Meta-analysis in agricultural and food sciences

Evidence synthesis in agriculture and food

Agricultural science depends on replicated experiments. A practice such as no-till farming or a feed additive is tested in many places and seasons, and the results vary with soil, weather and management. Meta-analysis is a natural way to summarize this evidence and to find out under what conditions a practice works. It has been used for studies of crop rotation, organic and conventional farming, fertilizer use, irrigation, livestock nutrition, post-harvest storage and food processing.

Philibert, Loyce and Makowski assessed the quality of meta-analyses in agronomy in 2012 and found weaknesses in reporting and in the handling of heterogeneity and dependence. Their findings, and the book by Makowski, Piraux and Brun on methods for experimental networks, guide how reviews should be done. In animal science, Lean and colleagues have described meta-analysis for animal health and production.

The approach we use follows meta-analysis and systematic review, with the adjustments for agricultural data.

Yield ratios and a worked pooling example

Yield is on a ratio scale and responses are usually proportional: a treatment raises yield by a percentage that is similar in high and low-yielding sites, more than by a fixed number of tonnes. The log response ratio, the log of the treatment mean divided by the control mean, is therefore the standard effect size, and it back-transforms to a percentage change. Absolute differences in the original units are useful when the units are common and the baseline varies little.

The table shows four invented trials, with treatment and control mean yields in tonnes per hectare, the log response ratio and its approximate variance. The numbers show the method and are not data from a real review.

With inverse-variance weights the pooled log ratio is 0.089 (standard error 0.050), which corresponds to a yield ratio of 1.09, an increase of about 9 percent. In a real review a random-effects model would normally be used because the true response varies between sites, and the prediction interval, not only the confidence interval, would be given, since a grower wants to know the range of responses to expect at a new site.

Trial networks, site-years and mixed models

Agricultural trials are usually reported as a table of site-years, with several treatments and several blocks. A single paper may include a dozen site-years that share a protocol and a research team. A fixed-effect or simple random-effects model that treats each as independent overstates the information. Mixed models with random effects for study and site, and robust variance estimation, address this. When trials share a control or a common check variety, the correlation between effects should be modeled.

A second feature is the network of experiments. Many studies test different subsets of treatments, such as varieties or fertilizer rates. Network meta-analysis, adapted to this setting, estimates all comparisons at once when the trials are connected. Makowski and colleagues discuss mixed models for such networks. The same assumptions as in medical networks apply, including similarity of the trials in factors that change the response.

Dose is also central. Response to fertilizer, water or feed rates is usually curved, with diminishing returns at higher rates, so dose-response models (for example, a quadratic-plateau or a log-linear form) fit better than a single comparison of high and low. The review should state how it chose the functional form.

Context: soil, climate, management

The practical question for a farmer is whether a practice will work on their land. Moderator analysis relates the response to soil texture, organic matter, pH, climate, crop type, tillage and the starting yield. Meta-regression is the tool, with the usual caution about correlated moderators and about study-level data. A finding that no-till reduces yield in humid regions but not in dry ones may reflect crop type or soil as well as climate. Showing the distribution of studies across moderator values lets readers see where the evidence is thin.

Responses also depend on the yield level of the control. Studies with low-yielding controls often show larger proportional gains, and the log response ratio may be related to the control mean. Including the control yield as a covariate, or analyzing the relationship, helps.

For food science, the same logic applies to processing parameters such as temperature, time and ingredient concentration. Experimental protocols differ in ways that matter, and the review may need to harmonize units and methods before pooling.

Animal science and food studies

Livestock nutrition and health trials often have small groups per treatment, with the pen or herd as the experimental unit. Using the number of animals, not the number of pens, as the sample size inflates precision. Reviews should check the unit of analysis and adjust where needed. Many studies report only the standard error of the mean, not the standard deviation, so conversions need care. Outcomes include growth, feed conversion, milk yield, disease incidence and reproductive performance, and the effect size follows the outcome: ratios for continuous outputs and odds or risk ratios for events.

