Causal Effect Analysis of the Relationship Between Relative Bird Abundance and Deforestation in Mexico

Simple Summary

Deforestation can negatively affect birds, but its causal effects are not well understood. In this study, we analyzed how bird populations are affected by high levels of deforestation in Mexico. We used public bird observation records across the country to focus on ten bird species. We compared areas with high deforestation to areas with low or no deforestation over the past years. Our results show that five of the ten bird species studied tend to become less common in places where deforestation is more intense, although one species exhibited an increase in abundance. This research helps us better understand how deforestation influences bird numbers, which can be useful for developing conservation actions that are based on solid scientific evidence.

Abstract

In this study, we used a causal analysis approach to assess the impact of deforestation on bird abundance in Mexico. Based on records in the eBird and GBIF databases, ten species were selected in 807 grids on the mainland. Relative abundances by species were estimated using a fixed-effects panel data regression for the period 2016–2018. Deforestation was used as a quasi-natural experiment, classifying treatment and control groups according to the distribution of relative abundances by quintiles of gross deforestation rates during the period 2001–2018. The treatment group was defined as relative abundances of birds present in grids in the last deforestation quintile (≥4% to 12%); the control group included relative abundances of birds present in grids of the first four quintiles (<4%). Extended regression models were used to estimate the impacts of high deforestation rates on the relative abundance of birds, finding mixed causal effects: five showed statistically significant declines in abundance (Ruddy Ground Dove (Columbina talpacoti), Black Vulture (Coragyps atratus), Melodious Blackbird (Dives dives), Bewick’s Wren (Thryomanes bewickii), and Rufous-backed Thrush (Turdus rufopalliatus)), while one specie Yellow-winged Cacique (Cassiculus melanicterus) exhibited significant increases. These findings highlight the importance of causal effect studies in contributing to empirical evidence-based conservation decision-making.

1. Introduction

Deforestation is the permanent loss of induced or natural forest vegetation, implying the shift of forest land use to other uses (agricultural, pastures, human settlements, or wetlands). This is not limited to human-induced activities but can also include natural events exacerbated by human influence [1]. It is not only one of the main drivers of biodiversity loss worldwide, but also disrupts ecosystem processes, ecological connectivity, and species interactions, particularly in megadiverse regions, including Mexico [2]. In recent decades, Mexico has experienced alarming deforestation rates due to the expansion of agricultural activities and urbanization, which have led to habitat fragmentation and loss of ecological connectivity. Between 2001 and 2023, this country lost an average of 207,665 hectares of forest cover per year, with warm-humid and warm-dry forests being the most affected [1]. This landscape transformation has direct consequences on biodiversity, particularly affecting birds, a key taxonomic group in the ecological balance and functioning of ecosystems [3].

Birds perform essential ecological functions, such as pollination, seed dispersal, and pest control [4]. Their sensitivity to habitat changes and wide geographical distribution make them effective environmental quality bioindicators. Previous studies have documented that habitat loss and degradation in tropical forest from deforestation adversely impact the diversity and abundance of many bird species [5], including those that depend on the understory for food and shelter [6]. Brown & Sullivan [7] documented that in isolated forest fragments, medium-sized species with flexible diets tend to be more resilient, while highly specialized species inhabiting forest habitats suffer greater abundance reductions. In contrast, generalist species can benefit from deforestation due to their ability to exploit open landscapes, as evidenced in several studies [5]. This suggests that the effects of deforestation can vary widely, depending on the ecological characteristics of each species.

Although Mexico harbors more than 1100 bird species and ranks among the most avifauna-rich countries worldwide [8], the magnitude and variability of deforestation effects across biomes—such as tropical, temperate, and xeric forests—and bioregions remain poorly understood. This is particularly relevant in transition zones between temperate and tropical forests, where soil conversion rates are higher than in other areas [9]. Vargas-Cárdenas et al. [3] evaluated the relationship between forest cover loss and bird diversity in “La Montaña”, Guerrero, reporting that specialist forest species decrease in deforested landscapes, while those adapted to disturbance show an increased abundance.

Most studies on the relationship between deforestation and bird populations rely on local correlations, which do not provide robust causal evidence. Although these may show association patterns between bird populations and deforestation, they do not allow inferring that deforestation is the cause of the changes observed in bird diversity and abundance, either locally or globally. Making conservation decisions based on causal relationships rather than correlations is essential to ensure the effectiveness of the measures implemented. Conservation actions based solely on correlations pose the risk of directing resources to strategies that fail to address the actual causes of the issue. Therefore, the conservation of biodiversity and abundance of birds warrant studies focusing on causal effect analysis.

The objective of this study was to estimate the causal effects of deforestation between 2001 and 2018 on the relative abundance of ten bird species representing different ecological traits in different bioregions of the Mexican territory. To this end, we used large databases with information on local, regional, and national scales from open-access databases (eBird and GBIF). These databases compile records from collections, studies with probabilistic or systematic sampling, and citizen science data. Using these databases involves a series of methodological challenges to address selection bias issues since they derive from the interest of birdwatchers in particular sites, resulting in sites with and without information on the presence or abundance of birds. An additional challenge is to manage sampling effort bias caused by differences in birdwatching intensity (number of visits, number of watchers, watching time, watcher skills) between sites.

