When data lie: correcting a conflict variable reverses a significant finding in Chad's agricultural time series (1961–2024)

Abstract Data sometimes lie — not out of malice, but through the silent accumulation of errors in files passed from researcher to researcher without documentation. This paper offers a concrete, quantified demonstration of this problem through the Chadian case. By systematically verifying four data sources used to analyse the effect of climate variability and armed conflict on agricultural yields in Chad (1961–2024), we discovered that the conflict variable was miscoded for 32 of the 64 years studied — exactly half the period. Once corrected against the UCDP/PRIO Armed Conflict Dataset v26.1, a result that appeared robust — a positive and highly significant conflict effect on millet yields (β = 0.213, p = 0.007) — disappears entirely (β = 0.032, p = 0.472). The correctly specified analysis instead reveals a robust and differentiated SPI effect on the main rainfed crops: positive and significant for millet (β = 0.123, p < 0.01) and sorghum (β = 0.062, p < 0.05), null for maize — consistent with agronomic theory. Estimation relies on Newey-West standard errors, appropriate for the mono-country time series framework. These findings carry two conclusions: one empirical, on the differentiated climate sensitivity of Chadian agriculture; the other, perhaps more valuable, on the necessity of systematic data verification in applied Sahel research.