ARTICLE IN PRESS

Journal of Forest Economics 14 (2008) 159176

www.elsevier.de/jfe

Analyzing determinants of forest owners decision-making

using a sample selection framework
Sta( le Strdala,, Gudbrand Liena,b, Sjur Baardsena,c
a

Eastern Norway Research Institute, P.O. Box 223, NO-2601 Lillehammer, Norway
Norwegian Agricultural Economics Research Institute, Oslo, Norway
c
Department of Ecology and Natural Resource Management,
( Norway
Norwegian University of Life Sciences, As,
b

Received 13 December 2006; accepted 19 July 2007

Abstract
A sample selection framework that simultaneously takes into account the two-step decisionmaking of forest owners (rst whether or not to harvest, second the level of harvesting) is
applied on representative cross-sectional data for forest properties and owners in Norway.
Forest management plans, property size, forested area and income from agriculture are found
to increase both the propensity to harvest and the harvesting levels. Income from engagement
in other outeld-related productions and debt burden increase the propensity to harvest only,
while increased age impact negatively on the harvesting decision. Wage income affects both
propensities to harvest and harvesting levels negatively. The results suggest that other onproperty productions may stimulate harvesting decisions, while off-property income impact
harvesting decisions and levels negatively.
r 2007 Elsevier GmbH. All rights reserved.
JEL classification: D13; D24; Q23
Keywords: Forest owner behavior; Forestry; Harvesting decision; Simultaneous equations;
Translog production function

Corresponding author.

E-mail address: Stale.Stordal@ostforsk.no (S. Strdal).

1104-6899/$ - see front matter r 2007 Elsevier GmbH. All rights reserved.
doi:10.1016/j.jfe.2007.07.001

ARTICLE IN PRESS
160

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

Introduction
Numerous studies have analyzed the harvesting behavior of nonindustrial private
forest owners (NIPFs) the last decades (see for reviews, Amacher et al., 2003;
Cubbage et al., 2003; Beach et al., 2005). A common interpretation is that since
NIPFs operate in incomplete markets, their subjective preferences and idiosyncratic
characteristics will alter the harvesting decision, commonly modeled within utilitymaximizing frameworks (e.g. Max and Lehman, 1988; Hyberg and Holthausen,
1989; Kuuluvainen et al., 1996). Prices and costs, though still important, will
therefore be only two out of many variables affecting the timber supply.
Many studies have concerned to what degree nonmarket timber (and nontimber)
amenities impact negatively on harvesting as they are assumed to be utilityincreasing for the forest owner (e.g. Pattanayak et al., 2002). In addition, previous
studies have also shown an impact from nonforest income on harvesting. As other
income sources on- and off-property have gained importance, possible negative
impacts of timber supply have been a major concern arising from the fact that forest
income has become more marginal to NIPFs household economies. Kuuluvainen
and Tahvonen (1999) found that nonforest income had a negative impact on
harvesting. They note that wealthy forest owners can afford more nancial loss in
order to enjoy nontimber benets than owners with low nontimber income. Other
empirical studies (e.g. Kuuluvainen et al., 1996) have found no signicant impact
from the owners exogenous income on harvesting.
There is probably a difference between the impacts on harvesting from offproperty income and income from other sources at the property. Off-property (wage)
income may be assumed not to conict timber production as they are not
interrelated, while on-property productions might. However, evidence from previous
studies of NIPF behavior suggests that harvesting decisions are inuenced both by
nontimber production on the properties (Kuuluvainen et al., 1996), and off-property
(wage) income (Lyland et al., 1995; Kuuluvainen and Tahvonen, 1999). Conway
et al. (2003) pointed on the negative impact on both harvesting and nontimber
activities from landowners who do not live on the property (absentee owners), and
explain this with absentee landowners possessing perhaps less information with
regard to harvesting than resident owners. This is also in accordance with the
ndings of Lyland et al. (1995). Conway et al. (2003) conclude that variables like
debt, nontimber activities and the owners harvesting behavior are related. Previous
studies have, however, not attempted to test the impact on harvesting from various
types of on- and off-property exogenous income.1
Most previous empirical studies on NIPFs harvesting behavior have applied
various versions of discrete choice models. However, in line with the decision
whether or not to harvest, also the harvest intensity depends on various property and
1

Bolkesj and Baardsen (2002) may be viewed as an exception. However, they found diverging results in
their different models as other income and agricultural income affected harvesting behavior negatively in
one model, while in another model they could not reject the null hypothesis that these variables had no or
positive impact. Further they did not distinguish between other income on-property and off-property.

ARTICLE IN PRESS
S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

161

owner characteristics (Vokoun et al., 2006), and both factors are likely to impact
timber supply. Wear and Parks (1994) state that the timber supply consists of two
parts: a timber production function (which describes how inputs are converted into
outputs) and the forest managers objective function, which is often guided by utility
rather than prot maximization. We posit that NIPF decision-making should be
modeled in two, probably simultaneous, steps; rst decide whether or not to harvest,
second decide upon the level of harvesting. Acknowledging that the analyst observes
only owners that harvest, many studies have applied censored regression (tobit)
models in order to analyze impact on harvesting levels from idiosyncratic
characteristics (e.g. Kuuluvainen and Tahvonen, 1999; Bolkesj and Solberg,
2003). In the present paper, it is attempted to explain why owners choose to harvest
in the rst place, and second what determines the level of harvesting. By doing this
one runs into the problem of truncated data. In the rst step a forest owner will
choose to harvest if the timber price, or utility from harvesting, exceeds a certain
threshold level (for example the reservation price, which again may be inuenced by
owner and property characteristics). In the second step, the amount of timber
supplied (to a given price) depends on both the owner-specic production function
and various owner and property attributes. Since the analyst is only able to observe
the actual production function of owners who are harvesting (and thus inferring that
the utility from harvesting exceeds a certain threshold value) the sample is truncated.
One way to cope with this, and which is chosen in this paper, is to apply a sample
selection framework. Thus, we account for the fact that the harvesting level depends
on the propensity to harvest, at the same time as we account for the joint distribution
of disturbances which may occur from sample selection bias.
In the present paper we address the questions: How does income from various
sources outside forestry affect the decision of whether or not to harvest, and second,
how does such income affect the level of harvesting once the forest owner has
decided to cut? Two equations are estimated simultaneously in a full information
maximum likelihood (FIML) sample selection framework: rst a selection (probit)
equation explains the decision to harvest and, second, a production function explains
the level of harvesting. The paper is organized as follows. In the next section, we
explain the materials used in the study and derive a simple model for forest owners
harvesting behavior in a sample selection framework and present the econometric
model. Then, we present the estimation results utilizing a large set of cross-sectional
data of Norwegian forest owners, before we discuss our ndings and give some
policy implications of the study.

