Features
Causal Wizard has a wide range of features to help you explore causal effects in your data. This will enable you to draw causal conclusions from historical, observational data, and predict the results of interventions (changes to the system). The features are designed to help you understand and gain confidence in your results.Exploratory data Analysis
It's always important to look closely at your data and it also helps to be able to quickly check the
contents of particular columns (variables) when drawing your Causal Diagram. Causal Wizard provides
tabular and graphical views of your data
file.
The table view allows sorting and filtering to find particular values or ranges.
Graphical views display selected variables as histograms and a scatter plot allows you to
visualise the
relationship between pairs of variables.
Tabular data views
Univariate data plot - histogram
Bivariate data view - scatter plot
Choose from a range of popular Causality methods
CausalWizard offers you several popular Causal Inference techniques to choose from.
If your data is in Panel Data
format, which tracks the characteristics of
many entities over time, you can use our Fixed-Effects and Difference-in-Differences
models.
You can also use our more powerful and general-purpose Causal Diagram &
Potential
Outcomes models to estimate causal effects on many types of data.
Modelling the system
Capturing your domain knowledge in the form of a Causal Diagram is a key part of Causal
Inference and one of the things that enables estimation of causal effects rather than just association or
correlation between variables,
which any predictive model could provide.
Our Causal Diagram editor allows you to draw a Directed Acyclic Graph (DAG) using all the variables
(columns) in your data file.
However, if you prefer not to upload any data, you can just draw the DAG and label the variables yourself.
Causal Wizard also provides tools to define your Outcome variable and treatment groups either by
thresholding a numerical variable or with specific values
(categorical variables).
Together, these configuration options define the experiment and effect you wish to explore.
After drawing the diagram, Causal Wizard will analyse it for you to
determine whether a causal effect can
be measured,
and the methods which are applicable to the system as described. The role of each variable in the analysis
will be illustrated; Causal Wizard will show you which variables you should include and which
variables you must exclude to obtain an unbiased estimate of the causal effect. If you have uploaded
data, Causal Wizard
will also offer to apply the selected method for you.
Causal Diagram (DAG) editor
Treatment groups identification widget
Exploring outcomes
If you have provided data, Causal Wizard can perform a causal effect estimation and analysis for you.
The results will include key findings, assumptions and a description of the modelling
performed.
The app will also provide graphical plots depicting the causal effect on the outcome,
broken down by treatment group (treated, or control). The exact plots shown depend whether your outcome is
numerical or categorical.
A key benefit of causal analysis is the ability to accurately predict counterfactual outcomes.
When suitable analysis methods are provided, Causal Wizard will give you a table listing various
counterfactual
outcomes for different sub-populations and treatment variable status (treated or control).
Outcome distribution plot
Counterfactual outcomes table
Understanding model behaviour
Stakeholders often want to confirm that the model uses covariates (input features) as expected, which gives confidence that the system is being modelled correctly. The feature-importance plot shows the sign and magnitude of regression coefficients, which indicate the direction and contribution of each covariate (and the treatment).
Feature importance plot
Refutation tests
The app uses a set of statistical tests from the
DoWhy module to estimate the statistical significance of
results
and whether they are robust to challenges such as randomized outcomes, a placebo treatment
or addition of a random, previously unobserved confounder.
Passing these tests provides additional confidence that the estimated effect is significant and robust.
Refutation test results
Understanding your results
Many causal inference techniques are based on the concept of Propensity scores.
These methods also enable us to understand the ways in which the observed population differs between treated
and control groups, which is a primary cause of bias in observational studies.
The propensity distribution plot shows these differences between groups, allowing you to verify that
this bias is likely to be manageable after appropriate weighting, matching or sampling.
The covariate
balance plot (also known as a Love plot) shows the differences between treatment groups
before and after weighting by propensity, confirming to what extent these differences have been controlled.
Propensity score distribution plot
Covariate balance plot
Validating model accuracy
In Machine Learning, it is common to validate model performance on a set of samples not available during
model training or fitting. This allow us to estimate how well the model generalizes to unseen data.
Causal Wizard keeps a user-defined percentage or specific rows of your data for this type of validation.
The results are displayed in a range of plots or tables, depending whether your outcome variable is
categorical or numerical.
For numerical outcomes, we show a scatter plot of predicted vs actual outcomes for individual samples
coloured by treatment group. We report key predictive metrics such as r-squared and
root-mean-square error.
For categorical outcomes, we provide a confusion matrix listing all predicted classification outcomes
and summary metrics including Accuracy, F1-score, Precision and Recall.
Actual vs predicted generalization performance plot
Generalization performance metrics
Categorical generalization performance - confusion matrix
Need a new feature? Tell us about it!