2.2 Approach and Design
This study employed a mixed method approach, drawing on both quantitative as well as qualitative research methods [46]. A mixed methods approach is valuable as it can draw from the strengths and minimize the weakness of both and it is now being widely used and recognized as a research paradigm in itself [47]. A cross-sectional quantitative household survey was used to elicit information on respondents’ characteristics, their behaviour in relation to on-farm tree planting, as well as the attitudes, subjective norms and perceived behavioural control in relation to tree planting. In addition, qualitative focus group discussions were conducted to explore some of the findings in more detail and as a way of triangulating the results of the survey.
2.3 Sampling and Sampling Procedure
A total of 288 respondents were randomly selected from a sampling frame of 540 households established by the National population and housing census of 2012 using the formula by [48]. The unit of analysis was the household and the subject of analysis was the household head.
s = ” x^2NP(1-P)
d^2(N-1)+ x^2P(1-P)
Where:
s = required sample size
N = Sampling frame (the given population size, this case N = 540)
P = Population proportion that for table construction was assumed to be 0.50 as this magnitude yielded maximum possible sample size required.
d = degree of accuracy as reflected by the amount of error that can be tolerated in the fluctuation of a sample proportion (P)
x^2 = Chi-square value corresponding to one degree of freedom relative to the desired level of confidence (95 percent)
Scholars have found that using mixed methodologies during in-field studies are particularly helpful because it allows a study to better capture the complexities seen on the ground [49]. This study combined household surveys, focus group discussions and field observations. The quantitative analyses provide an additional rigour to the research that is often lacking in agroforestry project evaluations [50].
A household survey was used to elicit information about respondents’ attitudes, perceptions and behaviour in relation to tree planting. Prior to the survey, informal visits and discussions with farmers and an exploratory survey were conducted in both study areas to elicit information about beliefs, attitudes, normative referents and control factors in relation to tree planting. In these interviews respondents were asked about their experiences with and opinions of planting trees and this information was used to develop the final questionnaire. The questionnaire comprised two parts. The first part contained questions about personal, household and farm characteristics, as well as questions on the extent of tree planting. Several socio-economic variables were extracted from this part of the survey and used in the analysis in this paper. These included age, sex, education level, employment, wealth, household size, estimated annual income (estimated by the respondent in the local currency) and farm experience. The questionnaire also asked respondents about any trees they have planted on their farms or on communal lands, making the behaviour studied reported rather than actual measured behaviour.
The second part of the questionnaire consisted of an attitude scale to assess the attitudes, subjective norms and perceived behaviour control towards tree planting. Based on the responses during the informal discussions and exploratory survey, items for an attitude scale were developed to measure the modified TPB constructs towards tree planting. The response format used in the attitude scale was a five-point Likert scale [51]. The components of attitude were each measured on a scale ranging from ‘strongly agree’ (5) to ‘strongly disagree’ (1). The components of subjective norm were evaluated on a scale ranging from ‘strongly agree’ (5) to 'strongly disagree’ (1). The control beliefs were also measured on a scale ranging from ‘strongly agree’ (5) to ‘strongly disagree’ (1).
In each district, 12 villages were selected using random numbers from a list of villages provided by District Lands Officers. In each village, 12 households were selected randomly from the lists of all farm households in each village. The household head was interviewed, in most cases this was a male, but in some cases, mostly due to divorce, death, separation or long term absence of the husband, the woman was the household head. If the head of the household was not available to be interviewed, another household was selected from the list using the random sampling procedure. In Nzega District, the household survey was administered to 65 male headed households and 79 female-headed households, whereas 86 male-headed households and 58 female-headed households were included in the survey in Sikonge. The final list was piloted to improve the order of the statements. The final questionnaire was administered to 288 respondents.
