FAMILY FOREST LANDOWNER BEHAVIOR IN THE SOUTHEAST 
 
Except where reference is made to the work of others, the work described in this dissertation 
is my own or was done in collaboration with my advisory committee. This dissertation 
does not include proprietary or classified information 
 
 
 
 
 
 
 
Indrajit Majumdar  
 
 
 
 
David. N. Laband 
Professor 
Forestry and Wildlife 
Sciences 
 
 
 
 
 Lawrence D. Teeter, Chair 
Professor 
Forestry and Wildlife 
Sciences 
 
 
 
 
C. Robert. Taylor 
Agricultural Economics and  
Rural Sociology 
 
 
 
 
 Yaoqi Zhang 
Professor 
Forestry and Wildlife 
Sciences 
 
 
 
 
Stephen. L. McFarland 
Graduate School 
Acting Dean 
 
FAMILY FOREST LANDOWNER BEHAVIOR IN THE SOUTHEAST 
 
 
Indrajit Majumdar 
 
 
A Dissertation 
Submitted to 
the Graduate Faculty of 
Auburn University 
in Partial Fulfillment of the 
Requirements for the 
Degree of 
Doctor of Philosophy 
 
 
Auburn, Alabama 
December 15, 2006 
 iii
FAMILY FOREST LANDOWNER BEHAVIOR IN THE SOUTHEAST 
 
 
 
 
Indrajit Majumdar 
 
Permission is granted to Auburn University to make copies of this dissertation at its 
discretion upon request of individuals or institutions and at their expense. The author 
reserves all publication rights 
      
      
      
 
 
Signature of Author 
 
Date of Graduation 
 iv
VITA 
Indrajit Majumdar son of Mrinal Kumar Majumdar and Subha Majumdar was 
born in Bihar, India. After finishing high school from Delhi Public School, India he went 
on to complete his BS. in Botany from Delhi University in 1997 and his MS. in Forest 
Economics and Management from the Forest Research Institute, Dehradun, India in 2000. 
After a small stint of work in a project funded by the International Crops Research 
Institute for the Semi-arid Tropics (I.C.R.I.S.A.T.) in Hyderabad, India, during December 
2000-January 2001 he joined Auburn University in the fall of 2002 for his graduate 
studies. He was married on 27
th
 November, 2004 to Anindita, daughter of Deba Prasad 
Dutta and Atasi Dutta from Durgapur, India. 
 v
DISSERTATION ABSTRACT 
FAMILY FOREST LANDOWNER BEHAVIOR IN THE SOUTHEAST 
 
 
Indrajit Majumdar 
Doctor of Philosophy, December 15, 2006 
(M.S., Forest Research Institute, Dehradun, India, 2000) 
(B.S., Delhi University, India, 1997) 
90 Typed pages 
Directed by Lawrence D. Teeter 
Forests and forestry have played a significant role in the economic development 
and psyche of the South. Forests, which pre-settlement occupied nearly all of the land 
area of the South, now occupy only 55 percent. More important, perhaps, is the change in 
the structure and composition of these forests. Some of these changes have resulted from 
harvest for conversion to agriculture, and the subsequent reversion to forest. Other 
changes have occurred as fiber demand increased and harvested lands were replanted 
with pines. More recent is the recognition that forests provide amenity and recreational 
values which may lead to reductions in harvest by family forest owners. With an 
unprecedented growth in the number of family forest landowners there is an increase in 
the surge for researchers to investigate the motives of these landowners to manage their 
land for timber and/or non-timber use. The diversity amongst these owners in terms of 
 vi
their objectives and their forest conditions coupled with the increasing urbanization 
pressure warrants need for a thorough investigation. This study presents result from 
investigating the complexities of family forest owner behavior in three separate chapters 
(2, 3 & 4) which can be read independently and is styled in journal publication format. 
Chapter 2 explores the impacts of population pressures on forestland use change 
in Alabama and results suggest initial forest type and population gravity index to be 
significant correlates to forest type transition and conversion o forest to non-forest use. 
Measures of anthropogenic influence such as population density and real per capita 
income had significant impact on conversion of forest to non-forest use. We also found 
that Nested Logit is a appropriate econometric technique to study the discrete choice 
behavior of private landowners. 
Chapter 3 investigates the diversity of motivations of family forest owners in the 
Southeast and tests the assumption of homogeneity (treating family forest owners as a 
single homogeneous group) that previous researchers have made. Our study using 
multivariate cluster analysis procedures suggest that family forest owner ?group? is in fact 
a diverse set of owners who can be grouped into three attitudinal types namely multiple-
objective, non-timber and timber. 
Finally, in Chapter 4 multivariate non-parametric discriminant procedures was 
used to discriminate the three attitudinal groups (Chapter 3) using bio-physical, socio-
economic and demographic variables. Analysis results indicate that 84% of landowners 
across all landowner groups were correctly classified. With all the variables used to 
develop the classification scheme in this study known, a-priori, that is before landowners 
 vii
on a Forest Inventory and Analysis (FIA) plot location is contacted for the National 
Woodland Owner Survey (NWOS), it may be possible to predict membership of a future 
landowner with known FIA and Census demographic attributes. 
 viii
ACKNOWLEDGEMENTS 
First of all I would like to express my sincerest of gratitude to my advisor, Larry 
Teeter, for his relentless support, encouragement and thoughtful guidance throughout the 
four years of my graduate studies here in Auburn. Thanks to my committee members, 
Drs. David Laband, Yaoqi Zhang and Robert Taylor for enriching me with our 
stimulating conversations and knowledge gained from the courses taken under them. 
A special word of thanks to my colleague and good friend, Maksym Polyakov, for 
his help throughout and his uncanny way of interpreting my often not so articulate 
queries and coming out with the appropriate solutions. 
It is indeed been a privilege studying and doing research at the School of Forestry 
& Wildlife Sciences and a special word of thanks to the Forest Policy Center for 
supporting me financially. I would also like to acknowledge my fellow graduate students 
for all their help and support. 
Words fail me to express my gratitude to my mother, father, brother and sister-in-
law who have always pampered me and showered me with their love and affection, I owe 
this to them who had dreamt long before me that I can do this. 
Last but not the least, I want to thank my dear wife, Anindita, for her unending 
love and support throughout this endeavor and for being my strongest motivation without 
which I could never have completed this. 
 ix
Style manual or journal used: Forest Science, Guide to Preparation and 
Submission of Theses and Dissertations, Forest Science. 
Computer software used: SAS ? 9.1, ArcGIS ? 9.0, MS Word ?, MS Excel ?.
 x
TABLE OF CONTENTS 
LIST OF TABLES ................................................................................................................. xii 
LIST OF FIGURES ...............................................................................................................xiii 
CHAPTER 1 INTRODUCTION .............................................................................................. 1 
CHAPTER 2 EFFECT OF URBANIZATION ON FOREST LAND USE CHANGE IN ALABAMA: A 
NESTED LOGIT APPROACH................................................................................................... 5 
2.1 Introduction......................................................................................................... 5 
2.2 Literature Review................................................................................................ 6 
2.3 Empirical Land Use Model................................................................................. 7 
2.4 Data and Variables............................................................................................ 13 
2.5 Population Gravity Index.................................................................................. 16 
2.6 Results............................................................................................................... 17 
2.6.1 No Harvest Vs Harvest ..................................................................... 22 
2.6.2 Non-forest Vs Forest......................................................................... 23 
2.6.3 Forest Types...................................................................................... 24 
2.7 Discussion......................................................................................................... 26 
2.8 Conclusion ........................................................................................................ 27 
2.9 References......................................................................................................... 28 
CHAPTER 3 FAMILY FOREST OWNERS IN THE SOUTHEAST: ARE THEY ALL THE SAME?.. 32 
3.1 Introduction....................................................................................................... 32 
 xi
3.2 Literature Review.............................................................................................. 34 
3.3 Landowner Model............................................................................................. 37 
3.4 Data and Methods ............................................................................................. 39 
3.4.1 Data................................................................................................... 40 
3.4.2 Statistical Methods............................................................................ 41 
3.4.2.1 Cluster Analysis......................................................................... 44 
3.4.2.2 Cluster validation....................................................................... 49 
3.5 Conclusion ........................................................................................................ 51 
3.6 References......................................................................................................... 53 
CHAPTER 4 DISCRIMINATING FAMILY FOREST OWNERS: A NON-PARAMETRIC APPROACH
........................................................................................................................................... 58 
4.1 Introduction....................................................................................................... 58 
4.2 Literature Review.............................................................................................. 59 
4.3 Data and Methods ............................................................................................. 62 
4.3.1 Data................................................................................................... 62 
4.3.2 Discriminant Analysis....................................................................... 64 
4.3.3 K-nearest neighbor method............................................................... 66 
4.4 Results............................................................................................................... 68 
4.5 Conclusion ........................................................................................................ 72 
4.6 References......................................................................................................... 75 
 xii
LIST OF TABLES 
Table 2.1 Matrix of number of FIA
a
 plots in observed starting use and end use by period
........................................................................................................................................... 14 
Table 2.2 Univariate statistics of the variables and their description ............................... 16 
Table 2.3 Nested Logit Parameter estimates for the three-level nested model................. 19 
Table 3.1 Survey questions from NWOS used for cluster analysis.................................. 42 
Table 3.2 Factor loadings representing the correlations between factors and the variables
........................................................................................................................................... 44 
Table 3.3 Socio-demographic and forest characteristics of family forest owners by cluster
........................................................................................................................................... 49 
Table 3.4 Confusion matrix for cluster validation............................................................ 50 
Table 4.1 Data sources and their descriptive statistics ..................................................... 64 
Table 4.2 Multivariate Normality and Homogeneity of variance-covariance test result.. 65 
Table 4.3 Classification results for the apparent-error-rate KNN method........................ 69 
Table 4.4 Classification results for the one-leave-one cross validation KNN method..... 69 
Table 4.5 Confusion Matrix of one-leave-one cross-validation method for k=2.............. 70 
Table 4.6 Producers and Users accuracy for leave-one-out cross-validation method (k=2)
........................................................................................................................................... 70 
 xiii
LIST OF FIGURES 
Figure 2.1 Three level-nested representation of landowner decision ............................... 11 
Figure 3.1 Landowner Behavior Model............................................................................ 39 
Figure 3.2 Mean response of reasons for owning forestland by cluster group, 1 denoting 
?Most important? and 7 denoting ?Least important? ......................................................... 47 
Figure 4.1 K-nearest neighbor classification based on k=1 or k=5 or k=10..................... 67 
Figure 4.2 Graphical representation of apparent-error-rate accuracy of KNN by k ......... 71 
Figure 4.3 Graphical representation of one-leave-one cross-validation accuracy of KNN 
by k.................................................................................................................................... 72 
 
