Poverty is a complex and multidimensional concept [16] Some authors have defined and viewed it from an economic perspective [9] [35] [36] while others have defined it from the wellbeing perspective [21] [24] . In fact, evidence exists to support the use of both perspectives in defining and measuring poverty [1] [2] .

Traditionally, poverty has been conceptualized as having insufficient income to meet the basic needs of a household [36] . It is not surprising that the definition and measurement of poverty has been largely limited to income, wealth, and consumption [2] . However, these measures of poverty are biased towards urban conditions and explain little about the realities of the rural poor [37] . From the perspective of a rural and remote areas of a developing country, poverty is a complex, dynamic, and multidimensional construct involving material deprivation, lack of access to basic needs, and social inequality [16] [37] . To understand poverty, there is a need to move beyond macroeconomic measures such as income, which are more applicable at a national, rather than a village or household levels.

The literature suggests that through the SL framework, poverty can be better understood from a household perspective [3] [12] [38] . Chambers’ work on sustainable livelihoods in the mid-1980s has generated debate over the years on the construction of poverty from a household perspective [16] [21] [24] [39] . This debate has significantly advanced the conceptualization of poverty from an income and/or assets-based phenomenon to a multidimensional concept of sustainable livelihoods encompassing, not just income, but also the capabilities of a household to maintain means of living. The SL approach, therefore, has evolved as a framework through which poverty can be better understood [21] . One of its key attributes is that it adds unquantifiable or non-material aspects of a household’s or an individual’s wellbeing to the poverty definition [2] [21] .

Reference [19] defined livelihood as a means of making a living, comprised of capabilities, assets, and resources used in daily activities. It has become an important approach for defining poverty, because of its emphasis on the means of maintaining a living, which allows for a broader definition of poverty [13] . The sustainable livelihoods’ framework, as defined above, has been further operationalized and widely applied in poverty alleviation discussions and intervention in developing countries. For example, CARE international has developed a Human Livelihood Security (HLS) Index to measure poverty for humanitarian and development intervention [39] . CARE’s HLS framework takes a multidimensional approach to measure livelihood security through indicators such as food security, health security, education security, economic security, and empowerment [39] . Oxfam has also operationalized the concept of sustainable livelihood from economic, social, institutional, and ecological perspectives [38] . While there are operational variations, the underlying multidimensional outcomes of SL, such as food, health, education and economic security, are commonly agreed upon. In this paper, we use these multiple dimensions and their validated measures from the HPFD Index.

Residents neighboring parks in developing countries have historically depended on forest resources for livelihoods in times of scarcity [4] [10] [20] . Forested areas, especially those in the tropics, provide multiple benefits, such as exploitation of the resources for commercial purposes, tourism services, harvests of non-timber forest products (NTFPs), harvests for subsistence livelihood needs, and ecological services [15] [41] [42] . For most poor households in proximity to forested areas, the forests harbor resources from which they derive their subsistence livelihoods [6] . For example, they hunt for meat and harvest resources such as wood, NTFPs, and medicinal plants for both household use and income [10] [11] [18] . In countries like Rwanda, with high population density and limited farmland, households adjacent to forests commonly rely on forest resources to supplement their livelihoods [4] . In some cases, it is the only source of livelihood for those households that have no farmland [27] .

While such forest-based livelihoods are important to the poor and vulnerable, they also are believed to be the primary drivers of biodiversity loss in developing countries [4] [18] [43] . In most developing countries, such forest dependent activities have become a source of conflict between local communities and government officials responsible for conservation [32] [42] . These conflicts are inevitable because dependence on forest resources is unsustainable in the face of increasing biodiversity loss in developing countries [8] . For this reason, forest dependence has become an important aspect of conservation in developing countries. As a result, several authors have called for a better understanding of drivers of forest dependence in order to manage it effectively [4] [20] .

