Identifying the MPC-Liquidity Gradient in High-Quality Data

New Estimator Reveals Sharp MPC Gradient Linked to Household Liquidity

How do people react when their income suddenly changes? Researchers have spent decades studying the Marginal Propensity to Consume (MPC)—the fraction of an extra dollar of income that a person spends rather than saves. Understanding this link is vital for predicting how monetary policy or tax rebates ripple through the economy.

Current models struggle to pin down exactly how this response changes depending on a person's financial cushion. While we know that someone with zero savings reacts differently than a millionaire, measuring this "gradient" has proven difficult. Previous methods often lacked the precision to see the fine-grained patterns of how liquidity dictates spending.

A new study using high-quality Swedish administrative data reveals a striking reality: the MPC isn't a steady number. Instead, it drops sharply as people gain more cash. The researchers report that the MPC falls from approximately 0.7 in the lowest decile of cash-on-hand to 0.3 in the top decile. This uncovered a steep, convex relationship that earlier, less efficient methods simply could not detect.

The inefficiency of robust estimators

For years, economists have relied on "semi-structural" approaches to estimate the MPC. These methods, such as the canonical BPP estimator (Blundell et al., 2008), are designed to be robust against measurement error. In many economic studies, income data comes from surveys where people often misreport or forget their earnings. To prevent this "noise" from ruining the results, these traditional estimators intentionally ignore parts of the income history.

While this robustness is a virtue in messy survey data, it becomes a liability when working with high-quality administrative records. Administrative data—collected by tax authorities and banks—is incredibly precise. By using "safe" but blunt instruments, traditional estimators throw away massive amounts of information.

The authors demonstrate this cost of caution in their comparison of methods. As shown in, the refined BPP-C estimator (Commault, 2022) is so imprecise when applied to this clean data that its standard errors are massive. Consequently, it fails to detect the gradient. The researchers argue that by prioritizing robustness to errors that do not exist in administrative sets, scientists have been missing the most interesting patterns in household behavior.

Recovering shocks with a Kalman smoother

The authors propose a different architecture: a Kalman-shock estimator designed specifically for data where income is measured with negligible error. Their approach moves away from picking isolated "windows" of income. Instead, it uses the entire household history to reconstruct the truth.

The mechanism works in three distinct stages:

1. Parameter Estimation: First, the researchers estimate the underlying "income process." This is the mathematical rule governing how permanent and transitory income fluctuate over time. They treat income as a mix of a permanent component (long-term earning power) and a transitory component (short-term spikes or dips).
2. Latent Shock Recovery: Using a Kalman smoother—a recursive algorithm that looks at an entire sequence of data to estimate hidden states—the authors recover the actual, unobserved shocks for each household. Think of the Kalman smoother like a high-fidelity audio filter. While the raw income signal might have jitter, the smoother uses the "past and future" of the signal to isolate the true underlying melody of permanent and transitory shifts .

3. State-Dependent Regression: Finally, they regress consumption growth on these recovered shocks. Crucially, they allow the relationship to change based on the household's "state." This state is measured as normalized cash-on-hand (cash-on-hand divided by permanent income).

By using the Kalman smoother, the authors achieve "minimum variance." In technical terms, the smoothed shock is the best possible linear prediction of the true latent shock given the available history. This means they extract the maximum amount of signal from every available data point.

A steep descent in spending response

The results of this high-precision approach reveal a dramatic landscape of consumer behavior. The authors report that the response to transitory income shocks is highly sensitive to how much liquid cash a person holds.

The primary finding is a "sharp and convex negative gradient" in the MPC .

For households in the bottom decile of liquidity, the MPC is high, around 0.7. This means if they receive a transitory boost in income, they spend most of it immediately. However, as cash-on-hand increases, this tendency evaporates. By the time a household reaches the top decile, the MPC has fallen to roughly 0.3.

The researchers further decompose this gradient to see what is actually driving the change. They find that the drop in MPC is not caused by changes in disposable income. Instead, it is driven by changes in liquid financial wealth . This confirms a core tenet of the "buffer-stock" model: households use liquid assets as a shock absorber. When they have plenty of cash, they can smooth out temporary income dips without changing their lifestyle. When they are cash-poor, they have no choice but to spend whatever comes in.

Interestingly, the response to permanent income shocks behaves quite differently. The authors find that the pass-through of permanent shocks is relatively flat and stays close to one across the entire distribution .

Whether you are rich or poor, a permanent increase in your lifetime earning power tends to lead to a proportional increase in your consumption.

Trade-offs in the presence of noise

Every methodological choice carries a cost. The authors are transparent about the limitations of their "high-precision" approach. Their estimator is a specialist tool. It is optimized for the clean, high-resolution world of administrative registers.

If this method were applied to standard survey data (like the PSID), it would likely suffer from significant "attenuation bias" (a systematic underestimation of effects). Because the Kalman smoother tries to use every bit of information available, it cannot easily distinguish between a genuine transitory income spike and a simple reporting error. In a noisy environment, the estimator would mistake the error for a signal. The authors address this in simulation exercises . They note that while traditional BPP-style estimators remain centered on the truth in noisy settings, their projection estimator can be biased.

Furthermore, the authors note that their MPC values are technically "bounds" derived from an elasticity. To get a literal dollar-to-dollar MPC, one must apply a conversion factor based on the ratio of consumption to income. While they provide these conversions, practitioners must be aware of the underlying model assumptions.

The verdict: use the right tool for the data

Is this method ready for widespread use? The answer depends entirely on your data source.

If you are working with survey data characterized by high measurement error, stick to the robust BPP or BPP-C estimators. They are designed for that environment. The loss in precision is a necessary insurance premium against bias.

However, if you have access to high-quality administrative or transaction-level data, the Kalman-shock estimator is a superior choice. It provides the statistical power necessary to detect the complex, state-dependent behaviors that drive modern macroeconomics. The authors have demonstrated that the "liquidity gradient" is a real, measurable phenomenon. But you need a high-resolution lens to see it.

Figures from the paper