 1
CHAPTER 1 
INTRODUCTION 
Forest ownership patterns are rapidly changing due to broad-based sale of forest 
land by the major forest product firms and the subsequent parcelization of those lands 
into smaller ownerships through successive rounds of purchase and resale by 
intermediate land brokers. At the same time family forest management in the Southeast is 
undergoing significant change due to the evolving motivations and values of new family 
forest owners, their reduced reliance on land-based occupational income, and general 
urbanization pressures on land use decisions. The economic and societal impacts of these 
changes still remain largely unknown.  
Substantial research has been conducted over the last few decades regarding the 
factors influencing the behavior of private forest owners with the underlying goal to 
identify more effective ways of communicating/encouraging family forestland owners to 
actively practice forest management. The non-consumptive uses of family forests and the 
psychic tradeoffs between the timber production (primarily for profit) and the non-timber 
(amenities) consumption (primarily for utility) for the family forest owners have often 
been neglected. Identifying these diverse motivations and the subjective preferences for 
managing their forest for timber and/or non-timber amenity values is warranted in the 
wake of their increasing numbers. 
 2
To address the general questions stated above this study was conducted and the 
results are presented in three chapters (2, 3 & 4). Chapter 2 investigates the effect of 
increasing urbanization pressures on choices made by private landowners in Alabama. 
Nested logit analysis was conducted to model the land use decision making behavior of 
private landowners. Private landowners are assumed to choose from amongst a discrete 
set of management alternatives the one that maximizes their utility. The alternative 
choice set includes either converting forest into non-forest use, regenerating into one of 
the three forest types (hardwood, softwood or mixed) and a no harvest decision 
maintaining initial forest type. Results show that the initial forest type and population 
gravity index are significant variables in explaining the variation in type transition. 
Consistent with previous research findings population gravity index, a proxy for the 
anthropogenic influence, favored forest land conversion to non-forest use. The 
probability that a forest plot will be converted to non-forest at the mean of all the 
explanatory variables in the model is 0.02. In the softwood, mixed and hardwood forest 
types those probabilities increased to 0.05, 0.17 and 0.06 following harvest. The 
probability of no harvest at the mean of the variables was 0.7. To our knowledge 
application of the nested logit technique to analyze the forest harvesting decision by the 
landowner has not been considered previously. 
The names used to describe the family forest owners have changed over time, but 
the inclination to treat/analyze them as a homogeneous one has been fairly common. 
Chapter 3 characterizes the family forest owners in the three Southeastern states of 
Alabama, Georgia and South Carolina based on their feelings about forest stewardship 
 3
and their stated reasons for owning forestland. Multivariate clustering technique was used 
to analyze the National Woodland Owner survey (NWOS) data on the family forest 
owners and results reveal the presence of three owner clusters (classes) based on their 
stated motivations for owning their forestland. These three owner classes were termed as 
multiple-objective, non-timber and timber respectively. The multiple-objective ownership 
type was found to be the largest group (533 owners, 49.1%) with almost every 1 out of 2 
family forest owners in the sample population belonging to this category. Owners 
belonging to the timber (319 owners, 29.4%) cluster had only the timber management 
and land investment as the strong motivating factors behind their forestland ownership 
while owners belonging to the non-timber (233 owners, 21.5%) cluster value the non-
consumptive uses of their forestland such as aesthetic values, biodiversity, recreation and 
privacy. 
After exploring the diverse motivations and classifying the family forest owners 
into three groups (Chapter 3) the question of interest next was whether we could identify 
certain characteristics that would discriminate the landowner groups and develop a 
classification scheme that will help in predicting the most likely group that a new 
landowner will be characterized into given these characteristics. Chapter 4 presents 
results of this investigation. Multivariate non-parametric discriminant analysis using k-
nearest neighbor (KNN) method was used in the study. It was found that bio-physical 
(Slope, Site productivity index, Forest type and Stand age), socio-economic (Median 
household income and Distance to nearest paved road) and demographic (Population 
gravity index and Population density) variables had a strong association with landowner 
 4
group profiles. The average accuracy of prediction across all the landowner groups was 
84% and it was 87%, 78% and 82% for multiple-objective, non-timber and timber owners 
respectively. This study indicates that landowners clustered into heterogeneous attitudinal 
groups (multiple-objective, non-timber and timber) can be accurately separated using 
non-parametric KNN technique.
 5
CHAPTER 2 
EFFECT OF URBANIZATION ON FOREST LAND USE CHANGE IN ALABAMA: A NESTED 
LOGIT APPROACH 
2.1 Introduction 
Land use change and respective change of land cover attributed to human 
activities on land is a common phenomenon associated with population growth, market 
development, technical and institutional innovation and policy action.  
Prompted by the re-allocation of scarce resources to satisfy growing human needs, 
land use change impacts numerous ecosystem services and functions such as watershed 
protection, biogeochemical cycling, soil degradation, and habitats for different species. 
Vitousek (1994) identified land cover changes by humans as the primary effect of 
humans on natural systems. Few forested areas on our planet have not been influenced by 
human actions, yet the effects of long-term human influences on land use/land cover 
changes from forestry are not well documented. Various models differing in temporal and 
spatial scales and quantitative techniques have been applied by researchers/scientists to 
uncover the determinants of land use change. A close look at past land use studies reveals 
that biophysical factors such as land quality and topography; economic factors such as 
population, market conditions, proximity to population centers and income; and 
institutional factors such as government policy are the major determinants of land use 
change. The objective of this paper is to explore the effect of increasing population 
 6
pressures on choices made by the private forest landowners of Alabama. Alabama ranks 
second in the nation in acres of forestland (excluding Alaska), ((NRI, 1997) 
http://www.al.nrcs.usda.gov/technical/nri/97highlights.html) and the forests of the state 
account for 13% of the total timber removals in the South (Smith et al 2002). The impacts 
on forestry land use including changes to non-forest uses viz. agriculture and 
urban/developed land or changes in forest types (land cover changes) will have 
significant effects on the ability of Alabama?s forests to provide both timber and non-
timber amenities in the future. 
2.2 Literature Review 
Empirical land use change models have been constructed using primarily two 
approaches. The first is the aggregated approach that models areas or proportions of land 
in different use categories such as forestry, agriculture and urban (Alig 1986, Hardie and 
Parks 1997) or different forest types such as softwood, mixed hardwood, hardwood, 
agriculture and urban land (Zhang et al 2005) within a well defined geographic region 
such as a county as a function of socioeconomic variables and land characteristics 
aggregated at the level of the geographic unit of observation. The second is the spatially 
explicit approach that explicitly models land use change on the basis of pixels, parcels, or 
sample points (Bockstael 1996, Chomitz and Gray 1996, Munn and Evans 1998, Wear 
and Bolstad 1998, Kline et al., 2001, Lubowski et al., 2002). While the aggregated 
approach has the disadvantage of averaging the physical land characteristics for the unit 
of study, the spatially explicit approach has often found it difficult to obtain spatial socio-
demographic data at scales finer than the census tract level which are virtually 
 7
nonexistent. Another distinction worth mentioning between the aggregate area-based 
approach and the spatially explicit approach is that, in the former approach, the 
coefficients of the model capture simultaneously both the spatial and temporal effects of 
the explanatory variables and though this methodology has helped a lot in identifying the 
factors affecting allocation of land to different uses, it generally has done a poor job in 
identifying the temporal component, or the effects of factors on land use change, as well 
as in projecting land use shares through time (Ahn et al., 2000). In contrast, the spatially 
explicit approach models the change directly by taking into account the dynamic nature 
of the land use change decision. 
Econometric studies on shares of land uses and forest types (Nagubadi and Zhang 
2005) and shares of forest type by ownership category (Ahn et.al. 2001; Nagubadi and 
Zhang 2005) have been done using the aggregate approach at the county level. In the 
absence of a spatially explicit model, the full power of human impact is hard to explore. 
Some of the spatially explicit studies on the impact of development on forest land use 
change have been done (Munn et al. 2002; Kline et al. 2003) but they have assumed 
similar impact on different forest types. We wish to fill this gap by developing a spatial 
explicit model to investigate the impact of population pressures on changes in forest land 
use for different initial forest types viz. softwood, mixed hardwood and hardwood. 
2.3 Empirical Land Use Model 
Researchers have extensively used multinomial logit models (Chomitz and Gray 
1996; Turner et al. 1996; Hardie and Parks 1997) for explaining landowners? choice of 
land use without taking into consideration the possibility of correlation between 
alternative choices. A feature of our study is the use of the less restrictive nested logit 
econometric framework as it relaxes the assumption of Independent and Irrelevant 
Alternatives (IIA)
1
 to account for the possible substitution patterns amongst alternative 
choices. 
Lubowski (2002) in his study on the determinants of land use change at the 
national level used a nested logit framework to consider the substitution patterns amongst 
land use alternatives, but the scope of the study did not take into account the substitution 
possibility amongst different forest types. Moreover he was interested in evaluating the 
effect of the economic and policy factors on land use change, not the effect of population 
pressure as is central to this study. 
We employ a discrete choice approach to model the land use decision making 
behavior of private forest landowners. It is assumed that a landowner starting with an 
initial forest type chooses between the five possible discrete alternatives the one that 
maximizes his utility. The alternative choice set includes either converting forest into 
non-forest use, regenerating into one of the three forest types (hardwood, softwood or 
mixed) and a no harvest
2
 decision maintaining initial forest type. A landowners? utility 
gained from choosing a particular alternative depends on the attributes associated with 
each forest plot. For models of land use change, the vector of plot characteristics, , 
typically consists of data on land quality, socio-demographic, socio-economic and rent 
x
 8
                                                 
1
 McFadden (1973) suggested that IIA implies that conditional and multinomial logit models should only 
be used in cases where the outcome categories can plausibly be assumed to be distinct and weighed 
independently in the eyes of each decision maker. 
2
 This study does not assume type transition if there is no harvest and considers the forest type as fixed until 
harvest occurs. 
(return) to alternative land use choices. In this discrete choice framework, a risk neutral 
landowner is assumed to choose for parcel i an alternative k from a set of J alternatives 
that maximizes his utility at time t.
3
 Assume that the landowner?s utility function for 
choice j is given by: 
 9
j
(,) (,)
jj
Vv?=? x ? x + (1)
where  is the vector of attributes of plot characteristics and  is a vector of preference 
parameters on the observable portion of the landowner?s utility function for the 
alternative j, . Finally, 
x
j
?
(,)
j
v ? x
j
?  is the unobservable portion of the landowner?s utility 
function and is assumed to be a function of certain forest plot characteristics and the 
characteristics of the decision maker. The landowner then compares all potential choices 
in his choice set ?J? and chooses the best land use alternative ?j? such that: 
(,) (,) , ,
jk
VV jJkJk>???? x ? x j? (2) 
The challenge is to take the model given by (1) and (2) and develop a statistical 
model that will enable the recovery of the parameters? . The structure of the model will 
depend heavily on the assumptions about the form of the distribution of error terms. 
Assuming error terms 
j
?  are i.i.d. with Type I Generalized Extreme Value distribution 
(GEV)
4
, (1) and (2) could be expressed as a multinomial logit model:  
exp( ' )
Prob( )
exp( ' )
k
j
jJ
k
?
=
?
? x
? x
 (3) 
                                                 
3
 For notational simplicity the subscripts i and t will be dropped from the equations. 
4
 Type I GEV also known as Gumbel distribution is based on simplifying assumptions such as independent 
and identical distribution (iid) of random components and the absence of heteroscedasticity and 
autocorrelation in the model (see Mcfadden (1974) for details). 
In order that this model be estimable, one of the  needs to be set to . A well-
known restriction associated with the model given in (3) is the Independence of Irrelevant 
Alternatives (IIA) due to the fact that  
j
? 0
exp( ' )Prob( )
Prob( ) exp( ' )
k
j
k
j
=
? x
? x
 (4) 
denoting that the ratio of probabilities of choices k and j would remain unchanged with a 
change in the parameters of choices other than j and k. In reality, that might not be the 
case. For example, a change in the stumpage price of hardwood might influence the ratio 
of probabilities of transition to pine plantation vs. probability of transition to agricultural 
land. A study by Lubowski (2002) on the economic and policy determinants of land use 
change using a nested logit model supports the need for exploring alternative nesting 
structures in land use studies by taking into consideration the similarities between forest 
types.  
To relax the IIA assumption in our analysis of the human impact on changes in 
forest land use we adopt a nested logit model, which groups similar alternatives into 
groups called nests thus allowing different variances for choices in different nests and 
correlation between choices within a nest. We use a three level nested logit model, which 
assumes that decisions are made at three hierarchical levels (Figure 2.1). 
 10
 
 
 
 
    
 
 
 
Harvest   No Harvest 
 
 
 
 
 
Non-Forest    Forest 
 
 
 
 
   
 
 
        Softwood               Mixed               Hardwood 
Figure 2.1 Three level-nested representation of landowner decision 
The decision at each of these three levels is modeled as an outcome of separate 
utility functions. In other words, the landowner is seen as making an independent 
decision at each decision node. The decision to harvest or not to harvest at the uppermost 
level of the nested tree can be modeled as a binary logit model. Assuming the landowner 
makes the decision to harvest, he has to make another decision at the medium level of the 
nested model, which is whether to keep the land in forest or convert it to non-forest use. 
This can also be modeled as a binary logit model. Finally, assuming the landowner 
decides to keep the land in forest use, he decides whether to regenerate it to a softwood, 
mixed or hardwood type of forest. Each of these decisions is taken with a view of 
 11
maximizing utility. The three level nested model decomposes the choice probability into 
three components, the marginal probability of choosing a particular subgroup (nest) s at 
the uppermost level, S=1,2 for harvest or no harvest, the marginal probability of choosing 
a particular sub-nest l within the nest s, where L=1,2 for non-forest or forest, and the 
conditional probability of choosing a particular alternative j at the lowest level within the 
alternative set J =1?J
l,s
 in the sub-nest ?l? and nest s conditional on the choice of that 
sub-nest and nest. Given this, the probability that a landowner i is observed choosing 
alternative j at time t in the nested logit formulation requires the decomposition of the 
choice probability in (3) into three components: the marginal probability  of choosing 
a particular nest s (s=1,2) and conditional probabilities  and  of choosing a 
particular sub-nest l (l=1,2) conditional on the choice of that nest s and choosing a 
particular alternative j from within the alternatives (j=1,2,3,4,5) conditional on the choice 
of that nest and sub-nest. The probability defined in (3) thus becomes: 
is
P
sil
P
| slij
P
,|
|
||,
|
exp( ' )exp( ' )
exp( ' )
exp( ' ) exp( ' ) exp( ' )
jili lsil
si sis
islj is il s ij l s
ki kik mi msim ni
kS mL nJ
I
I
PPPP
II
?
?
??
?? ?
+
+
=?? = ? ?
++
?? ?
? x? z
? y
? y ? z ? x
 (5) 
where 
s
?  and 
sl|
?  are the parameters associated with the Inclusive Value (IV) for nest s 
and sub-nest l defined as 
|
ln exp( ' )
is m i m s im
mL
I I?
?
=+
?
? z  (6)
and 
ln exp( ' )
il n i
nJ
I
?
=
?
? x (7) 
 12
where, 
i
y  are the observed plot attributes influencing the choice of the nest,  are the 
observed plot attributes influencing the choice of the sub-nest and  being the observed 
plot attributes influencing the decision to keep in an alternative forest type conditional on 
the choice of the nest and sub-nest. The inclusive value for nest s and sub-nest l defined 
in (6) and (7) is the log of the denominator of the conditional probabilities in (5) and 
measures the average utilities of the alternatives within that subset of alternatives for the 
choice of a particular nest s and sub-nest l. If the parameters  and  are zero and the 
inclusive value parameters
i
z
i
x
k
?
m
?
k
? , 
s|m
?  are jointly equal to one then the model will collapse 
into a multinomial logit model shown in (3). 
2.4 Data and Variables 
The data for this study comes from the Forest Inventory and Analysis (FIA)
5
 
program of the U. S. Department of Agriculture (USDA) Forest Service, USDA 
Economic Research Service (ERS), Bureau of Census and the Regional Economic 
Information System (REIS) of the Bureau of Economic Analysis (BEA). We used 
Alabama FIA data for the census years 1972, 1982, 1990 and 2000 and the Census 
Bureau data on population demographics for the same periods
6
. REIS provided us with 
the per capita personal income by county for the corresponding years. All the plots 
considered for the study were restricted to be in forest use at the beginning of the period 
and privately owned. The publicly owned plots were not considered for the study as they 
                                                 
5
 Historically FIA provides detailed data on forest inventory for all the states on approximately 10-year 
periodic cycle with each plot roughly representing a 3?3 mile grid pattern 
 13
6
 Census collects decennial data and so for the FIA counterpart of 1972 and 1982 we used its closest census 
counterpart which was 1970 and 1980. 
 14
are more predictable in terms of land use change since public land use decisions are 
determined mainly by forces other than the market. The total number of observations for 
the period (1972-2000) of the study that consisted of three transition periods was 10383. 
The matrix of plots by starting and end use is given in Table 2.1 for each of the three 
periods 1972-1982, 1982-1990 and 1990-2000. The FIA data for Alabama show a 
conversion of 94, 74 and 128 plots to non-forest use for the period 1972-1982, 1982-1990 
and 1990-2000 respectively (Table 2.1). 
Table 2.1 Matrix of number of FIA
a
 plots in observed starting use and end use by 
period 
  Ending Land Use   
Initial Land Use Non-forest Softwood Mixed Hardwood
  1972 - 1982   
Softwood 29 561 153 129 
Mixed 23 142 194 185 
Hardwood 42 38 102 723 
  1982 - 1990   
Softwood 25 739 199 123 
Mixed 17 179 282 185 
Hardwood 32 104 182 1161 
  1990 - 2000   
Softwood 38 843 275 305 
Mixed 24 334 259 403 
Hardwood 66 277 241 1499 
a 
excludes all plots under public ownership and contains only privately owned FIA plots 
 
All the explanatory variables in the model, associated with the FIA plots were 
lagged values based on the previous period ?t-1? to incorporate the general trends in the 
variable?s effect on the landowners? discrete choice as observed at the current period t. 
For example a FIA plot observed in a particular land use for the FIA survey year 1982 
had all the corresponding explanatory variables from the FIA survey 1972 and the 
 15
population census for the year 1970 and so on. From among the array of variables used in 
this study the key variable that represents the influence of humans on forest land use 
change is the Population Gravity Index (PGI). The PGI was constructed by utilizing 
information on the location of the FIA plots in relation to the location of Census 
populated places within a 100km. The geographic location of census places
7
 was taken 
from ESRI Data and Maps, 2005 
(http://www.esri.com/data/about/data_maps_media.html). Other variables in the model 
include the initial forest type dummy for the three classes of forest type denoted by the 
variable names SW (softwood), MX (mixed) and HW (hardwood) for each FIA plot. 
Volume in cubic feet of all the trees within a FIA plot denoted by VOL was included as a 
potential measure of the propensity of harvest for the plot. We also included the growing 
stock removals in cubic feet (from FIA county data) per unit of county land area in acres 
(from ERS) as a proxy for forest land use return (RET) hypothesized as one of the chief 
economic drivers of land use change in almost all of the previous land use models. 
SLOPE in percent for the FIA plots was included to examine the potential influence of 
topography on landowner choice. Finally, to explore the full potential of the urbanization 
pressures acting on forest land use change, county level estimates of per capita personal 
income (INC) from REIS of the BEA deflated by the Consumer Price Index (Urban 
South, 1982=100), and county level estimates of population density (PD) were also 
                                                 
7
 Bureau of Census definition for a place is ?concentration of population either legally bounded as an 
incorporated place, or identified as a Census Designated Place (CDP) including comunidades and zonas 
urbanas in Puerto Rico. Incorporated places have legal descriptions of borough (except in Alaska and New 
York), city, town (except in New England, New York, and Wisconsin), or village?. 
 16
included in the model. A list of the variables used in the analysis with their sources and 
standard statistical summary is given in Table 2.2. 
 
