Journal : FORSCI
Article Doi : 10.1093/forsci/fxz058
Article Title : Predictive Ability of Mixed-Eects Height–Diameter Models Fit Using One Species but Calibrated for Another Species
First Author : CurtisL. VanderSchaaf
Corr. Author : CurtisL. VanderSchaaf
We encourage you to use Adobe’s editing tools (please see the next page for instructions). If this is not possible, please list responses clearly in an e-mail, using line num-
bers. Please do not send corrections as track changed Word documents.
Changes should be corrections of typographical errors only. Changes that contradict journal style will not be made.
ese proofs are for checking purposes only. ey should not be considered as nal publication format. e proof must not be used for any other purpose. In particular
we request that you: do not post them on your personal/institutional web site, and do not print and distribute multiple copies (please use the attached oprint order
form). Neither excerpts nor all of the article should be included in other publications written or edited by yourself until the nal version has been published and the full
citation details are available. You will be sent these when the article is published.
1. Permissions: Permission to reproduce any third party material in your paper should have been obtained prior to acceptance. If your paper contains gures or
text that require permission to reproduce, please conrm that you have obtained all relevant permissions and that the correct permission text has been used
as required by the copyright holders. Please contact email@example.com if you have any questions regarding permissions.
2. Author groups: Please check that all names have been spelled correctly and appear in the correct order. Please also check that all initials are present. Please
check that the author surnames (family name) have been correctly identied by a pink background. If this is incorrect, please identify the full surname of the
relevant authors. Occasionally, the distinction between surnames and forenames can be ambiguous, and this is to ensure that the authors’ full surnames and
forenames are tagged correctly, for accurate indexing online. Please also check all author aliations.
3. Figures: Figures have been placed as close as possible to their rst citation. Please check that they are complete and that the correct gure legend is present.
Figures in the proof are low resolution versions that will be replaced with high resolution versions when the journal is printed.
4. Missing elements: Please check that the text is complete and that all gures, tables and their legends are included.
5. URLs: Please check that all web addresses cited in the text, footnotes and reference list are up-to-date, and please provide a ‘last accessed’ date for each URL.
6. Funding: Funding statement, detailing any funding received. Remember that any funding used while completing this work should be highlighted in a sepa-
rate Funding section. Please ensure that you use the full ocial name of the funding body, and if your paper has received funding from any institution, such
as NIH, please inform us of the grant number to go into the funding section. We use the institution names to tag NIH-funded articles so they are deposited
at PMC. If we already have this information, we will have tagged it and it will appear as coloured text in the funding paragraph. Please check the information
All commenting tools are displayed in the toolbar.
To edit your document, use the highlighter, sticky notes, and the
variety of insert/replace text options.
DO NOT OVERWRITE TEXT, USE COMMENTING TOOLS ONLY.
AUTHOR QUERY FORM
Journal : FORSCI
Article Doi : 10.1093/forsci/fxz058
Article Title : Predictive Ability of Mixed-Eects Height–Diameter Models Fit Using One Species but Calibrated for Another
First Author : CurtisL. VanderSchaaf
Corr. Author : CurtisL. VanderSchaaf
AUTHOR QUERIES - TO BE ANSWERED BY THE CORRESPONDING AUTHOR
Please ensure that all queries are answered, as otherwise publication will be delayed and we will be unable to complete the next stage of
the production process.
Please note that proofs will not be sent back for further editing.
e following queries have arisen during the typesetting of your manuscript. Please click on each query number and respond by
indicating the change required within the text of the article. If no change is needed please add a note saying “No change.”
AQ1 General comment: Alarge number of changes have been made to improve the English. Please check carefully throughout.
AQ2 Please give “HTs” and “HTCD” in full
AQ3 Please give “HTs” and “HTCD” in full
AQ4 What does the “T” in equation 3 denote?
AQ5 General comment: For all displayed and in-line equations, please ensure that variables and parameters have been correctly
formatted throughout for consistency–– roman, bold or italic as appropriate, according to standard mathematical style (a number
of changes have been made).
AQ6 “however it is suggested, for this particular topic of determining if the ability of calibrating mixed-effects models across a variety of
species improves model- fitting ability, is a very important research topic for this particular subject of study” does not make sense.
Please check and reword.
AQ7 Please provide publisher details for the reference ‘Bechtold, 2005’, ‘Greenhill, 1881’.
AQ8 Please provide accessed date for the reference ‘O’Connell, 2017’.
AQ9 The resolution of figures 1, 2, and 3 are lower in quality than usually printed. Please provide high resolution images for better
Forest Science • XXXX 2019 1
For. Sci. XX(XX):1–11
Predictive Ability of Mixed-Effects Height–Diameter
Models Fit Using One Species but Calibrated for
Mixed-effects individual tree height–diameter models are presented for important pines in the Western Gulf, USA. Equations are presented for plantations of loblolly (Pinus
taeda L.), longleaf (Pinus palustris P.Mill.), shortleaf (Pinus echinata Mill.), and slash (Pinus elliottii Engelm.) pine. To produce localized individual tree height estimates, these
models can be calibrated after obtaining height–diameter measurements from a plot/stand of interest. These equations can help answer an interesting question of whether
a model ﬁt for one species can be calibrated to produce reasonable height estimates of another species. In situations where mixed-effects models have not been developed
for a particular species, perhaps an equation from another species can be used. This question was addressed by calibrating these models using independent data of loblolly,
longleaf, and slash pine plantations located in South Carolina. For each calibration species, in addition to the models developed described above, previously published models,
but of the same model form, ﬁt using other species from across the USA were examined.
