How to Conduct On-Farm Research
A primer in designing your own field experiments
By Dr. Jochum Wiersma, University of Minnesota small grains specialist
In the past, I’ve explained in this space why researchers use replications across years and locations when conducting yield trials.
In this article, I will expand upon that and hopefully arm you with more guidelines that you can use on your farm, should you choose to conduct your own strip trials or comparisons between fields or parts of fields that received different inputs (e.g., you split a field in two and only applied a fungicide on half the field).
Such comparisons can be extremely useful in determining whether certain inputs, products or practices are practical or profitable for you.
However, your in-field comparisons should be conducted methodically to ensure results are as accurate as possible, and that the conclusions you find are valid.
If you are conducting on-farm trials, you want to determine whether what you measure in the field is the result of chance, or caused by the difference in treatments you applied to the field. Only
then can you come to a valid conclusion (inductive reasoning), and derive general principles that may help guide crop management decisions on your farm. Statistics are used to distinguish whether something occurs by
chance, or if something occurs enough that you are able to derive a general conclusion or principal from your collected data or observations.
For example, if in 2001 you split a quarter of land in two equal halves, and Variety A on one half of the quarter yielded 60 bu/acre, and Variety B yielded 65 bu/acre, does that mean Variety B always yields 5 bu/acre more than Variety A? Most of you would argue “probably not,” and you are right.
Statistics—the science of data analysis— is used to derive a general conclusion from data collected from field research. When dealing with data that cannot be predicted precisely, you need to use
statistics if you want to induct some general principle from it.
Observations or data you collect from your farm generally is influenced by a number of variables you did not have any control over (e.g. the weather) and thus the collected data deviates and varies from what you expected.
Most statistics courses start out by asking the following question: “If I toss a coin a hundred times, how many times do you expect to see tails up?”
Probably everyone will answer 50 times, which is correct. What is the chance that you throw 49 heads and 51 tails? That chance is pretty high and almost the same as tossing exactly 50 heads and tails. If you take this further, what are the odds that you toss only 1 tail and 99 heads? Well it is possible, but the chance would be pretty slim.
With this coin toss example, two very important concepts in statistics have been introduced: chance and the normal distribution. Chance is difficult to define, but the above example of
the repeated coin toss illustrates it well; it is possible to toss a coin 100 times and have 99 tails, but one would be awfully lucky (or unlucky). Should you decide to repeat the 100-coin toss several hundred
times, you will likely end up with a different outcome each time, but most often you would be approaching the 50/50 mark and in a few instances you would get a 99 to 1 ratio.
The expected percentage of times you would reach each of the possible solutions is called the normal distribution or bell shaped curve. In the case of the 50 heads or tails, you would be on top of the bell shape curve. The 99 versus 1 is found on the outside of the curve.
So what does coin tossing have to do with agriculture and on-farm yield trials? (Some feel farming is like a coin toss each year, but that’s beside the point). The point is, when you make
observations in the field and you want to extract or more correctly induce useful management information that you can use the next year, you will need to do some meaningful statistics to allow you to come to valid
Suppose you decided to split a field in half and apply fungicide to wheat on one half of the field and not on the other half of the field. The treatment
in this case is the application of fungicide and the unit we used to measure the effect of the fungicide is the half of the field. So the field has been split in two experimental units or plots
(the treated and untreated part of the field). The observations
in the field are the result of a directed input (the fungicide) and a slew of uncontrolled conditions (like small differences in soil fertility or soil type, presence of tire tracks, topography, etc.).
Hence, the observation you made is a function of the applied treatment and also all the other known (and unknown) but uncontrollable conditions. This introduces experimental error. The
effect of the treatment (the fungicide) is confounded with the other uncontrollable conditions in the experimental units, and as farmer or researcher you cannot determine whether the observation you made (e.g.
yield) is a function of the treatment or experimental error. This is why basic rules of conducting research (The three basic principles are replication, randomization, and blocking), even your
own on-farm research, are so important to follow.
On-Farm Research Tips
A few simple steps will allow you to design a good experiment, separate experimental error from the treatments’
effects and answer the questions you have. Following are tips for conducting on-farm research, courtesy of Emerson Nafziger, professor of crop sciences at the University of Illinois, and The Sustainable Agriculture
Research and Education (SARE) program, a USDA-funded initiative that sponsors sustainable agriculture research.
Nafziger has a backgrounder on how to conduct on-farm research that can be found online at: www.farmresearch.com/infoag/presentations/nafzigeremerson.pdf.
The SARE also has a publication on how to conduct on-farm research (the source for the preceding plot design examples, which also includes information on how to conduct livestock trials) online at www.sare.org/onfarm99/onfarm99.pdf
Keep it simple, especially at first: Limit your project to a comparison of two or three treatments. As you gain confidence, try something a little more challenging. Don’t try to answer a lot
of different questions with the same project: for example, it is probably not reasonable to try to figure out how six different varieties respond to both plant population and N rate in different parts of a variable
Seek help: Key times for professional assistance are at the design stage and then again when analyzing your data.
Extension educators, agronomists, and seed and chemical company representatives are good contacts for on-farm research help, and may even want to cooperate with you in providing inputs or material for you to analyze.
Replicate and randomize: Plan on enough field space to do more than one strip of each treatment being tested. Mix treatments within blocks.
Stay uniform: Treat all the plots exactly the same except for the differing treatments. If possible, locate your experiment in a field of uniform soil type.
Harvest individual plots: Record data from each individual plot. Don’t lump all treatment types together or you’ll lose the value of replication.
Remain objective: The results may not turn out as you hoped or planned. Be prepared to accept and learn from negative results.
Repeat the same research project multiple years: Weather is never the same from year to year. Repeat your experiment until you are comfortable with the results under varying conditions.
Don’t under-estimate qualitative or unexpected results: Look for changes outside your test parameters that may result from your experiment.
Manage your time wisely: Expect to devote extra time to your research during busy harvest seasons. Make sure you can carry out your experiment even though you are busy, or get extra help.
Look for ways to cooperate:
Cooperation with neighbors, organizations, companies, universities or ag professionals can make the work more productive, both by increasing the number of sites and also by encouraging discussion.
I would add that it is generally useful to choose the smallest experimental unit size that is workable for your research. Larger plots may mean that you introduce more, uncontrollable, experimental
error. Smaller experimental units replicated more often generally allow for better statistical comparisons and detection of smaller treatment differences. And I would emphasize again to ask for help.
If you plan to do on-farm testing, ask someone for help if you are not sure about how to design and analyze a good experiment, or not sure how to compile or analyze results.
Common On-Farm Plot Designs
Two experimental designs commonly used by researchers are the Randomized Complete Block design and the Split-Plot design, and both are well suited for on-farm research.
In the randomized complete block design, the experimental units or treatment plots are put together and randomized within replicated blocks. The following example shows how a trial testing three
nitrogen rates (0, 80, and 160 lbs/A), each replicated three times, might be laid out in a randomized complete block design.
Another popular and useful design for on-farm researchers is the split-plot design. This allows you to test different factors and how they might interact. For example, to determine how much you
can reduce nitrogen in wheat after a hairy vetch cover crop, the following split-plot design could be used in the following manner. As the name implies, split-plot design is a two-step process.
In this case, you set up your main plot (vetch, no vetch) then overlay a second treatment (N rates).
Treatments can be laid out in strips, with length of the plots determined by the length of the field and width by the equipment you use.
Randomized complete block design
0 160 80
160 0 80
80 0 160
0 80 160
80 160 0
160 0 80
160 80 0