Table 2.2 Univariate statistics of the variables and their description 
Variable Definition Source Mean Std.dev. 
PGI Number of persons/Km
2
 
around each FIA plot 
within 100 Km radius of 
each FIA plot 
FIA plot and 
Census Bureau 
136.03 99.60 
VOL volume in cubic feet of 
the FIA plots 
FIA plot data 1027.19 965.92 
SW initial forest type dummy 
for Softwood forest 
FIA plot data 0.35 0.47 
MX initial forest type dummy 
for Mixed forest 
FIA plot data 0.44 0.50 
HW initial forest type dummy 
for  Hardwood forest 
FIA plot data 0.20 0.40 
INC Real (1982=100) per 
capita personal income 
by county in $ 
BEA 111.95 23.48 
RET growing stock removals 
in cubic feet per acre of 
county land area 
FIA county data 
and ERS 
20.40 12.43 
SLOPE slope in percent for FIA 
plot 
FIA plot data 9.93 10.75 
PD number of persons per 
unit of land area by 
county 
Census Bureau 73.28 97.75 
 
2.5 Population Gravity Index  
Gravity models were initially developed by Reilly (1929) to describe the degree to 
which cities attracted retail trade from surrounding locations. In this study a population 
gravity index was created to explore the combined effects of population and proximity to 
city centers. A 100km
8
 buffer around Alabama incorporating the influence of Census 
places from the four contiguous states of Georgia (GA), Tennessee (TN), Mississippi 
(MS) and Florida (FL) in addition to all the designated census places within the state of 
Alabama was created. Population Gravity index (PGI)
9
 for a plot k was specified as 
2
: 100
pt
kkp
p
kp
P
PGI p D km
D
=?<
?
 (8)
where  is the population of populated place p at time t, and  is the distance between 
FIA plot k and populated place p.  
pt
P
kp
D
Location coordinates for measuring the Euclidian distance in (8) was computed 
using ArcGis 9.0. The gravity index itself using the population of the census place around 
FIA plots and distance of the FIA sample plots to census places was calculated in SAS 
9.1. PGI was previously found to be positively correlated with conversion to non-forest 
use from forest use (Majumdar et al. 2005).  
2.6 Results 
Acknowledging the fact that different factors may be important at different 
nesting levels, we define below the three utility functions representing the variables likely 
to influence landowners? decisions at the three decision nodes of the nested tree and the 
attribute vectors in 
i
y ,  and  above as 
i
z
i
x
Pr(no harvest relative to harvest) ( , , )f VOL SLOPE SW?  (9) 
                                                 
8
 100 km within an average 60-minute commute time from FIA plots was assumed as the threshold distance 
and varying this distance did not substantially affect the sign and magnitude of the estimated coefficients of 
the gravity index and other variables. 
 17
9
 Kline et al (2001) used a similar formulation of gravity index but with different exponents on the 
population and distance components of the index and they used three cities with population greater than 
5000 and greatest urban influence based on their gravity index on each FIA plot. 
Pr(Non-forest relative to Forest) ( , , , )f PGI INC RET PD?  (10) 
),,,(Hardwood)or  Mixedor  dPr(Softwoo PGIHWMXSWf?  (11) 
The reference category
10
 in (9) was harvest and in (10) forest. In (11) the 
reference category was hardwood (for PGI) and no change in forest type for the variable 
initial forest types (SW, HW or MX) respectively. The results are summarized in Table 
2.3. 
                                                 
 18
10
 With all the explanatory variables being characteristics of the FIA plot and not the alternative land use 
choices we used interactions of each variable with the dummy of choice alternatives and hence had to 
remove a particular choice and make it as a reference base for model identification. 
 19
 
Table 2.3 Nested Logit Parameter estimates for the three-level nested model 
Variable Coefficient Standard Error Odds Ratio t-statistic 
No Harvest Vs. Harvest 
SW ? C
NH
?0.5714 0.059* 0.56 ?9.61 
SLOPE ? C
NH
0.0217 0.002* 1.02 10.12 
VOL ? C
NH
0.0009 0.00003* 1.00 32.31 
Non-Forest Vs. Forest 
PGI ? C
NF
0.0011 0.0005** 1.00 2.16 
RET ? C
NF
?0.0172 0.004* 0.98 ?4.40 
PD ? C
NF
0.0019 0.0005* 1.00 3.99 
INC ? C
NF
?0.4266 0.0397* 0.65 ?10.75 
Forest Type 
MX ? C
SW
?1.9815 0.1225* 0.14 ?16.17 
MX ? C
HW
?2.3806 0.1032* 0.09 ?23.06 
HW ? C
SW
?0.2388 0.1184** 0.79 ?2.02 
HW ? C
MX
0.4010 0.0909* 1.49 4.41 
SW ? C
MX
?0.9931 0.0857* 0.37 ?11.59 
SW ? C
HW
?1.2661 0.0874* 0.28 ?14.48 
PGI ? C
SW
?0.0045 0.0005* 0.99 ?8.72 
PGI ? C
MX
?0.002 0.0003* 0.99 ?6.88 
IV
e
 (Forest) 0.87 0.1360*  6.42 
IV
e
 (Harvest) 0.85 0.1311*  6.48 
Log likelihood 10096    
McFadden?s LRI 0.39    
Observations 10383 
a 
C
SW
,
 
C
MX
, C
HW
, C
NF
, C
NH
  represent the dummies for the choice alternatives softwood, mixed, 
hardwood, non-forest and no harvest respectively 
e 
IV were constrained to be the same at each nest level for model identification, moreover for degenerate 
branches the such as Non-forest and No harvest IVs cannot be identified (for detail see Hunt)  
  * p < 0.10 
** p < 0.05 
We estimated a three level nested logit model in which the landowner decides to 
either harvest or not to harvest at the top level, then makes the decision to convert the 
harvested plots into non-forest use or keep them in forest use at the next level, and finally 
decides on whether the forested plot will be of softwood, mixed hardwood, or hardwood 
forest type at the lowest level (see Figure 2.1 for the nested tree depiction). 
 20
Land use choices are influenced at least in part by the environment of the 
landowner and with the increasing population and expanding metropolitan areas 
landowners are subjected to increasing anthropogenic pressures that influence decisions 
about their land use. We also think that depending on the initial use of the land this factor 
acts differently both in magnitude and in direction such that it has differential impacts on 
the decision of the landowner. 
The focal point of the research was to explicitly define the decision environment 
of the landowner using the location of their forest plot in relation to the location of the 
population center and the mass of the population and then analyze their effects on the 
landowner decision.  
The nesting structure in figure 2.1, together with equations (5)-(7) and (9)-(11) 
can be used to formulate appropriate log likelihood function to estimate the parameters of 
the model. Nested logit model was estimated using full information maximum likelihood 
estimation in SAS 9.1. Because one of the nests at each level contains only one 
alternative, this model is partially degenerate, and therefore over parameterized with 
respect to the inclusive value (IV) parameters (Hunt, 2000). According to Hunt (2000) 
?In general, the degenerate partition?s IV parameter can be restricted to any value and the 
estimates obtained for the model will be invariant?. Therefore, in order to estimate the 
model, we constrained IV parameters at each nest level to be equal. The hypothesis that 
the coefficients of the IV parameters are simultaneously equal to 1 was rejected at more 
than the 99% level of significance based on the likelihood ratio test confirming the 
validity of the proposed econometric framework. The results support the choice of a 
 21
nested logit model, over a more restrictive multinomial logit model that does not allow 
for correlation within nests. A major test of the correct specification of the nested model 
is that the IV parameters should lie within the range of zero to one. Maddala (1983, page 
73) states that if the IV parameters lie outside the range of zero to one then this should be 
considered as evidence for a specification error and warrants re-examination of the 
model. Further McFadden (1981) states that if the dissimilarity coefficients (IV 
coefficients) are larger than 0 and not statistically larger than 1, it can be concluded that 
the nested model is consistent with stochastic utility maximization. 
The estimated maximum likelihood nested logit model had a reasonable fit with 
McFadden likelihood ratio index statistic (pseudo-R
2
) being 0.39. There are several 
methods of interpreting estimates of probability models (Long 1997) but one of the 
easiest ways is that of the odds ratio (Liao 1994) which is the exponentiated logit 
coefficients estimated in the model (Table 2.3). Odds ratio values greater than 1 indicate 
that an increase in the independent variable translates into greater probability of that 
category relative to the base (reference) category, values less than 1 indicate a decrease in 
probability with an increasing independent variable and values at 1 indicate equal 
probability of the particular category and the base. For example, increases in the INC 
significantly decreased the conversion probability of forest to non-forest use relative to a 
reference category of stable forest practice with no change in land use. We employ the 
interpretation of the parameter estimates at each of the three decision nodes (in separate 
sections) of the landowner decision hierarchy (refer to Figure 2.1) by using the odds ratio 
in the following discussion. 
 22
                                                
2.6.1 No Harvest Vs Harvest  
SW, representing the initial forest type as pine, had the expected negative sign 
and indicates less likelihood of no harvest of a pine plot. The odds of a pine plot not to be 
harvested, ceteris paribus, is 0.56. In other words there is a greater probability of a 
harvested plot to be a pine plot relative to its being either a hardwood or a mixed plot. 
SLOPE had the expected positive sign with the statistically significant parameter 
estimate which indicates that with an increase in slope there is a greater likelihood of no 
harvest due to a possible hindrance to accessibility of logging equipment and associated 
increase in harvesting cost. Moreover steep slope also constrains excessive harvests to 
prevent erosion and increase in cost of maintaining land in productive use. The result is 
consistent with previous studies (Wear and Flamm 1993). The odds that the plot will not 
be harvested relative to it?s being harvested owing to a unit increase in slope is 1.02. 
VOL, denoting the average volume per acre in cubic feet for the FIA plot had a 
positive sign. This implies that higher the volume the less likely it will be harvested. This 
is contrary to our expectation that greater average volume would lead to a greater 
probability of harvest. A close examination
11
 reveals that most of the harvests took place 
in the softwood plantation type, which typically has lower average volume in comparison 
to the hardwood plots.  
 
11
 Separate models had to be estimated which could include interaction terms of VOL and the initial forest 
type keeping the VOL main effect for each forest type due to collinearity problem and results showed the 
coefficient of the interaction term of pine with the average volume as negative while that of the hardwood 
and mixed as positive 
 23
2.6.2 Non-forest Vs Forest  
The population gravity index (PGI), representing the developmental pressure on 
forestland, had a statistically significant positive coefficient indicating an increase in non-
forest use. This is an expected result since in general demand for developed land near the 
population centers with higher PGI is high in comparison to the demand for forests. 
Researchers have found other measures of urbanization like increase in population 
density (Nagubadi and Zhang 2005) and decrease in distance from the center of the 
county to the nearest city (Ahn et al 2002) to favor an increasing non-forest share of land. 
INC had a statistically significant negative coefficient suggesting that counties 
with higher real per capita income are more likely to maintain their forests with a less 
inclination for conversion to non-forest use, ceteris paribus. This is contrary to the 
expectation of a casual observer and inferences drawn from previous research (Zhang and 
Nagubadi 2005). Our results show though that for a one unit increase in county real per 
capita income, the odds that forest will be converted to non-forest use is 0.65. The 
intuitive explanation could be that with an increase in income the landowner may 
perceive the returns from the consumptive use (aesthetics, amenities) of his forestland as 
higher in comparison to the return that can be gained with conversion to a developed use 
(intuitively somewhat like an environmental Kuznets curve). 
RET, denoting the total amount removals of growing stock from all the FIA plots 
within a county adjusted for the difference in county land area in cubic feet per acre has a 
negative coefficient and is statistically different from 0 at the1% level of significance. 
This result is consistent with our expectation, since with an increase in growing stock the 
 24
                                                
potential of future returns to forest use are higher and so less conversion to non-forest 
occurs. The decrease in odds in favor of conversion from forest to non-forest use due to 
an increase in a unit of RET, ceteris paribus, is 0.98. This result is consistent with the 
landuse change based on the Ricardo-Th?nnen land rent theory. Positive forest use 
returns (denoted by higher RET) are expected to decrease the likelihood of forest 
conversion to non-forest use. 
PD had a statistically significant positive coefficient reflecting the increased 
likelihood that a plot will be converted to non-forest use when there is an increase in 
demand for land for residential purposes, a result consistent with past studies on land use 
change (Wear et. al 1998, Nagubadi and Zhang 2005). 
2.6.3 Forest Types  
The negative parameter estimate for five out of the six (except HW ? C
MX
)
12
 
initial forest type variables indicates less likelihood of a forest type transition from one 
type to another relative to its likelihood of remaining as the same type. The decrease in 
odds of a mixed type of forest transition to softwood (MX ? C
SW
) or hardwood (MX ? 
C
HW
) relative to its regeneration as the same mixed type are 0.14 and 0.09 respectively. 
The log odds of the HW ? C
SW
 transition has a negative sign denoting that a hardwood 
plot has less likelihood of being converted into softwood relative to its remaining in the 
same type, with the odds being 0.79. On the other hand the increase in the odds that 
hardwood plots will be regenerated into a mixed type (HW ? C
MX
) following harvest is 
1.49. This result is somewhat contrary to what might be expected, although Zhou et al. 
 