Various statistical methods have been used to estimate the effects of anthropogenic and environmental transformation on bird populations. Regression methods, such as the generalized linear mixed model (GLMM), are common in the analysis of bird population counts. The work of Kéry and Schaub [10] used GLMM to analyze trends in bird populations in different habitats, allowing the estimation of the variation in counts due to environmental and temporal factors. In another study, Liang et al. [11] used ordinary least squares (OLS) and instrumental variable methods to estimate the impact of air pollution on bird abundance in the U.S., highlighting the importance of robust approaches to endogeneity bias in bird abundance predictors. Propensity score matching (PSM) between treatment and control groups has also been used in ecological studies as an effective tool to control confounding variables and estimate causal effects. Examples are the work of Pearson et al. [12], who used PSM to assess the effect of agriculture on the biodiversity of aquatic ecosystems, and Claassen et al. [13], who applied PSM to assess the effectiveness of grassland conservation for migratory birds. Although recent studies in conservation ecology have evaluated causal effects, these types of studies are still scarce. Another relevant aspect is that PSM methods have limitations in addressing endogeneity issues caused by sample selection biases, omitted variables, and simultaneity (reverse causality), which results in biased and inconsistent estimators that invalidate contrasts of commonly used statistical significance hypotheses, such as Student’s t-test and the F test.

This type of bias may arise when estimating the causal effects of deforestation (treatment variable) on bird abundance (outcome variable) in natural experiments. This occurs particularly when observational data are not randomly selected, leading to selection or sampling effort biases. Endogeneity issues may also arise from measurement errors or omitted variables, resulting in regressors (predictors) correlated with unobserved factors contained in the perturbation term [cov(ε,x) ≠ 0]. An additional endogeneity problem in this type of study is the presence of reverse causality biases, which occur when deforestation (treatment) explains changes in bird abundance (outcome) and, in turn, bird abundance predicts deforestation. This phenomenon is also known as simultaneity endogeneity bias, which occurs when response and treatment variables have a bidirectional causal relationship, i.e., changes in vegetation cover due to deforestation affect bird abundances, and sites with higher bird abundances are associated with deforestation-prone vegetation cover. To address these issues when performing causal effect analyses between deforestation and bird abundance, we used the Extended Regression Model (ERM), characterized by its flexibility to address these types of endogeneity in the information available in large databases, such as eBird and GBIF. Additionally, this model facilitates the estimation of causal effects between bird abundance and landscape changes due to deforestation. Based on this approach, the study tested two main hypotheses: (1) high deforestation rates (≥4% to ≤12%) significantly reduce the relative abundance of analyzed bird species, and (2) the effect of deforestation varies between species, depending on the ecological and biogeographic traits of each.

It is worth mentioning that to date, we found no previous studies that have used this method in bird conservation biology or ecology. The results of the present study are expected to contribute to formulating evidence-based conservation strategies to support sustainable natural resource management and biodiversity protection in Mexico. Additionally, this study seeks to strengthen the global understanding of the impacts of deforestation on bird communities and their role in maintaining ecosystem services.

2. Materials and Methods

2.1. Study Area

The study was carried out in the mainland territory of Mexico using a 0.5° × 0.5° grid mesh generated in QGIS 3.32.1 (QGIS.org) [14], dividing the country into grid cells of approximately 50 km². These were the analytical units to evaluate the relationship between deforestation and bird abundance (Figure 1).

2.2. Description of Variables

2.2.1. Species Selection

The species selection strategy consisted of collecting bird records from the eBird and GBIF platforms for the period 2016–2018 [16,17]. Taxonomic data curation was performed using the American Ornithological Society (AOS) list [18], updating names, correcting synonyms, and geographically validating records; those that lacked coordinates were excluded.

Given Mexico’s high avian species richness and the large volume of available records, we conducted an initial reduction in the number of species considered to ensure analytical feasibility. To preserve ecological representativeness while maintaining manageable data, we implemented a stratified selection approach based on natural abundance levels. All recorded species were grouped into four quartiles according to their total national abundance, reflecting differences in species’ detectability and natural population sizes linked to their biological and ecological traits. From each quartile, the 60 most frequently recorded species were selected based on two criteria: (1) the number of grid cells in which the species was recorded and (2) the total number of records. This process yielded a final set of 240 species representing a wide range of abundance classes across Mexico.

This representative species pool was used for exploratory multivariate analysis to identify general patterns of association between bird communities and environmental gradients. We generated a database that included the percentage of vegetation cover, based on the national land use and vegetation map from INEGI [19] at a 1:250,000 scale, and the mean deforestation index for the period 2001–2018 from CONAFOR [15], with a spatial resolution of 1 hectare per grid cell. Relationships between species distribution patterns and deforestation intensity were evaluated using canonical correspondence analysis (CCA) performed in CANOCO 5 [20] (see Appendix A).

Following the CCA, we selected ten focal species based on the following criteria: strong association with deforestation gradients in the ordination space (canonical coefficient ≥ |0.35| on the first two axes) and ecological representativeness, including distribution, variation in habitat preferences and diet. This multi-criteria selection ensured the inclusion of species, both ecologically informative and statistically robust for further analysis. The selected species were: Cardellina rubra, Campylopterus hemileucurus, Cassiculus melanicterus, Cathartes aura, Columbina talpacoti, Coragyps atratus, Dives dives, Geranoaetus albicaudatus, Thryomanes bewickii, and Turdus rufopalliatus (Figure 2). All selected species are currently listed as Least Concern on the IUCN Red List, and all are classified as resident species within Mexico.

According to eBird [21], Cardellina rubra is restricted to humid montane conifer and pine-oak forests and adjacent shrublands in the central and western regions of the country (18° to 24° N); Campylopterus hemileucurus is found in humid evergreen forests in mountains and slopes (15° to 20° N); Cassiculus melanicterus inhabits deciduous forests and edges, as well as in plantations and living fences with tall trees (16° to 27° N); Cathartes aura and Coragyps atratus share a broad range throughout the country, being common in open and peri-urban areas; (15° to 33° N). Columbina talpacoti lives mainly in open and semi-open areas (15° to 27° N); Geranoaetus albicaudatus is distributed in open areas and agricultural land in southeastern, central, and northeastern Mexico (16° to 27° N); Thryomanes bewickii inhabits dry areas covered by shrubs, open areas, and open forests near water bodies—it is also commonly found in urban environments (17° to 33° N); Turdus rufopalliatus is found in wooded areas, gardens, and forests in tropical lowlands and slopes in western Mexico (15° to 28° N); finally, Dives dives is common in semi-open lowland tropical habitats (15° to 24° N).