Upon completion of all interviews and surveys, interactive focus group discussions were conducted. They were carried out according to the methodology described by [52]. In each district, two focus group discussions were carried out with female participants and two with male participants resulting into 8 focus group discussions in total. Each Focus Group Discussion (FGD) consisted of 7-9 participants and lasted approximately one hour. After the villages had been selected, participants were selected randomly from the list of all farming households provided by the Village Executive Officers (VEOs). Some participants of the focus group discussions had also participated in the household survey in the preceding cycle. A discussion guide was developed and the focus group discussions were conducted in the national language of Kiswahili. The focus group discussions included several open discussion questions about people’s experiences and opinions about tree planting.
2.4 Structured Equation Modeling
This study applied the Structural Equation Modelling (SEM) approach performed in Partial Least Squares (PLS), a path modelling technique, to analyse the survey data. The SEM approach, unlike other widely used methods (such as multiple regression, multivariate analysis of variance factor analysis and path analysis) which can only examine a single relationship at a time, combines factor analysis and multiple regression analysis which enables the investigation of a series of dependent relationships [53]. SEM techniques should not be operated without a strong theoretical foundation for specification of both the structural and measurement models [54]. SEM has been used to study environmental behaviour in a variety of fields including tourism [55], agriculture [56], and risk perception [57]. Therefore, the usage of SEM along with the TPB was best suited for this study.
2.5 Measurement Scale
A measurement scale was developed for each major variable consisting of multiple items (indicators) borrowed from previous studies. Guided by the theoretical understanding derived from the literature, questionnaire items that were relevant to the constructs in this study were identified. The principal constructs were developed based on existing measures where possible or were adapted from similar scales. Measures for attitude (A), perceived behavioural control (PBC), and subjective norms (SN) (societal norms and social influences) were based on empirical studies of [25], [24], and [18]. Although most items were based on previous empirical studies, actual measurement scales were developed to capture the context of this study. The questionnaire items were then modified to match this study of on-farm tree planting in Tabora region.
2.6 Analytic Framework
The quantitative survey data collected was analyzed using non-parametric statistical techniques to detect associations and differences between respondents of the two study sites and between male and female household heads. First, the psychometric quality of the measures was assessed by calculating their validity and reliability. Second, the theoretical relationship between the variables was tested by estimating structural models. To obtain more accurate results, the SEM technique using the PLS algorithms was applied to evaluate the measurement model and structural model simultaneously. To conduct quantitative research on practical problems, the SEM evaluates the theoretical model according to the extent of consistency between the theoretical model and the actual data. The use of SEM is commonly justified in the social sciences because of its capability to impute relationships between unobserved constructs (latent variables) from observable variables.
This approach lends itself to this research because SEM answers a set of interrelated research questions in a single, systematic, and comprehensive analysis [58]. It also accommodates latent variables that are unobservable and cannot be directly measured. Therefore, the use of LVs in this study has the potential to model theoretical constructs such as intentions, attitudes, and perceptions that are difficult to measure directly. Analysis of Variance (ANOVA) was used to test if attitudes, subjective norms and perceived behavioural control were different among the respondents. The test was performed in order to assess whether the TPB constructs explain significant variance in tree planting behaviour among farmers. This step was necessary because there were no prior knowledge of potential multicollinearity among variables. It was necessary to use this test to explain variations in tree planting behaviour. SEM techniques were performed with the aid of Analysis of Moment Structure (AMOS) version 19 [59] software package. Focus group discussions were transcribed verbatim and coded according to a thematic framework, and presented in narrative summaries. The software used for the transcriptions was f4 (Windows). Data were analysed using the Software Atlas.ti v 6.2.26.
Results
3.1 Response rate
Two hundred and ninety questionnaires were administered to household heads in selected villages of Nzega and Sikonge Districts of Tabora region, Tanzania. The response rate was 100 percent, with 288 valid questionnaires, which was considered adequate for testing the stated hypotheses. Sample size plays a major role in the estimation and interpretation of SEM results [60]. In general the literature suggests that sample sizes for structural equation models commonly run in the 200 to 400 range. This study sample size is reasonably enough to analyze descriptive statistics, multivariate analysis and structural equation model. There are several studies using less than 300 of sample size, such as seatbelt use (N=277) by [61], motorcyclists’ intention to speed (N=110) by [62], drivers’ decision speed (N=250) by [60], and truck driver behaviour (N=232) by [63].