Results show that models of a variety of species can be calibrated to provide reasonable predictions for a particular species. Predictions using this particular model form indicate
that model calibration is more important than species-speciﬁc height–diameter relations.
Keywords: Pinus, pine, ﬁr, growth and yield
H–D models are an integral component of forest inventories
used to reduce sampling times. Mixed-eects H–D models
have been developed for many species (e.g., Lappi 1991,
Lynch et al. 2005, Trincado et al. 2007, VanderSchaaf 2014,
Mehtätalo et al. 2015). e greatest advantage when using linear
mixed-eects models often is the ability to calibrate the model using
data independent of those used in model tting. Plot/stand-specic
H–D relations can be produced for trees from plots/stands not in the
model-tting dataset if H and D observations have been collected
from trees in that plot/stand. Hence, calibration for a specic stand
(or plot) in a sense “borrows strength across units,” supplementing the
information from the stand/plot by information from all the other
stands/plots used in model tting (Fitzmaurice etal. 2004). Unlike
ordinary least squares, which ignores individual plot trends when t-
ting an H–D curve of trees from several plots when estimating H only
as a function of D, a calibrated mixed-eects model better captures
the individual behavior of the H–D curve within a particular stand
(or plot). Mixed-eects models t using plot-level data can also be
calibrated at the plot level during operational inventories, providing
more localized H–D relations within a stand. To increase the e-
ciency of forest inventories in loblolly (Pinus taeda L.), longleaf
(Pinus palustris P. Mill.), shortleaf (Pinus echinata Mill.), and slash
(Pinus elliottii Engelm.) pine plantations in the Western Gulf, USA
region, individual tree mixed-eects models were developed.
Mixed-eects H–D models have been developed for loblolly
pine plantations across the southeastern United States (Trincado
etal. 2007) and for East Texas plantations (Coble and Lee 2011).
Additionally, Coble and Lee (2011) presented an equation for slash
pine plantations in East Texas. However, all models were devel-
oped exclusively using data from industrial lands. e equations
presented here use data from a range of ownerships including
nonindustrial private landowners (NIPF), state, federal, etc. lands,
and are developed from plots located throughout the Western Gulf,
USA region (Figure 1). ere has not been a mixed-eects H–D
model developed for longleaf pine and shortleaf pine plantations in
the Western Gulf. Mixed-eects H–D models have been developed
for naturally regenerated shortleaf pine stands for western Arkansas
and eastern Oklahoma (Budhathoki etal. 2008).
Manuscript received April 4, 2018; accepted July 31, 2019; published online June 17, 2019.
Afﬁliations: Curtis L. VanderSchaaf (firstname.lastname@example.org), Assistant Professor, 101 Reese Dr, School of Agricultural Sciences and Forestry, Louisiana Tech
University, Ruston, LA 71272.
Acknowledgments: Useful comments were received from Associate Editor Dr Scott Roberts, an Applied Research Editor, and three reviewers.
Copyright © 2019 Society of American Foresters
applyparastyle “g//caption/p” parastyle “FigCapt”
applyparastyle “body/p” parastyle “Text_First”
2 Forest Science • XXXX 2019
ese newly presented equations can help answer an inter-
esting question of whether similar levels of predictive ability can be
obtained for H when using a calibrated mixed-eects H–D model
t using data of a species other than the one of interest as compared
to a calibrated model t using data of that species of interest. When
a species lacks an existing regionally applicable H–D model, using
predictions from an equation of another species would be very
useful if adequate estimates could be produced. Hence, in a sense,
if mixed-eects models can be calibrated across species, it can be
thought that all species are in the same population and that a single
model would be applicable to all species.
VanderSchaaf (2008) showed that using a loblolly pine plan-
tation mixed-eects H–D model when calibrated using sweetgum
(Liquidambar styraciua L.) data produced reasonable results to
predict H of sweetgum plantations. However, H was not of indi-
vidual trees but of arithmetic mean height, whereas D was quad-
ratic mean diameter. Asimilar concept was proposed by Huang
(2016) in which he used the term “universal” model to describe
calibrating models across species. He found that “composite” H–D
equations, or equations t using several species, following calibra-
tion, often produced similar predictions to species-specic mixed-
eects H–D models. is “universal” approach is similar in nature
to Zeide’s (1978, 1993, 1994) two-point principle. He suggested
for several stand-level variables that growth curves common across
many species could be tailored for an individual stand by obtaining
measurements of that variable in that stand at only two ages.
Although applicable operationally for many stand-level variables,
the two-point concept would be very dicult operationally to apply
to individual tree H–D relations because the two points would need
to be quantied for each individualtree.
Russell et al. (2014) showed that including species-specic
random eects when modeling the change in height in north-
eastern United States and eastern Canadian mixed-species stands
showed little dierence from species-specic models. However,
their study did not directly examine whether models t using one
species could be calibrated to produce adequate predictions for
another species. Crecente-Campo etal. (2013) conducted a sim-
ilar analysis to that carried out by Russell etal. (2014) for uneven-
aged mixed-species forests in the state of Durango, Mexico. ey
mentioned the potential to apply their models through the cali-
bration process in neighboring states but cautioned against it and
stated that the models needed to be tested prior to conducting such
an analysis. Although including species-specic random eects into
the model is a valid approach, their analyses did not address the
predictive ability of their models if a species was not included in the
In addition to producing H–D models for the four pine species,
the aim was to determine the predictive ability of a calibrated H–D
model t for one species but calibrated for another. e previously
cited studies were limited in scope in that they addressed only a
small region of the world, and each only addressed H–D relations,
one at the stand-level (VanderSchaaf 2008) and the three others for
H–D relations of individual trees (Crecente-Campo etal. 2013,
Russell etal. 2014, Huang 2016). Calibrating across species may be
useful for many other variables such as basal area, volume, tree per
acre estimates, etc., and the ability to calibrate across species from
across the world may be applicable but of course must be tested.
is study examines the ability to calibrate across species from
throughout dierent parts of the United States. Other than than
VanderSchaaf (2008) and Huang (2016), no studies are known that
have examined the predictive ability of using mixed-eects models
of dierent species to predict for a species. e objective of this re-
search was to determine whether a linear mixed-eects H–D model
t using data of one species can be calibrated for another species to
produce adequate predictions.
