What is Causal Inference? What are the fundamental differences between it and traditional predictive modeling?
Okay, this is a very interesting question. I'll try my best to explain it to you in plain language.
What is Causal Inference? How is it different from Predictive Models?
Imagine you're a watermelon farmer, growing watermelons every summer. You notice a phenomenon: the more fertilizer you apply, the bigger the watermelons seem to grow.
At this point, two thoughts might cross your mind:
- Predictive Thinking: "Hmm, based on my records from previous years, 10 pounds of fertilizer usually yields an 8-pound watermelon. So, if I apply 12 pounds of fertilizer next year, can I predict the watermelon will grow to 10 pounds?"
- Causal Thinking: "Is 'applying fertilizer' the true cause of the watermelon growing bigger? Could other factors be responsible? For instance, in the years I applied more fertilizer, perhaps the sunlight and rainfall were also exceptionally good? If I can confirm that fertilizer truly makes watermelons grow larger, then how much should I apply to get the best return?"
See the difference?
- Predictive Modeling is concerned with "What". It makes predictions by finding patterns and correlations. As mentioned above, it observes a strong positive correlation between "fertilizer amount" and "watermelon weight," and then uses one variable to predict the other.
- Causal Inference is concerned with "Why". It aims to determine if there's a genuine "causal relationship" between an "action" and an "outcome." It seeks to answer: "If I take a certain action (e.g., apply more fertilizer), will the outcome change as a result?"
Fundamental Difference: Correlation vs. Causation
The most fundamental difference between the two can be explained with a classic example: "ice cream sales" and "drowning incidents".
-
Predictive Modeling Perspective: Data shows that every summer, when ice cream sales are higher, the number of drowning incidents also increases. There's a very strong positive correlation between them. Therefore, a good predictive model can confidently state: "Based on today's ice cream sales, I predict tomorrow's drowning incidents will reach XX people!" And this prediction might even be quite accurate.
-
Causal Inference Perspective: Causal inference would ask a profound question: "Does eating ice cream cause people to drown?" The answer is clearly no. It would seek a common, "hidden" underlying cause, which we call a "confounder". In this example, this factor is "hot weather".
- Hot weather -> More people buy ice cream
- Hot weather -> More people go swimming -> Increases the risk of drowning
So, ice cream sales and drowning incidents merely "happen" together; they are both driven by the same cause (hot weather), but there's no causal relationship between them. If we naively banned ice cream sales to reduce drowning rates, it would be completely useless.
In Summary
| Feature | Predictive Modeling | Causal Inference |
|---|---|---|
| Core Goal | Predict the future, find associations between variables | Explain the past, find causality between variables |
| Question Answered | "If X happens, what might Y be?" | "If I intervene on X, how will Y change?" |
| Typical Examples | Stock prediction, weather forecasting, recommendation systems | Evaluating new drug efficacy, analyzing marketing campaign effectiveness, formulating public policy |
| Focus | How high is the model's accuracy | The reliability of conclusions and their guidance for the real world |
| Way of Thinking | Correlation does not imply causation (but accuracy is sufficient for prediction) | Must exclude confounding factors to find the true causal chain |
To put it plainly, a predictive model is like an experienced stock trader who can guess tomorrow's stock price fluctuations based on various indicators (K-lines, trading volume), but doesn't necessarily know why. Causal inference, on the other hand, is more like an economist who wants to figure out if the action of "interest rate cuts" directly caused the stock market to rise.
In the fields of machine learning and artificial intelligence, we have long been better at prediction. However, increasingly, many scenarios (such as healthcare, economics, and refined operations) demand that we not only "know what happened" but also "know why it happened." This is precisely why causal inference is becoming increasingly important.