpretation of the data. This practice, however, can be misleading. Legal-decision makers may assume that CDRs and raw survey results provide definitive evidence of the location of the telephone of interest, when, in fact, there are multiple factors contributing to variability in CDRs and in survey results.

A cell-site typically consists of multiple directional antennae (that may be mounted on a visible mast or may be hidden from view). Each antenna may support multiple channels, each with a different cell ID. The area served by a cell (or simply the “cell”) is the area within which a telephone could (at least potentially) connect to telecommunications services associated with a particular cell ID. Cells associated with different channels on the same antenna will overlap substantially. The edges of cells associated with different antenna on the same mast may overlap. Cells associated with different cell-sites may partially overlap. The area served by a cell will be affected by terrain and buildings that block or attenuate radio signals. In an urban environment a cell may extend only ~100 m from the cell-site and overlap with many other cells, but in a rural environment a cell may extend over tens of kilometres and have little overlap with other cells. In the latter case, CDRs would be much less informative for inference as to the exact location of the telephone than in the former case.

Which cell a telephone actually connects to is dependent on multiple factors, including service accessed by the telephone (voice, SMS, data), elevation in buildings, and direction of past travel of the telephone: telephones tend to stay connected to the cell they are already connected to, and the cell they are already connected to can influence which cell they connect to next. Different telephones accessing the same service from the same network at the same time from the same location will not necessarily select the same cell, and the same telephone accessing the same service from the same network at the same location but on different occasions will not necessarily select the same cell. The most common survey problem is failing to detect a cell that serves a location: Tart et al. (2021) reported miss rates ranging from more that 20% to less than 1% depending on survey equipment used, with some cell IDs missed by all equipment. If a cell serves both the crime scene and alibi location, and is missed at one but not the other, without proper interpretation the raw data would be highly misleading.

Expert interpretation of cell-site data is therefore required, but so far (with one exception) no statistical models have been developed for calculating evidential likelihood ratios from cell-site data. The exception (Bosma et al., 2020) addressed the question of whether two telephones were travelling together, depended only on comparison of pairs of simulated CDRs, and did not address absolute location of telephones.

The proposed research addresses the question of the location of a telephone during a given time period. This question is overwhelmingly the most commonly encountered in casework, and, to our knowledge, there are no existing statistical models for calculating evidential likelihood ratios addressing this question.

– Methodology

Pilot data collected in preparation for this research (and anonymized past casework survey data) will allow us to start development of statistical models soon after the beginning of the proposed research project.

We plan to adopt a Bayesian approach using a model (e.g., a beta-binomial model) for the probability of a telephone connecting to a cell from a location specified in terms of its distance and angle relative to the mast location and antenna direction (azimuth) for that cell, and updating the model using successively more case-relevant data. A Bayesian approach will prevent a miss in a casework survey from resulting in an inference that the probability of connecting to a cell of interest from a location of interest is zero.

Data used to update the model will be selected based on available relevant information, e.g., whether a cell of interest is in an urban or rural location, the terrain and location of buildings between the mast and location of interest (using map data and visual inspection), direction of travel of the telephone if known or hypothesized, service accessed. Finally we will update the model using survey data from the specific location of interest in a case.

The expected values of the posterior-probability distributions from two models, each associated with the same cell but with different locations (e.g., crime-scene location and alibi location), will be used to calculate a Bayes factor.

Models will be validated using protocols based on those that are now standard in forensic voice comparison.

	Forensic Data Science Laboratory
Home	Cell-Site Analysis