0
$\begingroup$

I am new to the idea of causal inference or causality in statistic and in Python.

I have a dataframe test which looks as follows:

x   y
0   0.03    315.98
1   -0.03   316.91
2   0.06    317.64
3   0.03    318.45
4   0.05    318.99
... ... ...
58  0.92    406.76
59  0.84    408.72
60  0.97    411.66
61  1.01    414.24
62  0.84    416.45

test.to_dict() is given as:

{'x': {0: 0.03,
  1: -0.03,
  2: 0.06,
  3: 0.03,
  4: 0.05,
  5: -0.2,
  6: -0.11,
  7: -0.06,
  8: -0.02,
  9: -0.08,
  10: 0.05,
  11: 0.02,
  12: -0.08,
  13: 0.01,
  14: 0.16,
  15: -0.07,
  16: -0.01,
  17: -0.1,
  18: 0.18,
  19: 0.07,
  20: 0.16,
  21: 0.26,
  22: 0.32,
  23: 0.14,
  24: 0.31,
  25: 0.16,
  26: 0.12,
  27: 0.18,
  28: 0.32,
  29: 0.39,
  30: 0.27,
  31: 0.45,
  32: 0.4,
  33: 0.22,
  34: 0.23,
  35: 0.31,
  36: 0.44,
  37: 0.33,
  38: 0.46,
  39: 0.61,
  40: 0.38,
  41: 0.39,
  42: 0.53,
  43: 0.62,
  44: 0.62,
  45: 0.53,
  46: 0.67,
  47: 0.63,
  48: 0.66,
  49: 0.54,
  50: 0.65,
  51: 0.72,
  52: 0.61,
  53: 0.64,
  54: 0.67,
  55: 0.74,
  56: 0.89,
  57: 1.01,
  58: 0.92,
  59: 0.84,
  60: 0.97,
  61: 1.01,
  62: 0.84},
 'y': {0: 315.98,
  1: 316.91,
  2: 317.64,
  3: 318.45,
  4: 318.99,
  5: 319.62,
  6: 320.04,
  7: 321.37,
  8: 322.18,
  9: 323.05,
  10: 324.62,
  11: 325.68,
  12: 326.32,
  13: 327.46,
  14: 329.68,
  15: 330.19,
  16: 331.12,
  17: 332.03,
  18: 333.84,
  19: 335.41,
  20: 336.84,
  21: 338.76,
  22: 340.12,
  23: 341.48,
  24: 343.15,
  25: 344.85,
  26: 346.35,
  27: 347.61,
  28: 349.31,
  29: 351.69,
  30: 353.2,
  31: 354.45,
  32: 355.7,
  33: 356.54,
  34: 357.21,
  35: 358.96,
  36: 360.97,
  37: 362.74,
  38: 363.88,
  39: 366.84,
  40: 368.54,
  41: 369.71,
  42: 371.32,
  43: 373.45,
  44: 375.98,
  45: 377.7,
  46: 379.98,
  47: 382.09,
  48: 384.02,
  49: 385.83,
  50: 387.64,
  51: 390.1,
  52: 391.85,
  53: 394.06,
  54: 396.74,
  55: 398.81,
  56: 401.01,
  57: 404.41,
  58: 406.76,
  59: 408.72,
  60: 411.66,
  61: 414.24,
  62: 416.45}}

There are two variables in this dataframe x and y. x is the independent variable, and y is the dependent variable.

I can calculate the correlation between two using:

test.corr()

It returned:

x   y
x   1.000000    0.961354
y   0.961354    1.000000

This means, that the correlation between x and y is 96%. However, this does not show the causal relationship between the two variables.

How can I show statistically in Python that x causes y and show the effect by certain value?

$\endgroup$
2
  • 6
    $\begingroup$ You can't. Some way or another, you will have to use subject-matter or design knowledge of why the relationship between x and y is causal. E.g., because you were lucky enough to be able to run an experiment, or because you can, for some other reason, be confident that there are no confounding variables in your research design. $\endgroup$ Commented Aug 15, 2022 at 12:01
  • 1
    $\begingroup$ You need a causal model, e.g. x causes y and causes z, y causes z, like this pic. You might then be able to come up with some tests to check whether your model agrees with data. So you can use data to rule out causal models, I don't think you can uniquely identify causal model simply from data $\endgroup$
    – Cryo
    Commented Aug 15, 2022 at 12:31

1 Answer 1

6
$\begingroup$

As Christoph and Cryo have mentioned, you are asking the impossible unless you have more information. Christoph is absolutely correct in saying that you would need to have run an experiment to get the data you have, or for some other reason, you would need to be confident that you have no confounding variables.

Formally, Theorem 1.2.8 (Observational Equivalence) on page 19 of Pearl's Causality: Models, Reasoning, and Inference, 2nd Ed., states the following:

Two DAGs are observationally equivalent if and only if they have the same skeletons and the same sets of $v$-structures, that is, two converging arrows whose tails are not connected by an arrow.

The skeleton refers to the nodes and undirected arrows. So two graphs would have the same skeleton if you crunched all their directed edges down to undirected edges and found you had the same graph. The $v$-structures are mostly to do with colliders.

In your case, you have only mentioned two variables, so you can't even have $v$-structures. It follows from the theorem, then, that you cannot use any data to distinguish between the two graphs $X\to Y$ and $Y\to X.$

There are algorithms to detect causal models, but they are subject to the fundamental limitation of this theorem. Pearl writes:

Observational equivalence places a limit on our ability to infer directionality from probabilities alone. Two networks that are observationally equivalent cannot be distinguished without resorting to manipulative experimentation or temporal information.

$\endgroup$
1
  • 2
    $\begingroup$ For completeness: There are a lot of approaches to distinguish cause and effect from observational data alone (see e.g. this summary). However, they rely on strong assumptions that tend to be hard to justify in practice. $\endgroup$
    – Scriddie
    Commented Aug 18, 2022 at 8:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.