Search strategy

The following web databases were searched for relevant documents: ISI Web of Science, Science Direct, JSTOR, Google (100 first hits), Google Scholar (100 first hits). Searches were conducted in English, French and German using translations of the following logical search string: (mowing OR cutting) AND (meadow OR grassland) AND (biodiversity OR richness). The term “Europe” was not included in the search keywords as stated in the Review Protocol [41], because European studies that do not mention the term Europe may have been missed. Studies originating from extra European regions were later excluded from the review. Any apparently relevant citations or links were followed one step away from the original hit. In addition, national and international experts on the subject were asked for any relevant literature and unpublished data.

Article screening

All references retrieved from the web search were scanned at the title, abstract and full text filter level by a first reviewer. From the 367 initial references, 200 (randomly selected) were rescanned by a second reviewer in order to check for inclusion consistency. The following study inclusion criteria were used:

Inclusion consistency was checked with kappa statistics, and agreement between the reviewers was satisfactory (k = 0.81) [42].

Study quality assessment

All articles accepted met the requirements of category II-2 and above of the classification system of [43]. This allowed for both experimental and observational studies to be included, but excluded studies that provided only qualitative evidence.

Data extraction

Some studies reported more than one treatment (two or more delayed cuts) or more than one type of measurable outcome (e.g. species richness and abundance, or different taxonomic groups such as plants and invertebrates). In these cases, all comparisons were recorded as independent data points, and this is why there are more data points (units of analysis) than articles [44, 45] (Figure 1; Table 1).

The following information was extracted from the studies for each data point: (1) taxon, (2) species richness or abundance, (3) standard deviation, (4) sample size, (5) study duration in years, (6) plot size of vegetation relevés or sampling methodology for invertebrates, (7) ordinal days of the early cut and (8) delayed cut, and finally (9) meadow type, classified as dry, mesophilous or wet. Additional potential sources of heterogeneity were also extracted such as fertilizer application, number of cuts per year, grazing activity, and biogeographical region where the study was carried out. Diversity indexes such as the Shannon index were recorded when present, but did not lead to sufficient data points for a meta-analysis (MA).

Taxa were plants, invertebrates or a specific group of invertebrates. Standard deviations (SD) were usually retrieved from standard errors (SE) or variances. If no estimate of variance was provided, we requested it from the original authors. If original authors could not provide SD, or sample size was equal to one (i.e. no variance), the corresponding study was included only in the unweighted analyses (see statistical analysis section below). The ordinal days (day 1 = January 1^st) of the early cut (control) and delayed cut (treatment) were used to calculate the number of days between the two mowing regimes. If the exact date of the early or delayed cut was unknown, but only the month was given, then the 15^th of the month was used for calculations. If the terms “early” or “late” in a given month were mentioned, then the 7^th or 24^th, respectively, of that corresponding month were used.

Delaying cutting is often studied within a broader context of agricultural extensification for biodiversity, including reduced number of mowing events, changes in fertilizer inputs and/or type of fertilizer, oversowing, etc. Studies of cases in which delaying mowing occurred in the presence of such confounding factors could not be included in the MA as the effect of delaying the first cut cannot be separated from these other confounding factors [e.g. [32].

Statistical analysis

Meta-analyses were conducted on three groups of studies according to their measurable outcomes: 1) plant species richness; 2) invertebrate species richness; 3) invertebrate abundances. Studies on plant species richness lasted between two and 40 years, and if multiple time-points were available along the time series, only the data for the last year (longest time period) were considered. Studies on invertebrates were usually shorter, mostly three to four years, and due to a high inter-annual variation, these studies often reported biodiversity responses averaged across the years. Here we used these reported average values.

where ${\overline{X}}^D$ and ${\overline{X}}^E$ are the means of the delayed and early cut outcomes, S is their pooled standard deviation, while the term J corrects for small sample bias [47]. It was calculated using the function escalc of the R package metafor [48].

Random- and mixed-effects models (mixed-effects models are random-effects models with covariates) were chosen as it is now common practice for this kind of analysis [47]. Under random- and mixed-effects models, the true effect size, i.e. the effect size as if there were no sampling errors, can vary from study to study, but usually do so under a normal distribution [49, 50]. Here the Q test and I ² statistic were used to assess heterogeneity between studies. The Q test is the test of significance, and the I ² statistic estimates how much of the total variability in the mean effect size (composed of heterogeneity and sampling error) can be attributed to heterogeneity among the true effect size [48, 50].

First, the null model was generated. Then all univariate models including the following moderators (effect modifiers) were tested: ordinal day, time lapse (in days) between the early and the delayed cuts, study duration (in years), meadow type and plot size of the vegetation relevés. Multivariate models (various combinations of the above mentioned variables) were also explored. Further subgroup analyses were conducted to investigate the influence of key moderators separately. Models were ranked based on their AIC values (Akaike Information Criterion) and on the level of significance of the estimates [51]. Permutation tests were not always possible due to an insufficient number of data points, which limits the number of possible iterations. Therefore test statistics of the effect sizes and corresponding confidence intervals (CIs) referred to the normal distribution (Z test). Publication bias was assessed using funnel plots, by applying a regression test for funnel plot asymmetry [46, 48].

