Variability between individual transients in an MRS acquisition presents a challenge for reliable quantification, particularly in functional scenarios where discrete subsets of the available transients may be compared. The current study aims to develop and validate a model for removing unwanted variance from GABA-edited MRS data, whilst preserving variance of potential interest - such as metabolic response to a functional task. A linear model is used to describe sources of variance in the system: intrinsic, periodic variance associated with phase cycling and spectral editing, and abrupt changes associated with subject movement. We broadly hypothesize that modelling these factors appropriately will improve spectral quality and reduce variance in quantification outcomes, without introducing bias to the estimates. We additionally anticipate that the models will improve (or at least maintain) sensitivity to functional changes, outperforming established methods in this regard. In vivo GABA-edited MRS data (203 subjects from the publicly available Big GABA collection) were sub-sampled strategically to assess individual components of the model, benchmarked against the uncorrected case and against established approaches such as spectral improvement by Fourier thresholding (SIFT). Changes in metabolite concentration and lineshape simulating response to a functional task were synthesized, and sensitivity to such changes was assessed. Composite models yielded improved SNR and reduced variability of GABA+ estimates compared to the uncorrected case in all scenarios, with performance for individual model components varying. Similarly, while some model components in isolation led to increased variability in estimates, no bias was observed in these or in the composite models. While SIFT yielded the greatest reductions in unwanted variance, the resultant data were substantially less sensitive to synthetic functional changes. We conclude that the modelling presented is effective at reducing unwanted variance, whilst retaining temporal dynamics of interest for functional MRS applications, and recommend its inclusion in fMRS processing pipelines.