In complex particle-physics analyses, where signal and background overlap across multidimensional phase space, statistically consistent event-by-event weights are often needed to extract signal observables without bias from the background. A common strategy is to fit a mixture model for the contributing components in a discriminating variable, typically an invariant mass, and then use the fitted component shapes and normalizations to assign a signal weight to each event. In practice, widely used weighting methods can perform poorly when their assumptions are only approximately valid, in particular when the discriminating variable is correlated with the observables of interest, or when the calculation of the weights does not account for correlations among the fitted signal and background yields.

We assess the limitations of these standard techniques with particular focus on the $Q$-factor method, a local fitting method based on $k$-nearest neighbors. Although $Q$-factors offer greater flexibility than a single global fit, their local probability weights can bias spectator observables, variables not included in the discriminating-variable fit but whose signal distributions are to be determined, even when the underlying mixture fits appear well behaved. We introduce a corrected formalism, the $_sQ$-factor method, which retains the $Q$-factor neighborhood construction but replaces the local probability weight with an $_s\mathcal{P}$lot-style covariance-corrected weight computed from each neighborhood fit. Using toy Monte Carlo studies, we demonstrate that $_sQ$-factors improve the recovery of non-discriminatory signal distributions and reduce the bias in fitted physics parameters in comparison to standard probability-weight methods.