Interventions¶

Many of our Vivarium simulations are designed to measure the effects of a hypothetical intervention on a simulated population. Such interventions are described by an intervention model document. Like a cause model document, a risk exposure model document, or a risk effects model document, the goal of an intervention model document is to organize the complexity of the model so that others can understand it: The simulation researcher is responsible for communicating in writing how the intervention model should function, with sufficient detail so that (1) the engineer can implement the model in Vivarium; and (2) other researchers can understand what was done, including the strengths and weaknesses of the approach and how to verify and validate the results. This page describes in more detail what should go into an intervention model document.

General guidelines
Key questions to address early
The structure of an intervention model document
Examples of intervention model documents
Data sources for intervention models

General guidelines ¶

When writing an intervention model document (or any other component model document for Vivarium), you must include the following details to allow a software engineer (SWE) to implement the model:

Attributes of a simulant that are required for modeling the intervention (e.g. age, sex, and systolic blood pressure level for a model of anti-hypertensive medications);
Attributes of a simulant that will be added by this model component (e.g. current prescription, adherence status, last-measured SBP level, and untreated SBP level;
How to initialize the attributes added by this model when a new simulant is instantiated;
How to update the attributes (including both required and added attributes) during each simulation time-step.

We have developed a template for intervention model documents, which can help organize a complex model, and the remainder of this document describes what a researcher might include for each section when using this template. An example intervention model document that uses this template is the latent tuberculosis treatment intervention.

Key questions to address early ¶

There are many questions about an intervention that we need to answer early in the research process in collaboration with the client. This section lays out several such questions that will help clarify the goals of the project and ensure that basic aspects of an intervention are understood and written out. The goal is to give the SWE team a clear idea of expectations and to avoid the need for researchers and SWEs to go back and forth about clarifying these details throughout the project.

Research aims and objectives ¶

Write a one sentence aim — What does the client want to learn, specifically?
- Example: “To evaluate the effect of different interventions of population-wide nutritional fortification of staple foods on disability-adjusted life years in children under 5 in India, Nigeria, and Pakistan.”
Write a few SMART objectives, answering one question per objective. Think about: What is the client willing to ignore? Is GBD/Vivarium able to answer these questions well? How might we adapt our questions to questions that GBD can answer?
- Example objective 1: “To quantify the what-if scenario of increasing coverage of vitamin A from existing to X% on under 5 childhood mortality and morbidity.”
- Example objective 2: “To quantify the what-if scenario of increasing coverage of folic acid from existing to X% on under 5 childhood mortality and morbidity.”

Outcomes the intervention affects ¶

What outcomes does the intervention affect?

Effect sizes and definitions ¶

What are the exposure and control groups?
What is the definition of the effect size (prevalence or incidence or mortality, etc.)?
What units are the effect size(s) measured in?
Are the effects measured in shifts or RRs? Is this compatible with GBD risk factor distribution/definition? Can we convert or find alternative data source if not?
Do we need to alter our units on GBD side (using cross-walks) or intervention data source side?
NOTE: We need to understand the GBD structure for the relevant risk/cause enough to determine compatibility with our intervention effect sizes.

Potential confounders and mediators ¶

What are the potential confounders and mediators?
Might we need to control for them?

Limitations ¶

What are the limitations of our modeling strategy?
Are these acceptable to the client?

The structure of an intervention model document ¶

Intervention overview ¶

This section should include a brief description of the intervention to be modeled. It is often useful to distinguish between an “intervention technology” and an “intervention implementation”. For example, Vitamin A Supplementation (VAS) is an example of an intervention technology, and Scaling Up Coverage of VAS is an example of an intervention implementation. We often compare scenarios with alternative intervention implementations, to see how different strategies for implementing an intervention technology might result in different impacts on population health.

This section does not need to be long; one or two paragraphs with carefully selected links for a reader who needs to learn more will often suffice.

Baseline coverage ¶

An essential detail in any intervention model document is how to initialize the attributes added by the model when a new simulant is instantiated. In the case of a Vitamin A Supplementation model, this could take the form a single coverage percentage, applied independently to each simulant. Even this has some complexity, however, because the coverage percentage is not known precisely and its uncertainty might be quantified by some probability distribution; and it is sure to vary by location and over time.

In a more complex intervention model, the baseline coverage section would describe a relationship between the simulant’s attributes that are required from other models and the attributes that will be added by this model component. For example, an intervention model document for anti-hypertensive medications might specify a model for the joint probability that a simulant was at each point in their treatment ramp and was/was not adherent to their prescribed treatment, as a function of their age, sex, and measured SBP by providing the betas of a multinomial regression fit to NHANES data.

Vivarium Modeling Strategy ¶

This section will go into detail about how to represent the intervention technology and the intervention implementation in Vivarium. It can begin with a high-level summary of just one paragraph, e.g. “The treatment model links anti-hypertensive treatments to an additive shift in SBP level, which has a ripple effect on IHD and other causes which have DALYs due to SBP.”

Intervention effects ¶

This section must describe precisely what the intervention effects are and how the affected outcomes should be modified by the intervention. It is often complicated because there is some baseline level of treatment which is already present in the population-level estimates from GBD, and therefore some amount of “treatment deletion” is necessary. For example, if the population mean SBP (for a specific time, location, age, and sex) is 140 and 45% of the population is treated, then the intervention effect for untreated simulants would be to increase their SBP to counterbalance the intervention effect for treated simulants.

This is where any relationship between prescription, adherence, and outcomes can be described precisely, as well.

The treatment algorithm ¶

This section is where the researcher can describe how to update the attributes added by this model component. We have found that a decision tree can be a good way to communicate this with engineers and outside researchers. In a decision tree, each node represents a thing that might happen during a simulation time step (such as “visit clinic?” or “get SBP measured?”) and the arrows out of each node represent whether it did indeed happen, leading to a terminal node describing how the simulant’s attributes are changed because of the treatment. (Typically the changes are to the attributes added by this model, e.g. prescription and adherence, not the attributes required from other models.)

Example: somebody goes to the doctor (start of treatment algorithm) ; treatment changes or doesn’t change (end of treatment algorithm).

FlatIron

Published Literature

Cochrane Review