Theoretically, forest resource dependence has been conceptualized as a multifaceted construct with multiple temporal and spatial dimensions [15] . For example, forest use results in multiple forms of commercial exploitation, recreational services, harvests of NTFPs, and subsistence livelihoods [15] . These forms of dependence operate and react differently at multiple levels―regional, national, community, household and individual [15] . The implication for multiple forms of forest dependence on these different levels is that forms of forest dependence and the unit of analysis have to be clearly defined to avoid ambiguity. Reference [10] proposed a conceptual framework for assessing the link between livelihoods and conservation from which forest dependence can be understood. In Salafsky and Wollenberg’s framework, forest dependence can be viewed through multiple dimensions of livelihood dependence on species, on forest habitat resources, as well as space and time of use [10] . In the model of forest dependence hypothesized in this study (see Figure 1), we limit forest dependence to animal species and habitat resources dependence because they are regarded as primary dimensions in the Salafsky and Wollenburg framework [10] , and are key indicators of biodiversity loss in developing countries [5] [43] .

Indicators of forest dependence can be seen in multiple forms. For example, the poor residents who depend on forests for subsistence livelihoods engage in activities such as bush meat hunting as well as forestland encroach- ment for farming [44] [45] . Other subsistence-based forest activities of poor residents who live in proximity to forests include harvests of NTFPs such as fuel wood, non-wood, and wood construction and handicraft-making materials and traditional medicine [13] . Evidence also exists that tropical forests are common sources of honey and water [46] . In this paper, we use the forest dependence indicators identified in the HPFD Index as valid measures of forest dependence at VNP.

The relationship between poverty and forest dependence has been extensively explored [1] [2] [4] [5] [10] [11] . From these studies, there is evidence to suggest that a correlation exists between poverty and human forest dependence behavior, which, in turn, results in biodiversity loss [4] [9] [17] [18] . It is also believed that alleviation of poverty reduces the degradation of biodiversity [7] . In fact, reference [9] refers to this as the “vicious cycle of poverty and environment relationship”.

Forest dependence among the poor in developing countries is commonly attributed to limitations, such as farmland scarcity and fragmentation, poor productivity of land as soil quality declines, and a lack of alternative sources of livelihood [18] . In some cases, economically enabled individuals are also known to influence forest dependence through fiscal, and political power [2] [25] . This is common when interests in forest resources are of commercial value. In resource deficient and highly populated developing countries, especially in Africa, forest dependence is largely driven by the subsistence livelihood needs [11] [47] . In a study done by Masozera and Alavalapati [4] , for example, rural poverty was found to highly influence forest dependence at Nyungwe National Park in Rwanda.

Despite the significant efforts made in both theory and practice, the link between poverty and forest dependence is still unclear [13] . One of the reasons for this is the poor understanding and conceptualization of poverty as a multidimensional phenomenon. Attempts to address this gap have led to the adaptation of the SL approach to establish the linkage between poverty and forest dependence [3] [5] . SL has been widely used to understand the relationship between poverty and biodiversity conservation, because it adds non-material aspects of a household’s or an individual’s wellbeing to the poverty definition [3] [12] [13] . In fact, reference [13] posits that managing biodiversity conservation effectively requires paying close attention to the livelihoods of residents in areas neighboring the forest areas.

It must be understood, however, that the effect of household livelihoods on forest dependence varies. To understand the relationship between poverty and forest dependence and in recognition of the existing gaps in literature, we must assess how different dimensions of poor household livelihoods influence or interact to influence forest dependence [10] . To do this effectively, we must systematically define, measure, and provide direct linkages between poverty and forest dependence [10] . Using SL framework, a model of the poverty and forest dependence relationship is hypothesized in Figure 1. Conceptually, it is designed as a Multiple Indicator and Multiple Causes (MIMIC) model for identification purposes, given the formative nature of livelihood indicators [48] [49] . Primarily however, a MIMIC model allows a clear view of the direct causal effect of each dimension of household poverty and each dimension of forest dependence.

