Statistics and On-Farm Testing Peanut Notes No. 38 2019

— Written By

Many people are involved in on-farm testing and there are many views on how that should be done. John Havlin and I teach Soil-Crop Management Systems at NC State University and we discuss the role of statistics in making good decisions on whether or not to change a production practice. I have attached those slides here and at some point will record that lecture to go with the slides. It is extremely important to have an appreciation for the variation that is considered uncontrolled or random in trials so that one knows whether or not a difference among the treatments being compared occurred simply by random chance. Ultimately, with on-farm testing (any testing for that matter) we want to use the data collected to make a recommendation on a broader scale (which can have significant economic ramifications.)  For that reason, we must use care in how we interpret data and make a recommendation. And one of the most important ways to make sure we make a solid interpretation that carries real meaning and usability is to set up the experiment well and have adequate controls or checks (non-treated controls and the commercial standard, for example) in place. There are other important elements, but randomization and replication are extremely important. Without these one does not have an estimate of random variation (error variance is a term often used).

Comparing fields is the easiest way to conduct a trial. But in essence it is the least valuable way (unless you have many many fields.)  Splitting a field in half is also relatively easy to do but unless you do that in many many fields you are very much limited. In both of these cases, you might have results that reflect a difference in sides of fields or fields that have nothing to do with the treatments you are comparing. You could make an erroneous conclusion about treatments using that approach. One might even have results suggesting there are no differences among treatments but the variation from field history or sides of fields masks the differences that might actually be “real.”  One really needs 10 or more fields (I am sure this too can be hotly debated) if they are going to make any meaningful conclusions using these approaches. Having randomized strips (at least 3 of each treatment) across the field is the very best way to approach this (and fewer fields are needed – perhaps saving resources and time.)  This takes extra effort and there will always be pushback because of logistics. I have said “uncle” plenty of times with on-farm work and even research station work because of logistics and time required to do it correctly. When it is all said and done, the numbers I got were generally difficult to interpret and I always had a nagging feeling that I was comparing sections of the field (and their history) and not the treatments I initially set out to compare. In essence, I have really nothing and I have wasted time and resources. This is when the old adage (a Louisiana Proverb from my first job out of graduate school comes to mind) – “There is never enough time to do it right but there is always enough time to do it over.”

There is a lot in the attached items. Call anytime if you have questions about this topic.

2019 462 Statistics Overview

2019 462 Statistics Summary

2019 462 Statistics Hypothesis Testing