Visualization of the available data allows us to understand any additional transformations we'd have to consider before modeling the state of food deserts in NYC. To access the data, I made use of NYC's Open Data portal as well as third party websites to create my own datasets where necessary. I began by understanding the spatial distribution of foodstores in NYC, with use of a dataset containing information on Retail Foodstores throughout NY state[1]. This data was parsed, along with data on zipcodes pertaining to NYC and geospatial data to show the distribution of foodstores throughout the city. 

Categorizing stores

The label "food store" includes establishments such as bodegas (delis), small grocers, bakeries, supermarkets, and much more. As we are interested in understanding the number of both fresh-food markets and bodegas as a measure of fresh food availability, the first challenge is understanding which of these food stores can be categorized as either. A reasonable approach would be to consider the area of the location in question. To explore this, below I show the distribution in area of four different categories of stores: (1) those with the words "Deli" or "Grocery" in the name (Fig. 1, top left); (2) those with the words "Supermarket", "Market", or "Cooperative" in the name (Fig. 1, top right); (3) those with just the word "Supermarket" in the name; and (4) all food stores in NYC. I distribute the stores over 4 Square Footage bins, including 0-2,000 Sq. Ft., 2,000-5,000 Sq. Ft., 5,000-15,000 Sq. Ft., and above 15,000 Sq. Ft following an extensive study by Olivia Limone and Nadia Sanchez. It is generally true that bodegas/delis tend to be smaller in size than full-service supermarkets. However, a binary approach where we, e.g., "classify all food stores below a given square footage as delis and all others as fresh food markets" may misrepresent the number of stores in each category, due to the large disparity in the number of locations in different categories. For instance considering the area distribution of stores, I find that there are:

• ~3,700 bodegas below 2,000 Sq. Ft., suggesting that at least 600 (14%) bodegas may be misclassified as fresh food stores

• Only roughly 400 total stores which can definitively be classified as Supermarkets, 25% of which fall below 2,000 Sq. Ft. and may be misclassified as bodegas

• Roughly 700 (53%) potential fresh food stores may be misclassified as bodegas

• Roughly 100 (25%) supermarkets may be misclassified as bodegas

• If we consider all NYC food stores and classify by store size, we may expect a city-wide ratio for bodega:fresh-food-stores of roughly 2:1. However, considering names of stores as classification criteria instead, we may expect a city-wide bodega:fresh-food-store ratio of 3:1 or an even more dire bodega:supermarket ratio of roughly 10:1

The classification of food stores into fresh food stores or other is clearly a complicated problem, due to the limited data available from the NYC Open Data portal which is not necessarily informative for classification purposes. In this study, I will proceed with what I consider potentially more informative (words in the names of stores) as a way to classify stores. Full-service supermarkets will be categorized as stores which contain "Supermarket" in the name. Fresh food markets (which include supermarkets) will be categorized as stores which contain either "Supermarket", "Market", or "Cooperative" in the name; from visual inspection of the dataset this category also includes specialty markets that mainly produce one type of object (such as butchers, fish markets, and bakeries) as well as small grocers and mini-markets (delis that sell limited produce). Finally, bodegas will be categorized as stores which contain either "Deli" or "Grocery" in the name. I will present my analyses considering fresh food stores generally, but also just supermarkets as I believe the latter is potentially more informative in terms of food accessibility (I assume that, typically, people with limited time would prefer to visit a single full-service supermarket location to meet all of their food needs rather than visiting multiple smaller specialty markets).

[1] Retail Food Store Data: https://data.ny.gov/Economic-Development/Retail-Food-Stores/9a8c-vfzj/about_data

Fig. 1: Distributions of the sizes of stores in the dataset based on categrorization. I consider categorizing stores based on keywords in their names (blue figures), or by size alone (red figure). I find it optimal to categorize stores based on keywords in their names to avoid erroneous categorization.