Materials and methods

Model outline
Elements in the utility function of the forest owner are decisive whether or not to
harvest. However, when it comes to the determination of the level of harvesting, a

ARTICLE IN PRESS
162

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

production function approach that takes elements of owner and property characteristics into account must be employed. The modeling approaches to this decision
process are given in a large number of previous studies (e.g., see for reviews, Max
and Lehman, 1988; Hyberg and Holthausen, 1989; Kuuluvainen et al., 1996;
Prestemon and Wear, 2000; Beach et al., 2005). Most often the harvesting decision is
viewed as an intertemporal problem, i.e., the owner will harvest today if the marginal
utility is larger than the discounted marginal utility of harvesting tomorrow.
In this paper, we measure more specically the impact of having various sources of
on- and off-property exogenous income on timber supply using a sample selection
framework (Heckman, 1979; Greene, 2003). This approach assumes a joint normal
distribution between errors of the selection equation (harvest/no harvest) and the
regression equation of interest (the measure of harvesting level). Sample selection
frameworks have previously been used in studies of contracting in agriculture (e.g.
Bierlen et al., 1999; Key, 2005; Key and McBride, 2003) and in forestry by Kline
et al. (2004) in order to examine the likelihood of tree planting after harvesting.
In a rst step the nth forest owners decision whether or not to harvest can be
expressed with the latent variable
zn U 0n;T  1 r1 U 0n;T1 ,

(1)

where r is the (subjective) rate of interest. If the forest owner harvests, we assume
that the marginal utility to a forest owner from harvesting today (U 0T is larger than
the discounted marginal utility of harvesting tomorrow, (1+r)1 U 0T1 and the latent
variable is positive. However, we are not able to observe the magnitude of zn only its
sign. Therefore, in the rst step, we have to reformulate Eq. (1) into a probit
equation:
zn a0 wn mn ; where zn 1 if zn 40; 0 otherwise;
Probzn 1 Fa0 wn ; Probzn 0 1  Fa0 wn ;

(2)

where wn is a vector of property and owner characteristics, and F is the cdf of the
standard normal distribution. In the next step, the level of harvesting will depend on
available technology (combination of inputs which yields the most efcient output):
yn b0 xn c0 wn n ;

observed only if zn 1,

(3)

where xn is a vector of both exogenous and endogenous inputs in the chosen

harvesting technology. We assume that r Corrmn ; n , and thus the disturbances
(mn,en)bivariate normal [0,0,1,se,r].
Since the decision process is characterized by two steps the modeling approach
may favor a two-step procedure as the one used by Lyland et al. (1995). They
employed a twin linear probability model (Fisher, 1962; Cragg, 1971) by rst
estimating the propensity to harvest with a Probit-technique and next analyzing only
active forest owners by OLS. However, because the sample rule is that zn is observed
only if zn 40 we will face the problem of selection bias (since we can view the
problem as an omitted variable) and OLS would produce inconsistent estimates of
the parameters (Greene, 2003, p. 783). In order to overcome this bias Heckman
(1979) introduced a consistent estimator for the linear model (see also Greene, 1981),

ARTICLE IN PRESS
S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

163

which is a two-step procedure based on the fact that

Eyn jxn ; wn ; zn 1 b0 xn c0 wn En jzn 1 b0 xn c0 wn yM n ,

(4)

where y rs ; M n fa0 wn =Fa0 wn ; and f(  ) and F(  ) are the pdf and cdf,
respectively, of the standard normal distribution. In the rst step a in Eq. (2) is
estimated by probit followed by computation of Mn (i.e. the inverse Mills ratio) for
all observations where zn equals 1. In the second step (b,c,y) is estimated by linear
regression of yn on xn, wn and Mn. Heckmans model is consistent albeit not efcient
since the error term in Eq. (4) is heteroscedastic (Puhani, 2000; Greene, 2006).
Therefore, FIML of (a,b,c,r,s) can be employed based on the joint distribution of
observations (zn 0,wn) and (zn 1,wn, xn) where the Heckman estimates are used
as initial values. The FIML estimates are efcient compared to those obtained from
the Heckman procedure (Key and McBride, 2003).
Since some of the inputs may be endogenous, additional equations in a simultaneous equations framework have to be introduced (Amemiya, 1978). Thus, a system
of equations, involving yn and the endogenous inputs, is constructed:
yn b0 xn c0 wn yn ,
xn;i d0 xn;j by yn k0 wn xn ,
iaj.

