Linear equality restrictions derived from economic theory are frequently observation-varying. Except in special cases, Restricted Least Squares (RLS) cannot be used to impose such restrictions without either underconstraining or overconstraining the parameter space. We solve the problem by developing a new estimator that collapses to RLS in cases where the restrictions are observation-invariant. We derive some theoretical properties of our so-called Generalised Restricted Least Squares (GRLS) estimator, and conduct a simulation experiment involving the estimation of a constant returns to scale production function. We find that GRLS significantly outperforms RLS in both small and large samples.