Study area

The research was conducted across the four Western Southeast (WSE) states: Arkansas, Louisiana, East Oklahoma, and East Texas (Fig. 1). The eastern parts of Texas and Oklahoma were taken in the study because of the concentration of primary wood processing mills in these areas (USDA Forest Service 2021a, 2021b, Klockow et al. 2023). Generally, the forests within the WSE are dominated by loblolly/shortleaf pine, oak, and hickory forests (USDA Forest Service 2023).

View Figure

• Download as Powerpoint
• Download Figure

Arkansas has 19 million acres of forestland, with 39 percent covered by oak/hickory, 33 percent by loblolly/shortleaf pine, 21 percent by mixed species, and the remainder by others (USDA Forest Service 2023). Most of the forestland (80.3%) is owned by private landowners, although the remaining 19.7 percent is owned by the public. Arkansas has the second most forest-dependent economy in the United States, contributing 4.1 percent to the total state’s GDP (Arkansas Center for Forest Business 2023).

Louisiana has 15 million acres of forestland, where loblolly/shortleaf pine is the dominant species (37%), followed by oak-hickory (10%) (USDA Forest Service 2023). Private entities own over 86 percent of the forestland, with public entities owning the remainder. Louisiana ranks 12^th in forest-dependent economies in the United States, contributing 2.1 percent to the state’s GDP (Arkansas Center for Forest Business 2023).

East Oklahoma has six million acres of forestland, primarily composed of oak and hickory forests (52%), followed by loblolly and shortleaf pine (24%), and mixed species (13%) (USDA Forest Service 2023). Nearly 85 percent of the forest land is privately owned, with the remaining 15.6 percent owned by the public. Oklahoma has the 28^th most forest-dependent economy in the United States, contributing 1.5 percent to the state’s GDP (Arkansas Center for Forest Business 2023).

In East Texas, 11 million acres are covered by forest land, 46 percent of which is loblolly/shortleaf pine, followed by mixed species (24%), oak/hickory (21%), and other forests (USDA Forest Service 2023). Out of the forest area, 89.8 percent is owned by private owners and 10.2 percent by the public. Texas has the 32^nd most forest-dependent economy in the United States, contributing 0.9 percent to the state’s GDP (Arkansas Center for Forest Business 2023).

Data types and their sources

The dependent or predicted variable was the number of mills (count variable) in each county of the study area. The explanatory variables were chosen based on the existing literature on wood processing mills. The explanatory variables were grouped into socio-demographic characteristics, transportation infrastructures, forestland characteristics, and mill attributes, as listed in Table 1. Depending upon the nature of the data structure, both binary—and continuous-type explanatory variables were included in models.

The county-level forest characteristic was obtained from the Forest Inventory and Analysis (FIA) database using the EVALIDator tool. The FIA database, managed by the USDA Forest Service, provides data on US forests, including annual removal, growth, and growing stock. It collects information through a network of permanent plots under the supervision of the National Forest Inventory (NFI), resurveyed every 5-10 years (USDA Forest Service 2025). The EVALIDator tool helps to access information from the FIA database and provides estimates and sampling errors of forest statistics (Miles 2015). County-level forestland characteristics such as annual net growth (cu.ft.) (Growth), annual removal (cu.ft.) (Removal), and growing stock (cu. ft.) in accrete were collected from Forest Inventory and Analysis using the EVALIDator tool for softwood and hardwood, respectively. The values in cubic feet were converted into tons using the conversion factor of 0.030 for softwoods (pines) and 0.049 for hardwoods for ease of interpretation and uniformity. The conversion factor was calculated by dividing the growing stock in green tons by the growing stock in cubic feet for softwoods and hardwoods, respectively. Meanwhile, the data on transportation, including the presence of state highways, county roads, other local roads, and railways, was obtained from TIGER/Line. The US Census Bureau uses TIGER/Line to provide information about features in various formats, like shapefiles, on nationwide data (US Census Bureau 2024).

The socio-demographic characteristics of each county were obtained from the US Census Bureau and Bureau of Labor Statistics: the total population aged between 18 and 64 years with different education levels and the annual average weekly wage for the forestry and logging industry for the year 2023. The mill characteristics, like location, type, production capacity and capacity units, species used, wood demand at capacity, and status, were collected from the Forisk North American Forest Industry Capacity Database, 2024. Forisk Consulting has several objectives, one of which is to update its yearly database on primary mills, including their locations and other related data, to forecast market trends, demand, and other relevant information (Forisk Consulting 2024).

