Study Design and Method options

Categories → Study Design , Causal Wizard Concept , Data

The Design and Method of your Study defines the methodology and approach which will be used.

Introduction

Designing an observational, causal experiment requires you to make a number of choices. Causal Wizard simplifies this to two decisions:

• Method (or Methodology)
• Design

In short, Method allows you to choose the Causal Inference methodology which will be used, including the way you will provide your domain knowledge, the way Causal Effects are identified and the types of Causal Machine Learning models available.

The Design question is about how your data is structured, and whether you have a binary treatment design (e.g. treated and control groups), or a continuous treatment.

Both aspects are discussed below. You can change these decisions in the edit Study page.

Method(-ology)

We now offer two methodologies. The default methodology is Causal Diagram & Potential Outcomes. We recently added popular Econometric methods for Panel data, including Difference-in-Differences. These may be preferred (or used for comparison) if your data is compatible with Panel data format. All methods support use of observational data.

Causal Diagram & Potential Outcomes methodology

This is the default option, and should be selected if you want to draw and analyse a Causal Diagram. This methodology follows the DoWhy paradigm, which combines Graphical Causal Models (GCMs) and the Potential Outcomes framework.

• Draw and analyse Causal Diagrams with or without data.
• Binary treatment (case / control) study design.
• Explicit causal model defined by expert knowledge.
• Can estimate a range of Causal Effects, including Average Treatment Effect (ATE).
• Variety of Estimators, including regression models and Propensity Score methods.
• Does not explicitly model time; recommend present data as snapshots of pre and post-treatment per sample-unit. See also Datasets article.

Panel Data methods

Panel data format is commonly used when a group of entities are observed over two or more time-periods. Therefore, the Difference-in-Differences study design is a special case of Panel Data where there are only two periods (before and after intervention). Numerous methods have been developed to suit this data format.

• Treatment can be continuous / numerical or binary (case / control).
• One or Two-Way Fixed Effects (TWFE) models (Entity groups and / or Time).
• De-meaning approach for fixed effects, allowing a larger number of categorical values.
• Allows additional covariates.
• Estimates Average Treatment Effect on the Treated (ATT) using Ordinary Least Squares (OLS) regression.
• Suitable for Difference-in-Differences study design.
• Does not require Causal Diagram, but requires data.
• Not appropriate when time-varying confounders are present.

Study Design

You can select either a Binary treatment (case / control study design), or a continuous numerical treatment. The choice is usually determined by the data type of your treatment variable.

Binary (Case / Control) Treatment

• Use when your Treatment values are catgorical e.g. 0/1, True/False, or some other values which can be divided into Control and Treated groups.
• Can also be used when your Treatment values are continuous / numerical, but you want to use a threshold value to separate them into Treated and Control groups.
• Must be selected when Method is Causal-Diagram & Potential Outcomes, to allow group comparison.

Continuous, numerical Treatment

• Use when your Treatment variable has a continuous, numerical data type.
• Can only be used with Panel data methods, such as fixed-effects models.