The O2 uptake (V̇o2) data at the onset of an exercise are usually fitted with a mono-exponential function, after removal of the data pertaining to a conventional initial time period (ΔTr) lasting ∼20 s. We performed a thorough quantitative analysis on the effects of removing data pertaining to different ΔTr, aiming at identifying an objective method to establish the appropriate ΔTr. Breath-by-breath O2 uptake responses, acquired from 25 healthy adults performing a step moderate-intensity exercise, and 104 simulated biexponential responses, were analyzed. For all the responses, the kinetic parameters of a mono-exponential function and the corresponding asymptotic standard errors (ASEs) were estimated by nonlinear regression, removing the data pertaining to progressively longer initial periods (1 s each) up to 60 s. Four methods to establish objectively ΔTr were compared. The minimum estimated τ was obtained for ΔTr ≅ 35 s in both the V̇o2 and simulated data, that was about 30% lower compared with that obtained for ΔTr ≅ 0s. The average ASE values remained quite constant up to ΔTr ≅ 35 s, thereafter they increased remarkably. The τ used to generate the simulated response fell within the confidence intervals of the estimated τ in ∼85% of cases for ΔTr = 20 s ("20 s-w" method); this percentage increased to ∼92% of cases when ΔTr was established according to both the minimum τ and its narrowest confidence interval ("Mixed" method). In conclusion, the effects of removing V̇o2 data pertaining to different ΔTr are remarkable. The "Mixed" method provided estimated parameters close to those used to generate the simulated responses and is thus endorsed.NEW & NOTEWORTHY We propose a method to objectively establish the initial time period to be removed from the fitting window when, using a mono-exponential model, the kinetics of the fundamental component is determined on breath-by-breath O2 uptake data collected at the onset of a moderate-intensity exercise. Innovative statistical parameters ("Coverage" and "Concordance5%," applicable on simulated responses) were used to compare its performance with that of other three methods. The proposed method yielded the best "Coverage" and "Concordance5%."