For in vivo MRS spectral fitting, a baseline is often used to model background signals. The existing algorithms including LCModel rely on the linewidth to distinguish metabolite peaks from the background signals. In this work, we show that the fitted short-TE baseline strongly depends on metabolite linewidth due to large baseline-metabolite covariances. This dependence negatively affects metabolite quantification using short-TE MRS, resulting in large errors in metabolite concentrations. We also demonstrate that this baseline problem can be largely eliminated using 1D JPRESS which benefits from its substantially reduced background signals.