Meta-analysis and evidence synthesis for plant sciences

Evidence synthesis in plant sciences

Plant scientists ask how plants respond to light, water, temperature, carbon dioxide, nutrients, salinity, herbivores, pathogens and neighbors, and how traits vary within and between species. Meta-analysis has been used to summarize the response of growth and yield to elevated carbon dioxide, drought and warming; the effect of nitrogen and phosphorus addition; the benefits of mycorrhizal fungi; and the trade-offs between traits. Because experiments are small and use many species, pooling is often the only way to see general patterns.

The evidence has features that call for care. Experiments are often run in pots with limited soil volume, which restricts responses. Treatments range widely in magnitude. A single study may include many species, genotypes and treatment levels, with shared controls. Plant outcomes are measured at different stages. Our methods follow meta-analysis and systematic review practice, adapted to these features. This page builds on the general guidance for life sciences, and shares methods with the ecology page.

Treatment levels, doses and comparability

Because treatments are not standardized, effects cannot be interpreted without knowing the dose. A review should code treatment levels in common units, plot effect against dose and distinguish studies with a single level from those with graded levels. Where non-linear responses are plausible, such as optimum curves for nutrients, a linear meta-regression can mislead and flexible models are more appropriate. The control condition also matters: ambient conditions in a growth chamber differ from field conditions.

Shared controls, genotypes and dependence

A study that tests several genotypes against one control, or several treatment levels with one control group, produces effects that share the control sample. Treating these as independent effects overstates information. Approaches include splitting the control sample across comparisons, modelling the covariance between effects with a shared control, or using multilevel models. Effects for several species in one experiment are correlated, and species share evolutionary history. Multilevel models with random effects for study, species and phylogeny, and robust variance estimation, address these dependencies. We state the structure of the model and report the variance at each level.

Pots, chambers and field conditions

Most plant experiments are done in controlled conditions on young plants. Pots restrict roots and nutrients, chambers differ in light spectrum and humidity, and seedlings respond differently from mature plants. Field experiments, including free-air carbon dioxide enrichment sites, provide more realistic conditions but are fewer and costly. Reviews have found that responses of plants to elevated carbon dioxide are generally smaller in field settings than in pots. A review should code setting, plant age, rooting volume and the number of growing seasons, and compare effects by setting. Conclusions about crops in farmers' fields should rest on field studies, and findings from controlled environments should be labeled accordingly.

Crop yield and agronomic outcomes

Crop yield studies are field trials with replicated plots, where outcomes are yield per area, harvest index or quality traits. Effect sizes are often log response ratios or percentage differences. Yields are influenced by site, season, soil, cultivar and management, so heterogeneity is considerable, and pooled averages hide large differences between locations. Reviews should code crop, cultivar, site, climate, soil, management and year, and analyze them as moderators. Different studies compare to different baselines (conventional versus organic, tilled versus no-till), and the definition of the comparator is a key part of the review. See also the agriculture and food sciences area for farm-level questions.

Traits, trade-offs and comparative studies

Comparative studies examine relationships among traits across species, such as leaf mass per area, wood density and seed size. Species are not independent, so phylogenetic comparative methods are used. Meta-analyses of trait relationships pool correlations or slopes across studies, and they must handle differences in how traits are measured and the sets of species included. Trait databases offer standardized data, and when these are used the review should describe how records were screened. Relationships may differ between biomes, and extrapolating from temperate species to tropical ones needs caution.

Plant-microbe and plant-animal interactions

Studies of symbionts, pathogens, herbivory and pollination report effects on plant growth, defense and reproduction. Effects depend on the identity of both partners and on environmental conditions, so averages across many combinations can hide strong positive and negative effects. A review should show the distribution of effects, test host and partner traits as moderators and avoid concluding that an interaction is generally beneficial when the average is positive but a substantial share of effects are negative. Plant disease management and diagnosis are outside the service.

Publication bias and small-study effects

Plant experiments are often small, with three to six replicates, and studies with strong effects are more likely to be published. We use funnel plots adapted to multilevel data, tests that account for the structure of dependence and comparisons of published with unpublished studies. Reports from agricultural stations and theses are important sources of field trials that are not in journals.

Combined stresses and interactions between factors

Plants in nature meet several stresses at once, such as heat with drought, or elevated carbon dioxide with nutrient limitation. Factorial experiments test whether the combined effect equals the sum or product of the single effects, or departs from it. Meta-analysis of interactions needs the means of all four cells of a two-by-two design with their variances, and it is demanding because many studies report only main effects. A review should state which interaction model it uses, additive on the raw scale or multiplicative on the log scale, since the conclusion about synergy or antagonism depends on the scale chosen. With few studies reporting the full design, the evidence on interactions is thin, and the review should say so.

Combined-stress results from controlled environments are especially sensitive to the timing and order of the stresses, which are seldom reported in a standard way. We code them as moderators when available.

Developmental stage, duration and acclimation

Plant responses change with age and exposure time. Seedlings may respond strongly to elevated carbon dioxide in the first weeks, while the response of mature plants may be weaker as nutrients become limiting. Photosynthetic rates can decline after long exposure as plants acclimate. A review should code the stage at treatment and measurement, the duration of exposure and the number of growing seasons, and plot effects against duration. When most studies are short, conclusions about long-lived plants, such as trees, should be cautious. Perennial crops and forests are better informed by long-term field studies, which are few, and the review should tabulate them separately, noting the number of sites and years behind each result.

Common pitfalls we look for

• Pooling treatments of very different magnitude without dose analysis.
• Treating comparisons that share a control as independent.
• Applying pot-experiment results to field crops.
• Using linear models for optimum-type responses.
• Ignoring species relatedness.
• Reporting an average when effects split in sign.

Planning a plant sciences synthesis

We help define the plant group, the treatment and the outcome, plan searches in Web of Science, Scopus, CAB Abstracts and AGRICOLA, and set up coding of species, genotype, treatment level, setting, plant age, duration and outcome measure. Reporting follows PRISMA and, for ecology-style syntheses, the PRISMA-EcoEvo extension described under PRISMA extensions. See the meta-analysis service for scope and process.

An invented example of a response ratio

Suppose control plants produce an average of 10.0 grams of dry biomass and plants in a higher-carbon-dioxide treatment produce 12.5 grams in an invented experiment. The response ratio is 1.25, and its natural log is about 0.22, which corresponds to a 25 percent increase. If a second experiment on a different species finds 10.0 and 10.5 grams, the response ratio is 1.05 and the log about 0.05. Averaging the log values gives about 0.14 and back-transforming gives roughly a 15 percent increase. This hides the fact that one species responded strongly and the other barely, so the review would show both and test whether species traits explain the difference.

Coding and transparency

Plant data are often in figures, which we digitize and check with a second coder. We record sample sizes per treatment (replicates, not plants within plots, where plots are the experimental unit), the form of the variance reported, and the species name matched to a taxonomic database. Data and code are shared with the final report.

How we support research projects in this area

Boundaries of this service

A plant sciences synthesis describes average responses across published experiments. It does not provide crop management advice, breeding recommendations or plant disease diagnosis. Many studies are short, small and conducted in controlled environments, so results may not apply to farmers' fields or to species not studied.

Frequently asked questions