The procedure is rst to estimate the selection equation, then to estimate each
equation in reduced form and keep predicted (tted) values, and estimate the
structural equations with tted values of endogenous input (and y) as instrument
variables. Eventually, original data are used to compute the estimate of the
disturbance variance (Lee et al., 1980).
Data and econometric model specification
Assuming technical efciency and that a translog production function is suitable
to approximate the harvesting technology, the level of harvesting of the nth forest
owner can be specied as
X
X
1XX
log yn b0
bi log xni
bij log xni log xnj
gk wnk n ,
2 i j
i
k
(6)
where bij bji , xni and xnj refer to the inputs (i,j labor, area cut, forest capital), wnk
are exogenous intercept shifters idiosyncratic to the nth forest owner, and en is an
IID0; s2 error term.
Forest capital input is assumed to be exogenous to the harvesting level. This
variable will in the short-run be unaffected by the level of harvesting. However, in
the long run the maximum sustainable yield (MSY) and then forest capital may be
affected negatively by over-harvesting. Labor and forest area cut were, on the
contrary, supposed to be endogenous to the harvesting decision. A higher level of
harvesting will yield both more area cut as well as it will demand more labor, cet. par.

ARTICLE IN PRESS
164

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

These variables were therefore estimated in a simultaneous equations framework

with Eq. (2) as follows:
X
X
1XX
log yn b0
bi log xni
bij log xni log xnj
gk wnk ny ,
2 i j
i
k
X
l1k wnk nx1 ,
log xn1 d01 d2 log xn2 dy1 log yn
k

log xn2 d02 d1 log xn1 d2y log yn

l2k wnk nx2 ,

i; j 1; 2; 3.

All inputs and the output of the production function were divided by their sample
means. The corresponding rst-order coefcients are thus estimates of the respective
elasticities evaluated at means.
We have applied the model to individual cross-section data of 8542 forest owners
extracted from the 2004 Sample Survey of Agriculture and Forestry. The crosssection includes data from a postal survey linked to the tax register, to the harvesting
register, and to a register of active farmers (administered by the Norwegian
Agricultural Authority), the latter including each farmers stocking and cropping
details. All data represent the year 2003, though information about harvesting levels
is given for both years 2003 and 2004.
The specication of the variables in the selection equation (2) and the
simultaneous equation production function system (7) is given in Table 1.
Using only complete observations produces biased results unless observations are
missing completely at random. There is also a loss of precision as the sample size
is reduced (Hair et al., 2006). A thorough discussion on data imputation techniques
in the forestry literature are given in, e.g., Reams and McCollum (1999). Even
though we wanted to make the procedure for imputing data as simple as possible,
different variables needed different imputation procedures given various property
characteristics.
The labor variable (x1) was calculated as the sum of hours worked by contractors
and hours worked by the owner, his family or hired labor in 2003, the latter
calculated as contractor equivalent hours, i.e., the amount of hours an average
contractor would have used to cut and haul the specic amount of timber. The data
set gave no information of hours worked by contractors so these were estimated as
the total costs in NOK involved with contractor work divided by 1925 NOK/h. The
latter is an estimate of costs per hour of contractors given as the sum of 1175 NOK/h
for harvester and 750 NOK/h for forwarder (Eid and Hoen, 2005).
Forest area cut (x2) expresses the area in hectares of various types of nal fellings
in 2003. Area cut is closely related to the quantity cut on each property. Therefore, in
the cases where information of area cut were not given (457 properties) we replaced
missing values by dividing the quantity cut for the property with mean area per m3
cut from the total sample.
The capital input (x3) is an estimate of the value of the increment from the forest
and is an expression of the choice faced by the forest owner as the increment can

ARTICLE IN PRESS
S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

165

Table 1. Variable specication

Selection equation variables

Production variables

Total area of the property (ha)

Share of productive forest relative to total area
Forest owners age (years) for NIPF owners
Income from other outeld-related productions (NOKa
1000)
Income from agriculture (NOKa 1000) for NIPF owners
Wage income (NOKa 1000) for NIPF owners
Debt (NOKa 1,000,000) for NIPF owners
Dummy for forest management
Dummy for location near urban areas
Dummy for education at bachelors level or higher

x1 (labor input in hours)

x2 (hectares forest area cut)
x3 (forest capital input in NOKa)

1 NOK 0.155 USD per 1 November 2006.

either be cut today and its revenue spent on alternative investments (or on
consumption), or harvesting may be postponed if this is the optimal choice of the
owner. The variable was calculated as the mean weighted price of various timber
grades sold in 2003 multiplied by the MSY of the propertys forest resources (which
is the amount of timber that might be cut without decreasing future harvesting
opportunities). Using weighted mean price also adjusts for the timber quality (and
thus variation in valuation) of the property, which may vary signicantly. For a few
properties there was no information of value from timber sales from the ofcial
harvesting register. In those cases we chose to replace mean weighted price with the
mean value of the sample. Information about MSY was available from only 3983
(46.6%) of the properties. Replacing with mean values is not a good approach as the
MSY is closely related to the size of the property. Consequently, small wood lots will
get too high and larger properties too low MSY when replacing missing values with
arithmetic means. We therefore chose to estimate an OLS model on the forest
properties where we had information on MSY, and then on the basis of coefcient
estimates, we calculated a hypothetical MSY for the missing data. The model used
was MSY f(Productive forest area, Fellings in 2004, Fellings in 2003), (R2 0.87).
Previous empirical studies of NIPF behavior have shown that owner and property
characteristics have to be taken into account when modeling harvesting behavior. In
this study, we are especially interested in how various income sources on- and offproperty (wage income off-property, income from agriculture, and income from
other outeld-related productions at the property) impact harvesting. However, also
other variables are likely to impact harvesting levels and should be included in the
model.
Human capital inuence on harvesting behavior is represented by the variables age
of the forest owner, dummy variable for education and dummy variable for forest
management plan. Age of the owner is included in most micro-data studies, and in
several analyses found to have a signicant negative impact on harvesting, although