This study was done in three phases following an exploratory sequential mixed methods design [50] . Phase one was carried out in September 2012, and it involved an exploratory evaluation of perceptions of key stakeholders including local residents with access to tourism benefits, local residents without access to benefits and park officials. The aim of this phase was mainly to explore using focus group interviews, the meaning of poverty and forest dependence from the local context. The results from the focus group interviews were then used to inform the design of phase two and to interpret phase three results. In phase two, a survey was done in April 2013 to validate the HPFD Index. The validated measures of the HPFD Index were then used to design an instrument that was administered in phase three during the months of June and July 2013. This survey was intended to use the HPFD Index to evaluate the relationship between household poverty and forest dependence relationship, which is the focus of this paper [50] .

The target population of this study was the local residents in absolute poverty, whose households neighbor VNP. In Rwanda, poverty is classified into three categories-poverty, extreme poverty and absolute poverty [51] . The latter is the lowest categorization of poverty, where a household has no land or whose land is insufficient to meet the basic livelihood needs and relies on farm wage work to maintain a living [51] . The sample selection of participants was narrowed from all the households in absolute poverty adjacent to the park to only those in Kinigi sector. Kinigi sector was selected out of four administrative sectors neighboring the park because it has higher incidences of illegal forest resource use activities at VNP. The selection of participants was limited to four administrative cells in Kinigi sector that are adjacent to the park including Nyabigoma, Nyonirima, Kaguhu, and Bisoke cells. From each cell, a list of extremely poor household heads in the respective cells was obtained from local leaders, who have all community residents classified according to their social-economic status. Four lists were labeled, combined and numbered to form a sampling frame for this study. Following the suggestion of Dillman and colleagues, a systematic random sampling was used to select participants from the sampling frame using a random interval of 4 [52] . A sample of 322 participants was finally selected to participate in this study.

The instrument used in this study was developed from the phase two validated HPFD Index presented in Table 1. Following DeVellis’ guidelines for writing clear and unambiguous indicator statements, measurement indicators making up the HPFD Index were rewritten in clear and concise belief statements that local residents could understand [53] . For each belief statement representing a measurement indicator, a response format was created using a seven-point Likert scale where 1 represented Strongly Disagree and 7 represented Strongly Agree. The developed instrument was pretested on a group of 10 residents who were in the same category as the targeted participants. A few indicators were rewritten to address the uncertainty that was observed during pretesting. The instrument was implemented with the support of 10 assistants who reside in the same area. These assistants were trained and practiced using mock interviews among themselves to ensure that they all understood the content and potential problems they might encounter during interviews. The use of interviewers in data collection was selected as the most effective approach because most participants cannot read and write.

Use of interviewers in data collection allowed maximization of response rates, enabled the clarification of questions, and assured confidentiality for participants. However, the interviewer-administered approach to a sur- vey has the potential for interviewer bias [54] . To manage such potential bias, three steps were taken, following reference [52] guidelines. First, an interview protocol was developed to guide interviewers in data collection. This protocol was one of the tools used during interviewer training and the instrument pretesting exercise, and also served as a reference document during data collection. Second, interviewers were trained before the survey. Following their training, an instrument pretesting exercise allowed the researcher to select assistants who exhibited a good understanding of required tasks and aptitude to be able make good judgments. Third, the researcher closely worked with interviewers during the survey and monitored them during interviews. Additionally, completed instruments were reviewed overnight to identify anomalies that could be discussed the next day of data collection in order to limit recurrence.

Note: Each statement was measured on a seven-point Likert response scale, where 1 represents Strongly Disagree and 7 represents Strongly Agree. ^*Statements with an asterisk were added to the Index for variable identification purposes [55] .

Data analysis for the household poverty and forest dependence construct relationship was done in two stages of pre-analysis and model testing. The pre-analysis stage was aimed at stabilizing parameter estimates, and involved data screening using the SPSS software package to identify outlier effects and assessment of univariate and multivariate normality assumptions [56] . Missing values that are normally of concern at this stage, were not in this study because the instrument was interviewer-administered, which enabled full survey completion. Five cases were excluded from the analysis because they were found to be extreme multivariate outliers [56] . Upon data cleanup and stabilization, a total of 317 usable cases (samples) were then employed, through a Structural Equation Modeling process, for the hypothesized measurement model testing and to examine coefficients of the hypothesized causal and direct effect construct relationships [56] . The model testing was run in the EQS software (version 6.2) using CFA for measurement and structural modeling [55] .

