Food Protection Trends - July/August 2020 - 234

densities such as Toronto, they are relatively compact and
socioeconomically homogenous (34). This makes collected
census data representative and comparable among tracts.
Each tract has a unique identifier, allowing for other tabular
data to be joined by corresponding identifiers. The GIS
program ArcMap 10.6.1, from the Environmental Systems
Research Institute (ESRI, Redlands, CA, United States), was
used to consolidate and map data.
Census tract and public health unit digital boundary
files were obtained from Statistics Canada (35, 36). Any
census tracts falling outside the boundaries of Toronto
Public Health Unit (which follows the same boundaries
of the City of Toronto) were removed. The locations and
outcome types of each inspection were shown visually by
importing each year's outcomes as separate tables to ArcMap
and then geocoding the coordinates as separate map layers.
The number of first annual inspection outcomes per census
tract was then counted by spatially joining the restaurant
coordinates of each inspection outcome to its respective
tract. Since the number of inspections with closed results
was very low (2017: n = 2; 2018: n = 10), the results of
these were combined with the conditional pass results for all
further analysis. The percentage of CP/C results per census
tract was calculated for each year using the total count of first
inspections for each census tract as the denominator.
To visually display the rates of CP/C per census tract
across Toronto, two maps were created for each study year
(2017 and 2018). The first set of maps was created by using
equal interval classification and the second set by using the
Jenks natural breaks algorithm. Equal interval classification
was used to create five equally defined classes to facilitate
relative comparisons across years and to other maps, and
the Jenks natural breaks algorithm was used to calculate five
classes that best visually represent the data within each map.
The Jenks natural breaks calculation divides continuous data
into groupings that are most similar within groups, while
maximizing differences across groups. This allows for a more
meaningful representation of data values that are not evenly
distributed (10, 11).
Spatial analysis
The Global Moran's I statistic was used to measure the
overall spatial pattern of how higher and lower rates of CP/C
results were distributed by census tract. This statistic uses
the set of features and the associated attribute to calculate
whether the distribution pattern is clustered, dispersed (high
and low values evenly spread), or random (13). Moran's
index value is calculated, along with a z-score and P-value,
to identify the significance of the index. A significant and
positive index value suggests that adjacent observations are
highly likely to be similar.
To test for and identify local clustering, Getis-Ord Gi*,
also known as hot spot analysis, was used. This statistic
indicates whether a feature is a significant hot spot by

234

Food Protection Trends July/August

proportionally comparing the local sum of the target feature
and its surrounding features to the sum of all features (28).
If the difference between the observed local sum and the
expected local sum is too large to be due to chance alone, a
statistically significant z-score is obtained.
Regression modeling
Negative binomial regression analysis was conducted
to detect associations between selected census-tract
demographic variables and the rate of CP/C results per
census tract (24). The outcome for these models was the
number of CP/C instances per number of first restaurant
inspections across the two study years. Prior to this analysis,
census tracts with zero restaurants in either study year were
excluded, resulting in a sample size of 469. Negative binomial
models were calculated instead of Poisson, as preliminary
modeling indicated significant over-dispersion.
Census 2016 demographic data collected at the census
tract level was obtained from the City of Toronto (6). The
following demographic variables were selected for evaluation
in a series of bivariate models: population density per
square kilometer; median age; prevalence of low income
households. based on the low-income measure, after tax
(LIM-AT); prevalence of immigration status households;
prevalence of households using a nonofficial Canadian
language (i.e., other than English or French) most often at
home; and prevalence of labor households working in the
accommodation and food service sectors (North American
Industry Classification System, NAICS, 72). The LIMAT variable, calculated by Statistics Canada, considers
households to be low-income status if they fall significantly
below the median income of all households in Canada. A
combination of continuous P-values and 95% confidence
intervals were used to assess the statistical and practical
significance of each relationship. Regression coefficients were
expressed as incidence rate ratios (IRR). Given that this was
an exploratory analysis, we did not construct a multivariable
model. Pearson correlation coefficients were also calculated
between each pair of demographic variables. Regression
modeling was conducted using Stata IC (Version 14.2).
RESULTS
Descriptive analysis of inspection results
Of the first annual inspections performed in 2017 (n =
5,950), 92.6% resulted in a 'Pass,' 7.4% resulted in a
'Conditional Pass,' and 0.03% resulted in a 'Closed' result. Of
the first annual inspections performed in 2018 (n = 6,457),
91.5% resulted in a 'Pass,' 8.4% resulted in a 'Conditional
Pass,' and 0.15% resulted in a 'Closed' result.
Of the 572 census tracts located in the Toronto Public
Health Unit, 93 (16.3%) did not contain any inspected
restaurants in 2017, and 97 (17.0%) did not contain any
inspected restaurants in 2018. Among the 469 census tracts
with at least one inspected restaurant in both study years,

Food Protection Trends - July/August 2020

Table of Contents for the Digital Edition of Food Protection Trends - July/August 2020