3.2 Socio-demographic Characteristics of the Respondents
The socio-demographic characteristics of the sample are summarized in Table 1 and the parameters included are age, sex, education, household size, wealth and farming experience. Referring to the age of the respondents, most of them laid between 20 to 50 years (64 percent and 73.8 percent in Nzega and Sikonge districts respectively), then followed by those aged below 20 years and above 50 years by (2 percent and 0 percent) and (34 percent and 26.2 percent) for Nzega and Sikonge districts respectively. The mean age was 39.4 years (SD = 11.7, range = 18 ‘ 63) while only a few of the respondents were above sixty-five years of age. The larger number of a young population could imply increased pressure on agricultural land and therefore momentously affect its economic value. The chi-square tests indicated no significant (p>0.05) difference in age between villages in the two districts. As regards gender, out of the total sample, (45.1 percent) and (59.7 percent) of the respondents were males in Nzega and Sikonge districts respectively. The majority of the sample was male (52.4 percent). In this respect, this proportion
Table 1. Summary of socio-demographic profile of the respondents in study villages in two districts of Tabora region
Respondents characteristics Nzega Sikonge
_____________________________________________________
Frequency Percentages Frequency Percentages
Age
Below 20 years 3 2 0 0
20 ‘ 50 92 63.8 107 73.9
50 and above 49 34 38 26.2
”2 = 2.156, df = 2, p-value = 0.340
Gender
Male 65 45.1 86 59.7
Female 79 54.9 58 40.3
”2 = 0.102, df = 1, p-value = 0.749
Marital Status
Married 90 62.6 81 56.3
Singles 47 32.8 54 37.4
Widowed 7 4.6 9 6.3
”2 = 2.237, df = 2, p-value = 0.327
Education
None formal education 9 6 22 14.8
Primary school 127 88 109 75.4
Secondary school 9 6 12 8.2
Adult education 0 0 3 1.6
Post secondary 0 0 0 0
”2 = 10.125, df = 4, p-value = 0.038
Wealth
Economically poor 81 56 95 65.6
Very poor 55 38 40 27.9
Better off 9 6 10 6.6
”2 = 1.55, df = 2, p-value = 0.461
Household size
1-3 people 9 6 36 24.6
4 – 6 people 58 40 36 24.6
7- 10 people 75 52 66 45.9
More than 10 people 3 2 7 4.9
”2 = 11.750, df =2, p-value = 0.003
Farming experience
1-10 years 39 27 35 24.3
11-19 years 48 33 41 28.5
20 and above years 57 40 68 47.2
”2 = 0.276, df =2, p-value = 0.871
Notes: Alpha level or significance level set at 0.05
Source: Tabora Population-Agroforestry study, 2016
explains the fact that most of those who practice agroforestry and tree planting in general are mostly matured males though in reality, it is the women who are engaged in crop farming. On the other hand, female respondents were (54.9 percent) in Nzega and (40.3 percent) in Sikonge. These findings reveal the presence of more males than females when both districts are combined. The chi-square tests showed no significant difference (p >0.05) in gender between villages in the two districts. Educational wise, majority (88 percent and 75.4 percent in Nzega and Sikonge districts respectively) of respondents had completed primary school education, followed by those with none formal education (6 percent in Nzega and 14.8 percent in Sikonge), while those attained secondary education were 6 percent in Nzega and 8.2 percent in Sikonge. Very few of them have attended adult education (0 percent in Nzega and 1.6 percent in Sikonge). The chi-square test indicated significant difference (p<0.05) in education level between villages in the two districts. Additionally, a one-way ANOVA test was performed and presents that there were significant age differences among people reporting different education levels: F (4, 7.68) = 4.011, p = .047, ”2 = .01. The mean age of those who obtained a primary school education level certificate was (M = 42.24, SD = 12.52). These results show that majority of respondents attained primary education, thus indicating a low level of education in the study area. It further reveals minimal application of land management practices in the study villages, which partly could be caused by low level of education, amongst other factors. Regarding wealth, a large proportion of respondents were economically poor by 56 percent in Nzega, and 65.6 percent in Sikonge, followed by very poor (38 percent) in Nzega and 27.9 percent in Sikonge and better