Food studies include sensory panels, shelf-life studies and nutrient composition. Measurement methods vary between laboratories, which adds heterogeneity that should be modeled and not ignored. Observational nutrition studies in humans follow the methods described for the health sciences, with MOOSE reporting.

Bias and reporting

Publication bias occurs in agricultural science as elsewhere, and trials with negative or null results are less likely to be reported in journals. Reports from extension services, theses and trial databases are valuable sources. Industry-funded trials, common for feeds, pesticides and varieties, may differ from independent trials, and funding source should be coded and tested. Reporting follows PRISMA 2020, with a table of every trial, the data for each effect size and the code, so that others can reproduce the results.

Specialties and sub-fields

Sub-fields differ in designs and outcomes. Pages for sub-fields are added as they are completed.

Reading heterogeneity and the prediction interval

In agricultural meta-analysis, heterogeneity is not a flaw to be removed. Responses differ because soils, seasons and management differ, and the purpose of the synthesis is to describe that variation. The between-study variance on the log scale shows how much the true proportional response varies, and the prediction interval translates it into the range of responses a grower might see at a new site. A practice with a mean gain of 8 percent and a prediction interval from a 6 percent loss to a 24 percent gain is a different proposition from one whose interval runs from 5 to 11 percent, although the means are the same.

Reporting I-squared alone is not enough in this field. It depends on the precision of the studies and tends to be high when the studies are large, as in multi-site networks. Better practice is to report the between-study standard deviation, the prediction interval and the proportion of variation explained by the moderators in the meta-regression. When the moderators explain little, the review should say that site-level variation is not yet understood, which is itself useful for planning further experiments.

Sustainability trade-offs and multiple outcomes

Farming practices affect several outcomes at once: yield, profit, soil carbon, water quality, biodiversity, greenhouse gas emissions and labor. A review that looks only at yield may miss trade-offs. For example, a practice that raises soil carbon could slightly reduce yield, or a lower-input system could reduce yield and increase biodiversity. Syntheses of such trade-offs report the effect on each outcome with its uncertainty and, where possible, the studies that measured several outcomes together, so that the pattern of trade-offs can be examined within studies and not only across the literature.

The conclusions should avoid combining outcomes into a single score without a clear weighting rule. Weighting is a matter of values, and different stakeholders will weight differently. Presenting the outcomes side by side lets users apply their own priorities. Cost and profitability data, where reported, are much more variable than yield data, because prices and labor costs differ, and are best presented with their local context.

Searching agricultural literature

Searches use CAB Abstracts, AGRICOLA, Web of Science, Scopus and FSTA (Food Science and Technology Abstracts), with PubMed for health-related food studies. National research institutes and extension services publish reports that are not indexed in these databases, and theses from agricultural universities hold many trials. Language matters, because much agricultural research is published in Chinese, Spanish, Portuguese, French and other languages; the review should state its language limits. Searches use many synonyms for crops, practices and outcomes, and a documented strategy is needed for repeatability.

Data are often in tables of site-years or in figures, and extraction needs care for units (tonnes per hectare, kilograms per hectare, bushels per acre), moisture content for grain and the definition of the control. Two extractors and a plan for resolving differences are standard. Where a trial reports the standard error of the mean, the review converts it to a standard deviation with the number of replicates, and where only a least significant difference is given, a stated conversion is applied and tested in a sensitivity analysis, so that readers can see whether the conversion affects the pooled result.

How we support research projects in this area

Boundaries of this service

A meta-analysis summarizes trial results across sites. It does not predict the yield or animal performance on a particular farm, and responses depend on local conditions and management. We do not provide farm-specific recommendations, veterinary advice for individual animals or regulatory opinions.

If trials are too different to pool, we recommend a systematic review or a systematic map and say so at the start.

Frequently asked questions