Alternate Supermarket Dataset

For an alternative look at the of supermarkets in NYC, I cross-referenced the data filtered by the name criterion (see above) for categorizing supermarkets with data obtained from a third-party website which lists supermarket locations around the country: supermarketpage. To acquire the data from supermarketpage, I scraped each page with location information on NY state supermarkets. I then parsed the scraped html to extract the full street address of each location, including zipcode. I cross-referenced each extracted street address zipcodes with the NYC Open Data census tracts to find which locations are within NYC. Finally, I used Bing Maps API to query Bing maps for the geolocation of each full address. The result is show in the image below (Fig. 2, left). I find a total of 596 supermarkets throughout NYC using supermarketpage, which is close in number to the 414 supermarkets found using the name criterion filtering of data obtained using the NYC Open Data portal discussed above (see Fig. 1, bottom left). I also find that the geolocations of the supermarkets from supermarketpage coincide well with those from the NYC Open Data filtered dataset (see Fig. 2, right). However, each dataset appears to contain location information missing in the other. For example, the supermarketpage data shows that the aread directly south of Central Park is densely populated with supermarkets, whereas the NYC Open Data filtered data shows a relative dearth of supermarkets in that area. To cover all potential NYC supermarket locations given these two data sets, I combined both datasets and classified locations which are within 200 ft. of another location as a duplicate entry, to avoid overcounting as much as possible. The combined dataset of NYC supermarkets includes a total of 902 locations, resulting in a city-wide bodega:supermarket ratio of nearly 5:1.

Fig. 2: Geospatial locations of supermarkets from a third-party website (supermarketpage.com, left) and those in the NYC Open Data dataset (right). The full list of supermarkets is the combination of each of these with duplicates removed.

Other Relevant Quantities

One of several metrics I use to quantify the extent of food insecurity in NYC is physical access to fresh food. To account for New Yorkers that have access to vehicles and thereby supermarkets which are further away, I consider the estimated percentage of households with at least one vehicle available throughout the city using consumer market data from SimplyAnalytics. Below I show this as a heat map (Fig. 4, right) for The Bronx as an example. In order to account for population density and affordability, I also consider the estimated population and per capita income for each census tract in NYC using data from SimplyAnalytics.

Fig. 3: Heatmap of several quantities relevant to food accessibility and insecurity. I focus on The Bronx as an example. Shown from left to right are the percentage of households with at least 1 vehicle available, the population, and the per-capita income.

As I am also interested in understanding the public health outcomes related to food insecurity, I use census-tract level data available from the CDC on chronic diseases. These data show an epidemiological model-based prevalence of several chronic illnesses. Below I showcase an example of the geospatial distribution of some relevant illnesses (including High Blood Pressure, Diabetes, and Obesity).

Fig. 4: Heatmap of the prevalence of several illnesses that may be connected to food insecurity. Shown from left to right are the prevalence of high blood pressure, diabetes, and obesity.

With some understanding of the relevant quantities and how they change around the city, I decided to look for correlations between them to try to get a quantitative understanding of the extent of food insecurity in NYC. I quantify each relevant quantity using two metrics. The first metric assesses the nature of food accessibility in each census tract, by enumerating the number of supermarkets within walking and driving distance. The second metric assesses food expense, by considering the ratio of estimated food costs to per-capita income.

First Metric: food accessibility

In this study, I came across two terms commonly used when considering this problem:

• Food Desert: formally defined in this USDA study as a region with at least 500 people and/or 33% of the population with zero supermarkets within 1 mile. Given the dense population of NYC, and the difficulty of walking 1 mile for groceries, I instead consider 0.5 miles a walkable distance and the reference population as 1,000 people. I interpret this to mean that for a region in NYC to be considered a food desert, the population:accessible-supermarket ratio (i.e., the Food Desert Index) should be above 1,000:1. An area with a ratio above this threshold meets my criterion for being categorized as a food desert.

• Food Swamp: generally defined as urban environments where unhealthy food options (snack foods, fast food, etc.) outnumber fresh-food options.

To quantify the food accessibility throughout the city, I first discretize each NYC census tract into equidistant points on a grid. Gridpoints are separated by a distance of 0.15 miles, which I found to be a suitable resolution to ensure each census tract is sampled at least once. Below (Fig. 4) I show an example of what this discrete grid of points looks like for census tracts in The Bronx, with different gridpoint colors used for each census tract. To account for New Yorkers that have access to vehicles (and thereby supermarkets which are further away), I also consider the estimated percentage of households with at least one vehicle available (see Fig. 4, left).

Fig. 5: Discretized grid of sample points used to calculate the average quantitative indices considered in this study. I show The grid in The Bronx as an example. 

For each of the above sample points (Fig. 5), I ask:

• How many bodegas are within walking distance? (0.5 miles)

• How many supermarkets are within walking distance? (0.5 miles)

• How many supermarkets are within driving distance? (2 miles - typically a 10-15 minute drive in the busier areas of the city)

I then consider different metrics meant to assess the extent of each of these terms. To quantify the extent of food swamps and deserts in NYC, I calculate, for each census tract

where

I then consider the average value of the above metrics for each census tract which I show below (Fig.5). 