2.2.2. Outcome Variable: Relative Abundance of Birds

Information regarding species records and sampling efforts is available as raw records, number of observers, and observation dates in the eBird and GBIF databases [16,17]. As mentioned above, such records involve potential selection and sampling effort biases. These issues were addressed using the procedure of Liang et al. [11], who applied a fixed-effects panel data regression to estimate the relative abundance of birds. Assuming bird-count data follow a Poisson distribution, the probability of occurrence in recorded sightings can be expressed as follows:

$$f\left(h,\lambda \right)=P\left(y=h\right)={\displaystyle \frac{exp(-\lambda ){\lambda }^h}{h!}},h=0,1,2,\dots$$

(1)

where h is the number of occurrences in the bird record; and λ is a positive parameter that indicates the number of times the bird record is expected to occur during a year, which can be expressed as follows:

$$E\left(y_{it}\right)=\lambda =exp\left({\beta }_1+{\beta }_2{X}_{it}\right)$$

(2)

where $y_{it}$ is the bird count recorded in grid cell $i$ during the period $t$ and the parameters $\beta$ to be estimated measure the relationship between the number of records and the birdwatch effort represented by the vector $X$, which contains the variables number of birdwatchers per grid cell $i$ for the year $t$. The number of occurrences in the bird record is estimated using the maximum log-likelihood function:

$$lnL\left(\beta \right)=\sum _{i=1}^N\left\{-\mathit{exp}\left({\beta }_1+{\beta }_2{x}_{it}\right)+y_{it}\times \left({\beta }_1+{\beta }_2{x}_{it}\right)-\mathit{ln}\left(y_{it}¡\right)\right\}$$

(3)

where ${\beta }_1$ represents the occurrence of bird records that are unaffected by birdwatcher abilities during visits to each grid cell over a given year, and ${\beta }_2$ measures the effect of the number of watchers on the occurrence of records per grid cell during a year.

The Poisson equation was linearized using the natural logarithm of the records per grid cell for each year, which allowed for the estimation of records that are independent of the sampling effort, such as the watchers’ skills. The specification to be estimated to capture the heterogeneity of bird records that is time-invariant, specific to each grid cell, is expressed as follows:

$$ln\left(y_{it}\right)=\alpha +X_{it}^{\prime }\beta +{\mu }_i+v_{it}$$

(4)

where $ln\left(y_{it}\right)$ is the natural logarithm of bird records for grid cell $i$ in the year $t$; $\alpha$ is a common constant; $\beta$ is a vector of parameters to estimate that measure the relationship of sampling effort variables represented by the vector $X_{it}^{\prime }$ containing the number of watchers per grid cell; and ${\mu }_i$ is a parameter that controls the specific characteristics to the i-th grid cell that are invariant in time, called fixed effects, which represent the relative abundance of bird species per grid cell.

2.2.3. Treatment Variable: Deforestation

The impact of deforestation on bird abundance in Mexico was evaluated using the gross deforestation rate, which represents the permanent loss of forest cover (10% canopy cover and trees at least 4 m high), excluding land that could regenerate naturally. The resolution of the forestation rate during the period 2001–2018 was 1 hectare, obtained from CONAFOR [15]. The average index was calculated by grid cell and then divided into quintiles. This produced a binary variable that takes a value of 1 in grid cells belonging to the highest quintile (≥4% to ≤12%) and 0 in grid cells belonging to the first four quintiles (<4%).

2.2.4. Control Variables

Two variables were included as control: (i) Natural Protected Areas (PA), the information of which was obtained from their location and delimitation at a 1:250,000 scale (CONANP [22]); this was used as it is a public policy measure that controls changes in the landscape due to deforestation, which, in turn, affects the relative abundance of birds; and (ii) Bioregion as a variable that allows controlling the influence of regional ecological characteristics on the distribution of species. This variable was generated from the National Biodiversity Information System Geoportal (CONABIO [23]) classification at a 1:4,000,000 scale.

Table 1. Description of variables.

Variable	Description	Source	Measurement Unit
Records	Raw record, which represents the number of individuals observed in the period 2016–2018 for each selected species (Cardellina rubra, Campylopterus hemileucurus, Cassiculus melanicterus, Cathartes aura, Columbina talpacoti, Coragyps atratus, Dives dives, Geranoaetus albicaudatus, Thryomanes bewickii, and Turdus rufopalliatus.)	eBird and GBIF [16,17] 2016–2018.	Continuous variable
Whatchers	Number of watchers/collectors per selected species present in each grid cell during the period 2016–2018.	eBird and GBIF [16,17] 2016–2018.	Continuous variable
R_abundance	Relative abundance estimated by fixed-effects panel data regression for each selected species (Campylopterus hemileucurus, Cathartes aura, Columbina talpacoti, Dives dives, Geranoaetus albicaudatus, Thryomanes bewickii, and Turdus rufopalliatus)	Own estimate based on eBird and GBIF data [16,17] 2016–2018.	Continuous variable
Def	Mean gross deforestation rate per grid cell during the period 2016–2018.	CONAFOR, 2020 [15]	Continuous variable
Wdef	Spatially lagged deforestation rate based on a first-order queen contiguity matrix.	CONAFOR, 2020 [15]	Continuous variable
Defo	Binary variable that takes a value of 1 if the gross deforestation rate in the grid cell is ≥4% to ≤12% during the period 2001–2018 and 0 otherwise.	CONAFOR, 2020 [15]	Dichotomous variable
PA	Binary variable that takes a value of 1 if the grid cell has at least one federal protected area, and 0 otherwise (updated to 2024 with 204 PA).	CONANP, 2024 [21]	Dichotomous variable
Bioregion	Indicates the bioregion to which each grid cell belongs (Nearctic = 1, Mexican transition zone = 2, and Neotropical = 3). If the grid cell comprises more than one bioregion, it is classified as the bioregion with the highest percentage of area.	CONABIO [23]	Categorical variable
Landscape	Percentage of land use (xeric, forests, agricultural, tropical forests, grasslands) related to vegetation cover per grid cell.	INEGI [19]	Continuous variable

contains the description of the response (relative abundance), treatment (deforestation), and control (bioregions and PA), and

Table 2. Descriptive statistics of variables.

