A Deep Dive into Distributed Lag Models

Ever wondered why a simple hike in the interest rate by the Reserve Bank of India doesn’t immediately curb inflation, or why a new infrastructure project takes years to fully boost GDP? The answer lies in the concept of lags-the often-overlooked delay between an economic cause and its full effect. As economists, we can’t just assume that all consequences happen instantly. That’s where the Distributed Lag Model (DLM) steps in, offering a powerful, elegant way to measure how an event’s impact ripples through time. Think of it as throwing a pebble into a pond; the effect isn’t just one splash, but a series of widening ripples or waves.

Defining distributed lag models 🔗

A Distributed Lag Model is an econometric tool used to capture the idea that the impact of a change in an independent variable ($X$) on a dependent variable ($Y$) is spread out, or “distributed,” over several subsequent time periods. Instead of a single, immediate relationship, the model assumes that the current value of $Y$ is a function of the current and past values of $X$.

The basic form: understanding the ripple effect 🔗

The standard representation of a DLM is quite intuitive, even if the math looks complex initially:

Y_t = β₀X_t + β₁X_{t-1} + β₂X_{t-2} + ... + β_sX_{t-s} + u_t

Let’s break down this powerful equation:

• Y_t: This is the dependent variable (the effect) at time $t$ (e.g., current GDP growth).
• X_t: This is the independent variable (the cause) at time $t$ (e.g., current government spending).
• X_{t-1}, X_{t-2}, …, X_{t-s}: These are the lagged values of the cause. $X_{t-1}$ is the spending from one period ago, $X_{t-2}$ from two periods ago, and so on, up to the maximum lag length, $s$.
• β₀, β₁, …, β_s: These are the distributed lag coefficients (or weights). They represent the magnitude of the impact of $X$ at each specific time lag. For instance, $β₂$ tells us the effect of spending from two periods ago on the current GDP.
• u_t: The error term, capturing everything else not explained by the model.

The coefficient $β₀$ is known as the impact multiplier (the immediate effect), while the sum of all coefficients ($\sum_{i=0}^{s} β_i$) is the long-run multiplier (the total eventual effect). This sum is vital because it provides a complete picture of the policy’s final magnitude.

A relatable example: The Policy Lag. Consider the Indian government raising the Minimum Support Price (MSP) for a crop like wheat. The effect on farmer income ($Y_t$) isn’t just immediate. It’s distributed:

1. Immediate (β₀): Minor effect, maybe better farmer sentiment/stock-up.
2. Next Season (β₁): Significant effect as farmers respond to the price signal by increasing acreage and investment in the next planting cycle.
3. Harvest (β₂): Peak effect as the boosted yield hits the market, affecting prices, supply, and farmer revenue.
4. Later (β₃, β₄, etc.): Residual, smaller effects on related industries like storage and transport.

The DLM allows us to statistically estimate exactly how large each of those $\beta$ weights is, rather than just guessing. This level of precision is why DLMs are fundamental in central bank and fiscal policy analysis across the globe.

[Image: A simple line graph showing a single spike in an independent variable X_t at time t=0, followed by a gradual, declining bell-shaped or inverted V-shaped response in the dependent variable Y_t over subsequent periods t=1, t=2, t=3, illustrating a distributed lag.] —

Challenges in estimation 🔗

While the DLM framework is intellectually satisfying, turning theory into a robust, reliable econometric model is fraught with practical difficulties. Researchers must navigate several critical obstacles when estimating the equation above.

Choosing the correct lag length ($s$) 🔗

The first and most critical decision is determining $s$, the maximum relevant lag length. How far back in time does the cause still matter? Is the impact of the 2008 financial crisis still affecting current investment decisions 17 years later, or did the effect taper off after five years?

If you choose $s$ too small, you commit a specification error, excluding relevant historical effects. The long-run multiplier will be underestimated. If you choose $s$ too large, you include irrelevant lags (padding the model with zeros) and inflate the risk of the next major challenge: multicollinearity.

Econometricians use statistical tests like the Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) to help guide this choice, but ultimately, economic theory must dictate a plausible range.

Dealing with multicollinearity 🔗

The most notorious problem in DLM estimation is multicollinearity. This occurs when the lagged independent variables ($X_t$, $X_{t-1}$, $X_{t-2}$, etc.) are highly correlated with one another. Why does this happen? Economic time series often move together: if government spending ($X$) was high last year, it’s often high this year and was probably high the year before. They share similar trends.

High multicollinearity makes the ordinary least squares (OLS) estimates of the individual $β$ coefficients unstable and their standard errors large. You might find that your long-run multiplier ($\sum β_i$) is statistically significant, but none of the individual $\beta$ coefficients are. This is problematic because the entire purpose of a DLM is to estimate the shape and magnitude of the individual lag coefficients-how the effect is distributed.

In essence, OLS struggles to accurately attribute the shared effect to $X_t$, $X_{t-1}$, or $X_{t-2}$ because they are all moving in lockstep.