12
 C
SW
, C
MX
, C
HW
, C
NF
, C
NH
 refer to the choice alternatives: softwood, mixed, hardwood, non-forest and no 
harvest respectively. 
 25
                                                
(2003) found a significant percentage of FIA plots in the South (Upland hardwood), 
which were not harvested, that transitioned to a mixed type in the subsequent survey and 
considering that there were a large percentage of plots (53.2 %) in our study that were not 
harvested, this result seems reasonable. Also depending on the FIA classification
13
 of a 
forest type, it is possible that a stand classified as hardwood could be retyped as a mixed 
type in the subsequent census. The decrease in odds of softwood (pine) plot to be 
regenerated as a mixed (SW ? C
MX
) or hardwood (SW ? C
HW
) following harvest relative 
to its remaining in the same type are 0.37 and 0.28 respectively.  
Results on forest type transition are consistent with our a priori anticipation that 
the landowner will most likely maintain the same forest type due to the costs of 
conversion, a result consistent with previous studies by Zhou et.al. (2003) and Alig and 
Butler (2004). Zhou et.al. (2003) projected that most of the hardwood and mixed oak-
pine stands will remain as the same type in future surveys following harvest. Similarly 
Alig and Butler (2004) found that the forest types having the highest probability of 
remaining as the same type in subsequent forest inventories following partial and final 
harvest are the lowland hardwood and the oak-pine.  
The negative significant parameter estimate for PGI ? C
SW
 and PGI ? C
MX
 reveals 
landowners? (who are closer to population centers) preferences for regeneration of 
hardwoods. The decrease in odds of a forest plot to be regenerated as softwood or mixed 
type relative to hardwood due to an unit increase in PGI, ceteris paribus, is 0.99. This 
indicates that hardwoods with their higher non-timber amenity values may be preferred, 
 
13
 A classification of forest land is commonly based upon, and named for, the tree species that forms the 
plurality of live-tree stocking. 
 26
albeit marginally, by landowners near urbanized centers relative to other forest types with 
relatively low aesthetic values. 
2.7 Discussion 
The nested logit model seems to be an appropriate choice for studying the discrete 
choice behavior of the private forest landowner. It is superior to multinomial logit, an 
econometric technique widely used to model land use, and allows for correlation of the 
error terms within a nest of similar choices. To our knowledge application of the nested 
logit technique to analyze the forest harvesting decision by the landowner has not been 
considered previously. Our results show that the initial forest type and population gravity 
index are significant variables in explaining the variation in type transition. Consistent 
with previous research findings population gravity index, a proxy for the anthropogenic 
influence, favored forest land conversion to non-forest use. 
The probability that a forest plot will be converted to non-forest at the mean of all 
the explanatory variables in the model is 0.02. In the softwood, mixed and hardwood 
forest types those probabilities increased to 0.05, 0.17 and 0.06 following harvest. The 
probability of no harvest at the mean of the variables was 0.7. In summary, given the 21.7 
million acres of private timberland (Hartsell and Brown 2000) our model projects 
434,000 acres to be converted from forest to non-forest use over a period of the next 10 
years. For the same period the acreage of non-harvested forest plots is projected to be 
15.19 million acres with 1,085,000 acres, 3,689,000 acres and 1,302,000 acres of 
timberland projected to be regenerated following harvest as softwood, mixed and 
 27
hardwood forest types respectively. These results are consistent and can be used for 
short-term predictions. 
2.8 Conclusion 
We used a discrete choice random utility model to investigate landowner choices 
among a discrete number of land use alternatives using a nested logit model. We took 
into consideration the characteristics of the decision environment of the landowner, 
considering primarily the forest?s location relative to urban influences and certain forest 
characteristics. This paper illustrates the importance of considering possible correlations 
between similar alternative choices in modeling landowner decisions.  
In the model we assume that the utility functions for all the private landowners 
have the same correlations amongst the functions? random components for all the choice 
alternatives. In reality these correlations might vary widely among decision makers 
Future studies can explore these differences by analyzing the data separately for private 
ownership categories like non-industrial private owners and industrial owners. We do not 
make adjustments for the potential spatial correlation of the model error terms and 
assume that the systematic sampling of plots (roughly on a 3-mile grid with each plot 
approximately representing 6000 acres) minimizes the likelihood that plots fall under the 
same ownership. 
 28
2.9 References 
Ahn, S., Plantinga, A.J. and R.J. Alig 2000.Future forest land area: A comparison 
of econometric approaches. For. Sci. 46(3):363-376. 
Ahn, S.C., R.C. Abt and A.J. Plantinga 2001. Land Use in the South Central 
United States: A further investigation on land use practices by forest land ownership. P. 
165-171 in Proc. of conf. on Forest law and Economics. D. Zhang and S. Mehmood, 
(eds.). Southern Forest Economics Workshop. 
Ahn, S., A.J. Plantinga and R.J. Alig 2002. Determinants and Projections of Land 
Use in the South Central United States. Southern Journal of Applied Forestry 26(2):78-
84. 
Alig, R.J. 1986. Econometric analysis of the factors influencing forest acreage 
trends in the Southeast. For. Sci. 32:119-34. 
Alig, R.J. and B. Butler 2004. Projecting large-scale area changes in land use and 
land cover for terrestrial carbon analyses. Environmental Management. 33(4):443?456. 
Barlow, S.A., I.A. Munn, I.A. Cleaves and D.L. Evans 1998. The effect of urban 
sprawl on timber. J. For. 96(12):10-14. 
Bockstael, N.E. 1996. Modeling economics and ecology: the importance of a 
spatial perspective. American Journal of Agricultural Economics 78:1168-80. 
Chomitz, K.M. and D.A. Gray 1996. Roads, land use, and deforestation: a spatial 
model applied to Belize. The World Bank Econ. Review 10(3):487-512. 
ESRI Data and Maps 2005. 
(http://www.esri.com/data/about/data_maps_media.htm). 
 29
Hardie, I.W. and P.J. Parks 1997. Land use with heterogeneous land quality: an 
application of the area base model. American Journal of Agricultural economics. 79:299-
310. 
Hartsell, A.J. and M.J. Brown 2002. Forest statistics for Alabama, 2000. USDA 
For. Serv., Southern Research Station Resource Bulletin SRS-67, Asheville, NC 28802. 
Hunt, G. L. 2000. Alternative nested logit model structures and the special case of 
partial degeneracy. Journal of Regional Science. 40(1):89-113. 
Kline, J. D, A. Moses and R. J. Alig 2001. Integrating urbanization into 
landscape-level ecological assessments. Ecosystems 4(1):3-18. 
Kline, J.D., D.L.Azuma and A.Moses 2003. Modeling the spatially dynamic 
distribution of humans in Oregon (USA) Coast Range. Landscape Ecology 18(4):347-
361. 
Liao, T. F. 1994. Interpreting probability models: Logit, probit and other 
generalized linear models. Sage publication, Thousand Oaks, CA. 
Long, S. 1997. Regression models for categorical and limited dependent 
variables. Sage publication, Thousand Oaks, CA. 
Lubowski, R.N. 2002. Determinants of land-use transitions in the United States: 
econometric analysis of changes among the major land-use categories. Ph.D dissertation, 
Harvard University, Cambridge, Massachusetts. 266 p. 
Maddala, G.S. 1983. Limited-dependent and qualitative variables in 
econometrics. Cambridge University Press, Cambridge. 
 30
Majumdar, I., M. Polyakov and L.D. Teeter 2005. Landuse change determinants 
in Alabama. P. 67-70 in Proc. of conf. on Emerging issues along urban/rural interfaces: 
Linking science and society, Laband, D.N. (eds.) Auburn, AL, August 2005. 
McFadden, D. 1974. Conditional logit analysis of qualitative choice behavior. P. 
105-142 in Frontiers in Econometrics, (P. Zarembka eds.) Academic Press. 
McFadden, D. 1981. Econometric models of probabilistic choice. in C.F. Manski 
and D. McFadden (eds.), Structural Analysis of Discrete Data with Econometric 
Application, MIT Press: Cambridge, MA. 
Munn, I.A. and D.L. Evans. 1998. The Southern commercial timberland base: 
changes and projections. Presented at the first international conference on geospatial 
information in agriculture and forestry, June, Florida. 
Munn, I., S.A. Barlow, D. Evans and D. Cleaves 2002. Urbanization?s impact on 
timber harvesting in the south central United States. J. Environ. Manage. 64:65-76. 
Nagubadi, V.R. and D. Zhang 2005. Determinants of timberland use by 
ownership and forest type in Alabama and Georgia. Journal of Agricultural and Applied 
Economics 37(1):173-186. 
Smith, W.B., P.D. Miles, J.S. Vissage and S.A. Pugh 2002. Forest resources of 
the United States. USDA For. Serv. Gen. Tech. Rep. NC-241. St. Paul, MN. 
Turner, M.G., D.N. Wear and R.O. Flamm 1996. Land ownership and land-cover 
change in the Southern Appalachian Highlands and the Olympic Peninsula. Ecol. Applic. 
6(4):1150-1172. 
 31
Vitousek, P.M. 1994. Beyond global warming: ecology and global change. 
Ecology 75:1861-1876. 
Wear, D.N. and P. Bolstad 1998. Land-use changes in southern appalachian 
landscapes: spatial analysis and forecast evaluation. Ecosystems. 1:575-94. 
Wear, D.N. and R.O. Flamm. 1993. Public and private forest disturbance regimes 
in the southern appalachians. Natural Res. Modeling 7(4):379-397. 
Zhou, X., J.R. Mills and L.D. Teeter 2003. Modeling forest type transitions in the 
South Central region: Results from three methods. Southern Journal of Applied Forestry. 
25(3):190-197. 
Zhang, D. and R.V.Nagubadi 2005. The influence of urbanization on timberland 
use by forest type in the Southern United States. Forest Policy and Economics. 7:721-
731. 
 32
CHAPTER 3 
FAMILY FOREST OWNERS IN THE SOUTHEAST: ARE THEY ALL THE SAME? 
3.1 Introduction 
Forests and forestry have played a significant role in the economic development 
and psyche of the South. Forests, which in pre-settlement times occupied nearly all of the 
land area of the South, now occupy only 56 percent ((Economic Research Services of the 
USDA 2002)http://www.ers.usda.gov/Data/MajorLandUses/MLUsummarytables.pdf). 
More important, perhaps, is the change in the composition and use of these forests. Some 
of these changes have resulted from forest conversions to agriculture, and subsequent 
reversions back to forest (Healy 1985). Other changes occurred as fiber demand 
increased over time and harvested lands were replanted with pines. More recent is the 
recognition that forests provide significant amenity and recreational values, which may 
lead to reductions in harvest by non-industrial private forest landowners.  
While forests provide both market and amenity outputs, these outputs are not 
necessarily complementary. The dominant market output is timber, the harvest of which 
often conflicts with production of high-quality amenity benefits. Thus, the values held by 
private landowners for amenities play a role in influencing private forest management by 
changing the harvest date or amount of timber produced from any given stand.  
With the unprecedented recent growth in the number of private forest landowners 
there is an increased need to research and investigate the motives of these landowners to 
 33
manage their land for timber and/or non-timber use. Three broad categories of ownership 
constitute what we consider private forestlands: family owned or individual owners, 
industrial ownership, and Timber Investment Management Organizations (TIMOs) or 
Real Estate Investment Trusts (REITs). While of these latter two are considered to be 
primarily in the business of forestland management for profit and invariably their 
management actions focus on timber harvests, the objectives of the former, individual 
forestland owners, still remain largely unknown. The individual and family forest 
landowners hold 42% of the nations timberland ((261.6 million acres) and 59% (127.6 
million acres) in the South (Butler and Leatherberry, 2004). Given their numbers it is 
important to study their diverse objectives, goals and intentions for managing their lands 
for timber and/or non-timber purposes.  
Substantial research has been done over the past few decades focusing mainly on 
ways to motivate the family forest landowners to practice active forest management to 
boost timber supply. The relationship between harvesting decisions and the 
characteristics of landowners (Binkley 1981) has been the focus of most studies on 
private forest management behavior. The relationship between forest amenity 
characteristics and private forest harvest, however, has not been well established. One 
feature of all of the studies, with some exceptions (Butler 2005; Kluender and 
Walkingstick 2000; Finley 2002; Green and Blatner 1986; Gramann et al. 1985; Young 
and Reichenbach 1987) is to consider individual private landowners as a homogeneous, 
single group with similar motivations. In reality, the validity of this assumption is 
questionable. 
 34
                                                
This article tests the hypothesis that family forest landowners form a 
heterogeneous group with differing motivations and goals for forest management, and 
that even when they face the same market environment their actions differ. 
3.2 Literature Review 
The two most widely researched aspects of NIPF behavior, which have been at 
the center of studies over the past few decades, are their harvest (Binkley 1981, Boyd 
1984, Dennis 1990, Hyberg and Holthausen 1989, Pattanayak et al. 2003) and 
reforestation behavior (Alig 1986, Newman and Wear 1993, Kline et al. 2002)
14
. Family 
forest owners
15
 are often considered as a single entity, while their increasing numbers and 
ownership statistics suggest that they are a heterogeneous group with diverse objectives 
for managing their land. Almost 4 out of 7 family forest owners in the South owns less 
than 10 acres of land and constitute 5.6% of all family forests in the South. Very few 
studies in the US on NIPF behavior in general have explicitly considered categorizing 
forest owners according to their diversity in attitudes, objectives and non-timber values. 
Using data from a survey of 146 Finnish landowners in Southern Finland, 
Kuuluvainen (1996) employed K-means cluster analysis to empirically identify four 
groups of non-industrial private forest landowners (NIPFs) owners based on their 
objectives as multiobjective owners, self-employed owners, recreationists and investors. 
Lewis (1979) and Kurtz and Lewis (1981) utilized Q-methodology to construct a 
 