off were 6 percent in Nzega and 6.6 percent in Sikonge. The chi-square test indicated further no significant difference (p>0.05) in wealth between the two districts. These results imply that majority of respondents in both districts are economically poor. Furthermore, as discernible from Table 1, the average annual income of respondents was the equivalent of Tshs 885,132.05 per annum ($1=Tshs 2,228 as per exchange rate of 2017). Farmers in the lowest range of annual income (the very poor) might, however, need to augment their income earnings as they were all leaving below the poverty line by earning below Tshs 2,300 which is approximately $1 per day. This confirms the prevalence of poverty among rural farmers in Tabora region and Tanzania in general.
On household size, the majority of respondents in Nzega (52 percent) and Sikonge (45.9 percent) had a household size ranging between 7 to 10 household members. 1 to 3 households’ members were 6 percent in Nzega and 24.6 percent in Sikonge, while 58 (40 percent) in Nzega and 36 (2.6 percent) in Sikonge had 4 to 6 households’ members. Few respondents by 2 percent in Nzega and 4.9 percent in Sikonge had more than 10 households’ members. Overall 59.52 percent of the respondents have an average household size of 5 people with a Standard Deviation (SD) of ” 2 in both areas. Chi square tests indicated high significant difference (p<0.05) in the size of households between the two districts.
As regards marital status, the majority of the sample included couples (62.6 percent), followed by singles (32.8 percent), and widowed (4.6 percent) for Nzega and 56.3 percent for the married, followed by singles (37.4 percent) and widowed (6.3 percent) for Sikonge. The chi- square tests for marital status indicated no significant (p>0.05) difference in age between villages in the two districts. The farming experience of the respondents ranged from 14 – 51 with a mean of 31.19 (+18.09) years. Furthermore, 35.72 percent of farmers had an average total farm size of 0.74 (SD 0.64) hectare in Nzega and 2.37 hectares (SD 2.29) in Sikonge. This goes to confirm that land holdings in the rural areas are usually small and is obtained mostly through inheritance. Almost all households (98 percent) own land, and some respondents (11 percent) said they rented additional land for farming. In the bivariate (chi-square) test, farm experience and tree planting yielded insignificant results at p = 0.871.
3.3 Assessment of validity and reliability of the measurement items
As a first step, construct reliability and validity was assessed. The interpretation of the resultant coefficient takes into account the actual factor loadings rather than assuming that each item is equally weighted in the composite load determination. In this study, construct reliability was measured using [64], with a value of 0.7 or higher being recommended [65]. Construct reliability for all the factors in this study’s measurement model were above 0.7 an acceptable threshold representing strong reliability. [53] recommended a factor loading of 0.5 and above to be an acceptable indicative of validity at the item level. Construct validity for the measurement scales was assessed from their convergent and discriminant validity values. Convergent validity which indicates how each measurement item strongly correlated with its specific theoretical construct was determined from the constructs’ respective Average Variance Explained (AVE) values. Convergent validity was evaluated for the measurement scales using three criteria suggested by [66]: (1) all indicator factor loadings should be significant and exceed 0.7, (2) construct reliabilities should exceed 0.7, and (3) the square root of the average variance explained (AVE) by each construct should exceed the variance due to measurement error for that construct (i.e., AVE should exceed 0.50). All values in the Confirmatory Factor Analysis (CFA) model exceeded 0.7 and were significant at p = 0.001. Composite reliabilities of constructs ranged between 0.78 and 0.93. AVE ranged from 0.58 to 0.87 indicating that on average, all Latent Variables (LVs) were able to explain more than half of the variance of their respective indicators and thus demonstrated sufficient convergent validity. Therefore, all three conditions for convergent validity were met.