Data Used in Model Fitting
e data used in model development were obtained from USDA
Forest Service Forest Inventory and Analysis (FIA) annual surveys
for Louisiana, Mississippi, and Texas (Figure 1). For Louisiana,
surveys completed between 2002 and 2015, for Mississippi,
surveys completed between 2009 and 2015, and for Texas, surveys
completed between 2009 and 2015, were included, respectively.
Data were obtained from the FIA database website (O’Connell
etal. 2017, US Forest Service 2017). Aplot consists of a cluster
of four points arranged such that point 1 is central, with points
2–4 located 120 ft from point 1 at azimuths of 0, 120, and 240°
(Bechtold and Scott 2005).
Each point is surrounded by a 24-ft xed-radius subplot where
trees 5.0 in. in diameter at breast height and larger are measured.
e four subplots combined total approximately 1/6th acre. Each
subplot contains a 6.8-ft xed-radius microplot where only saplings
(1.0–4.9 in. in diameter at breast height) are measured. Combined,
the four microplots total approximately 1/75th of an acre. e rel-
ative probabilities associated with a particular tree in relation to
diameter have no impact in terms of estimating H as a function
of D. For individual tree H–D measurements, they were simply
considered individual tree measurements when model tting and
calculating average tree-level summary statistics. However, the rel-
ative probabilities have some inuence when determining average
plot-level summary statistics and the number of trees of a particular
D used in model tting (Figure 2). e relative probabilities have
dierent associated-tree-per-acre expansion factors that will impact
summary statistics of per-unit-area estimates such as trees per acre
and basal area per acre estimates.
Individual tree mixed-eects models were developed for loblolly,
longleaf, shortleaf, and slash pine plantations in the Western Gulf,
USA region. Data from only those plots where a particular pine
species comprised at least 60 percent of the total basal area were
Management and Policy Implications
This paper presents mixed-effects height–diameter models for southern
yellow pine species in the Western Gulf region that can be applied during
forest inventories to help reduce inventory costs. Additionally, this study
demonstrates that a single mixed-effects height (H)–diameter (D) model can
be calibrated across a range of species. An example of where this could be
useful is for a species such as sand pine, Pinus clausa (Chapm. ex Engelm.)
Vasey ex Sarg., in the Lower Atlantic Coastal Plain, USA. Since it has less
commercial value and is located on far fewer acres relative to loblolly pine
(Pinus taeda L.) or slash pine (Pinus elliottii Engelm.), few models have been
developed for this species. This study shows that a mixed-effects individual
tree H–D model developed for loblolly pine could be calibrated for sand pine
if H–D measurements have been conducted in a sand pine plot/stand to pro-
duce adequate and plot/stand-speciﬁc H predictions.
Forest Science • XXXX 2019 3
included in the model-tting datasets. For the model of a partic-
ular species, only H and D measurements of that particular species
were modeled. Before model tting, all trees with broken stems
were removed from the model-tting data set, and only trees whose
heights were actually measured (as opposed to visually estimated or
predicted using equations) were used in model tting (only HTs
with HTCD= 1 within FIA database were included when mod-
eling, or those trees whose heights were actually measured in the
eld). Individual tree and plot-level summary data are presented in
Tables 1 and 2.
Model Development and Parameter Estimation
Total tree height was predicted as a function of diameter at
lnHki =(β0+u0k)+(β1+u1k)lnDki +εki
where: ln=natural logarithm; Hki=individual tree total height (ft)
for tree i within plot (or stand) k; Dki=individual tree diameter at breast
height (in.) for tree i within plot (or stand) k; β
1=parameters to be
estimated; u0k, u1k=plot/stand-specic random eects, assumed to be
0) and N(0, σ
1), respectively; (β
Figure 1. Height–diameter models were created using Forest Inventory and Analysis data obtained from Louisiana, Mississippi, and Texas,
USA. Calibration data were obtained from South Carolina.
Figure 2. Height–diameter relations by species for the model-ﬁtting and validation datasets. Sample sizes for the model-ﬁtting data:
loblolly pine, 64,415; longleaf pine, 1,049; shortleaf pine, 623; slash pine, 3,978. Sample sizes for the model validation data: loblolly
pine, 14,548; longleaf pine, 938; slash pine, 30. For the model-ﬁtting dataset, the gray curves are the “population average” estimates
from Equation 1 where individual plot-level factors creating correlations among individual trees from plots (created through the inclusion
of plot-speciﬁc random effects) were accounted for when model ﬁtting overlaid over the H–D measurements, and the black curve are
ordinary least-squares estimates overlaid over the H–D measurements where individual plot-level factors creating correlations among
individual trees from plots (created through the inclusion of plot-speciﬁc random effects) were ignored when model-ﬁtting.
4 Forest Science • XXXX 2019
1+u1k)=plot/stand-specic slope; and ε
where it is assumed ε ~N(0, σ
Additionally, a covariance, σ
01, can be assumed to exist between
u0k and u1k, and thus it is assumed the plot/stand-specic random
eects have a bivariate normal distribution with mean 0.In this
particular case, linear mixed-eects models produce an ecient es-
timate of plot/stand-specic parameters because only six parameters
are estimated using the model-tting algorithm (β
2). When using the variance and covariance estimates, plot/
stand-specic random eects (u0k, u1k) can be predicted and then
added to the “population average” intercept and slope (β
estimates to obtain plot/stand-specic parameter estimates.
