Motivation

“The poor are the ones who can never afford to have any bad luck. They can’t get an infection because they don’t have access to any medicine. They can’t get sick or miss their bus or get injured because they will lose their …job if they don’t show up for work. They can’t misplace their pocket change because it’s actually the only money they have left for food…Of course the bad news is, everybody has bad luck. It’s just that most of us have margins of resources and access to support that allow us to weather the storm…”

The Locust Effect by Haugen and Butros

In theory, addressing hunger seems simple…

but the civic infrastructure is complicated by many different programs and circumstances.

So we focused on one program…

• Summer Meals Program operated by United States Department of Agriculture
• Meals are served free to children and reimbursed by the state
• Local organizations serve meals to children
• The program is not used at capacity because organizations often can’t break even

with another deceptively simple solution to hunger.

But bad luck (uncertainty) affects summer meals too

Enter Decision Analysis with Bayesian Networks

• Purpose: Can be used to determine the coarse grain policy models (DBNs for decision support and sugar food insecurity Barons, Zhong, and Smith)
• Structure: Nodes (chance, decision, and utility nodes), and edges
• R package HydeNet for implementation of hybrid Bayesian networks using rjags

Constructing the Bayesian Network

• Nodes elicited from program administrators, state officials, local organizations, community advocates, and families at-risk of summer hunger
• Edges elicited from domain experts by populating and averaging an adjacency matrix to show causal realtionships
• Discrete Conditional Probability Tables were compiled from survey data, publicly available data, and expert elicitation workshops.

The Network

plot(summer_net)

Set the Decision Nodes

Decision nodes are all the program operations an organization could change.

summer_net<-setDecisionNodes(summer_net, Menu.Quality, Public.Transport, Media)
plot(summer_net)

Set the Utility Nodes

Total Meals served measures utility for this particular problem

summer_net<-setDecisionNodes(summer_net, Menu.Quality, Public.Transport, Media)
summer_net<-setUtilityNodes(summer_net,Total.Meals)
plot(summer_net)

Explore different policies

summer_net<-setDecisionNodes(summer_net, Menu.Quality, Public.Transport, Media)
summer_net<-setUtilityNodes(summer_net,Total.Meals)
trackedVars <- "Total.Meals"
        policies=policyMatrix(summer_net)
        invisible(compiledNets <- compileDecisionModel(summer_net, policyMatrix = policies))

## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 0
##    Unobserved stochastic nodes: 15
##    Total graph size: 194
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model
## 
## Compiling model graph
##    Resolving undeclared variables
##    Allocating nodes
## Graph information:
##    Observed stochastic nodes: 3
##    Unobserved stochastic nodes: 12
##    Total graph size: 178
## 
## Initializing model

print(policies)

##   Public.Transport Menu.Quality Media
## 1                1            1     1
## 2                2            1     1
## 3                1            2     1
## 4                2            2     1
## 5                1            1     2
## 6                2            1     2
## 7                1            2     2
## 8                2            2     2

Now sample and find maximum expected utility

The maximum expected utility indicates the best policy for the organization to adopt.

samples <- lapply(compiledNets,
                          HydePosterior,
                          variable.names = trackedVars,
                          n.iter=10000, trace=F)
        lapply(samples, function(l) mean(as.numeric(l$Total.Meals)))

## [[1]]
## [1] 1.5744
## 
## [[2]]
## [1] 1.5946
## 
## [[3]]
## [1] 1.595
## 
## [[4]]
## [1] 1.6061
## 
## [[5]]
## [1] 1.5335
## 
## [[6]]
## [1] 1.5501
## 
## [[7]]
## [1] 1.5329
## 
## [[8]]
## [1] 1.5423

In this instance the maximum expected utility corresponds to available public transport, high meal quality, and low media.

Communicating to Stakeholders

• Shiny site enables policymakers and organization leaders to test different programmatic changes for both urban and rural locations.
• Interactive website well received at both federal and state levels.
• Shiny site supplemented with site on local hunger data to provide resources for community organizations and advocates
• Government is currently considering funds for additional transportation funds or heat waivers.

All models are wrong, but some are useful.

George Box

Acknowledgements

• Texas Hunger Initiative staff for facilitating data collection
• The University of Warwick Food Insecurity Group for inspiration
• Walmart Foundation for funding
• R community for troubleshooting and support

Contact Information

• Texas Hunger Initiative website
• More about food insecurity and R on my personal website