Bias when predicting crude protein, dry matter digestibility and voluntary intake of tropical grasses by near-infrared reflectance

Abstract

A study was made of the bias (non-random error) of regressions for predicting crude protein (CP) content, dry matter digestibility (DMD) in vivo and voluntary intake (VI) of forage from its near-infrared reflectance (NIRR) properties measured at 19 wavelengths. The regression was derived from 97 samples including six tropical grasses, leaf and stem fractions, pelleted and chopped forms and regrowths varying from 28 to 150 days. Bias associated with differences in feed type was proved to exist in regression equations for predicting nutritive value when it was shown that the residual errors of these regressions were significantly reduced by including the combined effects of grass species, fraction, form and age. Measures of the bias which might be expected were calculated for two contrasting situations. The internal measure of bias (IMB) was ‘estimated’ as a possible average bias of a feed type when predicting nutritive value for samples of similar types to those used when deriving the calibration equation, while the extrapolation measure of bias (EMB) was ‘estimated’ as that associated with prediction of the nutritive value for feed types not included in the calibration equation. Crude protein was estimated with a residual standard deviation (RSD) of ± 0.77% when regressed against the 19 reflectance values, but when information on grass species, parts and regrowth was also included in the regression the RSD was reduced to ± 0.46%. The maximum difference in the IMB between grass species was 0.80% crude protein, while for EMB it was 3.35% crude protein. When separated leaf and stem were considered the differences were larger. Dry matter digestibility was estimated with an RSD of ± 2.7% which decreased to ± 1.8% when grass species, age, form and study were taken into account. Grass species had only a small effect on IMB but the maximum difference for EMB between species reached 5.6%. The EMB indicated that the largest differences between deviations were associated with age of regrowth but this effect was absent with the IMB. Voluntary intake was estimated with an RSD of ± 7.2 g/kg W 0.75/day, which decreased to ± 2.5 g/kg W 0.75/day when grass species, form, part and age were taken into account. The largest bias was for pelleted and chopped forage with mean differences of 14.5 for IMB and 28.8 g/kg W 0.75/day for EMB, respectively. It was concluded that when nutritive value is predicted by NIRR the estimates may often be biased according to the grass species, plant part or physical form of the sample tested and that the pattern of bias may be complicated. In order to obtain accurate predictions from NIRR full account must be taken of this bias. This might be achieved by correcting the nutritive value estimates made from an NIRR regression, based on a large number of feed types, by addition of the bias associated with the feed type being estimated.