Scope	Variable	Mean	Std. Dev.	Min.	Max.	No. of Grid Cells
Records	R_C.rubra	36.24	59.71	1	287	24
	R_C.hemileucurus	11.83	11.4	1	59	14
	R_C.melanicterus	39.97	90.23	1	588	43
	R_C.aura	36.62	73.77	1	828	393
	R_C.talpacoti	37.92	86.85	1	1124	135
	R_C.atratus	42.22	84.17	1	799	269
	R_D.dives	51.04	94.53	1	685	113
	R_G.albicaudatus	9.49	14.06	1	93	42
	R_T.bewickii	46.33	198.05	1	2061	90
	R_T.rufopalliatus	53.85	138.7	1	1183	82
Collectors (Colle)	C_C.rubra	25.96	44.29	1	162	24
	C_C.hemileucurus	18.88	21.8	1	116	14
	C_C.melanicterus	28.12	57.56	1	239	43
	C_C.aura	44.04	85.35	1	828	393
	C_C.talpacoti	48.1	178.41	1	3183	135
	C_C.atratus	33.77	59.62	1	441	269
	C_D.dives	65.92	132.28	1	1699	113
	C_G.albicaudatus	9.37	13.85	1	93	42
	C_T.bewickii	34.79	104.11	1	992	90
	C_T.rufopalliatus	41	92.74	1	651	82
Relative abundance	A_C.rubra	41.04	14.58	29	108	8
(R_abundance)	A_C.hemileucurus	13.29	3.17	8	18	5
	A_C.melanicterus	29.09	6.98	19	68	43
	A_C.aura	6.39	15.63	1	130	131
	A_C.talpacoti	11.74	9.52	1	81	45
	A_C.atratus	18.66	32.69	1	239	90
	A_D.dives	22.99	15.87	1	117	38
	A_G.albicaudatus	1.17	0.78	1	4	14
	A_T.bewickii	12.61	59.48	1	556	30
	A_T.rufopalliatus	7.82	19.34	1	145	27
Landscape	Xeric	1.95	4.88	0	47.38	807
	Grasslands	8.6	14.15	0	77.73	807
	Forests	7.93	14.39	0	80.26	807
	Agriculture	4.34	8.53	0	78.84	807
	Tropical forest	4.41	9.22	0	71.06	807
Anthropogenic	Def	0.02	0.03	0	0.12	807
	Defo	0.2	0.4	0	1	807
Protection status	PA	0.54	0.5	0	1	807
Biogeographic	Bioregions	1.71	0.76	1	3	807

shows the descriptive statistics of these variables.

2.3. Specification of Causal Effect Models

The empirical strategy in this work consisted of using information on deforestation rates that explains the impact on the relative abundance of birds through a quasi-natural experiment, considering treated grid cells as those with deforestation rates ≥ 4% to ≤12% (last quintile) and the rest as a control group with deforestation rates < 4%.

The selection bias and reverse causality issues present in the data of relative abundance of species per grid cell were addressed using an extended regression model with a main equation and two auxiliary equations, from which the potential selection and endogeneity biases of the treatment variable are addressed.

The first step was to address the selection bias inherent in the open-access databases (eBird and GBIF) by estimating the probability of participation in the databases on the presence and records/relative abundance of species. To this end, we estimated the probability that the species $j$ has records in the grid cell $i$ using the following Probit model:

$${selection}_i={\beta }_0+{\beta }_1{watchers}_i+{\beta }_2{bioregion}_i+{\epsilon }_i$$

(5)

where selection is a binary variable that takes a value of 1 if the species j has relative abundance observations in the grid cell i and 0 otherwise; watchers is the total number of birdwatchers per grid cell in the period 2016–2018; bioregion is a categorical variable that takes a value of 1 if the largest proportion of the grid cell i belongs to the Nearctic region, 2 if the largest proportion belongs to the Mexican transition zone, and 3 if the largest proportion belongs to the Neotropical region; ${\beta }_1,{\beta }_{2}y{\beta }_3{\epsilon }_i$ are parameters to be estimated; and ${\epsilon }_i$ is the disturbance term.

Due to the endogeneity of the defo treatment, the spatially lagged gross deforestation rate Wdef was used as an instrumental variable. This variable was created from the first-order queen-type spatial weight matrix $w_{ij}$:

$$w_{ij}=\left\{\begin{array}{cc}1,\hfill & when \mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d} \mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l} i is adjacent to \mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d} \mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l} j\\ 0,\hfill & when \mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d} \mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l} i is not adjacent to \mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d} \mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l} j\end{array}\hspace{1em}i\ne j\right.$$

(6)

In the second stage, the variable Wdef was used as an instrument in the following Probit:

$${defo}_i={\beta }_0+{{\beta }_1{bioregion}_i+{\beta }_2\mathrm{P}\mathrm{A}+Wdef\beta }_3+{\epsilon }_i$$

(7)

where defo is a binary variable that takes a value of 1 if the grid cell i recorded high deforestation rates ranging between ≥4% and ≤12% during the period 2001–2018, and 0 if it recorded low deforestation rates (<4); bioregion is the variable previously described in Equation (5); $PA$ is a dichotomous variable that takes a value of 1 if the grid cell i has at least one Natural Protected Area (PA) and 2 otherwise; Wdef is the instrumental variable described previously (VI), which satisfies the exclusion restriction, since it is only associated with changes in the relative abundance of birds through the local deforestation pathway; ${\epsilon }_i$ represents unmeasured environmental factors and/or measurement errors.

