The peer-reviewed literature was systematically reviewed ( Figure 1 ) following published guidelines (41, 43) to select publications that included laboratory measurement of denitrification and N₂O production potentials in soil. The Scopus search was conducted in August 2021 with the search term ‘nitrous oxide’ and any combination of the terms ‘denitrification potential’, ‘denitrif* enzyme assay*’, ‘denitrif* enzyme activit*’, ‘DEA’, and ‘DNP’ in the title, abstract, or key words. This search yielded 270 results when limited to those available in English. These articles were screened by reviewing abstracts and methods to remove unrelated studies (e.g., biochemistry studies of diethylamine nonoate, abbreviated DEA/NO) or studies that did not include data necessary to calculate the N₂O ratio. Included studies met the following minimum criteria: (i) laboratory experiments on soils, (ii) nutrient (C and/or N) addition, and (iii) N₂O and N₂ measurements, either using acetylene inhibition or direct measurement. In addition, seven studies were identified and included although they did not show up in the original search.

The following data was recorded for each paper: location (country), soil source (e.g., grassland), soil type, soil texture class, soil bulk density, soil pH, soil amount used in experiment, added C and N types and amounts, inhibitors used, headspace gas, incubation time and temperature, and the N₂O ratio and error for treatments and controls. Standard conversions were used to report all data with the same units. Data was extracted from tables and text or from figures using GetData version 2.26 (44). From each study, different field treatments, sites, sample locations, dates, experimental treatments, or incubation times were counted as separate observations when replicates and standard deviations were also reported. Missing data for standard deviations and replicates were obtained via author communications or estimated (range of study observations divided by four, with replicates reduced to one observation). When multiple depths were reported, only samples from the topmost soil layer (approximately 0-15 cm) were considered. Soil textures were often estimated using any available information from publications, such as soil series, order, or description.

A total of 60 studies (657 observations) were included in the final database ( Supplementary Table 1 ). Of these, 14 studies (100 observations) also included controls (defined as experiments without nutrient additions) for complete inclusion in the meta-analysis. Studies without controls were still included in the database for use in selection of moderators for the meta-regression. The final database is available as detailed in the “Data Availability Statement” section below.

Moderators for use in model improvement testing (meta-regression) were selected based on a literature review, tree-based modeling, and correlation analysis. All three approaches assessed the effect of categorical and/or numerical study variables on the N₂O ratio. The literature review was completed by reviewing results and discussions from all publications in the database to identify factors identified by authors as influencing the N₂O ratio. Tree-based modeling and correlation analysis were performed using the full database (n=657) in R version 4.2.2 (45) with the added ‘randomForest’ package for tree-based modeling. With tree-based modeling, a random forest model was fitted to the data and the categorical and/or numerical variables that were the most significant features were assessed through variable importance analysis. For correlation analysis, the nonparametric Spearman’s Rank Correlation test (‘corr.test’ function in R) was used to measure the association between the outcome and moderators. As this procedure is based on ranks, it can handle variables with distributions other than normal.

The meta-analysis was performed in R version 4.2.2 (45) using the ‘metafor’ package (46). A random effects mixed model was used with and without moderators. The effect size was calculated for each observation using the raw mean difference and heterogeneity (tau-squared, τ²) was calculated using the DerSimonian-Laird estimate. Treatment was defined as the experimental addition of C and/or N, and control was defined as identical experiments without any nutrient addition. The variable of comparison between studies was the N₂O ratio, a value ranging between 0 and 1, where 0 indicates no N₂O production (all of the denitrification is complete) and 1 indicates only N₂O is produced. Therefore, the treatment effect (or effect size) was calculated as the mean difference between the N₂O ratio for the treatment and control.

The mean difference could range from -1 to 1, with positive values indicating that nutrient addition (treatment) increased the N₂O ratio and negative values indicating nutrient addition decreased the N₂O ratio. Bias of this meta-analysis was assessed using Spearman’s Rank Correlation between effect size and variance (47).

A meta-regression model was fitted to test moderators, both factors and covariates, that accounted for variability between studies and moderate treatment effect. The model was improved stepwise by adding one moderator at a time and assessing improvement to model fit ( Figure 2 ). Improvements to model fit were assessed using likelihood ratio tests (LRT) between the model with and without the added moderators.

The complete database included studies from twelve countries, encompassing a variety of climates, soils, field managements, and experimental conditions ( Table 1 ). Most studies and observations were based on the acetylene-inhibition technique, with only three studies and 11 observations in the full database from direct measurement (48–50). While the meta-analysis database was able to capture some of the variability of the full database, it was biased towards some sample types. For instance, sandy soils and those from wetlands were more dominant in the meta-analysis database compared to the full database. In addition, mean experimental C addition was an order of magnitude lower in the meta-analysis database.

