Regression coefficients: distinguish association from intervention

A regression coefficient describes a fitted relationship under a model. Interpreting it as the effect of intervening on a variable requires additional causal assumptions about assignment, confounding, measurement and the variables included in the model.

Good predictive fit does not establish those assumptions. A model can predict an outcome well using a proxy that has no direct causal effect in the underlying process.

Construct a confounded relationship

This synthetic process has a common cause z. It influences both x and y, while x has no direct term in the y equation. Independent noise leaves variation in x beyond z.

import numpy as np

rng = np.random.default_rng(158)
n = 10000
z = rng.normal(size=n)
x = z + rng.normal(size=n)
y = 3*z + rng.normal(size=n)
unadjusted_design = np.column_stack([np.ones(n), x])
adjusted_design = np.column_stack([np.ones(n), x, z])
unadjusted = np.linalg.lstsq(unadjusted_design, y, rcond=None)[0]
adjusted = np.linalg.lstsq(adjusted_design, y, rcond=None)[0]
assert 1.3 < unadjusted[1] < 1.7
assert abs(adjusted[1]) < .1
assert 2.8 < adjusted[2] < 3.2
assert np.linalg.matrix_rank(adjusted_design) == 3
print({'unadjusted_x_coefficient': float(unadjusted[1]),
       'adjusted_x_coefficient': float(adjusted[1]), 'adjusted_z_coefficient': float(adjusted[2])})

The unadjusted x coefficient is positive because x carries information about z. After including the correctly measured common cause in this correctly specified linear simulation, the x coefficient is near zero.

The code runs in the analyst statistics lab. NumPy's least-squares reference documents the numerical fit. The causal structure is supplied by the authored equations, not inferred automatically by least squares.

Explain the intervention in the simulation

Changing x while holding z and the outcome noise fixed does not change y in the generating equation. That is the intervention effect defined by this model.

Observationally, higher x is associated with higher y. The two statements are compatible because the common cause creates the association. A coefficient from the unadjusted regression would misrepresent the intervention effect here.

In real data, the generating equations are not handed to the analyst. The causal claim needs design knowledge and assumptions that can be defended beyond the fitted output.

Do not control for every available column

A variable measured before exposure and causing both exposure and outcome can play a different role from a mediator created by exposure or a selection variable affected by multiple causes. Adding all columns indiscriminately can change the estimand or introduce bias.

Specify which effect is of interest and why each adjustment variable belongs in the model. A causal diagram can make those assumptions inspectable, but drawing a diagram does not prove it is correct.

Keep model diagnostics in their proper role

Residual patterns, influential observations, functional form and validation performance matter for whether the fitted statistical model is useful. NIST's model validation discussion describes diagnostic assessment beyond a single fit statistic.

Passing residual checks does not prove the absence of unmeasured confounding. Conversely, a randomized design still needs an appropriate analysis and reliable measurement.

Report the coefficient with its conditions

State the outcome unit, predictor unit, adjustment set, target population and uncertainty method. For repeated customer observations or time series, ordinary independent-error standard errors may be inappropriate.

This example checks coefficients against the known synthetic process but does not estimate or report inferential intervals. A production analysis must choose uncertainty methods that match its sampling and dependence structure.

Exercise: add a direct term 2*x to the y equation. Verify that the adjusted x coefficient is near two under the same model, then introduce measurement error in z and examine whether adjustment still recovers the intended effect.

NeuraPath's Data Analytics with Generative AI course connects regression output with causal reasoning. A useful coefficient interpretation says what is associated, what intervention is contemplated and which assumptions connect the two.

Continue learning

This article is part of the Statistics for analytical decisions sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Missing data mechanisms with an analyst's decision checklist.
• Continue with Build a statistical analysis plan before opening the results.