The aim of this study was to highlight how variability in roller compacted ribbon quality can impact on NIR spectral measurement and to propose a simple method of data selection to remove erroneous spectra. The use of NIR spectroscopy for monitoring ribbon envelope density has been previously demonstrated, however to date there has been limited discussion as to how spectral data sets can contain erroneous outliers due to poor sample presentation to the NIR probes. In this study compacted ribbon of variable quality was produced from three separate blends of microcrystalline cellulose (MCC)/lactose/magnesium stearate at 8 Roll Force settings (2–16 kN/cm). The three blends differed only in the storage conditions of MCC prior to blending and compaction. MCC sublots were stored at ambient (41% RH/20 °C), low humidity (11% RH/20 °C) and high humidity (75% RH/40 °C) conditions prior to blending. Ribbon envelope density was measured and ribbon NIR spectral data was acquired at line using a multi-probe spectrometer (MultiEye™ NIR). Initial inspection of the at-line NIR spectral data set showed a large degree of variability which indicated that some form of data cleaning was required. The source of variability in spectral measurements was investigated by subjective visual examination and by statistical analysis. Spectral variability was noted due to the storage conditions of MCC prior to compaction, Roll Force settings and between individual ribbon samples sampled at a set Roll Force/Blend combination. Variability was also caused by ribbon presentation to probes, such as differences in the presentation of broken, curved and flat intact ribbons. Based on the subjective visual examination of data, a Visual Discard method was applied and was found to be particularly successful for blends containing MCC stored at ambient and low humidity. However the Visual Discard method of spectra cleaning is subjective and therefore a non-subjective method capable of screening for erroneous probe readings was developed. For this data set a Trimmed Mean method was applied to set a limit on how data is cleaned from the data set allowing for the removal of a faulty probe reading (25% of data) or a poor sample (33% of data). The 33% Trimmed Mean reduced the impact of spectral variation or misreads between samples or probes and was found to be as successful as the Visual Discard method at cleaning the data set prior to development of the calibration equation.