14
 For an excellent review of literature on NIPF studies see Beach et al. 2005. 
15
 Family forest owners are defined as ?family forests include lands that are at least 1acre in size, 10% 
stocked, and owned by individuals, married couples, family estates and trusts, or other groups of 
individuals who are not incorporated or otherwise associated as a legal entity? (Butler and Leatherberry 
2004) 
 35
taxonomy of family forest owners in the USDA Forest Service Eastern Ozarks region of 
Missouri and identified four attitudinal types which were identified and described as 
timber Agriculturists, timber conservationists, forest environmentalists and range 
pragmatists.  
More recently, a survey of 866 family forest owners in Arkansas and subsequent 
cluster analysis identified four distinct groups of family forest owners: timber managers, 
resident conservationists, affluent weekenders and poor rural residents (Kluender and 
Walkingstick 2000). The types described by Kluender and Walkingstick (2000) were 
constructed using a combination of demographic and management characteristics. Using 
objective demographic characteristics of the landowners as variables to classify them fail 
to take into account explicitly the subjective attitudinal or psychic constructs of 
landowner motivations, which can be considered as latent qualities of the landowner. 
While forest management actions may differ from landowners? philosophies about 
forestland stewardship based on certain external factors, it is expected that their 
perceptions and motivations determine the nature of their forest management activities in 
the long run. 
Kittredge (2004) suggests that market segmentation may provide a superior 
approach to outreach compared with the traditional methods that assumed a single more 
homogeneous group of family forest owners. Market segmentation allows the audience to 
be broken down into relatively homogeneous similar classes, and the needs and desires of 
each class can then be ascertained. With the ownership class identified, certain groups 
can be chosen as priority targets for specific outreach programs. For example, Broderick 
 36
et al. (1996) grouped family forest owners in Connecticut based on their intentions 
concerning forest stewardship planning. The groups consisted of those who intended to 
sell their land (sellers), those who had a stewardship plan or had protected their land 
(planners), those who intended to develop a stewardship plan (intenders), and those who 
showed little inclination towards stewardship planning (non-intenders). 
Kendra and Hull (2005) used cluster analysis to group family forest owners who 
had recently purchased forestland in rapidly growing counties in Virginia. In this case, 
the typology was based solely on the owners? responses to survey items measuring forest 
ownership motivations. The resulting six types were then described on the basis of 
demographic, land ownership and management characteristics and labeled absentee 
investors, professionals, preservationists, young families, forest planners, and farmers. 
This study serves as a very recent example of a typology of family forest owners for 
which the classification was based on purely psychological variables. Though this study 
is significant in exploring the motivations of new owners and their reasons for acquiring 
forestland it fails to validate the results due to the absence of data on any of the past 
actions of the owners and as such the connection between landowner attitudes and their 
probable management actions in the future could not be made.  
Various studies have been done on family forest owner attitudes in the South. 
Although these have not explicitly considered grouping landowners into homogeneous 
attitudinal groups, they do give some insight to perceptions of an average forest owner. 
Bliss and McNabb (1992) found that 43 percent of Alabama family forest owners 
believed that forestry should be regulated on private lands to protect the environment. 
 37
Contrary to this result a recent study by Kennedy and Roche (2003) revealed that 55 
percent of Alabama family forest owners believed, ?Providing timber and wood products? 
was the most important role of their forests. Birch (1997) found that in Alabama 27 
percent of landowners felt that ?residence? was the primary objective of land ownership 
and 56 percent reported to having done timber harvests in the past. 
In a study on the NIPFs of Florida, Jacobson (1998) found 64% of landowners to 
be ?absentee owners? not living on their forestland and concluded that such owners are 
more likely to hold land for aesthetic beauty, wildlife habitat and recreation rather than 
timber. Newman et al (1996) reached a similar conclusion from a mail survey of the 
NIPFs of Georgia. Lorenzo and Beard (1996) found a significantly high correlation 
between the participation of NIPFs in governmental assistance programs and acreage of 
ownership in Louisiana.  
Summarizing this section on the review of past studies, specifically those on 
family forest owners in the South, we see a lot of variation regarding their motivations 
and the management strategies they employ. Emphasizing the diversity of family forest 
owners in the South, Wicker (2002) stated, ?available research information is insufficient 
to define an average private southern forest landowner.? 
3.3 Landowner Model 
A typical rational forest landowner is assumed to maximize his utility from his 
forest holding by equating his preferences for timber and non-timber values to the total 
capacity of the land to provide these two benefits given resource and budget constraints. 
Based on Vincent and Binkley?s (1993) model for a single stand, the optimal point where 
 38
the landowner will maximize his utility depends on the interplay of the production trade-
offs (the combinations of timber and non-timber units that the stand can produce) and the 
consumption (psychic) trade-offs which are determined by the landowners? perception of 
the relative value of timber and non-timber products of the forest. Binkley argues that for 
a single stand, unless the relative price line is either ?too? steep or ?too? flat, the multiple 
use option is always superior and rejects the possibility of a corner solution where the 
landowner chooses either to produce only timber or only non-timber. We support 
Vincent?s and Binkley?s argument that the most plausible option for family forest 
landowners in general is to practice multiple-use forest management in absolute terms. 
We argue, however, that based on the psychic price (value) that individual landowners? 
perceive from non-timber benefits, which typically do not have any market price, the 
slope of the relative price (value) line can differ to such a degree that it may be possible 
to group/classify landowners? based on their motivation to manage for either only timber 
or only non-timber or both. 
To illustrate our point consider three family forest landowners? A, B and C who 
own single forest stands where each stand can produce two products, timber (T) and Non-
timber (NT). We assume a strictly concave production possibilities frontier (PPF) for 
each of the three landowners consistent with the usual microeconomic assumption of 
increasing opportunity costs as one produces more units of a product (see Figure 3.1). 
The landowners maximize their utility at the tangential point between the PPF and the 
relative price (value) line such that landowner A produces A
T 
and A
NT
, landowner B 
produces B
T
 and B
NT
 and Landowner C produces C
T
 and C
NT
 quantities of timber and 
non-timber (Figure 3.1). The object of this paper is to test the validity of the existence of 
similar family forest landowner groups in the Southeast as represented by landowners A, 
B or C using multivariate statistical techniques. 
Non-timber 
   
 
 
 
 
 
     A
T
          B
T
        C
T
  Timber 
C 
B 
A 
A
NT
B
NT
C
NT
Figure 3.1 Landowner Behavior Model 
3.4 Data and Methods 
This study is based on the analysis of National Woodland Owner Survey (NWOS) 
data on the family forest owners in three Southern states: South Carolina, Georgia and 
Alabama. NWOS is conducted under the Forest Inventory and Analysis (FIA) program of 
the United States Department of Agriculture (USDA) Forest Service (USFS). The data 
used in this study was collected during the period 2002-2004.  
The NWOS used a self-administered questionnaire distributed to family forest 
owners by the U.S. Postal Service as the 
interviews conducted sometimes to augm
questionnaire included 30 questions concerning: 
 
prim
en
ary survey instrum
t response rates (Butler et al 2005). The 
39
ent with telephone 
 40
                                                
? Forest land characteristics 
? Ownership objectives 
? Forest use 
? Forest management 
? Sources of information 
? Concerns and issues 
? Demographics 
The questions in the survey were prepared using a comprehensive questionnaire 
review process which included expert reviews, pretesting of the survey instrument at 
several forest land-owner conferences and professional meetings, input from state 
forestry agencies, expert opinion and review by the clearance office of the USDA forest 
service
16
. 
3.4.1 Data 
The total number of private landowners responding to the NWOS during the 
survey period in South Carolina (SC), Georgia (GA) and Alabama (AL) was 1854 
(SC=753, GA=813 and AL=290).  Out of these private owner responses, we discarded 
forest industry (FI) + TIMOs + REITs since we were interested in exploring the diverse 
set of motivations of the family forest owners. We assumed that the motivations of FIs, 
TIMOs and REITs were to generate profit from timber management. We also excluded 
all owners with parcel sizes less than 25 acres due to the economic inefficiencies 
 
16
 For a detailed description of the development and implementation of the survey instrument (NWOS) read 
?Design, Implementation, and Analysis Methods for the National Woodland Owner Survey? (Butler et al 
2005). 
 41
associated with managing such smaller parcels for timber, and assumed that a rational 
owner with the aim of maximizing his utility from the forestland had to be motivated 
mainly by the non-timber amenity values of the forest for a parcel smaller than 25 acres. 
This resulted in reducing the number of respondents included in the analysis to 1339 from 
1854. 
3.4.2 Statistical Methods 
This study is related to the identification of family forest landowner groups based 
on their similar motivations to manage their land and the attached values and interests of 
these owners in their forestland. The questions in Table 3.1 (question number 9 in 
NWOS) form the basis for identifying the landowner typologies and consist of 10 
questions, each emphasizing the perceived importance of various benefits that may be 
important to the forest owners. All questions were rated by the respondents using an 
ordinal Likert-type scale of 1-7 where 1 reveals the strongest motive corresponding to 
?Very Important? and 7 reveals the weakest motive corresponding to ?Not Important? for 
owning the land. The distribution of answers is given in Table 3.1. The responses to 
questions related to Non-timber Forest Products (NTFP) and Firewood (9f and 9g) were 
removed from further analysis due to their small variance as most of the respondents gave 
similar ratings to these questions. Question related to the perceived importance of having 
the woodland as part of their home or vacation home (part of question 9) was also not 
included in the analysis due to dissimilarities in the framing of this question across 
different versions of the questionnaire used in the three different states for the years 2002 
 42
to 2004. In effect we used 8 attitudinal questions (9a to 9j excluding 9f and 9g) for further 
analysis. 
Table 3.1 Survey questions from NWOS used for cluster analysis 
                                      Percentage of answers 
Question 
no. 
Question 1 2 3 4 5 6 7  
  Very 
Important
?? ?? ?? ?? ?? Not 
important
No 
answer
9 ?People own 
woodland for many 
reasons: How 
important are the 
following as reasons 
for why you own 
woodland?? 
        
9a To enjoy beauty or 
scenery 
46.23 13.97 10.83 10.38 4.18 1.72 4.71 7.99 
9b To protect nature or 
biologic diversity 
36.07 14.04 12.47 12.62 5.38 4.78 4.33 10.31 
9c For land investment 50.78 11.58 10.75 6.87 3.29 2.02 6.87 7.84 
9d For privacy 38.31 9.86 8.14 7.77 4.26 4.26 15.91 11.50 
9e To pass land on to 
my children or other 
heirs 
58.18 12.02 8.14 6.65 2.69 1.27 5.75 5.30 
9f For 
cultivation/collection 
of non-timber forest 
products 
7.54 4.63 6.35 11.80 7.24 11.95 37.27 13.22 
9g For production of 
firewood or biofuel 
(energy) 
4.41 2.91 6.05 8.66 7.47 14.64 42.94 12.92 
9h For production of 
sawlogs, pulpwood 
or other timber 
products 
44.51 13.74 10.90 7.84 3.29 2.84 10.31 6.57 
9i For hunting or 
fishing 
35.10 14.34 12.77 11.05 3.81 3.36 12.32 7.24 
9j For recreation other 
than hunting or 
fishing 
19.64 9.26 11.13 12.92 6.42 6.87 22.48 11.28 
 43
                                                
Principal Components Analysis (PCA) is the most important statistical routine for 
dimensional reduction and seeks to transform a larger set of correlated variables into a 
smaller set of uncorrelated variables or factors without losing much information. PCA 
with varimax rotation was used to reveal the latent constructs (factors) of forest owner 
motivations based on the 8 questions mentioned above by utilizing the variance-
covariance matrix of responses. Varimax rotation
17
 was used to facilitate interpretability 
of factors by maximizing the variance of loadings (correlation coefficients between the 
factors and the variables) on each factor for use in the subsequent clustering procedures. 
In other words Varimax solution ensures that each factor tend to have either large or 
small loadings of any particular variable. Items loading 0.4 or greater (Garbin and Teng, 
1988) on a factor were used to interpret the factors. Two factors were identified as 
economic and non-economic with the former denoting a strong timber interest related to 
timber harvests and land investment and the latter denoting the non-timber amenity 
values (biodiversity, aesthetics, recreation) of the forest perceived as the most important 
reasons for owning the forestland by the landowner. The overall Kaiser-Meyer-Olkin 
(KMO)
18
 measure for factor suitability was 0.72 confirming the factorability of the 
indicator variables (NWOS questions). The two factors together explained 55% of the 
variance in the responses to the reasons for owning forestland by the landowners. 
Reliability analysis was conducted by computing Cronbach?s alphas for each factor 
 
17
 Varimax is an orthogonal rotation of the factor axes to maximize the variance of squared loadings of a 
factor (column) and all the variables (rows) in a factor matrix (see Table 3.2). 
18
KMO is a measure of sampling adequacy and evaluates the appropriateness of the correlation matrix for 
factoring.  KMO values should be greater than 0.6 for a satisfactory factor analysis (Tabachnik & Fidell 
2001). 
 44
which ranged from .64 to .72, suggesting internal consistency for each of the factors 
extracted. Finally, a scores matrix of the order N ? 2 where N (1339) denotes the total 
number of NWOS respondents with a score on each of the 2 factors was computed by 
taking each respondents standardized score on each variable, multiplied by the 
corresponding factor loading of the variable for the given factor, and summing these 
products. The factor scores describing owner motivations to manage their forestland were 
used as criterion variables for the cluster analysis. The factor loadings denoting the 
correlations between the variables (rows) and factors (columns) are given in Table 3.2. 
The pattern loadings of the two factors are given in Table 3.2 below and represent 
the common variance in each of the variables (NWOS questions) explained by the 
factors. 
Table 3.2 Factor loadings representing the correlations between factors and the 
variables 
 Non-economic Economic 
Aesthetics 0.76 0.01 
Biodiversity 0.62 0.00 
Recreation 0.65 0.07 
Privacy 0.65 ?0.03 
Legacy 0.26 0.27 
Hunt 0.50 0.31 
Timber ?0.07 0.83 
Investment 0.03 0.46 
 