Variable selection

The Poisson regression assumes that the response variables are independent of one another (Hilbe 2014). In this study, we aimed to examine the effects of independent variables (socio-demographic factors, forest characteristics, and transportation infrastructure), which included several variables that were correlated with each other. To address this issue and reduce dimensionality, we performed Spearman rank correlation analysis (Mukaka 2012) and grouped the variables based on their correlations (Dormann et al. 2013). In group 1, there were five variables (population with a bachelor's degree, college or associate degree, graduate professional, studied high school, and studied less than high school). In group 2, there were two variables (frequency of state highways and frequency of local roads). In group 3, there were nine variables (area of timberland, average growth, average removal, surplus (equation 2), average growth (softwood), average removal (softwood), and surplus (softwood). Other variables, such as frequency of county roads, frequency of US highways, average railway proximity to PWPM, average railway proximity to SWPM, annual average weekly wage, and total annual average weekly wage (for all industries), were uncorrelated, meaning they were independent of each other. From the three correlated groups, we selected one representative variable per group and fitted several Poisson regression models along with the uncorrelated groups to assess their impact while minimizing multicollinearity (Graham 2003). Note that the variables mentioned above might not be reported in Table 1, as it only includes the variables that were included in the models reported in Tables 2 and 3.

Table 2.Poisson Regression results and goodness-of-fit measures for factors influencing the presence of Primary Wood Processing Mills (PWPMs) across Western Southeast (WSE). The table presents estimated coefficients, odds ratios with 95 percent confidence interval (CI), p-values, and multicollinearity diagnostics (Variable Inflation Factor and Tolerance). Statistically significant variables are marked with * (p-value < 0.05), ** (p-value < 0.01), and *** (p-value < 0.05). The model’s fit was evaluated using the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), deviance/df, and the likelihood ratio test (LRT).

View Fullscreen

Table 3.Poisson Regression results and goodness-of-fit measures for factors influencing the presence of Softwood-Using Primary Wood Processing Mills (SWPMs) across Western Southeast (WSE). The table presents estimated coefficients, odds ratios with 95 percent confidence interval (CI), p-values, and multicollinearity diagnostics (Variable inflation factor and Tolerance). Statistically significant variables are marked with *** (p-value < 0.00) and * (p-value < 0.05). The model’s fit is assessed using the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), deviance/df, and the likelihood ratio test (LRT).

View Fullscreen

To further assure that there was no multicollinearity among the variables, variable inflation factor (VIF) and tolerance were tested (Zainul Armir et al. 2022, Kyriazos and Poga 2023). For VIF, a value less than 5 was taken as a value greater than that was rendered problematic (James et al. 2013). For tolerance, a value less than 2.5 was considered problematic as well, and the variables having a value less than that were removed (Stataiml 2024).

Model selection

The study employed the Poisson Regression model (equation 1) to identify county-level factors influencing the establishment of PWPMs and SWPMs, suitable for count dependent variables (Hilbe 2014). For PWPMs, the mean (0.803) was less than the variance (1.744); similarly, for SWPMs (second dependent variable), the mean (0.552) was less than the variance (1.05). However, according to Cameron and Trivedi (1986), only a higher variance does not necessarily mean that the data are overdispersed. Therefore, by using the overdispersion test by Cameron and Trivedi (1990), the Poisson Model was not overdispersed, as for the primary wood processing mills (PWPMs) and softwood-using primary wood processing Mills (SWPMs), confirmed by the insignificance of the overdispersion test (p-value > 0.05). The possible reason for the higher variance for PWPMs was because of the high number of zeros, i.e., 61 percent of the counties had zero PWPMs. To see if there is evident zero-inflation or not, the ratio of the observed and predicted number of zeros (observed = 67; predicted = 65.512) was 1.022. The ratio of approximately one indicated that there was no strong evidence of zero inflation based on the ratio. Therefore, we decided to employ the standard Poisson regression.

