On track to achieve no net loss of forest at Madagascar's biggest mine

The study evaluates whether biodiversity offsets linked to the Ambatovy nickel and cobalt mine in Madagascar can deliver no net loss (NNL) of forest, a proxy for biodiversity, within a context where development and conservation goals must be reconciled. The authors situate the work within global NNL policies and lender requirements, noting widespread use of offsets but a scarcity of robust evaluations. They examine averted-loss offsets that aim to slow deforestation from shifting agriculture in high-biodiversity eastern rainforests, applying counterfactual methods to assess if avoided deforestation in offset sites compensates for mine-induced forest loss.

The paper reviews NNL policy proliferation (101 countries mandate or support biodiversity compensation) and widespread offset implementation (>12,000 globally) alongside limited rigorous effectiveness evaluations. It discusses controversies around offsets—permanence, equivalence, equity, and measuring gains against declining baselines—while justifying forest cover as a pragmatic proxy for biodiversity in forested ecosystems. It highlights advances in impact evaluation using statistical matching and robustness checks to mitigate dependence on arbitrary model choices, and situates the study within broader literature on conservation effectiveness and forest policy evaluations.

Study area: Eastern Madagascar (former province of Toamasina), encompassing Ambatovy’s four biodiversity offsets—Ankerana, Corridor Forestier Analamay-Mantadia (CFAM), Conservation Zone (within the mine concession), and Torotorofotsy. The offsets total 28,740 ha, selected as like-for-like with the impact site. The mine footprint and upper pipeline cleared or degraded 2,064 ha of natural forest. Units and sampling: 30×30 m pixels that were forested in 2000 (aligned with local smallholder plot size) formed the analysis units. A 150×150 m grid sampling ensured minimum 120 m spacing to reduce spatial autocorrelation. Control pixels were sampled from the wider forested landscape in Toamasina, excluding established protected areas and pixels within 10 km of offsets (to reduce leakage bias). CAZ (Corridor Ankeniheny-Zahamena) was included in controls due to late formal protection and limited management; some offsets overlap CAZ. Outcome: Annual deforestation rate per year (2001–2019) from Global Forest Change (GFC), restricted to 2000 forest, with alignment to reduce spatial error. The outcome was aggregated by treated status (offset vs matched control) and year to give annual counts of deforested pixels. Covariates for matching: Five essential covariates capturing accessibility, demand, and agricultural suitability—slope, elevation, distance to main road, distance to forest edge, distance to prior deforestation. Five additional covariates used in robustness checks—annual precipitation, distance to river, distance to cart track, distance to settlement, population density. Matching: Conducted separately per offset in R (MatchIt). Main specification: 1:1 nearest-neighbour, Mahalanobis distance, caliper 1 s.d., without replacement, achieving excellent balance (max standardized mean difference 0.05). Data were pre-cleaned to remove controls outside treated calipers. Matched datasets were aggregated by treated status and year. Difference-in-differences (DiD) per site: Ordinary least squares on log(y+1)-transformed annual deforestation counts with before–after (BA), treated indicator (CI), and interaction (BA×CI) to estimate the percentage difference in annual deforestation post-protection relative to counterfactual. Parallel pre-trends were tested via interaction of year and treated indicator in pre-intervention years; CFAM failed pre-trend parity and was excluded from site-level DiD. Portfolio-level effect: Fixed-effects panel regression on pooled data (n=152; 8 series—4 offsets and 4 matched controls—over 19 years): log(count+1) ~ treatment (treated years for offsets) + site FE + year FE. Coefficients were back-transformed to percentage differences. Robustness: 116 alternative matching specifications tested to assess sensitivity to modeling choices: combinations of distance measures (Mahalanobis, propensity score via GLM-logit, propensity via Random Forest), calipers (0.25, 0.5, 1 s.d.), control:treated ratios (1, 5, 10), with/without replacement (54 combos with essential covariates); all combinations of additional covariates with essential covariates (31 combos); and 31 random combinations. Models were assessed a posteriori for sufficient matches, covariate balance, and pre-trend parallelism. Additional temporal robustness used equal pre/post years and dropping individual years. Leakage analysis: Buffer zones (10 km) around offsets were partitioned by Voronoi polygons to assign buffers to each offset, excluding established protected areas. Matching and outcome regressions were repeated to test for displacement of deforestation into buffers. Conversion to avoided deforestation: Treatment effects were converted to hectares of avoided deforestation per site and overall between each site’s protection year and January 2020; comparisons were made to the 2,064 ha loss at the mine site. Time since protection: Ankerana 9 years, Conservation Zone 11 years, Torotorofotsy 6 years; overall 11 years since first offset in 2009.