ARTICLE IN PRESS
166

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

it has no a priori hypothesized effect (Kuuluvainen and Salo, 1991; Lyland et al.,
1995; Kuuluvainen et al., 1996; Kuuluvainen and Tahvonen, 1999). However, older
owners may have other bequest motives (Conway et al., 2003) or attitudes toward
debt (Fina et al., 2001) which may affect harvesting. Increasing age is also found
to decrease the owners technical efciency in timber production (Lien et al., 2007).
A dummy ( 1) for education is imposed for owners with Bachelors degree or
higher. Previous studies (e.g. Lyland et al., 1995) have found education to impact
positively on the harvesting decision. A dummy ( 1) is also imposed for properties
having a forest management plan. This is assumed to have a positive impact on the
harvesting behavior as it gives the owner larger information about the propertys
resources. A positive impact is also found in previous empirical studies (Lyland
et al., 1995; Kuuluvainen and Tahvonen, 1999; Lien et al., 2007).
Debt and location of the property are socioeconomic attributes that may inuence
the owners behavior. Debt is probably a good proxy for the owners subjective
interest rate as increasing debt induces owners to value money today more than
money tomorrow (because of imperfect capital markets), so that owners with large
debt have a higher subjective interest rate than owner without debt. This applies
directly to the intertemporal decision forest owners make (Eq. (1)), and thus forest
owners with a high debt burden are more likely to harvest. The effect of debt is
analytically derived in Fina et al. (2001) who also argue that owners with mature
timber are more sensitive to debt, though on the other hand the value of amenities
can work in the other direction. Previous empirical studies (e.g., Conway et al., 2003)
have found increasing debt level to impact positively on the harvesting decision. A
dummy ( 1) is imposed on properties located on or near urban areas (dened as
situated closer than 75 min of travel distance from a town larger than 50,000
inhabitants). Properties near urban areas are previously found to have different
harvesting behavior than properties located in more peripheral areas. Munn et al.
(2002) found for example that the probability of harvesting decreases with
decreasing distance to urban areas, while Strdal et al. (2004) found that remote
regions experience relatively higher harvesting levels in boom periods.
The property characteristics total area and share of productive forest relative to
total area are likely to impact the harvesting decision. The use of share in the latter
characteristic was done in order to avoid collinearity problems between the variables
during estimation. Hyberg and Holthausen (1989) found in an analysis of
landowners in Georgia, USA that both the size of the landowners property and
the amount of forested area were positively related to timber harvesting. A positive
relationship between the size of forestland and harvesting is also found by
Kuuluvainen and Tahvonen (1999) and Munn et al. (2002). Vokoun et al. (2006)
nd, however, that owners with larger forested parcels are likely to harvest relatively
less intensive when they enter the timber market.
Some of the owners lacked a personal number. These are most likely to be
industrial or public forest owners. For these owners there was no information about
age (explaining the missing value for this variable), wage income, income from
agriculture and debt. Information was, however, available for income from other
outeld-related productions at the property. In order to overcome the problem of

ARTICLE IN PRESS
S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

167

missing values for these owners we dened a dummy ( 1) for owners with a
personal number (i.e. NIPF owners) and interacted this with the four variables
discussed above. The variables are therefore slope shifters and their coefcients show
the impact on harvesting for these variables for NIPF owners only.
Regarding the variables age, debt, wage income and agricultural income we were
missing observations for those properties owned by industrial private or public.
Prior to estimation these variables were interacted with a NIPF-dummy so that they
have the value 0 during estimation.
As in several previous studies (e.g. Goodwin and Mishra, 2004; Lien et al., 2006;
Nehring et al., 2006) each observation is imposed a replication weight. Hundred and
forty-four strata were dened as each of the 18 counties was divided into 8 area
classes.2 Different samples were drawn from the population of forest owners,
ranging from 2.5% in the smallest area class to 100% in the largest ones. The
resulting gross sample yielded a total of 9136 forest owners (of which 8542 was the
net sample). On the basis of the share of population drawn in each stratum a
replication weight was calculated for each property. The replication weight, for each
stratum, takes into account the ratio between the number of properties in the total
population and the sample, the productive forest area of each property in the sample,
the sample and population mean forest area, and the variation of property size in the
sample, according to the following equation:
!
h
Ph
an;h  a h ah  A
rn;h
1
,
(8)
P
ph
1=ph n2h an;h  a h 2
where rn,h is the replication weight of property n in stratum h, Ph the total population
of properties in stratum h, ph the number of properties in the sample in stratum h,
h the arithmetic mean
an,h the productive forest area for property n in stratum h, A
productive forest area for the total population of properties in stratum h, and a h
arithmetic mean productive forest area for the sample of properties in stratum h.
Thus, in the largest area classes the data were imposed a replication weight equal
to one as all properties were included into the sample, while in smaller area classes
the replication weight was larger than unity.3 The sum of replication weights was
equal to the total population of 118,512 forest properties in Norway with a
productive forest area larger than 2.5 ha. A closer description of the data and the
sampling procedure is presented in Statistics Norway (2004).
Table 2 gives some descriptive statistics of the data. The sample with replication
weight imposed is supposed to approximate the total population of forest properties
in Norway and their characteristics. Before imposed a replication weight the sample
2
The area classes in hectares are: 2.59.9; 10.024.9; 25.049.9; 50.099.9; 100.0199.9; 200.0499.9;
500.01999.9; 2000 above.
3
The replication weight ranges from 0.3515 to 67.4974. The reason why some weights are less than unity
is that some 150 properties in the sample had the same owner and were located in the same municipality.
In order to get a better approximation the merged properties were given a replication weight equal to
calculated weight divided by the number of merged properties. Thus, a large property from a merger of
two properties might be given a weight equal to 1/2 0.5.

