Overview

This chapter traces Mauritania’s confidential monthly bank records through separate lending and profitability first stages, residual construction, size weighting, and the aggregate credit equation. It explains why low within-bank lending fit leaves variation for a shock but does not make it causal, while diagnostics offer limited reassurance rather than instrument strength. The 15-bank, June 2020–June 2024 panel and 48 aggregate observations support a short-run case diagnosis, not a long historical or completed IV claim.

Learning outcomes

First-stage residuals are constructed measures whose meaning depends on controls, fixed effects, and the weighting rule.
Low within-bank lending fit leaves unexplained movement, not an automatically clean causal supply shock.
The short 15-bank panel and direct-predictor design limit the evidence to conditional short-run associations.

Keywords

Mauritaniaevidenceresidualsdiagnosticsregressionsidentification

What the Evidence Is Actually Measuring

The Mauritanian case is now a question about evidence rather than intuition: how did the researchers turn confidential bank records into lending and profitability shocks, and what confidence should those measurements command? This chapter follows the study from bank-month variables through first-stage regressions and diagnostics to the aggregate credit equation. The goal is not to reproduce every specification, but to understand which parts are observations, which are constructed measures, and which conclusions remain conditional.

From bank-month records to two shock series

The study begins with monthly supervisory data collected by the Banque Centrale de Mauritanie. The panel covers June 2020 through June 2024 and reports 15 banks over 49 monthly observations, or 735 bank-month observations before lagged variables reduce the usable count. Two banks were excluded: one began operations in November 2023, and another was removed because of inconsistencies and a high incidence of outliers. The first-stage tables use 720 observations where a lag is required for lending growth. Later aggregate regressions use 48 monthly observations.

Those denominators tell us what kind of evidence this is. It is a short, high-frequency panel with many bank-month observations but only a limited number of system-wide months. The bank dimension helps the researchers compare institutions and extract residual movement. The time dimension lets them relate that movement to monthly aggregate credit. It does not provide a long historical record across different credit cycles, regulatory regimes, or political conditions.

The bank-level variables are chosen to connect balance-sheet activity to the central question. Lending growth is the proportional change in total loan volume from the previous month. Profitability is net income divided by total bank capital; the paper uses the level of this ratio rather than a growth rate because small or negative denominators could make changes difficult to interpret. Liquidity is cash and cash equivalents divided by total assets. Top-five borrower and top-five depositor ratios capture concentration on the lending and funding sides. The lending–deposit spread is the average loan rate minus the average deposit rate.

The macro controls are inflation, the policy rate, M3 growth, and the change in the USD/MRU exchange rate. These variables represent common conditions that can affect many banks at once. The bank-level controls represent measured differences that may make one institution more likely than another to expand or restrict lending or to report a particular level of profitability. Bank fixed effects then absorb persistent differences that do not change month to month.

The researchers run separate first-stage exercises for lending growth and profitability. For lending, eight fixed-effects specifications progressively combine the macro variables with liquidity, borrower concentration, profitability, depositor concentration, and the lending–deposit spread. For profitability, the corresponding specifications use macro variables along with liquidity, lending growth, borrower and depositor concentration, and the spread. Each specification produces residuals: the part of the bank-level outcome not accounted for by the included controls and fixed effects.

The contrast in fit is informative. The lending models have overall R-squared values of about 0.1246 to 0.1254, but within-bank R-squared values of only about 0.0039 to 0.0048. The overall figure is influenced by persistent differences between banks, which the fixed effects absorb. The much smaller within-bank figure says that the observed controls explain little of the month-to-month movement in a given bank’s lending growth. For a forecasting model, that would be a warning.

Published by DeepMox

FAQ

What does within-bank R-squared measure here?

It shows how much month-to-month variation within an institution is organized by the included controls; the lending models’ 0.0039–0.0048 values leave much movement unexplained.

Do residual diagnostics prove the shocks are causal?

No. Breusch–Pagan tests, residual–fitted correlations, and within-bank F-statistics check limited statistical properties; they do not test instrument strength or all omitted events.

Why did the study use eight first-stage specifications?

They vary the included bank observables, so stability of the later relationship shows it is less dependent on one control choice, not that the underlying event is directly observed.

Previous10 minHow Mauritania’s Bank Shocks Reach Credit

Next9 minMauritania’s Lending Shocks and Aggregate Credit

Measuring Mauritania’s Bank Shocks

What the Evidence Is Actually Measuring

From bank-month records to two shock series

FAQ