Food Protection Trends - January/February 2020 - 57

One model had some of the predictor variables in the earlier
models already cited but also had some important differences, including extensive use of existing inspection records. The
other model went in a different direction, without having any
analogies to risk.
Classification and Regression Trees (CART)
A project in Columbus, Ohio, proposed using Classification and Regression Trees (CART) (1) software with
food safety inspection data provided by the City to develop a new predictive model. Early applications of such a
"tree-structured approach" to data analysis (1) included a
project in which an airplane flying in large circles around
ships of six different structural types produced continuous fluctuating radar signals that were used to classify the
ships. Another application was analysis of mass spectra of
airborne contaminants to characterize specific pollutants.
Both of these applications are actually examples of artificial
intelligence: machine learning without specification of the
variables to use. The class intervals of the continuous random variables were selected by the computer software.
This graduate student project (8) used the aforementioned
automated nonparametric statistical methodology in 1992
in an attempt to show that the outcome of the most recent
inspection of a food service operation could be predicted
through use of the preceding series of inspections between
1986 and 1990. The outcomes of interest were time-tempera-

ture violations or inspection failure (a score below 90 out
of 100 and/or a "critical violation," one capable of causing
an outbreak of foodborne illness). This was similar to the
previous examples in that a variety of independent variables
(Table 1) were available, but the computer selected the ones
with the most predictive power.
The results of one classification tree are shown in Fig. 1. In
a "learning sample" of 1,000 full-menu restaurants, 528 had
failed a standard inspection at least once in 5 years. CART
made its first split with the question "Was the standard
deviation of scores in the previous year above 1.95?" 398
of the 573 establishments for which the answer was "yes"
had failed at least once. The bar at the right end of the box
indicates that it is a "terminal node," and it was not split
further. Of the 427 for which the first answer was "no," 159
establishments that did not have an extra inspection included
86 failures and could not be split further. A split of the 268
establishments that had had an extra inspection identified
41 failures by asking whether the average interval between
inspections was 241 days or more. Finally, of those 41, 36
had had 2 or 3 extra inspections in the previous year.
The results of the analyses generally make sense. Restaurants with a variable score are not under effective control and
therefore could be expected to have problems. If not getting
an extra inspection predicts failure, the inspections are generally doing what they're supposed to be doing. Also, perhaps
going without an inspection for 241 days has a protective

TABLE 1. Potential predictor variables; adapted from references 8 and 22
AVAILABLE TO CART IN COLUMBUS

USED IN CHICAGO FORECASTING MODEL

Average number of days between inspections, prior year

Length of time since last inspection

Standard deviation of scores, prior year

Three-day average high temperature

Number of extra inspections, prior year

Nearby garbage and sanitation complaints

Days since last regular inspection
Average duration in minutes, prior year

Has a tobacco or alcohol consumption

Average income in zip code

Nearby burglaries

Type of operation (food service or retail food establishment)

Type of facility

Vending location or regular

Length of time operation has been operating

Frequency of the outcome of interest, prior year
Any food safety violation, prior year

Had a previous critical or serious violation

Commercial versus non-commercial
Any violation, any year
Number of previous violations
Purpose of the index inspection

Inspector assigned

January/February Food Protection Trends

Food Protection Trends - January/February 2020

Table of Contents for the Digital Edition of Food Protection Trends - January/February 2020