Fig. 6: Food Desert (left) and Food Swamp (right) Indices throughout the city. Higher values for each index signifies a higher chance of the region being a food desert or swamp. I highlight regions which may be classified as food deserts and swamps using arrows.

In the left panel of Fig. 5, regions which appear as pink meet the criterion for being categorized as food deserts (i.e., with Food Desert Index < 0.001). Areas which appear as near-white or green are borderline food deserts (i.e., with Food Desert Index ~ 0.001) or do not meet my criterion for being food deserts, respectively. I find that there are a significant number of areas around NYC which may be considered food deserts, with the following neighborhoods potentially housing them:

• Manhattan:

  • Clinton, Hell's Kitchen, and Chelsea

• Bronx:

  • Hunts Point, Soundview (partially), Eastchester, Co-op City (partially), Mott Haven, Highbridge, Throgs Neck (borderline food desert)

• Brooklyn:

  • small parts of East Flatbush, Sunset Park, and Windsor Terrace

• Queens:

  • Blissville, Long Island City (partially), South Corona

  • Howard Beach, Breezy Point, Far Rockaway (these neighborhoods are along shores of different kinds so my assessment may not be suitable here as the buffers I consider include water!)

  • Neighborhoods along the Easternmost part of Queens may be misrepresented as food deserts because I don't account for supermarket locations on nearby Long Island

• Staten Island:

  • Large swaths of Staten Island appear as food desert or borderline food deserts. However, closer inspection of these census tracts reveals that large parks, golf courses, and cemeteries cover the midsection of the island but are not categorized as such in the NYC Open Data portal. Despite this, there appear true food desert locations on Staten Island due to such few supermarket locations. Moreover, the Southern Tip of Staten Island does not include data from nearby supermarkets in Perth Amboy, NJ (a 15 minute drive away). For these reasons, I exclude SI from the remainder of the discussions.

In the right panel of Fig. 5, regions which appear as purple meet the criterion for being categorized as food swamps (i.e., with Food Swamp Index < 1). Areas which appear as near-white or green are borderline food swamps (i.e., with Food Desert Index ~ 1) or do not meet my criterion for being food deserts, respectively. There are larger swaths of NYC that may be categorized as food swamps, where snack and processed food options may significantly outnumber fresh food options. I identify the following areas of NYC as potentially housing food swamps:

• Almost the entirety of The Bronx (with the exception of the wealthier reaches in the Northern parts of the borough)

• all of Upper Manhattan

• Large parts of North Brooklyn and South Queens

Based on Fig. 5, Food Swamps may be a more relevant public health problem in NYC. Although there are a significant number of potential food deserts throughout the city, the area corresponding to food swamps is much larger. 

Second Metric: food expense

Another important factor for food access is food expense. To assess the expense of food throughout the city, I consider similar dimensionless parameters to the food desert and swamp indices, but instead considering per-capita income and food price data. Factors I consider include:

• Per-capita income data acquired from SimplyAnalytics

• Food price data acquired by scraping the data on this site: Northeast region food pricing from the Bureau of Labor Statistics

• According to the US Census Bureau, there are on average 2.56 members per household on average in NYC - I use this to determine the average cost of groceries per month. Assuming a cost of $400/month for a single person, this would come out to $1,024/month for a household. This estimate is higher than the national average, but lower than other estimates of grocery costs in NYC. I leave this estimate as a free parameter in my model to understand how modifying this affects my conclusions.

Given these assumptions, I calculate the food expense index by considering the ratio of monthly per-capita income to approximate monthly food costs:

where

Fig. 7: Food Expense Index throughout the city. Higher values of the index correspond to regions where estimated fresh food costs are a significant monthly financial burden.

Given the set of metrics considered in the "Building A Model" section above, I considered the potential correlations that exist in the data. In particular, I am interested in which census tracts have low combined food availability and food affordability, and whether auxiliary data can inform us about these census tracts. To begin this exploration, below I show how the Food Expense Index correlates with the Food Desert and Food Swamp indices.

Fig. 8: Relationship between different indices that contribute to food insecurity in NYC. Left: Relationship between the Food Expense Index and the Food Desert Index. Points that fall above the vertical line can be classified as food deserts. There is no significant correlation between the Food Desert and Food Expense indices, indicating that Food Deserts, if they exist, can be found across income ranges. Right: Relationship between the Food Expense Index and the Food Swamp Index. There is no significant correlation between the Food Swamp and Food Expense indices, indicating that Food Swamps are likelier to be found in regions where fresh food is a significant financial burden (i.e., low income regions).

