Interpreting Long-Run Propensity in a Distributed-Lag Regression

Summary

The document explains the long-run propensity in a regression that includes current and lagged values of a predictor. It is calculated as the sum of the predictor’s coefficients across those time periods. The example shows coefficients whose combined effect is smaller than the immediate response, because later lagged effects partly offset earlier ones.

The intuition comes from tracing a one-time change in the predictor through the model over time. In the example, the dependent variable is a change in GDP: the lag coefficients describe successive changes relative to baseline, while their sum describes the eventual level shift after those effects have accumulated. This interpretation depends on how the dependent variable is defined and on the model’s lag structure; it is not a general measure of forecasting accuracy or proof of a causal effect. The document offers an illustrative explanation, not an empirical analysis or guidance on estimating uncertainty around the sum.

Key ideas

• Long-run propensity is the sum of coefficients on current and lagged predictor values.
• A one-time predictor change can affect the dependent variable over several periods.
• The sum of lag coefficients represents the accumulated eventual response in the example.
• Interpretation depends on whether the dependent variable is a level, change, or another quantity.
• The coefficient sum alone does not establish causality or predictive performance.

Tags

# Intuition behind the definition of “long run propensity” equaling sum of coefficients?


# Intuition behind the definition of “long run propensity” equaling sum of coefficients?












In a multiple linear regression model with independent variables x(t), x(t-1), x(t-2), etc. and dependent variable y, the “long run propensity” is defined to be the sum of beta coefficients in the entire model.

I don't understand this definition, can someone explain the logic behind this please?

## Answer by Bob Jansen (score 1)

https://quant.stackexchange.com/a/67790

I believe you're looking at models such as

$$y_t = \beta_0 + \beta_1 x_t + \beta_1 x_{t-1} + \ldots + \beta_n x_{t-n} + \varepsilon.$$

Suppose you gather some data and estimate a model and find:

$\beta_0 = 0,\, \beta_1 = 1,\, \beta_2 = -\frac{2}{3},\, \beta_3 = \frac{1}{3}$ and $\beta_n = 0$ for $n > 3$, that is $$\hat{y}_t = x_t - \frac{2}{3} x_{t-1} +\frac{1}{3} x_{t-2}.$$

According to the definition, the Long Run Propensity is $$1-\frac{2}{3}+\frac{1}{3}=\frac{2}{3}.$$

Why is this number interesting? That depends on the model, a simple example:

Suppose that $y_t$ is the change in GDP. This implies that in the long run a one time increase of $x_t$ of one unit implies an increase in GDP of $\frac{2}{3}$, an increase of $1$ on $t$, an increase compared to baseline of $\frac{1}{3}$ on $t+1$ and an increase of $\frac{2}{3}$ on $t+2$ and ever after.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.