Identify effect not showing backdoor variable

[asha24choudhary]

Hi there. I was referring to this issue. I am having a dataset with both observed and unobserved confounder as described below

Create the graph describing the causal structure

graph = """graph[directed 1 node[id "U" label "U"]
node[id "X" label "X"]
node[id "Y" label "Y"]
node[id "Z" label "Z"]
edge[source "U" target "X"]
edge[source "X" target "Y"]
edge[source "U" target "Y"]
edge[source "Z" target "X"]
edge[source "Z" target "Y"]]""".replace('\n', '')

# # Generate the data

U = np.random.randn(N_SAMPLES)
Z = np.random.randn(N_SAMPLES)
X = np.random.randn(N_SAMPLES) + 0.3U +0.2Z
Y = 0.65X + 0.2U+ 0.3*Z

df = pd.DataFrame(np.vstack([Z,X, Y]).T, columns=['Z','X', 'Y'])
print(df.head(10))

Create a model

model = CausalModel(
data=df,
treatment=['X'],
outcome=['Y'],
common_causes=['Z'],
graph=graph
)
model.view_model()
plt.show()

I expect to have backdoor variable but

as you can seen the estimate says 'Backdoor identification failed'. I don't know what is wrong and how can I resolve this?

Could you please help me?

[github-actions]

This issue is stale because it has been open for 14 days with no activity.

[amit-sharma]

The error you are seeing is unrelated to the linked issue.
In your case, the only valid backdoor set is $[U, Z]$, but since U is unobserved, identify_effect method returns that backdoor identification is not possible.

Note that graph argument takes precedence in CausalModel. So if you only want to condition on Z, you have can do so if by initializng CausalModel directly, without the graph.

model = CausalModel(
data=df,
treatment=['X'],
outcome=['Y'],
common_causes=['Z']
)

[asha24choudhary]

Thank you for your reply @amit-sharma, I reason why i linked the previous issue is because I wanted to include unobserved confounder. But don't you think I should include the graph which contains the info about the unobserved confounder 'U', which is also done in the issue I linked?

I was assuming that in order to have unobserved confounder, I should include it in the graph which is used while creating the model and exclude it in the dataset.

Yes if I exclude the graph while modelling, then the valid backdoor path includes Z. However, my question to u now is that should I not include the graph & why, because don't you think if I do so then I lose the info about the unobserved confounder in the model, of course it is still present in the data?

Would be really helpful if you could explain a bit more in detail.

[github-actions]

This issue is stale because it has been open for 14 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 7 days since being marked as stale.