Frequently Asked Questions (FAQ)

What is Causal Wizard?

Causal Wizard is an app to help you detect and measure causal effects - that is, changes in a variable caused by another variable. Using a causal model, you can more accurately predict counterfactual or hypothetical “what-if” scenarios. Most common statistical tools only measure association, or correlation. This often gives misleading results, especially when trying to understand the effects of an intervention (i.e. changing something).

For example, eating ice-cream is correlated with sunshine. But does eating ice-cream cause the sun to shine, or does sunshine cause people to eat ice-cream? Or maybe both sunshine and ice-cream sales are both caused by another variable.

Eating ice-cream on a hot day at the beach

By adding your domain knowledge of the direction of causation, the correct results can be predicted. We predict more ice-cream sales on hot days, but also that increasing ice-cream consumption would have no effect on the weather.

Does sunshine cause ice-cream or does ice-cream cause sunshine. Arrows indicate the direction of causation.

How does it help me?

Causal Wizard:

  • Helps you to measure causal effects in your existing, historical data
  • Recognises and incorporates your subject-matter expertise into its Machine-Learning models
  • Allows you to more accurately predict the outcomes of interventions (i.e. what happens if you change something)
  • Can confirm or refute the existence of a significant causal effect, allowing you to exploit this new knowledge in further work
While the methods used in Causal Wizard are not new, they are usually only used by statisticians, economists and data scientists. Causal Wizard makes these methods available to everyone, guiding you step-by-step. The app tries to avoid using maths terminology where possible, or explains it in context. You can find explanations of everything in our help articles.

Who can use Causal Wizard?

You can use Causal Wizard. We find the tool is particularly useful for people who manage complex systems, such as asset managers, product owners, researchers and manufacturers. Asset managers might want to understand the causes of asset failures. Product owners will want to understand what drives purchasing decisions. And researchers want to understand … how the world works! You don’t need advanced statistical knowledge, programming skills or other abilities - the Wizard will tell you what to do, how to interpret your results and potential gotchas to watch out for. Try some of the tutorials to learn more.

How does it work?

1/4: Data
Causal Wizard is basically a statistical tool, so it needs some data. Upload a CSV or Excel file.

2/4: Domain Knowledge
Use our Wizard tool to add your subject-matter expertise. You'll draw a diagram including all causal relationships which might exist, and their direction.

3/4: Validation & Estimation
Causal Wizard will check your diagram and data all make sense, and provide feedback. If the causal effect can be measured, choose a model and Causal Wizard will go away and calculate everything.

4/4: Review & download results
You'll receive a detailed report describing:

  • The estimated causal effect and confidence interval
  • Some graphical plots depicting the effect in the context of your data
  • The results of a suite of robustness and validation checks
  • A list of assumptions made
  • Details of all modelling choices and parameters, including your provided causal diagram

Example causal inference result

How much data do I need?

Many people ask “how much data is enough?” The answer depends on the strength of the effects, and the complexity of the interactions between the variables. Typically, users will provide at least hundreds of samples, and ideally thousands. Sometimes, millions of samples are used.

What data format is needed?

Causal Wizard accepts .CSV and Excel spreadsheet data files. Contact us if you have data that doesn’t fit this format - we can help transform it.

Samples and Variables

A Sample is one individual data point from the population being studied. For example, a sample might be a person, one subject in an experiment, an event, or an object, such as an infrastructure asset... Each sample has a number of properties, which we call Variables. The Variables are the properties of the Samples.

The data must be structured as a matrix - a big table with many rows, and columns. We assume that each row contains the properties of a single individual sample. So, one cell of the table contains the value of a specific variable, for that one sample.

Structure of data files used in Causal Wizard

Datasets and Studies

Once you’ve uploaded your data, it becomes a Dataset. You can share and reuse a Dataset between multiple Studies, allowing you to see how different modelling affects the result. Using CausalWizard is a form of experimentation, to explore the data. Each experiment with a different causal diagram, research question or hypothesis is called a Study.

Treatment & Outcome: Defining an Intervention

A Causal Wizard Study explores the causal relationship between two variables: a Treatment and an Outcome. The wizard will attempt to estimate the effect of changes in treatment on the outcome, using your data.

More specifically, it models what happens when the treatment variable changes from a Control value to a Treated value. The control value represents some sort of default scenario. The treated value represents an intervention (i.e. change) you could make. For example, Control might mean "no treatment" and Treated might mean "take a new medicine". Or, "control" value might be some default quantity, and Treated value might be a different dose or rate. The treatment variable doesn't have to be a numerical quantity - it can be categorical, such as "Control=Policy A" and "Treated=Policy B".

Measuring Causal Effects

What is a causal effect? There are a number of related quantities, but by default Causal Wizard estimates the Average Treatment Effect (ATE). This is the average change in outcome value caused by a change in treatment from the control value to the treated value.