In the third stage, the predicted value of Equation (7) was used to address the endogeneity issue by estimating the causal effect of high deforestation rates on the relative abundance of species j using Equation (8) in the Extended Regression Model (ERM):

$${abundance}_{ij}={\beta }_0+{\beta }_1{\widehat{defo}}_{ij}+{\beta }_2{bioregion}_{ij}+{\beta }_3{PA}_{ij}+{\beta }_4{IMR}_{ij}+{\epsilon }_{ij}$$

(8)

where abundance represents the relative abundance of the species j in the grid cell i; bioregion is a categorical variable that takes values from 1 to 3 for Nearctic, Mexican transition zone, and Neotropical, respectively; $\widehat{defo}$ is the predicted value of the treatment variable in dichotomous form, which takes a value of 1 for treated grid cells with deforestation rates within the range ≥ 4% to ≤12%; IMR is the inverse Mills’ ratio $\left[{\displaystyle \frac{\mathsf{Ø}\left(\mathrm{x}\right)}{\mathsf{\Phi }\left(\mathrm{x}\right)}}\right]$ composed of the standard normal density and the cumulative distribution function, which was obtained by estimating the probability that species $j$ has sighting records in grid cell $i$ (Equation (5)). The inclusion of IMR and statistical significance in Equation (8) allows correcting for sample selection bias. Finally ${\epsilon }_i$ corresponds to the disturbance term. After estimating the models for each species, endogeneity tests (Durbin–Wu–Hausman) and, consequently, instrument validity tests (Montiel–Pflueger weak instrument test) were performed. Based on these results, extended regression models were estimated with an exogenous or endogenous treatment variable (defo) to obtain the average treatment effect on the treated ATT.

The ATT, which represents the impact of high deforestation rates on the relative abundance of the bird species studied, was estimated with the following equation:

$$\begin{array}{cc}\hfill ATT=& E\left({abundance}_{1i}|selection,defo,x,z\right)\hfill \\ & \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}-E\left({abundance}_{0i}|selection,defo,x,z\right)\end{array}$$

(9)

where ${abundance}_{1i}$ is the expected relative abundance of the species j in grid cells with high deforestation rates and abundance_0i is the expected relative abundance of the species if the grid cells where it is distributed had not experienced high deforestation rates; x is a vector of independent variables included in Equations (5)–(8); and z corresponds to the Wdef instrumental variable. These estimates were obtained using the Stata version 18 SE-Standard Edition software (the databases used for the ATT estimators are available in the Supplementary Materials).

3. Results

The taxonomic and geographical curation yielded 255,109 records of 1038 different bird species. In the canonical correspondence analysis (CCA) (Figure 3), the first two dimensions accounted for 69.3% of the adjusted variance, capturing most of the relationship between species and environmental variables. The canonical axes were statistically significant (p = 0.001), which confirmed non-random associations.

Ten bird species with different abundance levels were selected in grid cells with deforestation: Cardellina rubra, Campylopterus hemileucurus, Cassiculus melanicterus, Cathartes aura, Columbina talpacoti, Coragyps atratus, Dives dives, Geranoaetus albicaudatus, Thryomanes bewickii, and Turdus rufopalliatus. The results on relative abundance obtained from the fixed-effects panel data regression were classified by treatment groups according to their presence in grid cells with deforestation of ≥4% to ≤12% and in control groups if they were present in grid cells with deforestation rates < 4%.

Table 3. Relative species abundance and grid cell representation in the control and treatment groups.

	Relative Abundance		Grid Cells
Species	Treatment	Control	Treatment	Control
Cardellina rubra	119 (12.08%)	866 (87.92)	3 (12.5%)	21 (87.5%)
Campylopterus hemileucurus	95 (51.08%)	91 (48.92%)	7 (50%)	7 (50%)
Cassiculus melanicterus	358 (28.62%)	893(71.38%)	11 (25.58%)	32 (74.42%)
Cathartes aura	477 (19%)	2033 (81%)	98 (24.94%)	295 (75.06%)
Columbina talpacoti	576 (36.34%)	1009 (63.66%)	56 (41.48%)	79 (58.52%)
Coragyps atratus	1423 (28.56%)	3559(71.44%)	95 (35.58%)	172 (64.42%)
Dives dives	1268 (48.8%)	1330 (51.19%)	60 (53.1%)	53 (46.9%)
Geranoaetus albicaudatus	316 (34.35%)	604 (65.65%)	53 (30.81%)	119 (69.19%)
Thryomanes bewickii	39 (3.44%)	1096 (96.56%)	11 (12.22%)	79 (87.78%)
Turdus rufopalliatus	68 (10.61%)	573 (98.39%)	16 (19.51%)	66 (80.49%)

shows the relative abundances by species for both groups.

Results of Causal-Effect Models

Each extended regression model (ERM) estimated for the ten selected species is composed of three steps: The Selection step (Equation (5)), which estimates the probability that there are abundance records for a species in a grid cell, addressing the selection bias arising from the sampling effort; the Treatment step (Equation (7)), which models the probability that a grid cell exhibits high deforestation rates (≥4% to ≤12%)—applied only in the case of Dives dives, a species for which deforestation was found to be endogenous according to the results of the Durbin–Wu–Hausman test—using the deforestation rate estimated or predicted by the model as an independent variable; the Outcome step (Equation (8)), which estimates the causal effect of deforestation on the relative abundance of species. In addition, correlations between the error terms of the equations are included, along with log-likelihood values and model fit statistics. These results are presented in

Table 4. Estimated effects of deforestation on the relative abundance of the first five selected bird species.