County-wise LTHS, AvgWW, AvgR, AvgG_Sft, Frequency_C, Frequency_LR, Frequency_US, AvgRP_PWPM, and AvgRP_SWPM were selected as the explanatory variables (Table 1) (Hagadone and Grala 2012, Aguilar et al. 2012, Zhang et al. 2013, Sasatani and Zhang 2015, Griffin 2020). All the explanatory variables listed in Table 1 remained consistent for Model 1 (PWPMs); however, for Model 2 (SWPMs) and Model 3 (HWPMs), all variables remained the same except for forest characteristics, where softwood was used for Model 2, and hardwood was used for Model 3. Different combinations of models were fitted, and only the best model was reported (Tables 2 and 3). Also, the railway proximity variable differed for PWPM and SWPM as the distances were calculated separately. Before fitting, the explanatory variables were transformed using log (1 + x) as it handles zeros to avoid undefined values and to even out the parameter approximates (Swihart et al. 2012, Cameron and Trivedi 2013, Hilbe 2014). However, the number of hardwood-using primary wood processing mills (HWPMs) was small, i.e., 37 counties had mills out of 203 counties (18%). Although the HWPMs had an average of 0.241 with a variance of 0.331, the fitted Poisson regression models exhibited underdispersion and had no significant variables. Further generalized linear regression was employed; however, the model was unsuitable to fit because of a lower number of HWPMs.

The study performed model fit test statistics such as AIC, BIC, deviance test, and Log-likelihood ratio test (LRT) to select the final models reported in Tables 2 and 3. And chose the models having the least AIC and BIC (Kuha 2004). We calculated the dispersion after the final model fitting to ensure no overdispersion. For equidispersion, the value should be approximately equal to 1, whereas a value greater than that is considered overdispersed (Nwakwasi 2018). Additionally, to test the model significance against the null model, we calculated the log-likelihood ratio test statistic (LRT) (Kent 1982). Odds ratio was also calculated for the better interpretation of the coefficient. This study did not report the model for HWPMs since it had no significant variables. The analysis was employed using R-studio.

The general mathematical equation of the Poisson Regression is (1) $$\mathit{log}\left(y\right)={\beta }_0+{\beta }_1{x}_1\dots \dots +{\beta }_{p-1}{x}_{p-1\mathrm{,}}$$where

y = Response variable,

β₁, β₂, …, β_p−1 = Numeric coefficients, β₀ being the intercept,

x_p−1 = Predictor variable up to the p−1, and

p = Number of parameters β in the model including the intercept.

Resource availability

The study estimated the availability of sustainable raw materials using the equation 2, which gives the surplus of the resources available. We used equation 2 to estimate the total surplus, as well as separate surpluses for softwood and hardwood. The merchantable net growth and removals utilized in equation 2 were averaged for five years using data available in Forest Inventory and Analysis database. For Arkansas, the data was available up to 2024; for Louisiana and East Texas, up to 2023, and East Oklahoma, up to 2018. (2) $$\mathit{Available}\mathit{resources}=\mathit{Growth}-\mathit{Removal},$$where

Growth = Annual net growth of merchantable bole volume of growing-stock trees (at least 5-inch d.b.h) on timberland (tons), and

Removal = Total annual removal of merchantable bole volume of growing-stock trees (at least 5-inch d.b.h) on timberland (tons).

Distribution of mills

Altogether, 205 primary wood processing mills (PWPMs) (open: 171, closed: 31, and announced to open: 2) were in the study area. Among open mills, the majority (65%) utilized softwood, only 29 percent utilized hardwood, and the remaining utilized mixed species (Fig. 2). Nearly 80 (39%) of 203 counties had PWPMs. The distribution of PWPMs in each county ranges from zero to six. Counties like Clark (AR), Columbia (AR), and Nacogdoches (ETX) had the highest number of PWPMs.

Of 203 counties, 63 (31%) had softwood-using primary wood processing mills (SWPMs). The distribution of SWPMs ranged from 0 to 5, with Polk County (ETX) and McCurtain County (EOK) each having five mills. At the same time, only 37 counties (17%) had hardwood-using primary wood processing mills (HWPMs) ranging from 0 to 4. Notably, Clark (AR) County had four hardwood mills.

The overall PWPMs had an average current annual production capacity of approximately 287,991 tons, ranging from 17.22 tons to 1,820,000 tons. When running at full capacity, the average wood demand of primary wood processing mills exceeds their current production capacity by a factor of 2.08. The difference between average current capacity and average wood demand at capacity implies that the PWPMs were underutilizing the raw materials and could meet more forest product demand if operated at their full potential. In the meantime, the SWPMs had an average of 289,567 tons of current capacity, while having an average demand at a capacity of 2.09 times greater than the current average capacity. HWPMs had a mean capacity of 220,035 tons, and mean wood demand was at a full potential of 2.66 times greater than the current average capacity. The higher average wood demand at capacity for SWPMs compared to HWPMs suggests that the production potential was much more favorable for SWPMs.

