MATERIALS and METHODS

Literature search strategy and data collection

To identify candidate studies, we screened a dataset included in a mapping review carried out by Ratto et al . (2022) that described the existing literature on biocontrol interventions for insect pests of crops in SSA. Ratto et al. (2022) systematically searched Web of Science All Databases and Scopus, using a combination of search terms relating to a wide range of biocontrol techniques and insect pests (e.g., “biocontrol”, intercrop*”, “armyworm”), agricultural settings (e.g., “agri*”, “farm*”) and the target geographical location (e.g., “sub-Saharan Africa”, “Southern Africa”)(Table S1). Grey literature was captured by conducting additional searches on Google and Google Scholar and by searching websites of relevant institutions. This mapping review covered a period between 2005 and April 2021 and was summarised narratively, with no quantitative analysis performed.
We integrated this initial dataset (149 articles) (Ratto et al. 2022) with a follow up search of relevant papers published between April 2021 and December 2021. Overall, we found 146 eligible articles. Only articles published after 2005 were included to reflect modern biocontrol practices and to determine biocontrol effectiveness within a short timeframe. We focused on the sub-Saharan region, which has a large population of smallholder farmers who depend on local food production, and who suffer substantial incidences of insect pest outbreaks and crop damage that threatens their food security.
We included in the definition of biocontrol interventions any practice that utilises natural enemies of pests, or chemical products derived from nature, for the control of pest populations. These include the augmentation, introduction, or inoculation of natural enemies (i.e., predators, parasitoids and entomopathogens, such as bacteria, viruses and fungi), and conservation biocontrol (Table 1). Conservation biocontrol was defined as the manipulation of habitat to enhance natural enemy abundance and diversity (Amoabeng et al. 2020) and included push-pull, intercropping and field margins. Botanical pesticides, defined as substances derived from natural materials (e.g. plant extracts), were also included.
To ensure biologically meaningful comparisons, we applied further inclusion criteria to all articles in Ratto et al. (2022). Only articles that quantitatively measured biocontrol performance on the outcome measures were included in the analysis. Only studies with replicated treatments at one or more sites were included. We screened studies wherein pest abundance (PA), crop damage (CD), crop yield (Y) or natural enemy abundance (NEA) (hereafter “outcome measures”) were compared between crops following the implementation of a biocontrol intervention and untreated crops. We also extracted, where available, data on the outcome measures in crops treated with synthetic pesticides. Measures of crop damage included dead hearts (i.e., drying of the central shoot), damage to stems (e.g., stem tunnelling), pods, leaves, fruits, shoots that were specific to the target pests. Crop yield was reported as either kg/ha or tonne/ha, which was standardised to the latter for analysis.
We categorised the sites that had been exposed to a biocontrol intervention as “treatment”, with those that were left untreated as “negative control (-)” and those treated with synthetic pesticides as “positive control (+)”. The mean, standard deviation (SD) and sample size of outcome measures were recorded for both the treatment and controls. When data were presented only in figures, we extracted data using ImageJ software (Schneider et al. 2012). We contacted the lead authors of the studies that had incomplete data and abandoned these studies if we could not obtain the missing statistics.
For articles that presented multiple years of data sampling at the same site, we used the most recent data to control for non-independence of temporal data (Gurevitch & Hedges 1993). When the study was conducted in two or more spatially independent sites, we recorded them as independent observations. When a study presented outcome measures for several successive weeks, we averaged the means and recorded it as a single effect size. When different concentrations or different types of biocontrol agent were applied (e.g., entomopathogens, botanical pesticides), we used the highest concentration and recorded each biocontrol type as an independent observation. The screening resulted in a total of 99 articles and 512 studies included in the analysis (Supplementary information S2).

Statistical analysis

In our meta-analysis, the log of the response ratio (lnRR ) represents the influence of biocontrol interventions on the outcome measures and expresses the proportional difference between the treatment and the control groups (Hedges et al. 1999):
lnRR = ln (x1 ) – ln(x2 )
where x1 is the mean of the outcome measure when biocontrol is applied (treatment) and x2 is the mean of the outcome measures under the untreated condition (control -) or after synthetic pesticide application (control +). The use of the natural logarithm linearizes the metric, treating changes in nominator and denominator equally, and produces a normalised sampling distribution (Hedges et al. 1999).
All outcome measures were analysed separately (pest abundance, crop damage, crop yield, natural enemy abundance). Fitted random effects models were used to calculate the overall means and 95% confidence intervals for each outcome measure to determine if biocontrol interventions significantly affected the outcome measures when compared to control areas (both untreated and pesticide treated). Random effect models do not assume that any variation in the effect size is due only to sampling error, and, instead, allow for a real random component of variation in effect size between studies (e.g., regional differences in study location). An effect of biocontrol intervention was considered significant if the 95% biased-corrected bootstrap confidence intervals (C.I.) of the effect size did not overlap zero (Koricheva et al.2013).
Meta-regression was used to explore sources of heterogeneity across each dataset. Our analysis focussed on the following ecological, environmental, and experimental parameters: (1) biocontrol technique; (2) crop type; (3) target pest taxon; (4) farming system. However, we could not use landscape complexity as a moderator as we found too few studies that investigated landscape context. To elucidate the variability of biocontrol efficacy across biocontrol techniques, we grouped studies according to whether they applied botanical pesticides, intercropping, field margins (border planting including legumes, sorghum or wild grasses), push-pull or augmentation/introduction methods. To determine if the effectiveness of biocontrol was dependent on crop type, we classified the study focus crops into cereal, fibre, fruits, vegetables, and pulses. We did not include stimulants (e.g., coffee, cocoa) and nuts due to small sample sizes. To establish whether biocontrol effectiveness varied across different pest insect taxa, we classified studies according to taxon of the targeted pest (Coleoptera, Hemiptera, Lepidoptera and Blattodea). Lastly, we classified studies into two field types: small farm (real smallholder farming conditions) and research farm (experimental field within a research centre), to identify any difference between these systems. Large commercial horticulture farms were not included in the meta-analysis as we primarily focussed on smallholder farmers and their food security. The above parameters were tested one by one as a sole moderator (i.e., fixed effects) for each outcome measure. To account for multiple comparisons from the same article, each model included “Study” nested within “Article” as random effects. The mean log response ratios and upper and lower bounds of 95% confidence intervals around the mean were back-transformed with the formula (elnR-1) *100 and expressed as percent change relative to the controls to facilitate interpretation.

Publication bias

We assessed publication bias in three ways. We first visually assessed funnel plots for strong asymmetries (Fig S1). Visual inspection of the funnel plots revealed symmetrical distribution of effect size around the meta-analytical mean of all outcome measures apart from pest abundance. We then ran Egger’s regression test (Egger et al. 1997; Nakagawa & Santos 2012) and the trim-and-fill test (Duval & Tweedie 2000). Egger’s test indicated that publication bias was significant for the pest abundance (z= -2.1065, p=0.0352), crop damage (z= -2.3886, p=0.0169), and NEA datasets (z=-2.4708, p=0.0135), but not significant for the crop yield dataset (CY: z= 0.0362, p=0.9711). This was inconsistent with the trim-and-fill tests showing no missing studies for all datasets. All statistical analysis was performed using the “metaphor” package in R (version 4.1.2) (Viechtbauer 2010).