Variable	Cardellina rubra	Campylopterus hemileucurus	Cassiculus melanicterus	Cathartes aura	Columbina talpacoti
Outcome Equation (8): Deforestation	−1.03	−1.35	4.71 **	0.82 *	−2.94 **
	(9.96)	(2.92)	(1.49)	(0.47)	(1.16)
ZTM	−17.50	-	-	−4.01 ***	-
	(17.24)	( - )	( - )	(1.4)	( - )
Neotropical	−36.53 **	-	-	−4.38 ***	-
	(17.24)	( - )	( - )	(1.67)	( - )
PA	−0.88	2.05	2.12	0.82 ***	-
	(7.72)	(3.15)	(2.26)	(0.27)	( - )
Inverse Mills ratio	−22.76 **	−13.04 **	8.63	−8.08 ***	12.24 ***
	(10.80)	(5.13)	(6.05)	1.54	(2.46)
Constant	100.79 ***	40.1 ***	1.02	14.26 ***	20.28 ***
	(18.88)	(9.68)	(6.83)	(1.00)	(3.52)
Selection Equation (5): Watchers	0.00006 ***	0.00007 ***	0.00004 ***	0.001 ***	0.0002 ***
	(0.00002)	(0.00001)	(0.00001)	(0.00005)	(0.00001)
ZTM	0.12	-	-	0.18 *	1.53 ***
	(0.53)	( - )	( - )	(0.09)	(0.18)
Neotropical	1.68 ***	-	-	0.31 ***	0.84 ***
	(0.43)	( - )	( - )	(0.11)	(0.21)
Constant	−2.48 ***	−2.92 ***	−2.48 ***	−0.93 ***	−2.27 ***
	(0.23)	(0.29)	(0.24)	(0.06)	(0.16)
corr(e.selection, e.abundance)	-	-	-	-	−0.94 ***
	( - )	( - )	( - )	( - )	(0.02)
Pseudo R2	0.33	0.22	0.08	0.54	0.38
AIC	155.60	262.54	316.55	527.0	457.90
BIC	179.06	281.31	335.33	551.03	481.37
Log likelihood	−847.30	−2402.86	−1014.4	−1781.59	−679.9
Wald Chi2 (dof)	973.11	3.18	15.44	40.88	6.70
Prob > Chi2	0.00	0.36	0.00	0.00	0.03
Durbin score (p-value)	0.70	0.95	0.81	0.67	0.21
Wu–Hausman (p-value)	0.74	0.95	0.82	0.67	0.22
Treated observations	24	42	43	393	135

Values in parenthesis are standard errors, * p < 0.1, ** p < 0.05 *** p < 0.01. ZTM = Mexican transition zone.

and

Table 5. Estimated effects of deforestation on the relative abundance of the rest of the selected bird species.

Variable	Coragyps atratus	Dives dives	Geranoaetus albicaudatus	Thryomanes bewickii	Turdus rufopalliatus
Outcome Equation (8): Deforestation	−14.34 ***	−4.23 **	0.07	−29.59 ***	−23.53 ***
	(5.38)	(2.1)	(0.26)	(5.13)	(5.28)
ZTM	14.78 ***	-	−0.09	-	−17.52 **
	(5.45)	( - )	(0.38)	( - )	(8.04)
Neotropical	3.9	-	−0.27	-	−44.56
	(6.15)	( - )	(0.39)	( - )	(6.59)
PA	-	-	-	−7.23	2.49
	( - )	( - )	( - )	(10.13)	(6.59)
Bioregion	-	−11.63 ***	-	-	-
	( - )	(3.72)	( - )	( - )	( - )
Inverse Mills ratio	−82.68 ***	−20.38 **	0.15	−129.52 ***	−79.91 **
	(17.07)	(4.31)	(0.23)	(44.31)	(33.55)
Constant	14.21 ***	−34.82 ***	0.96	115.07 ***	105.47 ***
	(4.58)	(10.53)	(0.64)	(1.61)	(7.83)
Treatment Equation (7): WDEF	-	25.7 ***	-	-	--
	( - )	(2.57)	( - )	( - )	( - )
Bioregion	-	0.36 ***	-	-	-
	( - )	(0.08)	( - )	( - )	( - )
Constant	-	−3.83 ***	-	-	-
	( - )	0.35	( - )	( - )	( - )
Selection Equation (5): Watchers	0.0002 ***	0.00008 ***	0.00008 ***	0.00006 ***	0.00006 ***
	(0.00002)	(0.00001)	(0.00001)	(0.00002)	(0.00001)
ZTM	1.07 ***	-	0.43 **	−0.37 ***	0.48 ***
	(0.16)	( - )	(0.19)	(0.18)	(0.19)
Neotropical	0.63 ***	-	0.35	0.2	0.64 ***
	(0.19)	( - )	(0.23)	(0.19)	(0.21)
Constant	−1.45 ***	−1.76 ***	−2.11 ***	−1.36 ***	−1.89 ***
	(0.11)	(0.15)	(0.15)	(0.1)	(0.14)
corr(e.selection, e.abundance)	-	-	-	-	-
	( - )	( - )	( - )	( - )	( - )
corr(e.Defo, e.abundance)	-	-	-	-	-
	( - )	( - )	( - )	( - )	( - )
corr(e.Defo, e.selection)	-	-	-	-	-
	( - )	( - )	( - )	( - )	( - )
Pseudo R2	0.23	0.23	0.18	0.12	0.08
AIC	797.4	385.37	281.19	506.97	496.46
BIC	820.86	401.57	304.66	530.43	519.93
Log likelihood	−2382.73	−2353.09	−185.05	−2750.73	−1397.42
Wald Chi2 (dof)	5404.01	12.18	0.61	648	928.01
Prob > Chi2	0	0.002	0.89	0	0
Durbin score p-value	0.1076	0.0337	0.8454	0.7319	0.7906
Wu–Hausman p-value	0.1104	0.0355	0.8557	0.7398	0.7976
Montiel–Pflueger weak instrument test (tau 5% = 37.418)	-	Effect. F statistic: 163.54	-	-	-
Treated observations	267	113	42	90	82

Values in parenthesis are standard errors, ** p < 0.05 *** p < 0.01. ZTM = Mexican transition zone.

