Understanding the significance of causal estimates

[tusharagarwal-zomato]

Hi,

I am new to causal learning and I am exploring dowhy currently.

I need some help in understanding the significance of the numbers returned during the third step of the process, i.e estimating the effect.

I am running my analysis on a simple graph with no confounders and instrumental variables and there are 4 variables that are all connected to the outcome. I am trying to understand the effect of one of them on the outcome variable.

my graph can be assumed to look like this:
{x1 -> y, x2 -> y, x3 -> y, x4 -> y }
Please note that Xi and y are all continuous variables which haven't been scaled.

Let's say I am trying to understand the causal effect of x1 on y.

The method I use is backdoor linear regression and my causal estimate is -0.46. I calculated the correlation between the x1 and y.. it came out to be a positive 0.16. Similarly, I tried to fit a simple LR on the data, the coefficient of x1 in that case is around 0.06.

1. With all these numbers, what can I infer about the causal effect and my previous understanding via coefficients and correlation? Does it refute the correlation as it is negative? or should we not compare both in any way at all?
2. I am unable to understand what does the magnitude(0.46) signify here? One unit of change in x1 causes negative 0.46 unit change in y? Should I compare it against the causal effects of other variables? i.e Is it for comparative analysis only?
3. Having said that, can we compare causal effects of two different variables obtained via different methods?
4. Do the variables need to be scaled? Will it affect the results?
5. What other methods can I use when both my treatment and outcome variables are continuous?
6. In model.estimate_effect, there are certain parameters like control_value and treatment_value. Since in this case, the treatment variable and the outcome variables are continuous, how do we input these values? What is the baseline assumption/values in case we skip these params?
7. Is it advisable to use continuous treatment variable or is it advised to break down treatment variable into categories and the create separate boolean features for the categories and then study the causal effects based on binary treatment features thus obtained?

(PS: I will add the confounding variables and other variables in my graph. These questions are purely from academic pursuit. Starting with an easier model so that it is easy to learn concepts)

[TalionSlade]

Hi, @tusharagarwal-zomato . I have been working on a very similar use case to as you seem to have mentioned here and these are almost the exact issues that I am facing. Please do share any insights that you might have had gathered meanwhile.