The CFA was used as a measurement model to determine the level of relationships and covariances in the poverty and forest dependence construct relationships [55] . CFA was also used to identify interrelationships between measured scales and latent constructs [56] . The Structural Equation Modeling procedure in this study therefore followed the recommended steps of model specification and estimation, testing model fit, and modification [55] [56] . Model estimation was performed using the Maximum Likelihood (ML) parameter estimation method [55] . The aim of using ML model estimation was to be able to observe the level of discrepancy shown by residuals between sample and population covariances implied by the hypothesized model [55] . Following review of residuals in covariance matrices for evidence of model misspecification [55] , the hypothesized model was then tested for goodness-of-fit was performed using Chi-square tests, Comparative Fit Index (CFI), Non- Normed Fit Index (NNFI) and Root Mean Squared Error of Approximation (RMSEA) [55] [57] . Both CFI and RMSEA are recommended in literature as considerably better indices for as assessing model fitness [55] . CFI is valued for the consideration of sample size effect [55] , while RMSEA is valued for the consideration of the number of estimated parameters in the model [55] [57] . The examination of the goodness-of-fit of a hypothesized model was performed and considered realistic when the CFI values were within the desirable range of 0.90 and above, with RMSEA values ranging below 0.08 [55] . The goodness of fit examination was aimed at determining whether the model is reflective of sample data. When this was confirmed, standardized parameter estimates were used to examine effect size in the hypothesized model [55] . Where the goodness-of-fit was found to be implausible (CFI < 0.90, RMSEA > 0.08), modifications were performed based on Lagrange Multiplier Test (LM Test) results that have strong theoretical justifications [55] [57] .

Gender, age, marital status, primary education attainment, income, livestock asset ownership, shelter availability and status, farmland ownership, and number of children in a household and their school attendance measured the characteristics of the sample. Using nominal scale responses, participants were asked to select a scale that best fit their response. A summary of demographic information on this sample is provided in Table 2. Demographic results, in brief, shows that a majority of participants were married (>93%), under the age of 30 years with about 4 children per household. A little over 18 percent of the sample had between 5 and 9 children in each household. Over 60 percent were women and more than 75 percent of participants had no basic primary education. The level of poverty was high among participants, as indicated in the table. For example, most of the participants (62.8 percent) earned less than 10,000 Rwandan Francs (approximately 15 US dollars) per month, and over 31 percent did not earn any income at all. Over 85 percent did not have basic livestock assets such as chickens or goats, and about 60 percent did not have land for cultivation. Almost 90 percent of participants live in incomplete shelters, and of these households, 62.5 percent did not send their children to school.

Confirmatory Factor Analysis was used to statistically assess the hypothesized measurement model presented in Figure 1. Goodness of fit for the model was examined to ensure that the hypothesized model was acceptable and consistent with sample data [56] . The initial assessment showed an indication of some degree of misfit in the hypothesized model, as CFI and RMSEA values were in the lower bounds of the plausible range of model fitness (CFI = 0.902, RMSEA = 0.08). To detect the potential source of misfit, LM Test results were examined to identify univariately the misfitting parameters with a sharp drop in Chi-square [55] . Following univariate test results, error covariances between harvests for crop support wood and harvests of bamboo wood for subsistence use, hunting bush meat for medicinal use and hunting bush meat for subsistence use, as well as between harvests of bamboo wood for subsistence use and hunting bush meat for income were specified. The model was re-esti- mated and indicated plausible fit (CFI = 0.95, RMSEA = 0.058).

Following the examination of model fitness, the hypothesized structural relationships were examined and the results are presented in Table 3. They indicate that food insecurity (H1a; β = 0.190, p < 0.05; H1b; β = 0.297, p < 0.05) and health insecurity (H2a; β = 0.187, p < 0.05; H2b; β = 0.264, p < 0.05) are two household poverty dimensions that influence both forest dependence dimensions of dependence on animal species and dependence on forest resources. Results also demonstrate that both education insecurity (H3a; β = 0.234, p > 0.05; H3b; β = 0.092, p > 0.05) and economic insecurity (H4a; β = −0.924, p > 0.05; H4b; β = −0.844, p > 0.05) do not significantly influence either forest dependence dimensions.