.

Durbin–Wu–Hausman tests were performed to elucidate the endogeneity/exogeneity of the treatment variable defo, finding that for nine models it was not possible to reject with 95% confidence the null hypothesis of exogeneity with any level of statistical significance (Cardellina rubra, Campylopterus hemileucurus, Cassiculus melanicterus, Cathartes aura, Columbina talpacoti, Coragyps atratus, Geranoaetus albicaudatus, Thryomanes bewickii and Turdus rufopalliatus). Only in the Dives dives species model was it possible to reject this hypothesis; therefore, the deforestation rate variable with first-order spatial lags (Wdef) was used as an instrument in the extended regression model. Additionally, the Montiel–Pflueger weak instrument test (tau 5% = 37.418) was applied to contrast the validity of Wdef, finding a value of the F statistic (163.54) of the first stage greater than the critical values (37.418) of both the two-stage least squares (TSLS) and the limited information maximum likelihood (LIML) estimators, considering a bias of 5% in the worst case, rejecting the hypothesis that Wdef is a weak instrument.

Furthermore, the statistical significance of the IMR variable in Equation (8) for abundance at the 95% confidence level indicates that selection bias was corrected in eight models (Cardellina rubra, Campylopterus hemileucurus, Cathartes aura, Columbina talpacoti, Coragyps atratus, Dives dives, Thryomanes bewickii, and Turdus rufopalliatus).

The coefficients of the deforestation variable in the main Equation (8) represent the average treatment effect on the treated (ATT), i.e., the effect of deforestation rates of ≥4% to ≤12% on the relative abundance of the selected bird species. ATT was statistically significant at the 95% confidence level (CL) for six species, five with a negative effect and one with a positive effect, indicating species-specific responses to deforestation (Figure 4). Specifically, we recorded negative effects on the relative abundance of Columbina talpacoti (β = −2.94, p < 0.05), Coragyps atratus (β = −14.34, p < 0.01), Dives dives (β = −4.23, p < 0.05), Thryomanes bewickii (β = −29.59, p < 0.01), and Turdus rufopalliatus (β = −23.53, p < 0.01). In contrast, Cassiculus melanicterus showed a statistically significant positive effect of deforestation on its relative abundance (β = 4.71, p < 0.05).

Differences in bird relative abundance across bioregions reveal consistent patterns aligned with species’ spatial distributions. Compared to the Nearctic bioregion, a significant decrease in abundance was observed in the Neotropical bioregion for Cardellina rubra (β = −36.53, p < 0.05), Cathartes aura (β = −4.38, p < 0.01), and Turdus rufopalliatus (β = −44.55, p < 0.001), as well as in the Mexican Transition Zone (ZTM) for Thryomanes bewickii (β = −2.91, p < 0.05), and again for Turdus rufopalliatus (β = −17.52, p < 0.05). In contrast, Coragyps atratus showed a significantly higher abundance in both the ZTM (β = 620.01, p < 0.001) and Neotropical bioregions (β = 505.69, p < 0.001), highlighting its adaptability to open and human-modified landscapes.

To correct for sampling bias inherent in open-access biodiversity databases, we employed the selection equation (Equation (5)) that modeled the probability of recording relative abundance data for a given species in each grid cell. This model included two predictors: the number of birdwatchers per cell and the dominant biogeographic region. Results showed that birdwatcher effort was consistently and significantly associated with higher probabilities of data presence across species—for example, Cardellina rubra (β = 0.00006, p < 0.01), indicating that higher observer activity strongly predicts data availability. In addition, biogeographic region had a substantial effect: species such as Cassiculus melanicterus and Turdus rufopalliatus showed significantly greater probabilities of data occurrence in the Neotropical region compared to the Nearctic baseline (e.g., T. rufopalliatus: β = 2.38, p < 0.001). Similarly, Coragyps atratus exhibited a pronounced increase in data presence in both the ZTM (β = 1.66, p < 0.001) and Neotropical regions (β = 2.64, p < 0.001), reflecting its widespread distribution and detectability. These results confirm a strong spatial bias in data collection, with tropical and transitional regions being more sampled, likely due to greater species richness and observer interest. The inclusion of the Inverse Mills Ratio from this equation in the outcome model ensures that estimates of deforestation effects are not confounded by uneven data availability across regions.

The graph shows a marked heterogeneity across the point estimators of the effect of deforestation on the relative abundance of species; five species exhibit significant negative effects, indicating a decline in abundance in areas with high deforestation. The strongest reductions were observed for Turdus rufopalliatus and Thryomanes bewickii. Dives dives, Columbina talpacoti and Coragyps atratus also show significant negative impacts, albeit of smaller magnitude. In contrast, Cassiculus melanicterus exhibits a positive ATT, suggesting a higher relative abundance in deforested areas.

4. Discussion

This study estimated the causal effect of deforestation on the relative abundance of ten bird species using ERM models. These findings highlight the importance of conserving the remaining forest fragments to mitigate the negative effects on bird communities. Deforestation translates into increased habitat fragmentation, leading to changes or loss of connectivity and ecological interactions, as well as alterations in environmental factors, such as humidity, sun exposure, temperature, and evapotranspiration; in turn, these affect the availability of resources, such as fruits, seeds, insects, shelters and nesting areas, on which birds depend [24,25,26,27].

The relative abundance estimate is in line with the study by Liang et al. [11], who demonstrated that the estimate from a fixed-effects panel data regression effectively captures site-specific unobservable heterogeneity, reducing biases associated with the sampling effort. Similarly, using ERM enabled control of some of these biases in open databases such as eBird and GBIF [16,17].