ARTICLE IN PRESS
168

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

Table 2. Descriptive statistics of variables used

Mean
of
sample
Harvesting, m3
Labor input in hoursa
Hectares forest area
cuta
Forest capital input
(NOK 1000)a,b
Total area of the
property, ha
Share of productive
forest relative to total
area
Forest owners age
(years)
Income from other
outeld-related
productions (NOKb
1000)
Income from
agriculture (NOKb
1000)
Wage income (NOKb
1000)
Debt (NOKb
1,000,000) for NIPF
owners
Dummy for forest
management
Dummy for location
near urban areas
Dummy for education
at bachelors level or
higher

Mean of
sample after
replication

Minimum
replicated
sample

Maximum
replicated
sample

Standard
deviation
replicated
sample

379.3
58.0
6.0

60.5
21.9
2.5

0
0.13
0.01

46,068
2632
229.0

496.5
64.2
7.0

313.7

105.2

1.48

16,186

359.8

160

2.5

970,406

1226
0.74

0.76

0.7  E5

1.0

4808
0.29

50.6

54.4

96

15.6

33.0

5.4

11,814

95.6

47.6

42.2

2488

104.7

250.4

256.6

3999

269.3

0.92

0.61

132.0

1.38

0.49

0.29

1.0

0.46

0.37

0.42

1.0

0.49

0.21

0.17

1.0

0.38

Note: Sample size 8542/118,512 before/after replication.

a
Figures for owners who conducted harvesting.
b
1 NOK 0.155 USD per 1 November 2006.

shows a bias towards larger properties which harvest more, owners that are
somewhat younger, have more income from other outeld-related productions and
who have slightly lower wage incomes. The sample has also, on average, a higher
debt burden, is more likely to have a forest management plan, less frequently located
near urban areas and is more likely to have higher education.

ARTICLE IN PRESS
S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

169

One limitation of the study is worth mention. In their model of harvesting

behavior, Conway et al. (2003) corrected for unknown types of heteroscedastistiy
using Whites method. Even if the FIML approach gives efcient results compared
to Heckmans model, estimation results may be awed by heteroscedasticity. Since
we are using a simultaneous equation approach, an algorithm adjusting for heteroscedasticity is, to the best of our knowledge, not readily available. The effect
of heteroscedastisity is, however, reduced to a certain degree as the translog specication uses logarithmic transposed data.

Estimation results
The sample selection modeling framework is applied on both unweighted data (i.e.
the original sample), and on data imposed a replication weight in order to
approximate the total population of forest properties in Norway.
The estimates of the translog production function in the simultaneous equations
framework are shown for unweighted and weighted data in Tables 3 and 4,
respectively. The coefcients to the left refer to the (rst step) probit selection
equations, i.e., the owners propensity to harvest, while the coefcients to the right
refer to the production function. Model tests show a signicant correlation between
the selection function and the production function. Also chi-squared statistics for
regressions were tested and found signicant for both models.
For most variables the sign and magnitude are approximately the same in the
unweighted regression compared with the one imposed a replication weight. The
exception is that the dummies for location are found negatively and education
positively affecting the harvesting level in the sample imposed a replication
weight.
Variables affecting the decision whether or not to harvest
The rst step, i.e. the selection function, contains the determination variables
affecting the decision of whether or not to harvest. The dependent variable equaled
one if the forest owners harvested in 2003, and zero otherwise.
The variables forest management plan, total area, share of productive forest area
relative to total area, income from other outeld-related productions, NIPF
agricultural income, debt level, and education on Bachelors level or higher all
increased the propensity to harvest. NIPF age and NIPF wage income decreased, on
the other side, the propensity. Location of the property close to urban areas was not
found to have any signicant impact in neither of the models.
Variables affecting the level of harvesting
In the second step of the procedure, we attempted to determine variables that
affected the level of harvesting. In doing so a translog functional form specication

170

Table 3. FIML estimates of selection and production functions, unweighted

Selection function

Production function

Constant

Constant

LX1

0.887*** (0.007)

1.5  E5*** (3.7  E6) LX2

0.098*** (0.006)

0.384*** (0.051)

0.059*** (0.005)

0.006*** (0.001)

LX3

LX11

0.049*** (0.006)
0.039*** (0.004)

Outeld
income

0.001*** (1.4  E4)

LX12

NIPF
agricultural
income
NIFP wage
NIPF debt
Location
Education

2.2  E4* (1.3  E4)

LX13

2.4  E4*** (5.3  E5) LX22

0.044*** (0.009)
LX23
0.001 (0.028)
LX33
0.160*** (0.036)

0.004 (0.005)

0.039*** (0.006)
0.012*** (0.005)
0.001 (0.005)

Forest
management
plan
Total area

0.066*** (0.013)

Share of
productive
forest
NIPF age

0.146*** (0.015)

Outeld
income
NIPF
agricultural
Income
NIFP wage

2.1  E6 (8.3  E6)

NIPF debt
Location
Education
s(1)
r(1,2)
F test w220;3228
Log-L

0.002** (0.001)
0.015** (0.007)
0.007 (0.009)
0.194*** (0.005)
0.398*** (0.087)
10,008.4***
474.4

1.5  E7 (1.4  E7)

1.7  E4* (2.0  E4)

8.6  E5*** (2.6  E5)

3.0  E5*** (0.1  E4)

Note: ***Signicant at 1%, ** at 5% and * at 10% level of signicance; number of observations: 8542 in selection equation and 3249 in production function.

