Novel statistical methods to improve analysis of laser methane detector data

Methane represents 16% of global greenhouse gas (GHG) emissions with enteric fermentation by ruminant animals such as livestock estimated to generate 28% of all anthropomorphic methane (Beauchemin et al, 2008). Diet digestibility and genetic merit can affect methane produced with estimates for dairy cows reported to range from 278 to 456g CH4/day (Garnsworthy et al, 2012). The LMD is a nonintrusive method used to measure CH4 from livestock in the field on the assumption that it is a sufficiently accurate approximation to more arduous methods such as use of closed chambers. Here we investigate the accuracy resulting from various statistical methods of analysis of LMD data by comparison with measurements taken simultaneously in respiration chambers. Material and methods Eighteen Holstein Friesian heifers were measured over 20 days, with each heifer placed in a chamber for 2 days during which methane in parts per million (ppm) was measured by the chamber once every 4 minutes. For 6-8 intervals per heifer, most of 5-6 minutes each (Bruder et al, 2017), methane was also measured (twice per second) using the LMD. This resulted in LMD data from 140 intervals totalling 12.9 hours over 20 days. For each interval, data from the LMD were log transformed (Figure 1) and various statistical methods were used to smooth the LMD data for each interval and remove the impact of 0 value data measurements. Some of these methods obtained smoothed estimates of the probability density function (pdf) for each interval whilst others were based on the time series (Figure 2). Each method resulted in a summary statistic for each interval (such as estimated mean or area under curve) adjusted for interval length. The resulting summaries were compared with estimates of g CH4/day from the chambers for each heifer. These methods were also applied to data from thrice-weekly LMD measurement of 28 cows over 12 weeks in another study, to see whether the LMD could be used to detect differences between two diets. Results Several of the alternative statistical methods gave very similar results, with Pearson’s correlation coefficient between the average over the LMD summaries and the average g/day measured over 2 days by the respiration chamber of 0.6 (P<0.01). Of the methods investigated there was no evidence to suggest that those that take into account the time series nature of the data outperform those that do not. Whilst these correlations are not particularly high it should be born in mind that they arise from just 6-8 LMD tests mostly between 5-6 minutes long. Furthermore, there was evidence to suggest that the LMD was sufficiently accurate to detect differences in methane emissions resulting from different experimental diets.