The effect of variables on the N₂O ratio were assessed through a literature review, tree-based modeling, and correlation analysis, to identify moderators for use in model improvement testing (meta-regression). For the literature review, we began by compiling a list of variables that influence the N₂O ratio, based on conclusions from studies in the database. We then identified studies that deliberately manipulated experimental conditions to test the effect of those variables on the N₂O ratio. The most common experimental consideration was pH, for which studies considered 2-8 different pH levels for the same soil sample (18, 22, 51–55). The next most common variable was experimental nutrient addition (C and/or N), where most studies only tested the presence or absence of added nutrients (26, 49, 56–58). However, there were two studies that tested multiple levels and combinations of nutrient additions (21, 59). Overall, the literature review highlighted soil pH and experimental nutrient addition as common variables tested for their influence on the N₂O ratio.

In parallel, we used a tree-based random forest model to identify variables that were the most important at explaining the variability of N₂O ratio in the full database ( Supplementary Tables 2 , 3 ). Three splits were sufficient to explain variability without increasing model complexity unnecessarily. The first split was based on soil pH (split at pH 6.65), and the following two splits were based on soil texture (categorical) and added N (split at 0.065 mg N g^-1 soil). Therefore, tree-based modeling highlighted soil pH, texture, and experimental N addition as important variables for the N₂O ratio. As such, all other considered variables (e.g., incubation temperature, headspace gas, experimental C addition) were considered less important in explaining variability between N₂O ratios from studies.

We also applied correlation analysis, identifying numerical variables that correlate most strongly with the N₂O ratio in the full database ( Figure 3 ). The strongest correlation was observed for soil pH, with a correlation coefficient of -0.44. The strong negative correlation indicated that the N₂O ratio decreased with increases in soil pH. Although other variables exhibited significant correlations, pH was the only variable with a strong correlation to the N₂O ratio.

For the meta-analysis, the treatment considered was experimental nutrient addition. The meta-analysis was first conducted using a random effects mixed model without moderators. The model showed a significant, positive treatment effect of 0.30 ± 0.03 (P<.0001), indicating that nutrient addition significantly increased the N₂O ratio. Since nutrient addition could include C and/or N addition, the conclusion cannot be directly extrapolated to either nutrient addition individually. However, since all observations in the meta-analysis included N addition and only half included C addition, the nutrient addition treatment is biased towards N addition. This reflects the literature as well since all the studies in the full database added N but fewer (63%) added C. The range of added C was smaller in the meta-analysis database (0-2 mg C g^-1 soil) than the full database (0-72 mg C g^-1 soil). Meta-analysis bias was further assessed using correlation analysis between effect size and variance, which had a correlation coefficient of -0.12 (P=.23), suggesting overall bias was low.

After finding significant effect of treatments on the N₂O ratio across studies, improvements to model fit were assessed using LRT between the model with and without the added moderators ( Table 2 ; Supplementary Figure 1 ). Since pH was identified as important through tree-based modeling and correlation analysis, it was selected as the first moderator. In addition, texture was also tested as the first moderator since it showed up as an important variable in the tree-based model. The model was significantly improved (P<.0002) by adding either pH or texture as the first moderator. Further improvement was found by adding texture to the pH model (P<.0001), but not when pH was added to the texture model (P=.37). Next, nutrient additions and total nutrient levels were added one at a time to the pH and texture model. The model was only significantly improved by using added C (P=.01) or added N (P=.01) as the third moderator. No further significant improvement was found through the addition of more moderators. Each incremental model improvement yielded an equally significant model (P<.0001), but negligible change in τ² was seen for moderator addition beyond texture ( Table 2 ). Therefore, the model with texture as the only moderator was selected as the final model with a τ² of.0033, which was a significant improvement from the model without any moderators (τ²=.0394). This outcome signifies that soil texture best explained the variability between studies.

For the best model (texture as the moderator), there was a significant overall treatment effect (mean difference: 0.17; P<.0001). Therefore, adding a moderator to the model did not change the overall trend observed in the base model—nutrient addition increased the N₂O ratio. We also considered the treatment effects within subgroups (i.e., texture class). When separated out by texture ( Figure 4 ), some soil textures (i.e., sandy loam, clay loam, silty clay loam) exhibited more positive treatment effects, indicating a stronger positive influence of nutrient addition on the N₂O ratio. All other soil textures had treatment effects near zero. It should be noted, however, that when separating out treatment effects by texture, some comparisons were based on only one or two studies (i.e., silty clay loam, sandy clay loam).

As the interactions between variables could lead to confounding results, we also considered soil pH and C addition within the texture class subgroups ( Supplementary Figures 2 , 3 ). However, no clear trends were observed.