Across all regions, SWPMs were more prevalent than HWPMs (Figs. 2 and 3). Overall, as shown in Figure 2, the PWPMs were more clustered around southern Arkansas, northern Louisiana, and East Texas. Especially, Figure 3 shows that Arkansas had the highest number of PWPMs, with a substantial portion of SWPMs remaining operational (43) and nonoperational (6), whereas East Oklahoma had only five SWPMs operational and one nonoperational. East Texas had a higher number of both open (36) and closed (4) softwood mills compared to Louisiana (34 open; 5 closed). Similar to the SWPM, HWPMs were abundant in Arkansas (33 open; 5 closed), followed by East Texas (13 open; 2 closed) and Louisiana (8 open; 4 closed). At the same time, East Oklahoma had no HWPMs.

Resource availability

The resource availability varied significantly across the study area (Fig. 4). As shown in Figure 4a, the southern part of Arkansas, northern Louisiana, and parts of East Texas had the highest availability of total available resources, ranging from about 605 Mtons to over 1,247 Mtons (thousand tons). In contrast, some areas in East Oklahoma and Northern Arkansas showed low value, ranging from about 1 Mton to over 285 Mtons. The counties with negative values indicated higher removal or over-utilization of resources.

As shown in Figure 4b, softwood resources were most abundant in southern Arkansas, some portions of northern Louisiana, western East Texas, and some portions of East Oklahoma, with availability ranging from approximately 327 Mtons to 1,076 Mtons. Central and eastern Louisiana, along with parts of East Oklahoma and northern Arkansas, had lower softwood availability, some as low as zero or even negative values.

As shown in Figure 4c, the hardwood resources were more concentrated in northern and central Arkansas, East Oklahoma, and a few parts of East Texas and northern Louisiana. The highest availability ranges between 235 Mtons and 391 Mtons. Southern Louisiana and parts of East Texas had lower hardwood resources, with some areas showing minimal to no availability.

Factors affecting the mill establishment

The final model for PWPMs, shown in Table 2, includes Frequency_C, AvgRP_PWPM, LTHS, AvgWW, and AvgR. The variable AvgWW showed a positive significance (p-value < 0.05), and an odds ratio of 1.101 indicates that for each unit increase in the log of AvgWW, the expected number of PWPMs rises by 10.1 percent. The effect of AvgR was also positively significant (p-value < 0.05), with an odds ratio of 2.054, suggesting that for every unit increase in the log of AvgR, the expected number of PWPMs increases by 105.4 percent. AvgRP_PWPM was negatively significant with the odds ratio of 0.213, implying that with an increase in log of railway distance from the PWPM, the likelihood of the presence of PWPM decreases by 78.7 percent. However, the effects of Frequency_C and LTHS did not have a statistically significant effect on the establishment of PWPMs. The dispersion parameter for the model was approximately one, indicating that the model had no evident overdispersion. The significant likelihood ratio test ( ${\text{χ}}_{0.05,5}^2$) with a test statistic of 112.99 (p-value < 0.000), demonstrates that the full model significantly enhances the explanation of the variation in the number of PWPMs compared to the null model.

The final model for SWPMs reported in Table 3 included Frequency_LR, Frequency_US, AvgRP_SWPM, LTHS, AvgWW, and AvgG_Sft. The variable LTHS was positively significant (p-value < 0.05), with an odds ratio of 1.282, indicating that for every unit increase in the log of LTHS, the expected number of PWPMs increases by 28.2 percent. The effect of AvgG_Sft was positively significant (p-value < 0.05), and the odds ratio of 2.121 suggests that for each unit increase in the log of AvgG_Sft, the expected number of PWPMs increases by 112.1 percent. AvgRP_SWPM was negatively significant with the odds ratio of 0.002, indicating that with a unit increase in the log of Railway distance from the SWPM, the chance of PWPM establishment decreases by 99.8 percent. However, Frequency_LR, Frequency_US, and AvgWW did not show a significant effect on the establishment of PWPMs, as the variables were insignificant. The dispersion parameter for the model was approximately one, suggesting that the model accounted for overdispersion. The significant likelihood ratio test ( ${\text{χ}}_{0.05,5}^2$) with the test statistic of 102.830, indicates that the full model significantly improves the explanation of the variation in the number of SWPMs compared to the null model.