The left panel of Fig. 7 shows the relationship between the Food Desert Index and the Food Affordability Index. I mark with a black line the criterion for a census tract to be considered a food desert. Anything above the black line qualifies as a food desert. The right panel note:

• Food Deserts are not signifincantly prevalent in NYC, with only ~10% of census tracts potentially qualifying as such.

• There is no significant correlation between food affordability in an area and the likeliness that the area is a food desert. Moreover, when they do appear, one can find them in both low- and high-income regions.

• However, there is a significant correlation between food affordability in an area and the likeliness that the area is a food swamp. Lower income areas where fresh food is a significant financial burden (typically ~30% the monthly income per person) also tend to be areas where the amount of unhealthy food options significantly outnumber fresh food options. This trend may be pointing to one of the realities of food insecurity for lower income residents of NYC. For lower income residents, fresh food is a significant financial burden and they tend to be surrounded by plentiful unhealthy and cheaper processed food options. These residents may be incentivized to spend their already limited income on these readily available, typically cheaper, unhealthy food options.

For a combined look at food insecurity - be it from accessibility (measured by the Food Desert Index), higher access to unhealthy foods (measured by the Food Swamp Index), or financial access to fresh food (measured by the Food Affordability Index) - I consider the combined, normalized food metric, defined as:

where each index is normalized to be at most equal to 1, such that 

• if the Food Desert Index is equal to 1, the area is very likely to be a food desert. The lower the value, the unlikelier it is to be a food desert.

• if the Food Swamp Index is equal to 1, the area is very likely to be a food swamp. The lower the value, the unlikelier it is to be a food swamp.

• if the Food Expense Index is equal to 1, food costs are a significant financial burden to the typical area resident. The lower the value, the lower the financial burden of fresh food.

I then consider how this metric varies around the city and how it correlates with quantities such as the per-capita income, access to vehicles, and rates of diet-related illnesses such as diabetes, heart disease, and cancer. Below, I show the geospatial distribution of the Food Insecurity Index. I highlight potentially troublesome areas with the Food Insecurity Index above 1, pointing to a significant amount of food insecurity stemming from either lack of access to fresh-food supermarkets, an overabundance of cheap unhealthy food options, or fresh-food costs being a significant financial burden (or all three). I label each of these areas using A-F for the sake of discussion.

Fig. 9: Left: Heatmap of the Food Insecurity Index, which accounts for fresh food access, prevalence of cheap and unhealthy food options, and financial burden of fresh food. Right: Currently open NYC FRESH stores (as of 06/22) (obtained from this site).

What is the impact of food insecurity on public health? 

To get an idea of how food insecurity may cause negative public health outcomes, I considered the correlation of the Food Insecurity Index with the prevalence of several chronic illnesses throughout the city. For each illness considered, I calculate the Pearson correlation coefficient, which gives a quantitative measure of the correlation between the disease prevalence and food insecurity. Below I show examples of the potential relationship between the prevalence of relevant illnesses and the food insecurity index. Above each figure I also show the Pearson correlation coefficient.

Fig. 10: Correlation between the prevalence of diet-related diseases and the Food Insecurity Index, shown in the title is the Pearson correlation for each case.

Below I also show a table showing the Pearson correlation coefficient between the Food Insecurity Index and the prevalence of other diseases. I highlight entries that show a positive correlation.

Tab. 1: Pearson correlation coefficient between the prevalence of chronic illnesses and food insecurity in NYC.

I typically find positive (albeit weak) correlations between the prevalence of certain diseases and the Food Insecurity Index. This suggests that NYC residents in areas with relatively higher food insecurity (as calculated by my model) are at higher risk of serious chronic illnesses. This presents a public health problem for the city that may be addressed by specifically targeting food insecurity. 

Clustering

Below I also use a kmeans clustering algorithm to consider potential clusters of census tract, categorized by their Food Insecurity Index, Per-capita Income, and Prevalence of several diet-related ailments. In the left panel of Fig. 11 I show an example of what these clusters look like, focusing on the prevalence of high blood pressure. I can identify the 3 following clusters:

1. Red: census tracts with relatively high food insecurity, relatively low income, and relatively high prevalence of diet-related ailments.

2. Light green: census tracts with relatively low food insecurity, relatively high income, and relatively high prevalence of diet-related ailments.

3. Dark green: somewhere in-between these two groups, with medium levels of food insecurity, income, and diet-related ailment prevalence.

In the right panel of Fig. 11 I show the spatial distribution of these clusters to examine where in the city we should foucs solutions to alleviate food insecurity.