Intervention Models¶

Introduction¶

Intervention modeling is intended to provide analysts concerned with potential intervention programming a guide to assess the efficacy and value of a range of potential options. The mechanism by which intervention modeling will do so is to provide “what-if” input scenarios and calculate the likely impact, as well as providing a rough order of magnitude cost estimate based on historical evidence of similar interventions. A fully functioning intervention modeling user experience will allow for cost or impact optimizations, A/B testing, and other similar experiments on potential intervention portfolio optimization without extensive training on how complex modeling works.

The intervention modeling helps customers to answer critical strategic questions of how best to prioritize resources. What interventions should be prioritized to achieve which impacts? In a single intervention use-case, the customer questions are:

• What is the likely cost of my planned intervention?

• What is the likely impact of my intervention on an impact indicator of interest? For example, how will the intervention change poverty rates, malnutrition rates or the risk of armed conflict in the project areas?

• What is the maximum scale of intervention I can achieve for a given budget?

As the intervention model develops to allow multiple interventions, we can help the customer to answer key strategic questions such as:

• Which combination of interventions will achieve the maximum impact for a given total budget? For example, what is the maximum reduction in poverty that could be achieved over a given time and space, with a US$100m budget and combining infrastructure investments, pro-poor cash transfers and other poverty alleviation interventions?

• What is the cost of achieving a given level of impact, using multiple intervention pathways? For example, what would be the dollar cost of a 10 percent reduction in malnutrition rates using the best available combination of feeding, deworming and water and sanitation interventions?

The intervention model also allows the customer to examine the costs and benefits of tactical and operational decisions. These are the more granular decisions on where and when to intervene. They are particularly relevant to interventions that require a supply chain, typically, the “provide class” interventions. Tactical and operational decisions include:

• Where should critical distribution points or points of service be located? For example, which health centers should be used for delivering vaccines?

• How should resources be allocated across space and time? For example, how should cash transfers be targeted between geographies and households?

• Which means of transport and routing should be taken? For example, what is the least cost feasible transport solution for delivering food assistance to all the required points of service?

Key Terminology¶

It is useful to distinguish between intervention modeling and intervention models, as the former is a workflow and user experience, while the latter is a specific capability:

• Intervention modeling is the approach of using intervention models to create “what-if” scenarios, (often including the subsequent use of additional domain models) to estimate the net effects.

• Intervention model is a light-weight estimation tool that allows a user to specify an intervention type and scale and calculate the immediate impact and estimated cost.

Furthermore, several other terms are used to describe the operation of intervention models that are useful to define upfront:

• Intervention class is a category of intervention types that share a common mode of operation. Intervention classes are the “verb” component of an intervention sentence. Examples include “provide,” “build,” “train,” etc.

• Intervention type is a subset of interventions within the class that share a noun, which is the object of the verb. Examples of interventions include “provide food,” “build roads,” or “train on mine-awareness” etc.

Reference Workflows¶

There is a standard workflow that intervention models use, illustrated in the following diagram.

Intervention Class Structure¶

One of the central principles in Kimetrica’s thinking about intervention modeling is that there is a limited number of forms that an intervention can take, and interventions within a form are fundamentally similar to each other. This implies that there are a limited number of intervention classes, and within each class there are shared cost and impact models. For each intervention type that is covered by the class, these cost and impact models are run using parameterized values drawn from historical data, or estimated by a user.

The following are the classes that we expect to cover the vast majority of humanitarian and development interventions:

• Provide The provide class is intended to capture interventions of the form where one party is granting another party some material object. This can take the form of cash handouts, food distribution, vaccine campaigns, and other similar efforts.

• Build infrastructure link The build infrastructure link class describes efforts to create connective infrastructure such as roads, internet connections, power lines, etc.

• Build infrastructure node The build infrastructure node class describes efforts to create key infrastructure facilities (public buildings, data centers, distribution hubs).

• Teach The teach intervention class describes efforts to share knowledge and practices to improve the efficacy and efficiency of other activities.

• Secure The secure intervention describes efforts to harden infrastructure links and nodes against conflict and human-caused disruption.

• Subsidize The subsidize intervention artificially lowers the cost of an intervention or any other cost-aware models to artificially lower the expected cost. Can be combined with other intervention models to make a “lower cost if” effect.

Additionally, there are modifiers to interventions that combine concepts on a “if/else” basis:

• Incentivize (Do if) The incentivize modifier creates a conditional statement around an intervention, such that an intervention is executed only if some condition is met.

Cost Models¶

Intervention models must contain a price estimate that accounts for intervention costs. This cost model must account for efficiencies of scale in both time, volume, and geography, must expose the key variables to define the scale of an intervention and must allow for scaling parameters that are tied to the key variables. To execute a cost model, the user must provide the intervention location, duration, and scale parameters (either user-entered or looked up).

A simplistic example of a general cost model would be the following, where P is a scaling parameter, and x is an input scale variable. In this toy example, for each key variable, the variable is multiplied by a scale factor and summed to get a cost.

The actual cost equations for each intervention are much more refined of course, accounting for efficiencies and taking input specific to that intervention class. The provide intervention class for example uses a logarithmic-linear approach, with modifications for time decay. Additional considerations include geography (for corruption), and date (for relative purchasing power).

Once this cost model has been created, historical data needs to be captured that provides the net cost and key variables required to run a regression and estimate the scaling parameters for that intervention type. A successful cost model will be shared across the entire intervention class, and each intervention type can be described using a set of scaling parameters that are accurate for estimating that intervention at both small and large scales.

Impact Models¶

The impact model component of the intervention models follows the same logic as the cost model, to take parameters that describe the intervention type, combine them with the user-provided scale parameters, and create a resulting dataset to represent a perturbed key indicator. A key difference is that the impact model also requires a starting dataset that represents the state of the key indicator variable before the intervention. The impact model output is thus the same indicator dataset, with changes of value that reflect the intervention in question. It is critically important that the impact model maintains dataset formatting, only changing the content, as this is the feature that allows intervention models to be used immediately prior to a downstream domain model.