Causal software for effect estimation
Quantify cause-and-effect relationships using Causal Inference and Machine Learning
You may have heard that you can't establish or measure causality from existing data, only through Randomized Controlled Trials (RCTs). That's not true.
Many peer-reviewed papers including Nobel prize-winning research* have proven that Causal Machine
Learning techniques can
detect, confirm and quantify cause-and-effect relationships.
But you must have 2
things:
-
Some data - to validate and quantify the model and its effects
-
A causal model - can be specified in advance by elicitation of expert domain knowledge, or learned from data (causal discovery)
Causal Wizard helps you use your knowledge to build causal models, and apply them to your data.
Observational experiments
Causal Wizard actually implements a type of scientific experiment, called an Observational or Natural
experiment. It's used where it's impossible, impractical, too expensive or simply unethical to
conduct a
Randomized Controlled Trial. For example, you can't force people to smoke cigarettes for your research!
Observational experiments don't control who gets "treated", but they benefit from observing the outcomes
of a
"treatment", wherever and however it occurs.
Using Causal Inference and Machine Learning, you can obtain cause-and-effect insights about your data and
predict the outcomes of an intervention before trying it.
But you've got to use the right techniques to avoid mistaking spurious correlations for genuine cause and
effect. Using modern Causal Inference techniques, and with your domain knowledge, you can get
accurate
results from existing data.
What's wrong with other Machine Learning methods?
With Causal ML you can use historical data to accurately predict the outcomes of hypothetical interventions (i.e. changes to a system), or counterfactual scenarios - things that didn't actually happen, but you want to understand "what would have happened, if ..."
Non-causal ML does learn accurate predictive models, but these models can be very wrong under changed conditions not well represented in the training data.
Causal models exploit prior knowledge of the system to give more accurate answers in changed conditions, such as after an intervention. So if you want to explore interventions and counterfactuals, it's important to use Causal models.
The limitations of learning from data alone
Data tends to contain a lot of spurious or misleading correlation or association, which might be coincidence, or the effect of other variables, maybe unobserved. Machine learning models will learn these correlations and include them in their predictions.
See Tyler Vigen's site for more.
If your training data is statistically identical to your use-case, this might not be a problem. But when you want to model the results of an intervention - a change to one or more variables - you're changing the statistics of the data. This is when the limitations of purely associative modelling become dangerous!
The figure above is an illustration of "Simpson's Paradox", in which the relationship between variables X and Y appears negatively correlated over the entire dataset (magenta line), but is in fact the opposite (positively correlated) for all sub-groups (black, red, green and blue).
Causal
Wizard has been independently tested against the