Plot/stand-specic random parameters produce a more localized
H–D equation, since the random eects account for local site
conditions such as soil type, genetic stock, site preparation,
midrotation silvicultural practices, spatial and time-specic cli-
matic conditions, etc. e prediction of plot/stand-specic random
eects is conducted outside the model-tting algorithm, and thus
degrees of freedom are not lost due to specic plot/stand random
eects. In terms of model tting, a less ecient method of obtaining
plot-specic parameter estimates would be to estimate parameters
separately for eachplot.
When compared to the traditional means of developing local
H–D equations, where H and D are measured and then a sepa-
rate equation is t for a stand (or in some cases a plot), a mixed-
eects model analysis is ecient because a model can be calibrated
without having to statistically t a model, and thus even small
sample sizes can be used (Lynch etal. 2005). In many cases, xed-
eects region-wide H–D equations are used to predict tree heights.
To account for growing-condition dierences among stands for the
same species, in addition to D, these models contain measures of
site quality and/or stand density, among others. Studies have shown
that a calibrated mixed-eects H–D model often produces better
predictions than a region-wide H–D model containing stand-
level regressors (e.g., Trincado et al. 2007, Temesgen etal. 2008,
Crecente-Campo etal. 2013).
e ln–ln transformation of the Power function was chosen be-
cause it (or its nonlinear untransformed form) has been shown to
provide reasonable predictions for a variety of species (e.g., Huang
et al. 1992, O’Brien et al. 1995, Lei etal. 2009, Sharma 2009,
Uzoh 2017, Subedi etal. 2018). eoretical and empirical studies
of the H–D relation suggest that it is an allometric function with
the power of diameter (or the slope in linear transformed form),
1, equal to 2/3 (Greenhill 1881, McMahon 1973, Norberg 1988,
O’Brien etal. 1995), and the slope is associated with the constant-
stress theory value, essentially that any diameter along the stem must
be of sucient size to support all weight above it, else the tree will
collapse under its own weight and other factors such as wind force
(e.g., Zeide and VanderSchaaf 2002). It is well known that the ln–
ln transformation often accounts for heterogeneity, and since it is
linear it will likely have fewer statistical convergence problems than
nonlinear equations, particularly when trying to include random
eects into the model. In addition, to determine whether this par-
ticular H–D mixed-eects model form (β
ki) could be
calibrated across species, previously published mixed-eects models
using only the same model form and structure when tting this
model form for the other species were included in the analysis. is
congruence between the t and selected published models helps
avoid confounding issues between species and model form/struc-
ture when calibrating Equation 1 across species. If dierent equa-
tions or structures (e.g., the direct modeling of heterogeneity) are
used, then any ability or the lack of ability to calibrate cannot be
Table 1. Tree-level summary statistics of trees used in model ﬁtting and model calibration/validation.
Species No. of trees D (in.) H (ft)
Min Mean Max SD Min Mean Max SD
Loblolly pine 64,415 1.0 8.0 59.1 3.5 6 53 148 18.17
Longleaf pine 1,049 1.0 9.7 28.1 4.8 8 65 112 22.34
Shortleaf pine 623 1.0 10.5 35.9 5.7 8 66 130 27.80
Slash pine 3,978 1.0 8.0 23.8 3.5 8 55 123 19.48
Loblolly pine 14,548 1.0 8.1 28.0 2.90 8 54 130 16.47
Longleaf pine 938 1.2 6.5 20.1 1.84 9 40 83 9.85
Slash pine 30 3.0 12.0 19.6 4.50 29 69 99 26.96
Note: Max, maximum; Min, minimum; SD, standard deviation.
Table 2. Model-ﬁtting and validation plot-levelmeans.
Species of interest No. of plots Species of interest Other species
TPA Dq (in.) BAA (sq ft) TPA Dq (in.) BAA (sq ft)
Loblolly pine 2,289 332 6.6 78.8 363 3.0 18.0
Longleaf pine 70 181 7.9 62.2 119 5.0 16.0
Shortleaf pine 49 167 8.4 63.8 259 4.3 26.0
Slash pine 157 349 6.1 71.7 298 3.2 16.3
Loblolly pine 440 328 6.9 86.2 300 2.7 12.0
Longleaf pine 27 416 5.4 66.9 155 4.0 13.7
Slash pine 3 83 11.0 55.0 156 5.8 28.4
Note: “Species of interest” refers to plot-level summary statistics for the species used in tting the height–diameter equations. “Other species” refers to hardwoods as well
as other conifers. BAA, square ft of basal area per acre; Dq, quadratic mean diameter (in.); TPA, trees per acre.
Forest Science • XXXX 2019 5
separated exclusively into the eect of species. Aquestion would
then arise as to whether the ability or lack of ability to calibrate was
due to species, model form and/or structure, or the combination of
species and model form/structure.
Before model tting, all trees with broken stems were removed
from the model-tting data set, and only trees whose heights were
actually measured (as opposed to visually estimated or predicted
using equations) were used in model tting (only HTs with
HTCD=1 within FIA database were included when modeling,
or those trees whose heights were actually measured in the eld).