ARTICLE IN PRESS

Share of
productive
forest
NIPF age

0.725*** (0.027)

0.231*** (0.027)

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

Forest
management
plan
Total area

0.713*** (0.066)

Table 4. FIML estimates of the selection function and the production function, weighted with replication weight
Selection function

Production function

Constant

Constant

LX1

0.887*** (0.010)

4.3  E5*** (8.5  E6) LX2

0.118*** (0.009)

0.246*** (0.035)

0.043*** (0.008)

0.007*** (0.001)

LX3

LX11

0.043*** (0.006)

Outeld
income

0.001*** (1.6  E4)

LX12

0.045*** (0.005)

NIPF
agricultural
income
NIFP wage
NIPF debt
Location
Education

0.001*** (9.4  E5)

LX13

0.017*** (0.006)

7.5  E5** (3.8  E5)

0.019** (0.009)
0.011 (0.019)
0.121*** (0.025)

LX22
LX23
LX33

0.043*** (0.008)
0.017*** (0.005)
0.001 (0.006)

Forest
management
plan
Total area

0.066*** (0.018)

Share of
productive
forest
NIPF age

0.107*** (0.015)

Outeld
income
NIPF
agricultural
Income
NIFP wage

2.7  E6 (1.6  E5)

NIPF debt
Location
Education
s(1)
r(1,2)
F test w220;3228
Log-L

0.001 (0.001)
0.012* (0.007)
0.016* (0.009)
0.180*** (0.006
0.268* (0.154)
10,942.7***
759.3

3.5  E8 (2.7  E7)

1.3  E3 (2.6  E3)

8.4  E5*** (3.0  E5)

2.1  E5* (1.2  E5)

171

Note: ***Signicant at 1%, ** at 5% and * at 10% level of signicance; number of observations: 8542 in selection equation and 3249 in production function.

ARTICLE IN PRESS

Share of
productive
forest
NIPF age

0.721*** (0.019)

0.179*** (0.047)

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

Forest
management
plan
Total area

1.134*** (0.049)

ARTICLE IN PRESS
172

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

was chosen to approximate the harvesting technology where (the logarithm of)
harvesting level in cubic meters formed the dependent variable.
In the production function labor showed the largest elasticity: 0.89 in both models;
area 0.10 in the unweighted model and 0.12 when a replication weight was imposed;
and capital 0.06 and 0.04 in the unweighted and weighted model, respectively. Once
possessing a forest management plan, the share of productive forest area and NIPF
agricultural income were both found to work positively, while NIPF wage income
and location near urban areas worked negatively, affecting the level of harvesting at
10% level in both models. NIPF debt level was found to affect the level of harvesting
positively in the unweighted model, while it was insignicant in the model imposed
a replication weight. Education level was found to affect the harvesting level signicantly positive in the weighted model, while it was insignicant in the unweighted
model.
Neither total area, NIPF age, nor income from other outeld-related productions
in the production model (step two) were found to affect the harvesting level
signicantly in any of the two models.

Discussion and policy implications

In this study, we applied a two-step sample selection estimation procedure which
simultaneously took into account the two-step decision-making of forest owners;
rst to decide whether or not to harvest and second to decide on the level of
harvesting. Since previous studies have provided knowledge of the inseparability
between production and consumption in NIPFs decision-making (Kuuluvainen
et al., 1996), idiosyncratic elements relevant for the harvesting decision were
incorporated both in the selection and in the production model. The results seem to
be consistent both for owners with relatively large properties (the unweighted
sample) and with the representative sample imposed a replication weight where the
properties are smaller on average. Dummies for education level and location near
urban areas are, however, found signicant at the 10% level for the weighted sample
indicating that the choice of harvesting levels of owners with larger properties
(unweighted sample) is not so affected by his or her education level and location as
the average forest owner.
The selection model showed similar results to previous studies as variables NIPF
age and NIPF wage income were found to affect the selection process as well as the
production level negatively. The negative impact from age is found in almost all
previous studies (exceptions are e.g. Bolkesj and Baardsen (2002) who found age to
be ambiguous to harvesting behavior), though not very much discussed. Different
bequest motives (Conway et al., 2003) compared to younger owners can be one
explanation for this result; another is that older owners may be semi-retired; further
is that younger owners also may have larger debt or facing large investments on the
property. Thus, increased harvesting may give these owners better liquidity. As debt
may also be viewed as a proxy for the owners subjective interest rate, owners with a

ARTICLE IN PRESS
S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

173

high debt burden may value harvesting today higher than owners with a lower debt
burden. This applies directly to Eq. (1), and thus debt will increase the propensity to
harvest, cet. par. The positive impact of debt is supported by the results and is also
found in previous studies (e.g., Conway et al., 2003; Kuuluvainen and Tahvonen,
1999).
Education was found to positively impact the harvesting decision, though more
weakly when it comes to the level of harvesting since this effect is only signicant for
the sample imposed a replication weight. Thus, the effect of education is not so clear
for larger properties. Except Lyland et al. (1995) the owners level of education has
in most previous studies of harvesting behavior not been used as an explanatory
variable. Still it is commonly used in analyses of agricultural production (e.g. Ahituv
and Kimhi, 2006; Lien et al., 2006). However, Aronsson and Carlen (2000) found
education level to be one important determinant of forest land prices. Moreover,
Lien et al. (2007) found education to impact positively on the technical efciency in
timber production. Reasons why education is positive regarding the owners
harvesting behavior may be that educated owners also are more informed about the
propertys resources or that they also might have more debt.
Location of the property near urban areas has previously been found as a variable
that reduces the level of harvesting (e.g. Munn et al., 2002). This is supported in this
study as the dummy variable for location near urban areas are negative in the
production model, i.e. properties located nearer than 75 min of travel distance from a
town with more than 50,000 inhabitants (commuting distance) have a lower
harvesting level than other properties. However, regarding the harvesting decision
no impact from location is found. One explanation is that properties located near
urban areas to a larger extent may adapt shelter-wood cutting rules rather than clear
cutting in order to reduce conicts between harvesting and multiple-use forestry.
This is not explicitly analyzed in this study and might be a line for further research.
Size is decisive to whether an owner chooses to harvest or not, though not
necessarily regarding the level of harvesting. However, in line with previous studies
(Hyberg and Holthausen, 1989; Kuuluvainen and Salo, 1991; Kuuluvainen and
Tahvonen, 1999), a larger degree of forested area relative to the propertys total area
is positive regarding the level of harvesting. Moreover, and a result also found in
previous studies (Kuuluvainen and Tahvonen, 1999; Lyland et al., 1995), forest
management plans are positive both for the harvesting decision and the choice of
harvesting level. Note that owners may have chosen a management plan for activityrelated reasons (e.g. property size), introducing some self-selection bias to the model,
but it might also be that information of the propertys resources stimulates owners
to harvest.
In this paper, focus has mainly been put on the impact of different types of onand off-property income, which effects have shown to be ambiguous when
comparing previous studies. One proposition might be that since exogenous income
could be used to satisfy the owners debt commitment, this will work in direction of
increasing harvesting (Fina et al., 2001). On the other hand, increased wealth (due to
increased exogenous income) will make owners put more weight on the utilityincreasing effect from for example amenities, and thus reducing harvesting levels and