3.4 Evaluation of the Measurement Model (outer model)
The measurement model specifies the relationships between the constructs and the associated indicators. The parameters in the SEM were estimated by maximum likelihood (ML) method using the computer software program AMOS version 19. A variety of indices was used in this study. These include absolute fit indices that measure how well the proposed model reproduces the observed data. In other words, the fit indices assess the overall discrepancy between the implied and observed covariance matrices.
Measures of overall model fit included absolute, incremental, and parsimonious indices of fit. The widely-known index of absolute fit is the Chi-square (”2). While the ”2 statistic has been found to be sensitive to sample size, two indices were used to assess the overall absolute fit of the proposed model: the Goodness of Fit Index (GFI) and the Comparative Fit Index (CFI). To assess the fit of the proposed model and for incremental fit measures, the Adjusted Goodness of Fit Index (AGFI), the Incremental Fit Index (IFI), and the Normed Fit Index (NFI) were used. Finally, the Root Mean Square Error of Approximation (RMSEA) was used to evaluate the parsimonious fitness of the proposed model (used in this study). The suggested minimum acceptance values and the observed values from this model of these indices are presented in Table 2. The test of the overall model fit yielded ”2 = 289.2 with 177 degrees of freedom and a p-value of more than 0.05. Thus, it is accepted that the model fits the data. The other indices are also higher than the suggested values. The recommended cut-off value for the goodness of fit indices was based on [67] recommendation. Based on the recommended values, the study concludes that the research model fitted the data quite well.
Table 2. Goodness-of-fit test results.
Fit index ”2 GFI CFI RMSEA AGFI IFI NFI
(p-Value)
Suggested value >0.05 >0.9 >0.9 <0.08 >0.8 >0.9 >0.9
Observed value 0.061 0.902 0.973 0.022 0.837 0.911 0.902
Conclusion Accepted Good fit Good fit Good fit Good fit Good fit Good fit
Source: Tabora Population-Agroforestry study, 2016
3.5 Evaluation of Structural Model (inner model)
The structural model represents the relationship between the constructs. It specifies the relationships between the latent variables. Latent variables can play the role of predicting. A latent variable which is never predicted is called an exogenous variable. Otherwise, it is called endogenous variable. This section presents results of the test of the structural model (in which research hypotheses are embodied). The structural model was tested using the structural equation modeling (SEM) approach performed in PLS. This approach is particularly appropriate for testing theoretically justified models [68]. Each indicator (manifest variable) was modeled in a reflective manner which means a variation of the construct yields a variation in the measures. As a result, the direction of causality is from the construct to the indicator. Each manifest variable represents the corresponding latent variable, which is linked to the latent variable using a simple regression model. The six constructs comprise four exogenous variables (attitudes, subjective norms, perceived behaviour control, and situational factors and two endogenous variables (intention and behaviour). All of these were linked as hypothesized (see Figure 4), and model estimation was done by assessing the path coefficients that indicate the strength of the hypothesized relationship between the exogenous and the endogenous variables and the variance explained (R2 value) by each path. Figure 4 shows the standardized path coefficients and path significance, as reported by PLS. The betas were used to determine the relative weights of each factor.
The sample size of n=288 was sufficient because the required number of cases for this PLS analysis is only ten times the number of indicators in the reflective constructs [69]. The modified model derived from the Theory of Planned Behavior, was made up of all reflective constructs that are influenced by the prime latent indicators [70]. These reflective latent constructs (attitude, subjective norms, perceived behavioural control, intention situational factors, and behaviour), are characterized by the fact that changes in the underlying latent construct will be reflected in changes in their corresponding measurement indicators. Since the indicators in a reflective construct represent the construct in a reflective model, a high degree of correlation between the indicators was expected to be seen.