SAS Proc MIXED (Littell etal. 1996) was used to estimate the
parameters of Equation 1.No attempt was made to include spa-
tial correlation in the models to reduce complexity when applying
these models. Although heteroskedasticity (nonconstant variance)
likely exists, we are interested in predicting heights of trees equally
across all diameters. Hence, the random error variance–covari-
ance matrix was assumed to be σ
2Ink. Using weighted least squares
would put more weight on smaller-diameter trees when tting the
regression line that would reduce the predictive ability of larger-
diameter trees. Other studies have shown the tradeo between
biological and statistical considerations when tting models and
the impacts they can have on predictions and ultimately manage-
ment (VanderSchaaf and South 2009, VanderSchaaf etal. 2011).
Although it is well known that ln–ln transformations often ac-
count for heterogeneity, residuals were examined for trends in the
data (Figure 3).
Data Used in Model Calibration
If indeed models can be calibrated across species such that
adequate predictions are obtained, to gain insight into the op-
timal calibration sample size, independent FIA survey data from
the state of South Carolina were used. Surveys completed be-
tween 2012 and 2016 were included. Individual tree- and plot-
level summary data are presented in Tables 1 and 2. Data from
South Carolina were chosen because this is a meaningfully dif-
ferent population than those trees used in model tting from the
A minimum sample size of 10 trees per plot was selected, and so
a plot (both trees from the FIA subplot or microplot) had to con-
tain at least 12 trees of a particular pine species to be included. Ten
trees would be used to calibrate the model, the 11th and 12th trees
would be predicted, and at least two predicted trees would allow
calculation of variance. Calibration sample sizes of 1, 2, 3, 4, 5,
8, and 10 were examined. However, for slash pine, to have at least
three plots used in calibration, a minimum sample size of 5 was
used (a plot had to have at least seven trees to allow the calculation
e tree used to calibrate the model for a sample size of one
was also the rst tree for a sample size of 2, the two trees used to
calibrate the model for a sample size of 2 were the rst two trees
for a sample size of 3, etc. To avoid the dependence of the calibra-
tion results on one particular sample for an individual calibration
sample size, calibrations were conducted 50 times for each calibra-
Validation analyses follow those presented in Trincado et al.
(2007). e dierence between the observed (Hobs) and predicted
height (Hpred) of all trees (i) whose heights were predicted for each
individual plot (k) and for each of the 50 replications (r) was cal-
culated (ekri= Hobs kri−Hpred kri), and trees used in calibration for a
particular plot and replication were not included in the validation
statistic calculations. For each plot (k) and replication (r) combina-
tion, the mean residual (ē) and the sample variance (v) of residuals
were computed and considered to be estimates of bias and preci-
sion; respectively. An estimate of mean square error (MSE) was
obtained for each combination by combining the bias and precision
measures using the following formula:
To obtain an average value of the three calibration sta-
tistics for each plot and calibration sample size, the 50 MSE,
ē, and v values of a particular plot for a particular calibration
sample size and species were then averaged. Average plot cali-
bration statistics were then averaged to obtain the nal calibra-
tion statistics for a given calibration sample size. e procedure
recommended by Baskerville (1972) was used to account for the
Calibration of the same model form for several species
throughout the USA was tested. In addition to calibrating the
model developed for each of the four species (loblolly, longleaf,
shortleaf, and slash) for itself and the three other species, inde-
pendent models presented in other papers for loblolly pine across
the southeastern United States (Trincado etal. 2007), Douglas-r
(Pseudotsuga menziesii [Mirb.] Franco var. menziesii) and lodgepole
pine (Pinus contorta Douglas ex Loudon) from the inland north-
western United States (VanderSchaaf 2014), and quaking aspen
(Populus tremuloides Michx.) from Minnesota, USA (VanderSchaaf
2013), were also calibrated using the model validation data from
South Carolina, USA.
Figure 3. Residuals from Equation 1 over standardized diameter at breast height for the model-ﬁtting datasets. Sample sizes for the
model-ﬁtting data: loblolly pine, 64,415; longleaf pine, 1,049; shortleaf pine, 623; slash pine, 3,978. Standardized diameter at breast
height is calculated as
where AMD is the arithmetic mean diameter (inches), and SD is the standard deviation of D (diameter at
6 Forest Science • XXXX 2019
Results and Discussion
Model-tting results (Table 3) for all t species (loblolly,
longleaf, shortleaf, and slash) showed that it is best to assume
that both β
0 and β
1 are random, or, essentially, that each plot
(or stand) has its own intercept and slope, and that a covariance
01) exists between u0k and u1k. Parameter estimates and model-
tting statistics are presented in Table 4. Residuals showed no de-
parture from model assumptions (Figure 3). When choosing an
optimal model calibration sample size at the plot or stand level
for a particular species, a suggested reasonable tradeo between
statistical measures (precision and accuracy) and sampling times
(e.g., costs) when using that species to calibrate is two or three
trees (Figures 4–6).
Example of Model Calibration
For clarity and ease of application, the methodology to predict
random eects at the plot or stand level is presented. Nomenclature
is based on Schabenberger and Pierce (2002, p.431). e expres-
sion used to predict random eects, estimated best linear unbiased
= predicted random effects of plot/stand k,
a 2 × 1 vector where 2 is the number of predicted random
= estimated variance–covariance matrix (2× 2) of
the random effects;
= matrix (nk×2) containing observed
values of D (natural log-transformed) from plot/stand k, and
a column of 1s;
= estimated variance–covariance random
error matrix expressed as
, since the variance is assumed
constant across all plots/stands and Hs are assumed tempo-
rally and spatially uncorrelated within a plot/stand;
vector of observed Hs (natural log-transformed) from plot/
=regressor matrix (nk×2) consisting of a column
of 1s and the observed values of D (natural log-transformed)
from plot/stand k; and
= 2 × 1 vector of estimated fixed-
To demonstrate model calibration, three individual longleaf
pine trees were randomly selected from an individual plot (or
stand) to be used in calibration. e heights for the three trees are
48, 50, and 44 ft, and the corresponding Ds are 5.8, 6.3, and 7.3
5.8 in .