ARTICLE IN PRESS
174

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

frequencies (Kuuluvainen and Tahvonen, 1999). Our results show that it is not
arbitrary, for the owners harvesting behavior, whether other income is earned offproperty or on-property. There seem also to be differences between impacts from
different types of on-property income sources. One clear tendency is that offproperty income (typically wage income) seems to impact negatively on the
harvesting decision as well as on the harvesting level. A conclusion drawn from this
is that wage-earners might adopt the same utility-maximizing rule discussed by
Kuuluvainen and Tahvonen (1999), namely that they can afford more nancial loss
in order to gain utility from nontimber benets. If this is true, the same effect is not
the case when the owners get a high level of agricultural income, neither to a certain
extent when it comes to income from other outeld-related productions. Both
income sources are found positive for the owners decision whether and not to
harvest, and increased agricultural income also impact positively on the level of
harvesting. Thus, it may not be the level of exogenous income which counts, but
rather differences in the utility functions between owners with high wage income and
owners where most income is found on-property. One effect not accounted for in this
study, but found to impact harvesting negatively, is whether or not the owner lives
on the property (e.g., Conway et al., 2003). Some of the owners with a high level of
wage owners probably live outside the property. Conway et al. (2003) argue that
absentee owners may have less information about harvesting than residence owners
and that such owners are more concerned about for example recreational values of
their property rather than (small) revenues from harvesting. As a byside, the possible
effect of information may also explain the positive impact from forest management
plans. Owners with a plan have larger information about the properties resources,
which might be positive regarding their harvesting behavior.
From the discussion above it seems that increased dependency on the property as
an income generator leads to a greater utilization of the propertys resources. Some
differences are however found between income from other outeld-related
productions and agricultural income since the latter also impact positively on the
level of harvesting. This may be explained by the fact that agricultural production
will to a minor degree conict with forestry, while productions such as cottage
renting and tourism might. Thus, in such productions owners may adopt other types
of harvesting (for example, a larger degree of shelter-wood harvesting). This is also
discussed in Lien et al. (2007) who nd that other outeld-related productions seem
to impact negatively on the owners technical efciency. The owners seek to
maximize the utility from a portfolio of activities, where forestry is one out of several
others, and thus conicting productions have to be weighted up against each other.
One important policy implication is found in our study. The Norwegian
agricultural authorities have over the last few years stimulated growth of other
outeld-related productions on the properties. Our ndings support this agricultural
policy also from a forest policy point of view. Forestry seems go hand in hand with
other activities on the property. Forest activities are more promoted by forest owners
who engage in outeld-related business than by those who also engage in offproperty work. Besides, in a forestry context it could be argued that the agricultural
authorities should encourage rather than work against the structural development.

ARTICLE IN PRESS
S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

175

The larger the property, the more intensive are the forestry activities and the
activities in other outeld-related businesses.
Acknowledgments
This study is a part of the project Private forest owners decision making under
multiple objectives and productions, nanced by the Research Council of Norway,
Grant 166455/I10.
References
Ahituv, A., Kimhi, A., 2006. Simultaneous estimation of work choices and the level of farm activity using
panel data. European Review of Agricultural Economics 33 (1), 4971.
Amacher, G.S., Conway, C., Sullivan, J., 2003. Econometric analyses of nonindustrial forest landowners:
is there anything left to study? Journal of Forest Economics 9 (2), 137164.
Amemiya, T., 1978. The estimation of a simultaneous equation generalized probit model. Econometrica 46
(5), 11931205.
Aronsson, T., Carlen, O., 2000. The determinants of forest land prices: an empirical analysis. Canadian
Journal of Forest Research 30 (4), 589595.
Beach, R.H., Pattanayak, S.K., Yang, J.-C., Murray, B.C., Abt, R.C., 2005. Econometric studies of nonindustrial private forest management: a review and synthesis. Forest Policy and Economics 7 (3),
261281.
Bierlen, R., Parsch, L.D., Dixon, B.L., 1999. How cropland contract type and term decision are made:
evidence from an Arkansas tenant survey. International Food and Agribusiness Management Review 2
(1), 103121.
Bolkesj, T., Baardsen, S., 2002. Roundwood supply in Norway: micro-level analysis of self-employed
forest owners. Forest Policy and Economics 4 (1), 5564.
Bolkesj, T., Solberg, B., 2003. A panel data analysis of nonindustrial private roundwood supply with
emphasis on the price elasticity. Forest Science 49 (4), 530538.
Conway, M.C., Amacher, G.S., Sullivan, J., 2003. Decisions nonindustrial forest landowners make: an
empirical examination. Journal of Forest Economics 9 (3), 181203.
Cragg, J.G., 1971. Some statistical models for limited dependent variables with application to the demand
for durable goods. Econometrica 39 (5), 829844.
Cubbage, F., Snider, A., Abt, K.L., Moulton, R., 2003. Nonindustrial private forests. In: Sills, E.O., Abt,
K.L. (Eds.), Forests in a Market Economy. Kluwer Academic Publishers, Dordtrecth, The
Netherlands, pp. 2338.
Eid, T., Hoen, H.F., 2005. Strategi ved valg av foryngelsesmetode noen forhold som p(avirker
lnnsomhet ved skjermstillingshogst [Policy for choice of regeneration method some factors affecting
protability of shelter wood cutting]. INA Fagrapport 5. Department of Ecology and Natural
Resource Management, Norwegian University of Life Sciences (in Norwegian).
Fina, M., Amacher, G.S., Sullivan, J., 2001. Uncertainty, debt, and forest harvesting: Faustmann
revisited. Forest Science 47 (2), 188196.
Fisher, J.A., 1962. An analysis of consumer goods expenditures in 1957. Review of Economics and
Statistics 44 (1), 6471.
Goodwin, B.K., Mishra, A.K., 2004. Farming efciency and the determinants of multiple job holding by
farm operators (Proceedings from the ASSA Winter Meetings, San Diego, CA, January 2004).
American Journal of Agricultural Economics 86 (3), 722729.
Greene, W.H., 1981. Sample selection bias as a specication error: comment. Econometrica 49 (3),
795798.
Greene, W.H., 2003. Econometric Analysis, fth ed. Prentice-Hall, Upper Saddle River, NJ.
Greene, W., 2006. Censored data and truncated distributions. In: Mills, T., Patterson, K. (Eds.), Palgrave
Handbook of Econometrics, Vol. 1: Theoretical Econometrics. Palgrave-Macmillan, London,
pp. 695736 (Chapter 19).
Hair Jr., J.F., Black, W.C., Babin, B.J., Anderson, R.E., Tatham, R.L., 2006. Multivariate Data Analysis,
sixth ed. Prentice-Hall, Upper Saddle River, NJ.