The assumption of monotonically declining coefficients 🔗

For many economic phenomena, a sensible assumption is that the effect of a change should eventually decline over time-a change today should have a larger impact than the same change five years ago. This is the monotonically declining coefficient assumption. While economically sound in many contexts, estimating a standard DLM (where all $\beta$ coefficients are unrestricted) often produces volatile, nonsensical estimates that jump up and down, which can violate this expectation, particularly when multicollinearity is present.

This led to the development of methods that impose structure on the lag coefficients, effectively forcing them to follow a smoother, more realistic pattern. This structure is key to solving the multicollinearity problem and ensuring more plausible estimates.

—

Popular lag schemes: De Leeuw and Almon 🔗

To overcome the estimation challenges, particularly multicollinearity and the wild fluctuation of coefficients, econometricians developed lag schemes. These schemes impose a mathematical structure on the coefficients ($\beta_i$) that allows the model to estimate the entire lag distribution using just a few parameters, greatly reducing the problem of multicollinearity.

The De Leeuw’s inverted V-shape 🔗

The De Leeuw scheme, sometimes referred to as the Inverted V-Lag, is a relatively early approach that imposes a very specific, simple shape on the lag coefficients. It assumes that the initial effect is small, rises to a peak, and then declines symmetrically or asymmetrically back to zero.

Imagine the effect of an economic stimulus: it takes time to kick in (low initial $\beta$), peaks as businesses and consumers fully react (highest $\beta$), and then the effect wanes as the economy adjusts (declining $\beta$). The DLM with a De Leeuw constraint forces the coefficients to follow this predictable pattern. While simple and effective at reducing parameters, its main limitation is its inflexibility; it cannot capture more complex, irregular lag shapes that might exist in reality.

The Almon polynomial distributed lag 🔗

By far the most famous and widely used structured lag scheme is the Almon Polynomial Distributed Lag (PDL), developed by Shirley Almon in 1965. Almon’s genius was realizing that the lag coefficients ($\beta_i$) could be approximated by a lower-order polynomial function of the lag index ($i$).

Instead of estimating $s+1$ individual coefficients, the Almon PDL estimates only the few coefficients of the polynomial (e.g., $a_0$, $a_1$, $a_2$ for a second-degree polynomial). For example, if you have 12 lags ($s=12$) you would need to estimate 13 coefficients ($\beta_0$ to $\beta_{12}$). With a second-degree Almon lag, you only need to estimate three parameters. This massive reduction in parameters effectively solves the multicollinearity problem and makes the estimates far more stable.

The polynomial function is mathematically defined as:

β_i = a₀ + a₁i + a₂i² + ... + a_d i^d

where $i$ is the lag number ($0, 1, 2, …$) and $d$ is the degree of the polynomial (usually $d \le 3$ is sufficient).

The Almon method is implemented by substituting this polynomial expression for the $\beta_i$’s back into the original DLM equation, transforming the model into one that can be estimated directly by OLS, using the polynomial coefficients ($a_0, a_1, …$) as the new independent variables. The key to its flexibility is the ability to select the degree of the polynomial, which controls the smoothness of the lag shape.

Boundary Conditions: Almon often includes endpoint constraints, typically assuming that the effect starts at zero before the first lag ($β_{-1}=0$) and/or ends at zero after the final lag ($β_{s+1}=0$). These zero constraints make the estimates of the lag coefficients more efficient and give the entire lag distribution a more realistic, hump-shaped look-a common pattern in economics.

—

Beyond the classics: modern distributed lag models 🔗

While Almon and De Leeuw are foundational, modern econometrics has moved towards more advanced techniques that are computationally simpler and less prone to specification errors. The introduction of the Autoregressive Distributed Lag (ADL) model and techniques like Error Correction Models (ECM), which integrate the short-run dynamics of a DLM with the long-run equilibrium relationships, have largely superseded the historical polynomial schemes in applied macroeconometrics today. These models are generally preferred as they allow the data to determine the lag structure more freely without forcing a specific polynomial shape onto the coefficients.

However, understanding the challenges faced by the original DLMs and the clever solutions provided by economists like Almon is essential for appreciating the evolution of modern time-series analysis. The core principle remains: economic consequences are distributed over time, and a good model must account for the delay.

What do you think? Can you think of a government policy (e.g., a change in tax rates or subsidies) where the long-run multiplier might actually be smaller than the immediate impact multiplier? How might the Almon PDL’s structure limit the discovery of a highly irregular, perhaps oscillating, lag pattern?

References

1. https://www.nber.org/system/files/chapters/c10811/c10811.pdf
2. https://www.investopedia.com/terms/e/errorterm.asp
3. https://www.rbi.org.in/Scripts/PublicationsView.aspx?id=17255
4. https://www.sciencedirect.com/topics/economics-econometrics-and-finance/akaike-information-criterion
5. https://www.jstor.org/stable/1913702?seq=1