Soil pH showed up as significant in all three moderator selection methods, indicating that soil pH is strongly related to the N₂O ratio in soils, with higher pH soils likely have the lowest potential to produce N₂O. The general negative correlation between soil pH and the N₂O ratio observed in this meta-analysis is consistent with several denitrification studies that explicitly tested at least two different pH levels (18, 22, 51, 52, 54, 55, 60). This relationship has been previously linked to assembly of functional N₂O reductase, which is prevented at low pH, ultimately limiting reduction of N₂O (61). At least one soil seems to behave differently, with a higher N₂O ratio at pH 7.41 than at more acidic pH conditions (53). However, that soil had a native pH of 3.71, and its denitrification activity was greatly diminished by adjusting the pH to 7.41, so it is probably not representative of most soils. The most comprehensive of the considered studies tested eight different pH levels between 3-11 for three distinct pasture soils (natural pH 5.3-6.9) (18). They found the inverse relationship between pH and the N₂O ratio could be fit with a single logarithmic function for all 24 data points. Overall, the inverse relationship between soil pH and the N₂O ratio was verified here across diverse soil samples and experimental conditions. It would be helpful to test additional soil types at multiple pH levels to determine if the logarithmic function described in that previous study is generally applicable (18). If so, incorporating pH might improve predictive models for N₂O emissions from denitrification. For example, a recent study applied regression modeling to determine drivers for N₂O production in semi-arid soils and included pH (along with water content and nitrate) in the final model (35).

The meta-analysis showed a significant positive treatment effect, indicating that experimental nutrient addition of N and/or C generally increased the N₂O ratio. Studies used in the meta-analysis were biased towards N addition, so this nutrient had more weight in the resulting treatment effect. Indeed, several studies have found that experimental N addition increased the N₂O ratio (21, 24, 25, 49, 57–59). The effect of C addition was not considered as often, with a few studies finding an inverse relationship between C addition and the N₂O ratio (21, 26, 56) and one finding no significant effect (59).

Interpreting the effect of C addition is complicated by evidence that the C:N ratio can also influence the N₂O ratio (28, 29). However, neither the tree-based analyses nor the correlation analyses reported here identified C:N as an important variable for the N₂O ratio. Furthermore, the meta-regression was not improved by adding C:N as a moderator, although adding either C or N addition individually slightly improved the model. Inconsistencies in the way soil C and N are measured and reported in publications could have contributed to this discrepancy. Although experimental N additions were usually in the form of nitrate, the carbon additions varied. Differences in C source complicate comparisons of the C:N ratio, as some forms may be more bioavailable than others. Some experimental glucose additions have even been shown to affect organic matter mineralization, potentially mobilizing C and N from soil and making them more bioavailable (62). Furthermore, interactions between these nutrients may complicate the overall effect, as one study noted significant interaction effects between C and N addition (59). Therefore, the impacts of C addition on the N₂O ratio need further investigation, with particular attention to the possible influences of the C:N ratio and the form of C added.

The identification of soil texture as an important moderator of the N₂O ratio is a key contribution of this work. While the importance of soil structure (and to a lesser degree texture) for emissions of N₂O and other greenhouse gases has been studied (63), the effects of soil structure are generally thought to occur through its influence on soil moisture, anaerobicity, and exchange of gases with the atmosphere. However, the laboratory assays are designed to be as uniform as possible in these properties, usually using soil slurries, so it is striking that soil texture is still an important moderator of the N₂O ratio. This could suggest the importance of textural components that would not be impacted by these laboratory manipulations, such as the proportion of clay, silt, and sand. Another possible mechanism could be the influence soil texture and/or structure had on the microbial community before the soil was sampled, leading to differences in their laboratory potential.

To our knowledge, no studies have explicitly tested soil texture effects on the N₂O ratio. One study observed a higher N₂O ratio in sandy soil compared to loam and clay soil, but the sandy soil also had a lower pH, which could be responsible for the difference (49). Another study found that sandy soils had lower potential to produce N₂O, but they did not measure total denitrification, so the influence on the N₂O ratio could not be compared (64). A recent study quantified the effect of several categorical variables on global forest and grassland annual N₂O emissions from field-based studies and similarly identified soil texture as a significant driver (65). Further, they identified soil clay content to be an important driver for field N₂O emissions from forest soils (65).Our meta-regression strongly suggests that soil texture merits greater scrutiny. Considering the low numbers of observations and studies for many soil textures, research priorities should include i) consistent measurement and reporting of detailed soil texture data (percentages of clay, silt, and sand) and ii) studies specifically designed to compare the N₂O ratio across soils with different textures, controlling for important variables such as pH and nutrient content. Fortunately, texture has the benefits of being more affordable to measure and less temporally variable than many soil properties, making its addition to planned and future studies more feasible.