ARTICLE IN PRESS
176

S. Strdal et al. / Journal of Forest Economics 14 (2008) 159176

Heckman, J.J., 1979. Sample selection bias as a specication error. Econometrica 47 (1), 153161.
Hyberg, B.T., Holthausen, D.M., 1989. The behavior of nonindustrial private landowners. Canadian
Journal of Forest Research 19 (8), 10141023.
Key, N., 2005. How much do farmers value their independence? Agricultural Economics 33 (1), 117123.
Key, N., McBride, W., 2003. Production contracts and productivity in the US hog sector. American
Journal of Agricultural Economics 85 (1), 121133.
Kline, J.D., Azuma, D.L., Alig, R.J., 2004. Population growth, urban expansion, and private forestry in
Western Oregon. Forest Science 50 (1), 3343.
Kuuluvainen, J., Salo, J., 1991. Timber supply and life cycle harvest of nonindustrial private forest owners:
an empirical analysis of the nnish case. Forest Science 37 (4), 10111029.
Kuuluvainen, J., Tahvonen, O., 1999. Testing the forest rotation model: evidence from panel data. Forest
Science 45 (4), 539549.
Kuuluvainen, J., Karppinen, H., Ovaskainen, V., 1996. Landowner objectives and nonindustrial private
timber supply. Forest Science 42 (3), 300309.
Lee, L.-F., Maddala, G.S., Trost, R.P., 1980. Asymptotic covariance matrices of two-stage probit and
two-stage tobit methods for simultaneous equations models with selectivity. Econometrica 48 (2),
491503.
Lien, G., Flaten, O., Jervell, A.M., Ebbesvik, M., Koesling, M., Valle, P.S., 2006. Management and risk
characteristics of part-time and full-time farmers in Norway. Review of Agricultural Economics 28 (1),
111131.
Lien, G., Strdal, S., Baardsen, S., 2007. Technical efciency in timber production and effects of other
income sources. Small-scale Forestry 6 (1), 12.
Lyland, K., Ringstad, V., y, H., 1995. Determinants of forest activities a study of private
nonindustrial forestry in Norway. Journal of Forest Economics 1 (2), 219237.
Max, W., Lehman, D.E., 1988. A behavioral model of timber supply. Journal of Environmental
Economics and Management 15 (1), 7186.
Munn, I.A., Barlow, S.A., Evans, D.L., Cleaves, D., 2002. Urbanizations impact on timber harvesting in
the south central United States. Journal of Environmental Management 64 (1), 6576.
Nehring, R., Barnard, C., Banker, D., Breneman, V., 2006. Urban inuence in costs of production in the
corn belt. American Journal of Agricultural Economics 88 (4), 930946.
Pattanayak, S.K., Murray, B.C., Abt, R.L., 2002. How joint is joint forest production? An econometric
analysis of timber supply conditional on endogenous amenity values. Forest Science 48 (3), 479491.
Prestemon, J.P., Wear, D.N., 2000. Linking harvest choices to timber supply. Forest Science 46 (3),
377389.
Puhani, P.A., 2000. The Heckman correction for sample selection and its critique. Journal of Economic
Surveys 14 (1), 5368.
Reams, G., McCollum, J., 1999. Evaluating multiple imputation models for the Southern annual forest
inventory. Proceedings of the Section on Statistics and the Environment of the American Statistical
Association, 1318.
Statistics Norway, 2004. About the statistics. Sample Survey of Agriculture and Forestry 2004. Available
from /http://www.ssb.no/english/subjects/10/04/20/skogbruk_en/arkiv/S (accessed 25 November
2006).
Strdal, S., Lein, K., rbeck, M., Hagen, S.E., 2004. Regional differences in harvesting levels and woodbased employment in Norway. Small-scale Forest Economics, Management and Policy 3 (1), 3547.
Vokoun, M., Amacher, G.S., Wear, D.N., 2006. Scale of harvesting by non-industrial private forest
landowners. Journal of Forest Economics 11 (4), 223244.
Wear, D.N., Parks, P.J., 1994. The economics of timber supply: an analytical synthesis of modeling
approaches. Natural Resource Modeling 8 (3), 199223.