Although less powerful than proper-weighted meta-analyses, this approach allows the inclusion of studies that did not report SD or where sample size was one, i.e. studies for which no Hedges’d could be calculated. Bootstrapping was used to calculate 95% confidence intervals (CI); if CI overlapped zero, the effect size was considered to be non significant. All statistics were performed using R version 2.13.0 [52].

Effects on plant species richness

Results based on the response ratio were qualitatively the same as the results based on the Hedges’d. Therefore, only the results of the weighted meta-analysis based on the Hedges’d are presented below due to their superior explanatory power. An additional file shows the results of the unweighted meta-analysis based on the response ratio [see Additional file 3].

In the null model, no overall significant effect of delaying the first mowing date was supported as regards plant species richness (mean Hedges’d = 0.017 with 95% CI −0.237 – 0.2716, z = 0.134, P = 0.882, Figure 2). However, heterogeneity between studies was significant (Q = 56.88, d.f. = 25, P < 0.001, I ²  = 54%), indicating that the true effect size does vary from one study to the next. With study duration (in years) included in the model as a moderator, no significant influence of that moderator on the effect size was discerned (slope = 0.016 with 95% CI -0.019 – 0.051, z = 0.878 P = 0.380, Figure 3a).

In further univariate models, a significant negative influence of the date of the early cut (control) was established (slope = −0.015 with 95% CI −0.025 – −0.005, z = −2.878, P = 0.004, Figure 3b). This means that the earlier the cut in the year, the more pronounced the effect on biodiversity of delaying the first cut. On the other hand, when the early cut occurred late in the season (July to August), delaying it had no, or even a negative, effect on plant species richness. Between studies heterogeneity was significant (Q = 43.12, d.f. = 24, P = 0.010), indicating again that other moderators may also influence the effect sizes. On the contrary, the date of the delayed cut did not significantly influence the effect size (slope = −0.007 with 95% CI −0.013 – 0.001, z = −1.805, P = 0.071), although it did explain some of the heterogeneity.

In order to further investigate this issue and to evaluate the extent to which heterogeneity can be explained by variation in this moderator (first mowing date), two subset MAs were conducted. The first included only the data points with an early cut in spring (before July 1) associated with a delayed cut in summer (July to September); the second included all other combinations of early and delayed cuts (spring to fall, early summer to late summer and summer to fall, but excluded one early spring to late spring study [58]). In the first case, mean Hedges’d became significantly positive (mean Hedges’ d = 0.388 with 95% CI 0.092 – 0.684, z = 2.569, P = 0.010, Figure 2b). Between studies heterogeneity was significant (Q = 24.88, d.f. = 14, P = 0.036), while I² (40%) was not. In the second case, mean Hedges’d became significantly negative (mean Hedges’d = −0.504 with 95% CI −0.763 – −0.246, z = −3.828, P < 0.001, Figure 2c). Heterogeneity was not significant (Q = 4.56, d.f. = 9, P = 0.871), indicating that these latter studies provided consistent results.

Note that none of the models including one or more moderators (study duration, mowing date, time interval between mowings, habitat type, and plot size of the vegetation relevés) performed better that the null model according to AIC values [Additional file 4]. In addition, no asymmetry was detected in any funnel plots, which rules out any publication bias effect [Additional file 5].

Effects on invertebrate species richness

A significant positive effect of delaying the first mowing date on invertebrate species richness was found (mean Hedges’d = 0.511 with 95% CI 0.129 – 0.893, z = 2.6217, P = 0.009, Figure 4). Heterogeneity was not significant (Q = 14.97, d.f. = 8, P = 0.060). No significant influence was found concerning the number of years during which a study was carried out (slope = 0.154 with 95% CI −0.074 to 0.382, z = 0.117, P = 0.186). No models including a moderator performed better than the null model according to AIC values [Additional file 4]. No asymmetry was detected in funnel plots [Additional file 5].