¹Hypothesis supported by consideration of direct effects; ²Hypothesis supported after consideration of indirect effects (see Table 4).

Results of the hypothesis testing for the relationship between household poverty and forest dependence in Table 3 above were based on direct effects. However, there is evidence that without a closer examination of indirect effects, understanding the relationship under investigation may be limited [56] . Conversely, indirect parameter effects associated with each measure of the four household poverty dimensions were examined to identify measures with significant indirect relationships with forest dependence.

Results in Table 3 indicate quantifiable measures of the four household poverty dimensions with potential to influence indirectly both forest dependence dimensions. As seen earlier, only two dimensions of household poverty (food and health insecurity) were found to have a direct influence on forest dependence, as indicated in Table 3. However, a closer examination of indirect effects in Table 4, reveals that two of the measures for the education insecurity construct (high adult illiteracy and limited access to scholastic materials for children from poor households) may also indirectly contribute to forest dependence. Additionally, results also indicate two quantifiable measures of food insecurity (insufficient annual production of food and skipping meals frequently) and two measures of health insecurity (inability to access healthcare centers and unavailability of clean water) significantly influence dependence on forest animals and resources for livelihoods among poor households neighboring VNP. Interestingly, the economic insecurity dimension of Household Poverty was not found to either directly or indirectly influence forest dependence at VNP, contrary to what is commonly believed locally.

From the above analysis, a structural model with the significant relationships was run to identify how each of the indicators―food insecurity, health insecurity and education insecurity―directly affect forest dependence dimensions as well as their indirect effects on the individual measures of both forest dependence dimensions obtained from the HPFD index. Results presented earlier reveal that the food insecurity, health insecurity and education insecurity influence forest dependence. When the indirect effects of household poverty on individual forest dependence indicators are examined, results presented in Table 4 reveal that food insecurity is the main cause of forest dependence, and this effect is attributed to both insufficient annual food production and the high frequency of skipping meals that are common among the poor households neighboring VNP. Additionally, results reveal that difficulty in accessing health care and high adult illiteracy rate are also among the key attributes of poverty that drives forest dependence among poor households neighboring VNP.

^*Total and indirect effects of household poverty on forest dependence is significant at 0.05.

To determine the household dimensions with greater influence on forest dependence, the structural model (Fi- gure 2) was used as a baseline model to test for invariance of mean structures between each of the quantifiable measures of food insecurity, health insecurity, and education insecurity, and forest dependence constructs they were hypothesized to predict. The model was estimated and produced a good fit (CFI = 0.985; RMSEA = 0.045). Following the guidelines for testing invariance of mean structures, a fully constrained model was estimated, and direct relationships between six measures of each dimension of household poverty and forest dependence were fixed to equivalence [55] [56] . To test for invariance, change in Santora-Bentler Chi-square was examined along with the change in CFI as indicated in Table 5. The consideration of the change in CFI was based on the rationale that the sole use of change in Santora-Bentler Chi-square to evaluate invariance is impractical and should be considered along with change in CFI [55] .

Both models were compared and the results presented in Table 5 reveal noninvariance (ΔS-Bχ2 30.857; Δdf = 10; p < 0.01). However consideration of change in Santora-Bentler Chi-square can be misleading [55] . When change in CFI is considered, results show the evidence of equivalence with a minimal change in CFI of 0.008, which is below the proposed level (ΔCFI > 0.01) for noninvariance [55] . Therefore, these results confirm that food insecurity, health insecurity, and education insecurity have equal effect on forest dependence and none of them has a greater influence or effect on either forest dependence dimensions.

¹Scaled differences were calculated using the Santorra-Bentler Scaled Chi-square adjusted difference tests (Santorra & Bentler, 2001); CFI = Comparative Fit Index; df = degree of freedom; S-Bc² = Santorra-Bentler Scaled Chi-square; Prob = test of significance at 95% confidence interval.