Hypothesis 1 examined the influence of socio-demographic variables on the TPB constructs. A One-way ANOVA test was performed to determine whether socio-demographic characteristics significantly and positively influence the respondents’ attitudes, subjective norms and perceived behaviour controls regarding tree planting behaviour. In this regard, the socio-demographic characteristics are statistically significant when the p-value is less than 0.05. Table 3 shows the statistics of the effects of the socio-demographic characteristics on tree planting behaviour variables. Results show that age, gender, wealth, and farming experience do not appear to be statistically significant except household size and education. Farmers of different education levels have significantly different separation behaviours, with the secondary education group demonstrating more positive behaviour than those with lower education. For situational factors, all sociodemographic variables were not significant.
Table 3. Analysis of variance of demographic variables.
Values of Significance (p)
Socio-demographic Attitude Subjective Perceived Intention Situational Behaviour
Variable Norm Behavioural Factors
Control
Gender 0.610 0.483 0.660 0.711 0.164 0.264
Age 0.589 0.391 0.372 0.703 0.314 0.405
Education 0.027 0.043 0.036 0.033 0.221 0.049
Household size 0.006 0.027 0.002 0.046 0.331 0.039
Income 0.803 0.344 0.534 0.360 0.522 0.662
Farming experience 0.069 0.211 0.257 0.077 0.311 0.682
Source: Tabora Population-Agroforestry study, 2016
Through the lens of the TPB, this study went further examining whether differences in gender exist within TPB constructs and whether these differences explain observed gender differences in agroforestry. T-test for independent groups was performed for each case.
Attitude toward tree planting was more favourable among females (M= 3.97, SD= .75) than males, (M= 3.65, SD= .83), t (3301) = -11.31, p < .001. Further, on each behavioural belief item comprising the attitudes construct, women reported more favourable beliefs than men, p< .001. Females also reported greater control beliefs and perceived facilitation over tree planting (M= 3.75, SD= .96) than males (M= 3.50, SD= 1.07), t (3291) = -6.91, p < .001. Across all items comprising the construct, women reported higher confidence than males, p < .001. Males on the other hand reported greater normative beliefs regarding tree planting (M= 2.29, SD= .83) than females (M= 2.13, SD= .81), t (3305) = -5.66, p < .01. As regards geographic locations of the study, respondents in Nzega district had more positive attitudes and subjective norms towards tree planting compared to respondents in Sikonge district. Irrespective of differences on their influence to TPB constructs, background factors held direct paths to antecedents of intention.
Hypothesis 2 examined the relationship between farmers’ attitude towards tree planting and their behavioural intention to adopt tree planting. In regression analysis, it yielded (”=0.421, t-value=17.64, p<0.001). This hypothesis was strongly supported and therefore not rejected. Hypothesis 3, examined the relationship between subjective norms and farmers’ intention to adopt tree planting (”=0.213, t-value=7.59, p<0.01). This hypothesis was also strongly supported and therefore not rejected. The fourth hypothesis examined the relationship between perceived behavioural controls and farmers’ intention to adopt tree planting. (”=0.138, t-value= 4.41, p<0.05). This hypothesis was also strongly supported and therefore not rejected. Hypothesis 5 examined the relationship between intention and behaviour to adopt tree planting (”=0.62, t-value=17.59, p<0.001). This hypothesis was strongly supported and therefore not rejected. Hypothesis 6 examined the relationship between situational factors and behaviour to adopt tree planting (‘=”0.55, t’=”9.12, p'<‘.001). Situational factors and tree planting behaviour were negatively correlated, but the influence was significant; thus, H6 was supported. Tree planting behaviour will only be restricted when farmers have barriers.