Irrigated agriculture accounts for about 80% of total agricultural production in Pakistan, an arid country home to the sixth largest population globally. Almost all irrigated production takes place in the Indus Basin Irrigation System, the world's largest contiguous irrigated area. The extensive system of canals transfers water from the river system to distributaries and from there to watercourses, farms, and eventually to individual farm plots. Given growing water shortages and the resulting increased variability in water flows in the system, it will be important to improve system efficiency to ensure continued increase in agricultural production. A key question remains, however, where to intervene in the system. This paper uses multilevel modeling to explore the relative importance of plots, farms, watercourses, and distributaries on the levels and variance of wheat yield, and identify entry points to improve agricultural water management in a way that can guide prioritization of investments across the different levels of the irrigation network. Our findings suggest that out of the total variation in wheat yield that is not explained by plot specific characteristics, 28% is explained by distributary level effects, 3% by watercourse level effects, and 7% by farm (household) characteristics. These findings suggest that investments that improve irrigation water delivery at the distributary level provide the highest impact per unit of dollar invested in terms of reducing yield gaps among farmers.