Our findings confirm that deforestation significantly impacts the relative abundance of six of the ten species analyzed in Mexico. The deforestation estimators in the ERM models showed a significant reduction in the relative abundance of Columbina talpacoti, Coragyps atratus, Dives dives, Thryomanes bewickii, and Turdus rufopalliatus, while one species (Cassiculus melanicterus) exhibited significant increases. These effects were estimated for grid cells experiencing deforestation rates between ≥4% and ≤12% during the 2001–2018 period. These findings are in line with previous studies reporting that various anthropogenic disturbances adversely affect bird populations. For example, our results are consistent with those of Pearson et al. [12], who documented negative effects of human-induced pressures on biodiversity, and Liang et al. [11], who analyzed the relationship between air pollution and bird abundance.

The confidence intervals of the ATT coefficients reveal a clear heterogeneity in the response of bird species to deforestation, which can be explained by considering the individual ecological characteristics of each species. Most of the species that declined due to deforestation share characteristics such as a high dependency on specific habitat structures (e.g., dense understory for Thryomanes bewickii, tall roosting trees for Coragyps atratus), limited dispersal capacity, or vulnerability to edge effects and predation.

Species that are not strongly dependent on forest systems showed significant reductions in abundance too. Thryomanes bewickii, although adaptable to open environments, relies on low tree vegetation and fallen logs for nesting and sheltering, so its decline could be related to the loss of these key structural elements. This is consistent with studies indicating that some species recorded on forest edges or semi-open areas may see their abundance reduced when deforestation removes key habitat elements [5].

Some birds reported to be ecologically flexible and able to exploit resources in disturbed environments, such as Coragyps atratus, Columbina talpacoti, and Dives dives, are not exempt from showing negative effects. In our results, ATT reveals that these species are negatively affected by deforestation. These results contrast with those reported by Brown & Sullivan [7] regarding which scavenger or granivorous species could be more tolerant to fragmented environments. Also, Turdus rufopalliatus, experienced one of the steepest declines among the focal species. Although it has been reported to be expanding its distribution range into urban areas [28], this species appears to be highly sensitive to habitat fragmentation and forest cover loss in its core deciduous forest habitat. Studies have shown that while it can utilize modified environments, its breeding success is compromised in areas with high anthropogenic disturbance [28].

In contrast, Cassiculus melanicterus seems to possess greater ecological plasticity, thriving in edge environments and secondary growth, and possibly benefiting from the increase in patchiness and resource heterogeneity that often follows deforestation. While this might appear surprising, the explanation could lie in local population dynamics and competition. Its population densities in deforested areas may be constrained by reduced nesting sites or increased predation and parasitism due to edge effects, as has been observed in other generalist species [29].

This study is the first effort to assess the large-scale impact of deforestation on bird abundance in Mexico. However, it is worth noting that only a subset of species was considered, implying that other species that were not included in the present analysis due to their lower representativeness in the sampling data could be experiencing significant declines. Future research could add more detail to these findings with techniques that would allow for the detection of more complex patterns between deforestation and biodiversity.

An additional contribution of this study is a methodological alternative to address the selection biases inherent in open-access databases, such as eBird and GBIF. Although these sources offer a large spatial and temporal coverage, their scientific use requires rigorous methods that reduce biases derived from the sampling effort, the heterogeneity in species detection, and the spatial representativeness of the data. Our results demonstrate that it is feasible to use these databases, provided appropriate statistical models, such as ERMs, are applied to control unobservable heterogeneity and enhance the validity of causal inferences. This approach allows for advancing toward a more robust and reliable use of citizen data and global repositories in large-scale ecological research.

The results highlight the importance of protecting mosaics of diverse habitats and that reducing deforestation and areas with anthropogenic disturbance will be essential to ensure the coexistence and sustainability of these species in the context of deforestation and habitat loss [30,31].

5. Conclusions

This study provides empirical evidence of the causal impact of deforestation on bird relative abundance in Mexico, highlighting that species responses to habitat loss are not uniform, but shaped by ecological traits and landscape context. Although most of the analyzed species exhibited a significant decline in relative abundance under high deforestation rates, the fact that even some birds previously reported as tolerant to disturbance were negatively affected suggest that severe and widespread deforestation can exceed the ecological thresholds of resilience. This challenges assumptions about the robustness of generalist species and underscores the need for caution when interpreting tolerance as immunity. Only one species showed a statistically significant increase in relative abundance. This may also be a result of context-dependent factors, such as increased resource availability, reduced competition, or niche expansion in altered landscapes.

Methodologically, the combination of fixed-effects panel regression and Extended Regression Models (ERM) allowed for robust estimation of causal effects using large-scale observational data. This approach explicitly corrects for two common sources of bias in ecological studies: selection bias—through the inclusion of the inverse Mills ratio in the selection equation—and endogeneity—through instrumental variable estimation when needed. While observational studies remain subject to inherent limitations, this framework strengthens causal inference and enhances the scientific value of open-access platforms like eBird and GBIF for macroecological analyses.

From a conservation perspective, the results emphasize the importance of designing adaptive management strategies that are not only evidence-based but also tailored to the functional traits and ecological requirements of species. For example, forest specialists with limited dispersal ability, such as Cardellina rubra and Campylopterus hemileucurus, may require stricter protection of core forest habitats and ecological corridors to prevent population collapse. In contrast, more adaptable species may benefit from maintaining landscape heterogeneity in areas with moderate transformation. Additionally, improving monitoring efforts in rapidly changing regions can support early detection of population shifts and inform dynamic conservation responses. Ultimately, large-scale causal analyses such as the present study provide essential insights to guide more targeted, flexible, and effective conservation strategies in the face of accelerating anthropogenic change.