1 ln(6.3 in .)
1 ln(7.3 in .)
1 ln(5.8 in .)
1 ln(6.3 in .)
1 ln(7.3 in .)
where, besides the column of 1s, all numerical values are
lnD (naturally log-transformed value of D in inches) for the
three observations randomly selected from this particular plot
ln(50 ft )
ln(44 ft )
where all numerical values are lnH (naturally log-transformed
value of H in feet) for the three observations randomly selected
from this particular plot (or stand).
3.912023 −(3.0542 +0.4990[ln (50 ft )]
3.78419 −(3.0542 +0.4990[ln (44 ft )]
0.01160 0 0
0 0.01160 0
0 0 0.01160
All numerical values for
were obtained from Table
4 for longleaf pine and are obtained from the model-tting results
of Equation 1.e dimensions of
, Z, X, y, and
change based on the number of observations used in the model
When performing matrix operations as shown in Equation 3,
the following predictions of the random eects, estimated best
linear unbiased predictors, for β
0 and β
1 of this particular plot (or
stand) were obtained:
ese predicted random eects for this plot (or stand) are added
to the “population average” parameter estimates,
, to obtain plot/
stand-specic parameter estimates for this particular plot/stand,
Table 3. –2Log likelihood (smaller is better) model-ﬁtting param-
eter estimates of the population average (β
0 and β
1) and random
effects variance (σ
1) and covariance (σ
01) parameter estimates
Parameters –2Log likelihood
Loblolly Longleaf Shortleaf Slash
1–23,841.6 –586.7 –213.0 –2,466.6
0–88,389.8 –1,175.3 –484.9 –5,418.5
1–83,179.4 –1,001.7 –426.6 –5,114.3
1–96,291.8 –1,211.7 –491.0 –5,623.6
β0, β1, σ
01 –100,037.0 –1,390.6 –581.3 –5,909.1
n2,289 70 49 157
Note: n, number of clusters, or plots.
Forest Science • XXXX 2019 7
It is well known that logarithmic transformations often linearize
data and produce homogeneity of variances; however, a transfor-
mation bias occurs, since additive errors in ln–ln models become
multiplicative when transformed back to the original scale. To ac-
count for the transformation bias, the procedure recommended by
Baskerville (1972) should be used when predicting heights:
2= mean square error (or residual variance) from the
model t (see Table 4).
To obtain H predictions in the original units (in feet), Equation
5 should beused:
Hki = exp
0k – (β
0+u0k) is the plot-/stand-specic intercept, and
1k – (β
1+u1k) is the plot-/stand-specicslope.
For the example above, Equation 5 would be expressedas:
Hki = exp[2.8405+0.5552l n D
Table 4. Population average (β
0 and β
1) and random effects variance (σ
1) and covariance (σ
01) parameter estimates by species.
Loblolly pine Longleaf pine Shortleaf pine Slash pine
Est SE Est SE Est SE Est SE
03.0005 0.01027 3.0542 0.06358 2.9697 0.09099 2.9399 0.03758
10.4605 0.003897 0.4990 0.02299 0.5336 0.03127 0.5250 0.01443
00.1997 – 0.2162 – 0.2869 – 0.1710 –
10.02686 – 0.02621 – 0.03071 – 0.02345 –
01 –0.06437 – –0.07319 – –0.08996 – –0.05754 –
–2LL –100,037 –1,390.60 –581.3 –5,909.1
AIC –100,029 –1,382.60 –573.3 –5,901.1
20.009853 0.01160 0.01643 0.01059
n2,289 70 49 157
2, estimated mean square error; 2LL, twice the negative log-likelihood (smaller is better); AIC, Akaike’s Information Criterion (smaller is better); n, number of
clusters, or plots; SE, standard error of the estimate.
Figure 4. Model calibration mean square error results for longleaf pine plantations located in South Carolina, USA. Calibration sample
sizes are the number of trees used in calibration. n=27 plots.
8 Forest Science • XXXX 2019
Assuming a tree has a D of 7.44 in., the predicted H would be
52.5 ft. Use of the noncalibrated, “population average” curve, on
the other hand, provides a predicted height of 58.1 ft, a 9.6 percent
Figure 7 clearly demonstrates that calibrating Equation 1 using
observed Hs and Ds from this particular plot (producing Equation
6) vastly improved predictive ability, consistent with many
other studies (e.g., Trincado et al. 2007, Temesgen et al. 2008,
VanderSchaaf 2014). Relative to the “population average” curve,
for a given D, most Hs for this particular plot are shorter (perhaps
because of genetic stock, stand-level species composition, aspect,
soil type, site productivity, etc.). Hence, Equation 6, through the
calibration process, provides an H–D curve that reects this be-
havior. However, using the so-called “population average” trend
fails to recognize the behavior of trees in this individual plot relative
to the average behavior across allplots.
Calibration Using Other Species
Figures 4–6 clearly demonstrate that mixed-eects models of a
variety of species can be calibrated to provide reasonable predictions
for a particular species. It is interesting to note for a particular cal-
ibration species that almost all curves regardless of model-tting
species converge after using just one tree in calibration. In general,
regardless of the model-tting species, for longleaf and slash pine
any model using two or more trees in calibration provided nearly
the same predictive ability. is is true even for model species such
as aspen. Loblolly pine showed more variability. However, the use
of three or more trees generally produced nearly the same predic-
tive ability among all the model species. For all three validation
species, in some cases models of other species actually produced
slightly better predictions. ese results indicate, at least for this
model form of a H–D equation, that when a species does not have
a species-specic model that models t using other species when
calibrated can provide similar, and in some cases even slightly
Figure 7 also shows that the H–D mixed-eect model form
evaluated in this study can be calibrated for a particular species
regardless of the model-tting species. In fact, for this longleaf
pine plot represented by these three trees when calibrating, the
calibrated aspen model produced the best prediction statistics.