3.4.2.1 Cluster Analysis 
In order to get meaningful groups of family forest owners based on their 
motivations for owning and managing their forestland, NWOS data was subjected to 
clustering analysis using the factor scores on the two factors extracted for each 
 45
respondent. Since all the clustering routines available through various mathematical 
software packages are biased towards identifying clusters with certain characteristics, 
once the data are input it is necessary to identify the algorithm which gives the best 
interpretable results and then test cluster validation. As a first step to clustering, the SAS 
procedure CLUSTER explored various hierarchical methods such as single linkage, 
complete linkage, average linkage, centroid and Ward?s method (SAS 2004, p. 955) to 
determine the best method for clustering the data. The hierarchical clustering method is 
exploratory in nature and assumes no a-priori information about the number of clusters. 
To get landowner clusters of reasonable proportions and exclude the possibility of 
producing groups that were too small, Ward?s minimum variance method was used. This 
method is based on least-squares criteria and tries to minimize the within-cluster sum of 
squares, thus maximizing the within-cluster homogeneity. The ?agglomerative 
dendrogram? that provides a visual representation of the step-by-step hierarchical 
clustering process wherein at each step the two closest clusters are merged into one 
bigger cluster, was not very useful to evaluate the cluster solution owing to the 
cumbersome interpretation of a large number of observations (respondents). Based on 
some of the most widely used statistics like root-mean-square standard deviations 
(RMSSTD), semi-partial R-squared (SPR) and R-squared (RS) a three cluster solution 
was found to be appropriate and supported our initial hypothesis. 
 46
                                                
Using a non-hierarchical (K-means) method to sort the observations to the nearest 
centroid through the procedure FASTCLUS
19
 in SAS we found similar results compared 
to the hierarchical method. The results discussed in the next section were obtained by the 
non-hierarchical clustering routine. 254 incomplete observations (no response on at least 
one of the 8 questions on reasons for owning forestland from Item 9 of the NWOS) were 
excluded from the cluster analysis and this resulted in reducing the number of 
observations to 1085 from 1339.  
The three clusters were named timber (319 owners), non-timber (233 owners) and 
multiple-objective (533 owners). The mean response for all the questions used in the 
cluster analysis by cluster groupings is plotted in Figure 3.2. As can be seen in Figure 3.2 
below, the multiple-objective owners valued both the economic and non-economic use of 
the forest as the driving forces for owning forestland. Surprisingly the multiple-objective 
owners gave higher importance to the non-economic reasons than the non-timber cluster 
and also a higher importance to the economic reasons than the timber cluster. It is evident 
that this owner group, which includes the largest percentage of all family forest owners 
(49.1%) are strongly motivated by both consumptive (hunting, timber harvest) use values 
and the non-consumptive (aesthetic beauty, biodiversity) use values equally. The timber 
cluster (29.4%), as expected, had only timber management and land investment as strong 
motivating factors behind their forestland ownership. Timber cluster owners are typically 
concerned about the consumptive use value of their forestland unlike the non-timber 
 
19
 FASTCLUS in SAS uses a nearest centroid sorting iterative method where a set of points known as 
cluster seeds is selected as the first guess of the mean of the clusters and each observation is assigned to the 
nearest seed to form temporary clusters, the seeds are then replaced by the seeds of the temporary clusters 
in an iterative manner until no further changes occur in the clusters (for detail see SAS 2004). 
cluster (21.5%) owners who value the non-consumptive uses of their forestland such as 
aesthetic values, biodiversity, recreation and privacy. 
0
1
2
3
4
5
6
7
Bi
o
di
ve
rsi
t
y
A
e
sthe
ti
cs
P
ri
v
ac
y
L
eg
ac
y
R
ec
re
ate
F
i
r
ew
d
NT
F
P
Hu
n
t
Ti
mb
er
Inv
e
st
Mean R
e
sponse
Multiple Objective
Timber
Non-timber
All Observations
 
Figure 3.2 Mean response of reasons for owning forestland by cluster group, 1 
denoting ?Most important? and 7 denoting ?Least important? 
The socio-demographic and forest characteristics of the three types of family 
forest owners are described in Table 3.3. Responses reveal that the non-timber cluster 
owners were the least educated and least wealthy in comparison to the multiple-objective 
cluster owners and the timber cluster owners. The majority of the non-timber type of 
owners was retirees though the mean age of all the owners across all ownership types was 
greater than 60 years, suggesting that the family forests are going to change hands and 
new owners are going to replace the present surveyed owners shortly. It remains to be 
seen whether these new owners will have similar motivations as their predecessors or if 
they will act differently. The longest average tenure of forestland ownership lies with the 
 47
 48
timber cluster owners reflecting that profit motivated owners generally have managed the 
same forestland for a longer time as compared to owners in other clusters. It also reflects 
that timber management is a long term decision of the owner belonging to the timber 
cluster when compared to maintaining forestland primarily for non-timber uses by the 
owners of the non-timber cluster. As expected the timber cluster tended to own the 
largest tracts of forest with a mean size of 1857.9 acres and the non-timber cluster owned 
the smallest sized tracts of 384.5 acres on an average. It suggests that owners of larger 
tracts are more likely to manage for timber quite different from owners of smaller 
parcels. The timber cluster owners were also found to have strong linkages to farming 
and owned on an average 412 acres of farmland. The multiple-objective owners, on an 
average, owned slightly more farmland with an average farm size of 444.7 acres. Not 
surprisingly, the non-timber group with strong non-timber ownership objectives had the 
smallest mean farm size of 229.7 acres. There was a stark contrast in the percent of 
owners belonging to the timber cluster (52.3%) who had leased their land relative to 
owners within the non-timber cluster (19.7%). Further empirical evidence amongst the 
single ownership objective groups (timber and non-timber) as expected reflected the 
difference in the behavior related to timber management with a sharp difference in the 
percent of owners within each group who had written management plans, had harvested 
timber or had done some site preparation to plant new trees. The majority of owners 
classified in the timber cluster had inherited their forest property while the non-timber 
owners were least likely to have inherited their forestland. This coupled with the fact that 
these owners have the maximum tenure show that timber motivated owners have high 
 49
legacy values relative to the non-timber type of ownership. However, owners belonging 
to the multiple-objective ownership class had relatively stronger preferences for both 
timber and non-timber products relative to the other two owner types (see Figure 3.2). 
Table 3.3 Socio-demographic and forest characteristics of family forest owners by 
cluster 
Characteristic Multiple-
objective 
Timber Non-timber 
Mean age (yrs) 61.5 64.2 62.2 
Men (%) 74.1 66.1 75.9 
Mean duration of 
ownership (yrs) 
28.6 31.2 22.4 
Income (1000$) 79.4 78.3 71.4 
Education 4.2 4.3 3.8 
Retired (%) 36.7 42.9 45.5 
Mean forest area 
(ac) 
1345.3 1857.9 384.5 
Farm area (ac) 444.7 411.6 229.7 
Management plan 
(%) 
32.1 25.7 10.3 
Site preparation 
(%) 
47.8 43.3 12.9 
Harvest (%) 89.3 86.2 56.6 
Leased (%) 44.1 52.3 19.7 
Inherit (%) 49.3 56.7 27.5 
 
3.4.2.2 Cluster validation 
While classification procedures using cluster analysis have been applied to family 
forest owners in a number of studies (Kurtz and Lewis 1981, Marty 1983, Kluender and 
Walkingstick 2000, Kittredge 2004, Kendra and Hull 2005), none of the studies reported 
results of any empirical cluster validity test. Based on the 5-step cluster validation 
technique as suggested by Lattin et al. (2002) we performed a validation test on the 
NWOS clustering results (Table 3.4). According to this technique at the first step the data 
 50
were randomly split in the ratio 1:1 using the RANSPLIT macro in SAS. The two 
samples thus formed are referred to as the calibration and the validation samples. At the 
second step the calibration sample was used for hierarchical cluster analysis and the 
appropriate number of clusters and their centroids were determined. In the third step the 
cluster centroid from the second step was used to assign each observation from the 
validation sample to the nearest centroid using non-hierarchical cluster analysis and the 
cluster solution was saved. In the fourth step the validation sample was used to perform a 
similar hierarchical cluster analysis as in the second step and the results were saved in a 
SAS database. Finally the cluster solutions obtained from the step-3 and step-4 were 
compared and a confusion matrix (Table 3.4) depicting the percent of observations in 
each of the three cluster groups incorrectly classified into another group was created. As 
can be seen, the percent-misclassification was pretty low and most of the observations 
that were clustered at both step-3 and 4 of the validation routine also were found to be in 
the same cluster group with the percent of correct classification for each of the three 
types of landowner to be above 95% (see table 3.4). 
Table 3.4 Confusion matrix for cluster validation 
 Multiple-objective Timber Non-timber % 
Misclassification 
 
Multiple-
objective 
322 * 8 2.4 
Timber 5 181 * 2.7 
Non-timber * 2 189 1.05 
* Denotes null or 0 number of observations 
 51
3.5 Conclusion 
Our study supports the presence of three groups of family forest owners in the 
three Southeastern states of AL, GA and SC as discussed in the theoretical model on 
landowner behavior above and also as reported by Butler (2005) in his study of family 
forest owners in five southeastern states. It also emphasizes the need to differentiate 
family forest owners into smaller well-behaved homogeneous entities. Contradictory to 
Kendra and Hull?s (2005) recent study on new owners in Virginia, the bulk of 
landowners in our study were found to be motivated strongly by the profit motive either 
through timber harvests as a source of income generation or choosing forestry as a better 
land investment option. As reported above, landowners have different objectives and 
motivations for managing their forestland and identification of those may be critical to 
developing better informed policy prescriptions. Policies can be targeted towards each 
owner group according to their needs and interests and thus policy implementation can be 
made more efficient. For example, timber harvests for owners within the non-timber 
group may be for wildlife habitat or to maintain a healthy forest which is quite different 
than for economic reasons.  
The multiple-objective ownership type was found to be the largest group with 
almost every 1 out of 2 family forest owners in the sample population belonging to this 
category. These owners derive utility from both economic and non-economic uses of the 
forest and also potentially could be the ones targeted by policy makers and resource 
managers to enhance their production of timber or non-timber outputs since they are not 
 52
devoted to any single management objective unlike owners in the timber and the non-
timber clusters.  
The above work is by no means complete and further analyses of the data by 
integrating the detailed forest characteristics, which complement the ownership NWOS 
data, along with linkages to the socio-economic Census data, could produce important 
information on family forest owner behavior. Also a large number of observations (223) 
excluded from the analysis due to incomplete responses warrants a closer look to check if 
there are enough similarities amongst them to be classified as a separate cluster or not. 
Finally, the average age of family forest owners is in the sixties and it remains to be seen 
if the future change of ownership will be associated with changing owner attitudes and 
motivations or not. This also suggests the dynamic nature of human behavior and one on 
which studies need to be updated from time to time. 
This study is based on the psychological responses of landowners to factors they 
think motivate them most for owning forestland and the results should be explored 
further to see if landowner attitudes are supported by their actual behaviors. In other 
words to draw conclusions, it is necessary to check if landowners do what they say or are 
the responses merely what landowners perceive as ideal. 
 53
3.6 References 
Alig, R., 1986. Econometric analysis of the factors influencing forest acreage 
trends in the Southeast. For. Sci. 32(1):119-134. 
Beach, R.H., S.K. Pattanayak, J.C. Yang, B.C. Murray, and R.C. Abt 2005. 
Econometric studies of non-industrial private forest management: A review and 
synthesis. Forest Policy and Economics. 7:261-281. 
Binkley, C.S. 1981. Timber supply from non-industrial forests: A microeconomic 
analysis of landowner behavior. Yale University Press, New Haven, CT. 
Birch, T.W. 1997. Private forestland owners of the Southern United States, 1994. 
USDA For. Serv. Resource Bulletin. NE-138, Radnor, PA. 
Bliss, J.C. and K. McNabb 1992. Landowners reveal some surprising attitudes 
about regeneration. Forest Farmer. 52(1):14-15. 
Boyd, R. 1984. Government support of non-industrial production: The case of 
private forests. Southern Economic Journal. 51(1):89-107. 
Broderick, S.H., L. Snyder, and B. Tyson 1996. Selling stewardship within the 
community: A social marketing approach, in Proc. of symposium on Nonindustrial 
private forests: Learning from the past, prospects from the future, M, Baughman, (eds.). 
Society of American Forestry, Washington D.C. 
Butler, B.J. 2005. The timber harvesting behavior of family forest owners. 
Unpublished Ph.D Dissertation, Oregon State University. 
 54
Butler, B.J., E.C. Leatherberry and M.S. Williams 2005.Design, implementation, 
and analysis methods for the national woodland owner survey. USDA For. Serv. Gen. 
Tech. Rep. NE-336, Newtown Square, PA. 43 p. 
Butler, B.J. and E.C. Leatherberry 2004. America?s family forest owners. J. For. 
102(7):4-9.  
Dennis, D. 1990. A probit analysis of the harvest decision using pooled time-
series and cross-sectional data. Journal of Environmental Economics and Management 
18:176-187. 
Finley, A.O. 2002. Assessing private forest landowners? attitudes towards, and 
ideas for, cross-boundary cooperation in Western Massachusetts. CFRF Final Report. 
Department of natural resources conservation. University of Massachusetts, Amherst.  
Garbin, C.P. and G.K. Teng 1988. Applied multivariate analysis. Springer-Verlag 
Publishing, New York. 
Gramann, J.H., T.D. Marty and W.B. Kurtz 1985. A logistic analysis of the 
effects of beliefs and past experience on management plans for nonindustrial private 
forests. J. Environ. Manage. 20:347-356.  
Greene, J.L. and K.A. Blatner 1986. Identifying woodland owner characteristics 
associated with timber management. For. Sci. 32(1):135-146.  
Healy, R.G. 1985. Competition for land in the American south. Washington, DC: 
The Conservation Foundation. 333 p. 
Hyberg, B. and D. Holthausen 1989. The behavior of non-industrial private forest 
owners. Can. J. For. Res. 19:1014-1023. 
 55
Jacobson, M.G. 2002. Factors affecting private forest landowner interest in 
ecosystem management: Linking spatial and survey data. Environmental Management. 
30(4):557-583. 
Kendra, A. and R.B. Hull 2005. Motivations and behaviors of new forest owners 
in Virginia. For. Sci. 51(2):142-154. 
Kennedy, V. and S. Roche 2003. Alabama landowner opinion survey, market 
research institute under contract with the Alabama Forestry Association. Unpublished 
report. 
Kittredge, D.B. 2004. Extension/outreach implications for America's family forest 
owners. J. For. 102(7):15-18. 
Kline, J.D., B.J. Butler and R.J. Alig 2002. Tree planting in the south: what does 
the future hold? Southern Journal of Applied Forestry 26(2):99-107. 
Kluender, R.A. and T.L. Walkingstick 2000. Rethinking how nonindustrial 
landowners view their lands. Southern Journal of Applied Forestry. 24(3):150-158.  
Kuuluvainen, J., H. Karpppinen and V. Ovaskainen 1996. Landowner objectives 
and nonindustrial private timber supply. For. Sci. 42(3):300-309. 
Kurtz, W.B., and B.J. Lewis 1981. Decision-making framework for nonindustrial 
private forest owners: An application in the Missouri Ozarks. J. For. 79:285-288. 
Lattin, J., J.D. Carroll and P.E. Green 2002. Analyzing multivariate data. Duxbury 
Press. 
 56
Lewis, B.J. 1979. Private non-industrial forest landowners in the Missouri Ozarks: 
A Q methodological establishment of types. Unpublished M.Sc. thesis, University of 
Missouri, Columbia. 251p. 
Lorenzo, A.B. and P. Beard 1996. Factors affecting the decisions of NIPF owners 
to use assistance programs, in Proc. of symposium on Nonindustrial private forests: 
Learning from the past, prospects for the future, Baughman, M.J., (eds.). Society of 
American Forestry, Washington D.C. 
Marty, T.D. 1983. An extension of nonindustrial private forest landowner 
typologies: A Wisconsin study. Unpublished M.Sc. thesis, University of Missouri, 
Columbia. 230p. 
Newman, D. and D. Wear 1993. Production economics of private forestry: a 
comparison of industrial and non-industrial forest owners. American Journal of 
Agricultural Economics. 75:674? 684. 
Newman, D.H., M.E. Aronow, T.G. Harris and G. Macheski 1996. Changes in 
forest land ownership characteristics in Georgia. in Proc. of symposium on Nonindustrial 
private forests: Learning from the past, prospects for the future, Baughman, M.J., (eds.). 
Society of American Forestry, Washington D.C. 
Pattanayak, S., K. Abt and T. Holmes 2003. Timber and amenities on non-
industrial private lands. In: Sills, E. and K. Abt, (eds.), Forests in a market economy. 
Forestry sciences series. Kluwer academic publishers, Dordrecht. 
SAS Institute Inc. 2004. SAS/STAT
?
 9.1 User?s Guide. Cary, NC: SAS Institute 
Inc. 
 57
Tabachnik, B.G., and L.S. Fidell 2001. Using multivariate statistics (4
th
 Ed.). 
Needham Heights, MA: Allyn and Bacon. 
Vincent, J.R. and C.S. Binkley 1993. Efficient multiple-use forestry may require 
land-use specialization. Land Economics. 69(4):370-376. 
Wicker, G. 2002. Motivation for Private Forest Landowners, in Southern forest 
resource assessment. Wear, D. and J. Greis (eds.). USDA For. Serv. Southern Research 
Station and Southern Region. 
Young, R.A. and M.R. Reichenbach 1987. Factors influencing the timber harvest 
intentions of nonindustrial private forest owners. For. Sci. 33(2):381-393. 
 58
                                                