• Site-level DiD: Two offsets significantly reduced deforestation relative to matched counterfactuals: • Ankerana: −96% per year (95% CI 89–98%; P<0.001; n=38). • Conservation Zone: −66% per year (95% CI 27–84%; P<0.01; n=38). • Torotorofotsy: no significant effect (−41% to +510%; P=0.28; n=38). • CFAM: excluded from site-level DiD due to non-parallel pre-trends.
• Portfolio-level fixed-effects panel: −58% average annual deforestation across all four offsets (95% CI 37–73%; n=152). Excluding CFAM yields −72% (54–83%; n=114), but not significantly different from −58% (Z-test P>0.2). Results robust to random-effects specification (−53%, 95% CI −27% to −69%).
• Robustness: Across 116 alternative matching specifications, 106/116 supported a significant overall reduction; no model showed a significant increase. For Ankerana and Conservation Zone, most specifications confirmed significant avoided deforestation; models with insignificant results were typically a posteriori invalid. Torotorofotsy remained mostly insignificant (78/79 valid specs). CFAM largely failed validity checks; among 7 valid models, 6 no effect, 1 increase.
• Avoided deforestation (to Jan 2020): • Site-level sums: Ankerana 1,922 ha (95% CI 669–5,260); Conservation Zone 26 ha (5–71); total 1,948 ha ≈ 94% of mine-caused loss (2,064 ha). • Portfolio-level: 1,644 ha avoided (95% CI 674–3,122), equal to 79% (33–151%) of mine-related forest loss. • Since 2014 (all offsets active): average 265 ha avoided per year; projecting this rate suggests full NNL of forest by end of 2021. Using bounds (674–3,122 ha) implies achieving NNL between 2018 and 2033; company’s 2014 estimate was 2022–2035.
• Leakage: No significant increase in deforestation within 10 km buffers around offsets (P=0.15).
• Comparative effectiveness: Cohen’s d effect sizes—overall −0.51 (medium), Ankerana −1.03, Conservation Zone −0.63 (large). Ambatovy’s offsets outperformed 97% of 136 compiled conservation interventions and all but one protected area intervention in reducing deforestation.

Findings indicate Ambatovy’s averted-loss offsets substantially reduced deforestation, nearly compensating mine-induced forest loss by early 2020 and likely achieving NNL of forest by 2021. The authors suggest offsetting’s focus on measurable NNL targets and sustained corporate funding may drive stronger impacts than many public protected areas. Variation across sites (strong success in Ankerana vs. non-significant effects in Torotorofotsy) may reflect differences in enforcement effectiveness and social impacts, including restrictions on local resource access. The results bolster the feasibility of mitigating environmental damage from large industrial projects even in weak governance contexts, but emphasize careful design, rigorous evaluation, and attention to broader biodiversity, permanence, and equity considerations.

The study provides robust, counterfactual-based evidence that a high-profile mining project’s biodiversity offsets can nearly achieve, and likely achieve, no net loss of forest within a decade of implementation. This demonstrates that well-designed, well-funded offset portfolios can mitigate development impacts on forests. The methodological framework—combining matching, DiD, fixed-effects panels, and extensive robustness checks—offers a template for future evaluations to strengthen the evidence base on NNL. Future work should: expand evaluations across contexts and taxa; better integrate species-level outcomes and small-scale degradation; account for indirect impacts (e.g., in-migration); address permanence post-project; and aim to go beyond NNL to contribute net gains where biodiversity is highly threatened.

• Causal inference depends on accurate counterfactuals: despite rigorous matching and pre-trend testing, omitted variable bias remains possible.
• CFAM failed parallel pre-trends at site level, limiting site-specific inference; it is only captured in portfolio-level estimates.
• Small effective sample for DiD (n=38 observations per site: 2 groups × 19 years) reduces precision of estimates.
• Methods for staggered treatment timing are evolving; potential biases may remain.
• Forest cover is an imperfect proxy for biodiversity; species-level impacts and small-scale degradation (selective logging, artisanal mining, NTFP harvesting) are not captured.
• Permanence is uncertain: sustained protection after Ambatovy’s involvement (expected to cease 2040–2050) is doubtful given chronic underfunding of protected areas; restoration at the mine site may be challenging.
• Indirect effects (e.g., in-migration increasing demand for food, fuelwood, charcoal, bushmeat) are not accounted for and may inflate counterfactuals if embedded in background rates.
• Equity and social costs: local communities experienced restrictions and insufficiently compensatory alternative livelihoods; benefits may not offset opportunity costs for the most affected.