Effects on invertebrate abundance

Delaying the first mowing date had no significant effect on invertebrate abundance (mean Hedges’d = −0.053 with 95% CI −0.889 – 0.783, z = −0.1249, P = 0.901, Figure 5). However, the resulting Q-Q plot was not satisfactory, while the funnel plot showed a significant asymmetry in the distribution of the data points due to the two outlying studies of Morris [59, 60]. Excluding Morris’s studies from the analysis resulted in model assumptions and funnel plot becoming satisfactory, with a significant positive effect of delaying the first mowing date (mean Hedges’d = 0.533 with 95% CI 0.222 – 0.844, z = 3.3564, P = 0.001, Figure 5a), even in the absence of heterogeneity (Q = 6.59, d.f. = 6, P = 0.360). The apparent generality of this result must be treated with caution, however, as it is based on only two independent experiments. Model ranking accounting for all studies, including Morris’s studies, showed that the model that included the dates of both early and delayed mowing had a lower AIC value, with a negative effect for early mowing (slope = −2.130 with 95% CI −3.017 – −1.241, z = −4.6989, P < 0.001) and a positive effect of delayed mowing (slope = 5.607 with 95% CI 3.283 – 7.930, z = 4.730, P < 0.001) [see Additional file 4. This means that effect size is greater the earlier the first mowing and later the delayed mowing. The influence of study duration was not investigated because all study durations were either 3 or 4 years.

Limitations of available information

The main limitation of this systematic review is the low number of data points stemming from an even lower number of studies (Table 1), which precluded investigations on specific invertebrate taxa, and on the influence of several moderators. As a consequence, only the main general effects of postponing mowing could be clearly investigated. Moreover, in the MA there was great heterogeneity in plant species richness, indicating that other factors (moderators) than delaying the first mowing probably influence the effect size. While the date of the first mowing was found to be an important factor, study duration was not (Figure 3). It was also expected that heterogeneity would be influenced by the great variety of meadow types involved. However, no analyses could be conducted on this factor due to the highly unbalanced distribution of the habitats among the data points (n = 36 mesophilous meadows; 16 wet meadows; 3 dry meadows). Moreover, from the sixteen wet meadow data points, nine could not be included in the weighted MA. Additional management factors such as fertilizer application, occurrence of a second cut, seed oversowing, and autumn grazing would also influence the effect size, but they could not be investigated for the same reasons. Note that the most common management practice (42 data points out of 55) was no fertilizer application, no grazing and a single cut per year.

Study design could also play a role. While most studies were experimental, three used a purely observational approach [31, 55, 56]. Experimental frameworks also differed greatly in sample sizes, plot sizes and sampling methodologies, which additionally affect the probability of detecting changes. Publication bias was not apparent from the funnel plots; however, some biogeographical bias might be present as most studies originated from the UK [see Additional file 2].

Evidence of effectiveness and management recommendations

This review confirms that postponing of mowing from spring (May-June) to summer (July-September) is appropriate to promote plant and invertebrate diversity. In contrast, postponing mowing from spring to fall (October-November) or from early summer (July) to late summer or fall may have a negative impact on the vegetation species richness. Invertebrates might still benefit from it but these two postponing schemes could not be differentiated due to small sample size. Regarding wet and litter meadows, a late cut (September or later) is usually recommended [72], but unfortunately we are not in a position to confirm this recommendation, in the absence of habitat specific analyses.

When postponing mowing cannot be done at the field scale, leaving uncut grass areas within the cultivated landscape matrix can be an alternative solution to favour plants and animals [see also [73–76]. At the landscape scale, creating a mosaic of different mowing regimes will increase species diversity, as there is no single appropriate mowing time that suits all organisms [33, 54, 77]. In addition to the date of first mowing, a low annual cutting frequency also promotes wild plants [78] and invertebrates [79, 80]. There was no single study on birds that complied with our selection criteria. However, all studies on ground-nesting birds recommend postponing mowing until after fledgings have left the nests [e.g. [81–85]. These management recommendations do not apply everywhere and must be related to the socio-economic context. For example, in highly fertilized systems (high intensity management) biodiversity is generally too low for these measures to have positive effects [e.g. [86].

Implications for further research

Our review focuses solely on the general effect of delaying the first mowing date upon plant and invertebrate species richness as well as invertebrate abundance. Some general trends could be extracted from the scientific literature, but there is still considerable uncertainty concerning the estimated effect sizes, since the influence of several moderators has barely been investigated. Altogether, invertebrates were far less documented than plants, with only seven studies of the impact of delaying mowing on species richness and/or abundance, and even these showed a major geographical bias (six studies from the UK and one from Finland). Clearly this is not sufficient to get the full picture: further long-term, experimental investigations of target taxonomic groups and species regarding responses to mowing regimes are needed. This lack of invertebrate studies is true not only for mowing but also for all factors that may influence grassland invertebrates, such as grazing, habitat fragmentation and management intensity [87]. Only experimental work can disentangle the effects of various, often concomitant management practices (e.g. mowing date and fertilizer application). We thus encourage experiments where management practices are investigated in a full factorial design or where a single management practice is tested against a control plot or field that differs only in regard to this practice. Field scale experiments should be preferred to plot scale, especially when investigating animals that can move from a plot to another. Additionally, landscape characteristics are known to influence communities of plants and animals within farmland, and should therefore be accounted for in any attempt to model the effects of management practices on those communities [88].