When calibrated using these three trees, the aspen, longleaf pine,
and Douglas-r models produced MSEs of 44.4419, 51.2244, and
59.0495, respectively. e aspen model may not always produce the
best predictions when applied to otherdata.
For this H–D model form, a calibrated model of another species
produces better predictions than an uncalibrated species-specic
model (Figures 4–7). Apparently, based on these model results,
the model calibration process exceeds the importance of species-
specic H–D relations, since, regardless in many cases of the partic-
ular species-specic parameter estimates used, improved predictive
Figure 5. Model calibration mean square error results for slash pine plantations located in South Carolina, USA. Calibration sample sizes
are the number of trees used in calibration. n=3 plots.
Forest Science • XXXX 2019 9
results were observed for a particular species using models t for
several other species. Often modelers and foresters believe that only
species-specic parameter estimates can apply to that specic spe-
cies. However, these results show that following calibration models
of many dierent species can be applied to predict variables of other
species; hence in many cases, for certain model forms, parameter
estimates and model forms may be universal across many species
rather than species-specic.
Mixed-eects H–D models were presented for loblolly, longleaf,
shortleaf, and slash pine plantations in the Western Gulf, USA.
By obtaining H–D measurements from plots/stands of interest,
Equation 1 can be calibrated to local site conditions. To calibrate
these models for specic plots/stands, an Excel spreadsheet is avail-
able upon request.
Mixed-eects models of a variety of species were calibrated for
individual species and were found to provide reasonable predictions
for that particular calibrated species. is was true, even for
groups of species as dierent as pines, rs, and in some cases even
hardwoods. Future work should concentrate on examining the pre-
dictive ability of other H–D model equationforms.
ese results, along with those of VanderSchaaf (2008) and
Huang (2016), suggest that models t for one species can be
calibrated across species. However, this process may not work for
all H–D model forms and for all variables. For instance, in order
Figure 6. Model calibration mean square error results for loblolly pine plantations located in South Carolina, USA. Calibration sample
sizes are the number of trees used in calibration. n=440 plots.
Figure 7. Observed longleaf pine total tree height and diameter at
breast height for n=22 trees. This plot resides in Charleston county,
South Carolina, USA. The bold solid lines are “population average”
estimates for three species (longleaf pine, aspen, and Douglas-ﬁr)
obtained by not calibrating Equation 1, and the lighter solid lines
are estimates for the three species following calibration of Equation
1.Black diamonds and black circles are paired height–diameter
observations. Black circles were trees used to calibrate Equation
1.Black lines are predictions using the longleaf pine model, dark
gray lines are predictions using the aspen model, and light gray
lines are predictions using the Douglas-ﬁr model.
10 Forest Science • XXXX 2019
to better constrain future predictions of basal area observed in
young stands for species with relatively longer rotations or of lower
value that do not justify the implementation of large-scale, com-
prehensive research studies, perhaps a mixed-eects model t but
calibrated using a species with a younger economic rotation or bio-
logical rotation age or that is valuable enough to justify large-scale,
comprehensive research studies, model calibration will be bene-
cial. By “borrowing” information from other species, expensive,
long-term, studies can be avoided for the species of interest, since
information on other species can be used to represent the temporal
growth dynamics of the calibrated species.
Future research should concentrate on determining whether
using parameters from a species where more complete data in terms
of stand ages, planting densities, site qualities, stand development,
etc., exist provides more reasonable extrapolations of stand develop-
ment for another species following calibration rather than using lim-
ited data of that desired species to t models (VanderSchaaf 2008).
is current analysis cannot denitively answer this question; how-
ever it is suggested, for this particular topic of determining if the
ability of calibrating mixed-eects models across a variety of species
improves model-tting ability, is a very important research topic
for this particular subject of study. Predicted plot-specic or stand-
specic random parameters depend on the amount of estimated
variability in the random eects and the “population average” pa-
rameter estimates. us, dierences in site requirements, structural
constraints, growth habits, etc., among species may not allow for
mixed-eects models t using one species to produce reasonable
predictions for another species.
Additionally, ecological and physiological studies can also be
conducted such as examining how calibrating across species and
determining its impact on the “population-average” and random
parameters (e.g., individual plot or stand parameters) of, say,
Equation 1 can provide an insight into the applicability of concepts
such as the constant-stress theory, etc. Perhaps mixed-eects model
calibration can be used to justify future research to determine the
inclusion or exclusion of certain species into various consistent
forms or functions of species that can lead to extensive amounts of
study to determine why the species have constant growth forms and
functions within categories or why they do not grow commonly
within categories by form and function.
B,G.L. 1972. Use of logarithmic regression in the estimation of
plant biomass. Can. J.For. Res. 2:49–53.
B,W.A., C.T.S. 2005. e forest inventory and anal-
ysis plot design. P. 27–42 in e enhanced forest inventory and anal-
ysis program—national sampling design and estimation procedures,
B,W.A., and P.L.P (eds.). US Forest Service Gen.
Tech. Rep. SRS-GTR-80.
B,C.B., T.B.L, J.M.G. 2008. A mixed-eects
model for the dbh–height relationship of shortleaf pine (Pinus echinata
Mill.). South. J.Appl. For. 32:5–11.