CHAPTER 4 
DISCRIMINATING FAMILY FOREST OWNERS: A NON-PARAMETRIC APPROACH 
4.1 Introduction 
Who owns the forest?  Ownership patterns have changed dramatically over the 
last 20 years, with Forest Industry (FI) selling most of its holdings to Timber Investment 
Management Organizations (TIMOs), Real Estate Investment Trusts (REITs) and private 
individuals, parceling their holdings to more, smaller ownerships. Within the last decade, 
while forest industry was divesting itself of its forest holdings, the number of family 
forest owners
20
 nationwide rose by 11%, from 9.3 million to 10.3 million, and these 
owners now own 42 % of the nation?s forestland (Butler and Leatherberry, 2004). Given 
their large numbers and the expectation that their numbers will continue to increase, a 
closer look at their motivations and reasons for owning forestland is warranted. Previous 
research has often treated family forest owners as a single homogeneous class when 
trying to analyze how they might manage their forests in the future. 
But given their diversity in terms of the characteristics of the forest properties 
they hold, their economic and social surroundings and their personal histories and 
characteristics, this assumption of homogeneity may, in reality, be totally inappropriate. 
 
20
 Family forest owners are defined as ?family forests include lands that are at least 1acre in size, 10% 
stocked, and owned by individuals, married couples, family estates and trusts, or other groups of 
individuals who are not incorporated or otherwise associated as a legal entity? (Butler and Leatherberry 
2004). 
 59
                                                
We investigated the notion that family forest owners have diverse objectives and that 
their subjective preferences for managing their forest for timber (primarily for income) 
and non-timber amenity values are wide-ranging (Majumdar et al 2006) and inferred that 
family forest owners can be grouped into three motivational types
21
 namely, Timber, 
Multiple-objective and Non-timber. 
The objective of this paper was two-fold, first to identify the characteristics that 
discriminate best the three above mentioned owner types using discriminant analysis 
procedures and second to develop a classification scheme that will help in predicting the 
objectives of new landowners given the vector of the discriminating variables. The results 
of this study will be helpful in making a connection between the policy makers and the 
family forest owners to help develop effective policy prescriptions and educational 
programs targeting at the forest stewardship goals of the landowner. Also, information on 
the implications of our research for future timber availability will benefit policy makers 
interested in finding ways to insure future timber supplies are sustained at current or 
enhanced levels. 
4.2 Literature Review 
A number of studies have looked into the family forest owners? heterogeneous 
motivations to own forestland and have explored the presence of owner typologies 
(Young and Reichenbach 1987, Gramann et al. 1985, Kulluvainen et al 1996, Kluender 
and Walkingstick 2000, Kendra and Hull 2005, Butler 2005, Finley et al. 2006), but apart 
from characterizing the landowner groups based on socio-demographic attributes and 
 
21
 For details of the data, techniques used for grouping the family forest owners of the southeast refer to the 
earlier work by Majumdar et al, 2006. 
 60
attributes of their forests few of the studies have actually investigated the differences 
between landowner groups subject to their exogenous social and economic environment. 
This article investigates how the Timber, Multiple-objective and Non-timber family forest 
owners in the Southeast differ from each other by linking their decision environment, 
which includes their social and economic surroundings, to their group membership, with 
a goal of being able to predict new landowner membership into one of the ownership 
categories. The multivariate discriminant function (DF) analysis technique was used to 
evaluate the predictive power of the explanatory variables. DF has been recently used in a 
number of studies related to private forest landowner behavior, for example, Arano et al. 
(2004) in their study of 829 Mississippi forest landowners consisting of two group, non-
regenerators (402) and regenerators (427), used canonical discriminant function analysis 
to discriminate the owner groups based on ownership characteristics, landownership 
characteristics, attitudes towards timber investment, awareness of assistance programs 
and attendance in educational programs. Finley et al. (2006) used multiple discriminant 
analysis to link the association of the predefined groups of Massachusetts private 
landowners, namely, general cooperators, conservation cooperators, neutralists and non-
cooperators, to forestland ownership reasons, attitudes and actions towards forest 
management and perceived barriers to cooperation. Though the studies mentioned above 
help in identifying factors or variables that best discriminate between the private 
landowner typologies identified, they do not report the classification results and so their 
reliability in terms of predicting landowner behavior is low. Greene and Blatner (1986) 
successfully classified nearly 80% of the respondents in a study of Arkansas private 
 61
                                                
forest landowner groups, timber managers and nonmanagers, based on a set of 47 
variables related to reasons for owning forestland, owner characteristics, forest 
characteristics and their attitudes towards forest management. This study suggests that the 
discriminating variables can be used to predict group membership, but since the entire set 
of variables comes from the survey instrument administered to the respondents, it is not 
possible to predict group membership for a new landowner who has not been 
administered the survey. This also reveals the importance of exogenous discrimination of 
landowner groups (where all the discrimination variables are exogenous, and not personal 
attributes of individual landowners). The objective of this study was to be able to use 
easily available secondary information on new landowners from sources other than the 
primary National Woodland Owner Survey
22
 (NWOS) and predict their membership to 
one of the landowner groups established using NWOS. This requires first to identify such 
exogenous variables that can discriminate the three family forest landowner groups 
(Majumdar et. al. 2006), timber, multiple-objective and non-timber identified, and second 
to test the accuracy levels of the classification for predicting landowner behavior in the 
future. Given their increasing numbers and the forestland acres they manage, family 
forest owners have become increasingly important stewards of productive timberland in 
the South. Thus it is important to critically examine their motivations for managing their 
land either for timber and/or non-timber forest products.  
 
22
 For details on the NWOS survey design, implementation and analysis methods see Butler et al. (2005) 
 62
                                                
4.3 Data and Methods 
This paper involves multivariate discrimination procedures of three predetermined 
family forest owner types, multiple-objective (765 respondents), non-timber (292 
respondents) and timber (459 respondents). The three groups of landowners identified 
from NWOS was linked to the Forest Inventory and Analysis
23
 (FIA) data on the forest 
plots owned and managed by the family forest owners. Multiple conditions (FIA 
condition
24
) on each FIA plot resulted in landowners who had multiple conditions in their 
forest plot to be represented multiple times and so each observation (landowner) was 
weighted by the proportion of the plot that the condition represented in the analysis. This 
grouping of family forest owners in Alabama, Georgia and South Carolina, is based on 
responses to the NWOS obtained during the period 2002-2004. The landowner clusters 
were based on the importance they assigned to various reasons for owning forestland, 
ranging from timber and investment objectives to non-timber objectives such as 
biodiversity, aesthetics, hunting and recreation. 
4.3.1 Data 
The data consists of information on the socio-demographic, economic and bio-
physical characteristics of each landowner belonging to one of the three family forest 
ownership types and was taken from various sources. The bio-physical data came from 
the FIA forest plots owned and managed by the family forest owners and included 
 
23
 FIA forest resources inventory collects forest resources data annually from a sample of standard plots 
each representing roughly 6000 aces in the east, the social counterpart of the FIA forest resource 
inventories is the NWOS which is conducted on a sample of private forest owners of FIA plot already 
inventoried 
24
 FIA Condition represents the multiple conditions and is defined by heterogeneity in reserved status, 
owner group, forest type, stand-size class, regeneration status and stand density within FIA plots (for 
details on FIA data description and collection methods see (Alerich et al. 2004)). 
variables such as slope (SLOPE), average stand age (AGE), volume per acre (VOL), 
diameter per acre (DIA), distance to the nearest paved road (DIST), forest type (FT) 
(either pine, hardwood or mixed hardwood), site quality (SITE), physiography 
(PHYSIO), and tree biodiversity indices
25
 characterizing forest management 
heterogeneity amongst the owner types. The socio-demographic and economic data were 
incorporated in the study from linkages of the Census Bureau data with the FIA plots. To 
represent the decision environment related to economic and socio-demographic factors 
the variables included were real median household income (INC), population gravity 
index (PGI)
26
, pulpmills gravity index (MGI)
27
 and population density (PD). The SAS 
procedure STEPDISC was used to select the variables that best discriminated the three 
groups of landowners and from the large pool of variables mentioned above the ones 
selected were, PGI, INC, PD (demographic and economic) and AGE, SLOPE, DIST, 
SITE and FT (bio-physical characteristics). The sources and the description of all the 
variables used in the discriminant analysis are given in Table 4.1. 
 63
                                                 
25
 Three indices were created to reflect the biodiversity of tree species in the landowner forest plot namely 
Shannon?s index, Simpson?s index and Richness 
26
 PGI was calculated by linking landowner forest parcel location with the census demographic data on 
populated places as : kmDP
D
P
PGI
kp
p
kp
p
K
100:
2
??=
?
 Where is the population of census 
populated place 
p
P
p and is the distance between FIA plot k  and populated place 
kp
D p  
27
 MGI was calculated by linking the FIA plot to the pulpmills located within the radius of 200km around 
the forest plot and was calculated as: kmDM
D
M
MGI
km
m
km
m
k
200:
2
??=
?
 Where  is the 
pulping capacity in million cords of pulpmill 
m
M
M  and  is the distance between FIA plot  and 
pulpmill  
km
D k
m
 64
 
Table 4.1 Data sources and their descriptive statistics 
Variable Description Source Mean Std. dev 
PGI Number of 
persons/Km
2
 
around each FIA 
plot within a 
100km radius 
FIA plot and Census 
Bureau 
522.34 1432.23 
INC Median household 
income by county in $$ 
Economic Research 
Services (ERS) unit 
of USDA 
32852.98 7164.54 
POPDEN Number of persons per 
square mile of county 
land area 
Census Bureau 106.96 179.34 
DIST Euclidean distance 
from FIA plot center to 
the nearest improved 
road 
FIA plot 4.22 1.44 
SLOPE Angle of slope in 
percent 
FIA Condition 6.69 9.26 
FT Forest type having 
value of 1 for softwood, 
2 for mixed hardwood 
and 3 for hardwood 
FIA Condition 2.01 0.92 
AGE Average stand age FIA Condition 31.87 23.86 
SITE Site productivity class 
code taking values from 
1 to 6 with 1 
representing the best 
site 
FIA Condition 4.41 0.90 
 
4.3.2 Discriminant Analysis 
Discriminant Analysis (DA) is a statistical technique that allows the researcher to 
study the differences between two or more groups of objects with respect to several 
variables simultaneously (Klecka 1988). The goals of DA are to classify cases into one of 
the several mutually exclusive groups on the basis of various characteristics, to establish 
which characteristics are important for distinguishing amongst the groups, and to evaluate 
 65
the accuracy of the classification. The aim of DA in this study was to investigate the 
accuracy of classifying a landowner into either a multiple-objective or a non-timber or a 
timber motivated group. The two assumptions of multivariate normality (tested using the 
Mardia?s Kurtosis and Skewness statistical tests) and homogeneity of variances (tested 
using the SAS procedure DISCRIM with POOL = TEST option) for parametric 
discriminant analysis was conducted (see Table 4.2) and results suggested deviations 
from both the assumptions.  
Table 4.2 Multivariate Normality and Homogeneity of variance-covariance test 
result 
Test Chi-square statistic p-value 
Mardia?s Skewness
a
197.6 <0.0001 
Mardia?s Kurtosis
a
328.5 <0.0001 
Levene Homogeneity
b
338.8 <0.0001 
a
H
0
: Multivariate normality   
b
H
0
: Homoscedasticity   
 
