Applied Statistics Case Studies

Working with Adjoined Consulting, Cox Associates has developed an accurate failure prediction model for quantifying the expected number and timing of equipment failures (with confidence intervals) and for deciding what specific preemptive equipment replacements can most reduce costs from failures.

HOW MANY NETWORK COMPONENTS WILL FAIL NEXT YEAR … AND HOW CAN FAILURE COSTS BE REDUCED?

Operators of large telecommunications networks typically have many aging, unreliable components and equipment assemblies, such as clocks for synchronizing signals and other equipment, installed throughout their networks. 

In 2005 and 2006, working with Adjoined Consulting, Cox Associates developed an accurate failure prediction model for quantifying the expected number and timing of equipment failures (with confidence intervals) and for deciding what specific preemptive equipment replacements can most reduce costs from failures.  The model is useful to both equipment and component manufacturers and to network operators, as it identifies cost-effective replacement schedules that lead to new sales opportunities for the manufacturers (for components no longer under warranty) while reducing replacement costs to the network operator.

DO GENETICALLY MODIFIED GRASSES THREATEN THE ENVIRONMENT, OR YOUR  LAWN?

In 2006, Cox Associates analyzed published gene flow data on the spread of pollen and  detection of glyphosate resistant seedlings at different distances and directions around an  experimental test area where resistant grasses. Based on the spatial density of compatible  recipient plants that might lead to resistant plants or hybrids, as well as on simple  mathematical models of gene diffusion in an established system such as a lawn or wilderness  area, we quantified upper bounds on the possible risk of resistance genes becoming  established at different distances remote from the experimental area. These risks are  typically extremely small (e.g., on the order of one chance in a hundred billion).

HOW CAN CONSTRUCTION DEFECTS BE ESTIMATED FROM LIMITED SAMPLES?

Construct defect litigation is often made more difficult by the fact that Plaintiff’s cannot  easily inspect all affected properties (e.g., because owners may not be available or may not  give permission when inspectors are available).  Defendants often suspect that the limited  sample information that can be obtained in practice may not provide a sound basis for  extrapolating defects to uninspected properties, especially if owners who give permission to  inspect tend to be those with visible construction defects.  In 2005, Cox Associates developed  new statisticalmethods for: (a)  Sampling for defects even when permission to inspect some  properties may be denied; and (b) Making valid statistical inferences about the extent of  defects from observed defect rates, despite practical limitations such as informative  censoring of samples based on owner permissions.  The new framework emphasizes making  correct decisions by conditioning on available information, rather than achieving  representative or random statistical samples.

WHEN WILL COPPER PIPES FAIL?

In 2006, Cox Associates analyzed failure time data on the dates of residential copper pipe failures due to thermogalvanic corrosion in the aggressive soil of a housing development. Based on these limited historical data, and by exploiting the fact that an exponential failure time distribution (constant failure rate) provides a lower bound on an entire class of non-decreasing failure rate processes, we were able to quantify plausible lower bounds on the cumulative number of future failures by each date over a 40-year forecast horozon

Selected Other Applied Statistics Projects

  • For the American Chemistry Council, quantified how well in vivo test results for rodent carcinogens could be predicted from much less expensive high-throughput in vivo test results. (2014-15)
  • For the American Chemistry Council, applied machine-learning algorithms to quantify how well in vivo test results for endocrine disruptors could be predicted from much less expensive high-throughput in vivo test results. (2013-14)
  • For TriZetto, assessed healthcare predictive analytics trends and vendor offerings, advised top management on predictive analytics technology acquisitions (2012)
  • For Rogers Communications, developed causal models of customer satisfaction; identified high-impact interventions for improving customer satisfaction; helped to develop achievable targets and strategies for improving customer experiences in different channels (2011)
  • Delivered a statistical analysis of the causal drivers of customer satisfaction to top executives at Comcast Cable; identified realistic targets and interventions for improving customer satisfaction (2010-11).
  • For a top cable company, worked in partnership with North Highland consulting company to deliver a predictive model that identifies which customers are most likely to drop accounts, well before the event and with much higher accuracy than previous models. (2008)
  • For an energy utility, worked in partnership with North Highland consulting company to deliver a predictive model of customer bad debt and account write-offs that greatly extended the lead time over which high-risk customers could be identified and targeted for intervention. (2007)
  • For a telecommunications company, worked in partnership with North Highland consulting company to develop a predictive model of customer marketing channel choice, and usage as a function of quality of channel experience (e.g., for web site, call center, retail store, and other channels.)   Used the model to quantify financial impacts of improving web-based customer care. (2006)
  • Also in partnership with North Highland consulting company, analyzed employee survey data for a major telecommunications provider and quantified patterns of internal communications (conference calls, managing e-mail, company news letters and bulletins, meetings, etc.); time spent on these activities by employees with different job roles and in different VP areas; and potential to reduce employee burden and improve the value and efficiency of internal communications. (2006)
  • For a European wireless telecommunications provider, analyzed customer data to help develop more predictive segments; held a one-day intensive course in Brussels on advanced statistical models and methods for quantifying customer value in the short and long runs, based on probability and statistics models of customer behaviors in response to company offers.   (2005)