C,D.W., Y.L. 2011. A mixed-eects height–diameter model
for individual loblolly and slash pine trees in East Texas. South. J.Appl.
C-C, F., J.J. C-R, B. V-L,
C.W. 2013. Can random components explain dierences in
the height–diameter relationship in mixed uneven-aged stands? Ann.
For. Sci. 71:51–70.
F,G.M., N.M.L, J.H.W. 2004. Applied longitu-
dinal analysis. Wiley, New York. 506 p.
G,A.G. 1881. Determination of the greatest height consistent
with stability that a vertical pole or mast can be made, and of the
greatest height to which a tree of given proportions can grow. P. 65–
73 in Proceedings of the Conference of the Cambridge Philosophical
Society Vol. 4.
H,S. 2016. Population and plot-specic tree diameter and height pre-
diction models for major Alberta tree species. Forestry Division, Alberta
Agriculture and Forestry, Edmonton, AB. 59 p.
H,S., S.J.T, D.P.W. 1992. Comparison of nonlinear
height–diameter functions for major Alberta tree species. Can. J.For.
L,J. 1991. Calibration of height and volume equations with random
parameters. For. Sci. 37:781–801.
L,X., C. P, H.W, X.Z. 2009. Individual height–di-
ameter models for young black spruce (Picea mariana) and jack pine
(Pinus banksiana) plantations in New Brunswick, Canada. For. Chron.
L, R.C., G.A. M, W.W. S, R.D. W.
1996. SAS® system for mixed models. SAS Institute, Cary, NC. 633 p.
L, T.B., A.G. H, D.J. S. 2005. A random-
parameter height-dbh model for cherrybark oak. South. J. Appl. For.
MM,T.A. 1973. Size and shape in biology. Science 179:1201–1204.
M, L., S. -M, T.G. G. 2015. Modeling
height–diameter curves for prediction. Can. J.For. Res. 45:826–837.
N, R.A. 1988. eory of growth geometry of plants and self-
thinning of plant populations: Geometric similarity, elastic similarity,
and dierent growth modes of plants. Am. Nat. 131:220–256.
O’B, S.T., S.P. H, P. S, R. C, R.B. F.
1995. Diameter, height, crown, and age relationships in eight neotrop-
ical tree species. Ecology 76:1926–l939.
O’C, B.M., B.L. C, A.M. W, E.A. B,
J.A. T, S.A. P, G. C, T. R,
J.M. 2017. e forest inventory and analysis database: Database
description and user guide version 7.0 for phase 2. USDA Forest Service.
830 p. [Online]. Available online at http://www.a.fs.fed.us/library/
R,M.B., A.R.W, J.A.K, J. 2014. Comparing
strategies for modeling individual-tree height and height-to-crown base
increment in mixed-species Acadian forests of northeastern North
America. Euro. J.For. Res. 133:1121–1135.
S, O., F.J. P. 2002. Contemporary statistical
models for the plant and soil sciences. CRC Press, Boca Raton, FL. 738 p.
S,R.P. 2009. Modelling height–diameter relationship for Chir pine
trees. Banko Janakari. 19:3–9.
S,M.R., B.N.O, S.S, S.C. 2018. Height–di-
ameter modeling of Cinnamomum tamala grown in natural forest in
mid-hill of Nepal. Intl. J.For. Res. 2018:1–11.
T,H., V.J.M, D.W.H. 2008. Analysis and com-
parison of nonlinear tree height prediction strategies for Douglas-r
forests. Can. J.For. Res. 38:553–565.
T,G., C.L.VS, H.E.B. 2007. Regional
mixed-eects height–diameter models for loblolly pine (Pinus taeda L.)
plantations. Euro. J.For. Res. 126:253–262.
US F S. 2017. FIA DataMart 1.6.1. Available online at
accessed August 10, 2017.
U,F. 2017. Height–diameter model for managed even-aged stands of
ponderosa pine for the western United States using a hierarchical non-
linear mixed-eects model. Aust. J.Basic Appl. Sci. 11:69–87.
VS, C.L. 2008. Stand level height–diameter mixed eects
models: Parameters tted using loblolly pine but calibrated for
Forest Science • XXXX 2019 11
sweetgum. P. 386–393 in Proceedings of the 16th Central Hardwood
Forest Conference, J, D.F., and C.H. M (eds.). USDA
Forest Service Gen. Tech. Rep. NRS-P-24, Northern Research Station,
Newtown Square, PA.
VS, C.L. 2013. Mixed-eects height–diameter models for
commercially and ecologically important hardwoods in Minnesota.
North. J.Appl. For. 30:37–42.
VS,C.L. 2014. Mixed-effects height–diameter models
for ten conifers in the inland Northwest, USA. South. For.
VS,C.L., D. S. 2009. Biological, economic, and
management implications of weighted least squares. South. J.Appl. For.
VS,C.L., Y.L, D.B.S. 2011. e weighted least
squares method aects the economic rotation age of loblolly pine—two
planting density scenarios. Open Forest Sci. J. 4:42–48.
Z,B. 1978. Standardization of growth curves. J. For. 76:289–292.
Z,B. 1993. A parsimonious number of growth curves. North. J.For.
Z,B. 1994. To construct or not to construct more site index curves?
West. J.For. 9:37–40.
Z,B., C.L.VS. 2002. e eect of density on the
height–diameter relationship. P. 463–466 in Proceedings of the Eleventh
Biennial Southern Silvicultural Research Conference, K.W. O
(ed.). USDA Forest Service Gen. Tech. Rep. SRS-48, Southern
Research Station, Asheville, NC. 622 p.