An alternative to linear discriminant analysis (used when both the normality and 
equal variance-covariance matrix assumptions are fulfilled) or quadratic discriminant 
analysis (used when the homogeneity assumption) or Fisher?s discriminant analysis (used 
when the normality assumption does not hold) is non-parametric discriminant analysis 
and logistic regression analysis. It is unusual to find examples where researchers consider 
the statistical limitations and assumptions required for parametric techniques. In these 
instances, it is difficult to know whether the predictions are reliable. Because the primary 
goal of this paper was prediction of landowner?s membership into one of the 
predetermined groups based on a vector of predictor variables and the relevant 
assumptions for linear discriminant analysis could not be met, the non-parametric 
analytical technique was adopted. 
4.3.3 K-nearest neighbor method 
K-nearest neighbor classification (KNN), also known as nearest neighbor 
discriminant analysis, was introduced by Fix and Hodges (1951) and is based on 
distances from 'immediate neighbors' eliminating the need for a probability density 
estimation based on some distribution assumption. It is used to predict the response of an 
observation using a non-parametric estimate of the response distribution of its K nearest 
(i.e., in predictor space) neighbors. Consequently, KNN is relatively flexible and unlike 
traditional classifiers, such as discriminant analysis and generalized logit models, it does 
not require an assumption of multivariate normality or a strong assumption implicit in 
specifying a link function (e.g., the logit link which assumes the distribution of the 
dependent variable to be within the exponential family of distributions, such as normal, 
Poisson, Binomial, gamma). KNN classification is based on the assumption that the 
characteristics of members of the same class should be similar and thus, observations 
located close together in covariate (statistical) space are members of the same class or at 
least have the same posterior distributions on their respective classes (Cover and Hart 
1967). To decide which group a test case belongs to, SAS calculates the squared distance 
(Mahalanobis distance) between the test observation and each remaining member of the 
training dataset and classifies based on majority of classes for the nearest (shortest 
distance) K-neighbors. To illustrate the point suppose an observation (landowner) whose 
group membership is not known a-priori has attribute vector  and , , , ?  x
1
x
2
x
3
x
4
x
n
x
 66
are the attribute vectors of ?n? landowners who are already assigned to a group ?i?. The 
squared distance between any two observations can be estimated as  
)()(),(
1
1
11
2
xxVxxxx ???=
?
ti
d (Distance between the observation vectors and 
) 
x
1
x
Based on the squared distance defined above and the specified parameter ?k?, a 
positive integer, which denotes the number of nearest neighbors to be considered, ?k? 
observations that are closest to  are identified.  is assigned to the group that the 
majority of the ?k? nearest neighbors belong to. For example, in Figure 4.1 below, a new 
member ?X? will be classified as ?White? when k=1, ?Black? when k=5 and cannot be 
classified based on majority votes when k=10. 
x x
 
 
 
 
 
 
 
 
 
Figure 4.1 K-nearest neighbor classification based on k=1 or k=5 or k=10 
 67
 68
4.4 Results 
This study focuses on whether it is possible to predict group membership of 
family forest owners in the southeast using variables which are not collected using a 
survey of landowners, in other words, is it possible to assign a new landowner to a 
broader group based on management objectives ex-ante to acquiring primary data from 
the landowners. The summary statistics of the variables selected as discriminators (using 
step-wise selection SAS procedure STEPDISC) between the three landowner groups 
namely, multiple-objective, non-timber and timber according to their preference for 
producing either both, non-timber or timber is given is Table 4.1. The socio-economic 
(INC, DIST), demographic (PGI, POPDEN) and bio-physical characteristics (SLOPE, 
SITE, AGE and FT) describing the three groups of landowners can be used to classify a 
new (previously unclassified) landowner into one of the groups. KNN classification 
performance is evaluated using two accuracy measures generated as part of the output 
from running the SAS procedure DISCRIM. These are referred to as the apparent error 
rate and the cross-validation error rate. Percentage of correct classification within each 
group (cluster) and the whole population based on predictions of KNN classification are 
reported in Table 4.3. Since in this case the same input data is used to define and evaluate 
the classification criterion, the resulting error-count estimate has an optimistic bias (SAS 
2004, p. 1163). One way to reduce the bias is use the one-leave-one (Lachenbruch and 
Mickey 1968) option to classify each observation based on the discriminant function 
computed from all other observations (Cross-validation). Results are reported in Table 
4.4. In order to identify the misclassifications we constructed the confusion matrix for the 
 69
one-leave-one cross validation technique in Table 4.5. In a confusion matrix each row 
represents a true class and each column represent the predicted class and results show 
(Table 4.5) that while the average accuracy of prediction across all the landowner groups 
was 68%, it was 74.4%, 64.5% and 58.6% for multiple-objective, timber and non-timber 
owners respectively.  
Table 4.3 Classification results for the apparent-error-rate KNN method 
Percentage Correct 
k multiple-
objective 
non-timber timber Total 
2 86.7 78.4 81.7 83.6 
3 74.4 58.6 64.5 68.3 
4 68.9 45.9 53.4 59.8 
5 76.6 46.9 54.9 64.3 
6 77.2 44.9 52.9 63.6 
7 78.8 41.1 49.7 67.7 
8 73.1 38.4 47.7 58.7 
9 76.3 39.4 47.7 60.5 
 
Table 4.4 Classification results for the one-leave-one cross validation KNN method 
Percentage Correct 
k multiple-
objective 
non-timber timber Total 
2 74.5 58.6 64.5 68.3 
3 68.9 45.9 53.4 59.8 
4 76.6 46.9 54.9 64.3 
5 77.2 44.9 52.9 63.6 
6 78.8 41.1 49.7 67.7 
7 73.1 38.4 47.7 58.7 
8 76.3 39.4 47.7 60.5 
9 76.9 38.0 42.7 59.0 
 
 70
 
Table 4.5 Confusion Matrix of one-leave-one cross-validation method for k=2 
 multiple-
objective 
non-timber timber Other Total 
multiple-
objective 
569 70 120 6 765 
non-timber 77 171 43 1 292 
timber 118 41 296 4 459 
Total 764 282 459 11 1516 
 
Accuracy measures (Table 4.3 & Table 4.4) can be calculated in more than one 
way as advocated by Congalton (1991) who presented two methods as users accuracy 
and producers accuracy. users accuracy calculates correctly classed from the trace 
variable (diagonal elements of the confusion matrix (Table 4.5)) over the row total and 
provides indication of errors of case omission. Similarly producers accuracy is the 
calculation of correctly classed from the trace value over the column total. producers 
accuracy gives an indication of the accuracy of what the model was able to itself predict, 
whereas users accuracy relates how well the training data was discerned. Table 4.6 below 
presents the results of users accuracy and producers accuracy calculated from the leave-
one-out cross validation results (Table 4.5) of KNN classification (K=2). 
Table 4.6 Producers and Users accuracy for leave-one-out cross-validation method 
(k=2) 
 producers accuracy users accuracy 
Classed Column 
Total 
Accuracy % Classed Row 
Total 
Accuracy % 
multiple-
objective 
569 764 74.5 569 765 74.4
non-timber 171 282 60.6 171 292 58.6
timber 296 459 64.5 296 459 64.5
Figure 4.2 & Figure 4.3 below are graphical representations of the correct percent 
of observations (landowners) that the discriminant analysis predicted they would be a part 
of for different values of k for the apparent error rate and the cross-validation methods 
respectively. As mentioned before, k=2 is the point where the figures (4.2 & 4.3) peaked 
consistently for all the landowner groups (except for the multiple-objective group of 
owners in the cross-validation method (see Figure 4.3) where it peaked at k=6)) 
suggesting the choice of the free parameter k in the model. 
0
10
20
30
40
50
60
70
80
90
100
23456789
k
P
e
r
c
e
n
t co
r
r
ec
t
multiple-objective
non-timber
timber
Total
 
Figure 4.2 Graphical representation of apparent-error-rate accuracy of KNN by k 
 71
0
10
20
30
40
50
60
70
80
90
23456789
k
P
e
rc
e
n
t
 c
o
rr
e
c
t
multiple-objective
non-timber
timber
Total
 
Figure 4.3 Graphical representation of one-leave-one cross-validation accuracy of 
KNN by k 
4.5 Conclusion 
This study indicates that landowners can be accurately classified into 
heterogeneous attitudinal groups (multiple-objective, non-timber and timber) using the 
non-parametric KNN technique. The high accuracy rate of classification (84% in 
apparent error rate and 68% in cross-validation) of landowner groups indicates a higher 
percent of accuracy than would be expected if they were due to chance alone. In other 
words, this means that the high prediction accuracy is not merely due to random 
occurrences that have happened by chance. We found that bio-physical (SLOPE, SITE, 
FT and AGE), socio-economic (INC and DIST) and demographic (PGI and POPDEN) 
variables had a strong association with landowner group profiles. 
Examples of the use of discriminant analysis in landowner studies are sparse and 
have concentrated on searching for variables that discriminate owner groups, eg., timber 
 72
 73
managers vs non-managers (Greene and Blatner 1986), regenerators vs non-regenerators 
(Arano et al. 2004), cross-boundary cooperators vs non-cooperators (Finley et al. 2006). 
All of the above studies have neglected exploring the predictive component of 
discriminant analysis. Moreover these studies have resorted to differentiate landowner 
groups based on survey questions and so ex-ante prediction of non-surveyed landowners 
into one of the groups is beyond the scope of their work. Since all the variables used to 
develop the classification scheme in this study are known, a-priori, that is, before 
landowners on a FIA plot location are contacted for the NWOS, it may be possible to 
predict membership of a future landowner with known FIA and Census demographic 
attributes. This piece of information can also help design the NWOS survey to better 
focus on issues of interest to landowners and can help improve communications and 
development of effective outreach and educational programs. Given the increasing 
number of family forest owners and the increasing proportion of timberland they own and 
manage as an ownership class, this study can effectively help in estimating the different 
adjustment factors for diversely motivated landowner groups in order to more accurately 
project future timber supply. 
To our knowledge KNN classification has not been used to study landowner 
behavior though the technique is relatively simple to implement especially since there is 
no need to meet the statistical assumptions inherent in parametric classification methods 
(e.g., variables characterizing the difference between the landowner groups must have 
multivariate normal distributions and equal variance-covariance matrix). Our study also 
extends research on ways to predict landowner behavior in the future by taking into 
 74
consideration their diverse set of motivations and attitudes towards forest management 
instead of treating the family forest owners as a single homogeneous group. 
We feel that in this paper we have just touched the ?tip of the iceberg? and it is 
necessary to explore in detail the diverse motivations of family forest owners and ways to 
differentiate them to be able to understand them, make a connection and develop in 
tandem with their needs specific educational programs to insure future timber and non-
timber supplies from family forests at current or enhanced levels. 
 75
4.6 References 
Alerich, C.L., Klevgard, L., Liff, C. and P.D. Miles 2004. The Forest inventory 
and analysis database: Database description and users guide version 1.7. 
http://www.ncrs2.fs.fed.us/4801/fiadb/fiadb_documentation/FIADB_v17_122104.pdf 
(Accessed on 05-07-2006). 
Arano, K.G., I.A. Munn, J.E. Gunter, S.H. Bullard and M.L. Doolittle 2004. 
Comparison between regenerators and non-regenerators in Mississippi: A discriminant 
analysis. Southern Journal of Applied Forestry 28(4):189-195. 
Butler, B.J. and E.C. Leatherberry 2004. America?s family forest owners. J. For. 
102(7):4-9. 
Butler, B.J., E.C. Leatherberry and M.S. Williams 2005. Design, implementation, 
and analysis methods for the National Woodland Owner Survey. USDA For. Serv. Gen. 
Tech. Rep. NE-336. 
Congalton, R.G. 1991. A review of assessing the accuracy of classifications of 
remotely sensed data. Remote Sensing of Environment 37(1):35-46. 
Cover, T.M., and P.E. Hart. 1967. Nearest neighbor pattern classification. 
Transactions on Information Theory 13:21-27. 
Finley, A.O., D.B. Kittredge, T.H. Stevens, C.M. Schweik and D.C. Dennis 2006. 
Interest in cross-boundary cooperation: Identification of distinct type of forest owners. 
For. Sci. 52(1):10-23. 
 76
Fix, E. and J. Hodges 1951. Discriminatory analysis, nonparametric 
discrimination: consistency properties. Tech. Rep., Randolph Field, Texas: USAF School 
of Aviation Medicine. 
Gramann, J.H., T.D. Marty and W.B. Kurtz 1985. A logistic analysis of the 
effects of beliefs and past experience on management plans for nonindustrial private 
forests. J. Environ. Manage. 20:347-356. 
Greene, J.L. and K.A. Blatner 1986. Identifying woodland owner characteristics 
associated with timber management. For. Sci. 32(1):135-146. 
Kendra, A. and R.B. Hull 2005. Motivations and behaviors of new forest owners 
in Virginia. For. Sci. 51(2):142-154. 
Klecka, W.R. 1988. Discriminant analysis. Sage University Paper series on 
Quantitative Applications in the Social Sciences, series no. 07-019. Beverly Hills and 
London: Sage Publications. 
Kluender, R.A and T.L. Walkingstick 2000. Rethinking how nonindustrial 
landowners view their lands. Southern Journal of Applied Forestry 24(3):150-158. 
Kuuluvainen, J., H. Karpppinen and V. Ovaskainen. 1996. Landowner objectives 
and nonindustrial private timber supply. For. Sci. 42(3):300-309. 
Lachenbruch, P.A. and M.R. Mickey 1968. Estimation of error rates in 
discriminant analysis. Technometrics. 10:1-11. 
Majumdar, I., L.D. Teeter and B.J. Butler 2006. NIPFs in the Southeast: A diverse 
group with diversified objectives. Presented at the Southern Forest Economics Workshop 
meeting 23-24 March 2006, Knoxville TN. 
 77
Young, R.A. and M.R. Reichenbach 1987. Factors influencing the timber harvest 
intentions of nonindustrial private forest owners. For. Sci. 33(2):381-393.