Once the effect is known, you can estimate counterfactual or hypothetical scenarios such as:

  • Counterfactual: If we had applied a different maintenance process to this equipment, what are the chances it would have failed when it did?
  • Hypothetical: If we applied a X% discount to product Y, what effect do we expect to have on sales next quarter?

Drawing a Causal Diagram

Use the Wizard tool to draw a graph, called a Causal Diagram, which represents the relationships you believe exist between your variables. A graph is a “network” of nodes and edges between nodes. Your diagram must have a directed path (following the arrows) from the Treatment to the Outcome, otherwise the causal effect is zero (which is a legitimate result).

Example causal diagram.

You do not have to include all your Variables in the graph, only the ones you think have a causal relationship to the ones you care about. You can also draw variables which you don’t have data for. These will be treated as unobserved variables.

Swap between drawing nodes and edges via the toggle buttons:

Example causal diagram.

In edit nodes mode:

  • Click anywhere to add a Node. A dialog will pop up. Choose the relevant variable from the data, or choose "Unobserved node".
  • Press and drag on an existing Node to move it around.
In draw edges mode:
  • Click and drag from one node to another to add an edge between them.
  • Click an existing edge to delete it.

Checking your input

You can press the big green CHECK button at any time. Causal Wizard will analyse your Study and report any issues. If the Study is fully defined, it will attempt to Identify the causal Estimand (see figure below). If this is not possible, Causal Wizard will tell you. This means the causal effect you're interested in cannot be estimated given the system described. However, you can adjust your domain knowledge to simplify it and try again. You should note these simplifications as additional assumptions. Keep in mind that the validity of the result is dependent on the validity of the domain knowledge you provide and depends on these assumptions.

The Causal Inference process with Causal Wizard

Understanding your results

After successfully passing the CHECK, the Calculate button will become available. Click Calculate and the Wizard will do its magic, by which we mean it will produce a statistical estimate of the causal effect of the Treatment variable on the Outcome variable, given the data and graph provided. The wizard will also attempt to refute the estimate provided, using several techniques. It will measure the stability of the answer on different subsets of your data, and provide a confidence interval around the estimated effect. This helps to you to understand how confident you can be in the answer, and what uncertainty remains.

You will receive a report describing:

The best way to use Causal Wizard is as a tool to help you understand your data and the system it represents. We suggest calculating causal effects several times with slightly different graphs and methods to see how the estimated effect changes.

Downloading and exporting results

There's a big DOWNLOAD button in the results page, which produces a PDF file containing all the results. The watermark is removed for paid accounts.

Where can I learn more about Causal Inference?

You can find out about many of the core concepts and ideas in our Help articles. They also explain some maths and stats terminology we can’t avoid using. We try to make Causal Wizard easy to use, but we don’t want to hide important words you might need later if you want to discuss your results with a statistician or data scientist.

You don’t need to read anything to start using Causal Wizard, but if you want to learn more about how it works and how to interpret your results, we maintain a list of recommended websites, papers, articles and courses.

What DOESN'T it do?

We have carefully avoided two significant, related fields: Causal Discovery and Causal Representation Learning. Causal Discovery aims to use Machine Learning to discover the Causal Diagram of a system, whereas Causal Wizard requires you to provide the Causal Diagram.

Causal representation learning aims to infer or detect causal variables and their relations from low level perceptual data such as images, video or other high-dimensional data.

Causal Wizard can't do these things for you. It requires your expert knowledge about a system. It's not because these topics aren't important - they are. But different tools and techniques are needed for each problem, which means a simple, friendly web app isn't possible.

Acknowledgements

This service depends on several Open-Source libraries from the Python Data-Science and Machine Learning ecosystem, most notably:

How to cite Causal Wizard

If you use Causal Wizard in your research, we would appreciate a citation. We recommend the following BibTeX entry: 

@misc{CausalWizard,
  title = {CausalWizard app},
  howpublished = {\url{https://causalwizard.app}},
  note = {Accessed: 2023-11-01}
}
Who created Causal Wizard?

Causal Wizard was created, and is maintained by, a small group of Data Scientists with backgrounds in Machine Learning who got fed up telling clients that statistics couldn't tell them about cause and effect. We were amazed there aren’t better tools to help people use Causal Inference in their work, given it’s so important! We believe causal inference is a crucial statistical tool that will change the way people reason about data, particularly to understand cause and effect rather than just correlation or association between variables. You don’t need to trust us, though. Everything Causal Wizard does for you is open and transparent using peer-reviewed and published methods, implemented using Open-Source Python software tools.

Registration

To allow you to control the retention and use of your data, we need a way to contact you. Therefore, users have to register (totally free) to upload data and create studies.

Future plans and features

We have an extensive roadmap of enhancements we want to make, including more powerful models, exploratory data analysis, more sophisticated advice and feedback on causal diagrams and additional analysis of results.

Contact us if there’s a feature you need, or a bug that needs fixing, or if something is confusing or somehow not right.