Abstract Various coalescence methods for Lagrangian microphysics schemes are tested in box and large-eddy simulation (LES) models, including the stochastic all-or-nothing (AON) superdroplet method (SDM) and a deterministic version of SDM (dSDM) that applies a fractional approach similar to the average impact method. In LES, variabilities driven by microphysics and by flow realizations are separated using the “piggybacking” technique. Rain initiation averaged over many realizations of the box model is delayed, and rain variability increases as the number of superdrops per collision volume NSD is decreased using SDM. In contrast, rain initiation time using SDM in LES is insensitive to NSD for 32 ≤ NSD ≤ 512. This is explained through the interaction between LES grid boxes, each acting as a separate collision volume. Variability across the ensemble of LES collision volumes using SDM results in rain quickly initiating in some of the LES grid cells at low NSD and leading to a similar overall timing of rain initiation from the cloud compared to simulations with high NSD. There is a ∼20% decrease in the total rain mass and mean rain flux as NSD is increased from 32 to 256, with little additional change as NSD is increased from 256 to 512. The fractional coalescence approach in dSDM leads to reduced microphysical variability and a 15–18-min delay in rain initiation compared to SDM. An additional LES ensemble with microphysical variability feeding back to the dynamics shows that flow variability dominates the impact of microphysical variability on rain properties. Thus, flow variability must be constrained to isolate impacts of microphysical variability. Significance Statement An emerging tool in cloud microphysics modeling represents drops by computational particles called “superdroplets” that evolve in the modeled flow. Recent studies have documented that different superdroplet models produce widely varying predictions of rain. Understanding these differences is an important step in the wider adoption of these models by the community. Here, we examine different methods for representing droplet collision-coalescence in a superdroplet model and sensitivity to the number of superdroplets employed. The timing of rain initiation is insensitive to the superdroplet number in 3D cloud simulations, but rain is substantially delayed using a coalescence method that limits random variability in droplet collisions. We also show that flow variability must be constrained to isolate impacts of microphysical variability on rain properties.