%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%Paper by Van Deusen and Roesch for 2012 FIA Symposium
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\documentclass[twocolumn,letterpaper,sort]{article}
\pagestyle{myheadings}
\usepackage{../mcfnsSep2012}
%\usepackage[OpenPeerReview]{McfnsDraftCopy}%\usepackage[opr]{McfnsDraftCopy}%\usepackage[preprint]{McfnsDraftCopy}
\usepackage{caption}
\captionsetup[table]{belowskip=2pt}
\usepackage{graphics}
\usepackage{amsmath,bm,url,amssymb}
\usepackage{graphicx}
\usepackage{blindtext}
\usepackage{setspace}
\usepackage[switch]{lineno}
\usepackage{natbib}
\usepackage{color}
\usepackage[breaklinks]{hyperref}
\widowpenalty=10000
\clubpenalty=10000
%\usepackage[breaklinks,pdfstartview={FitH -32768},pdfborder={0 0 0},bookmarksopen,bookmarksnumbered]{hyperref}
%\usepackage{bibtexlogo}
%______________________________________________________________________
% Define MCFNS variables:
% Moved to mcfns.sty like the \issueyear \publish {Feb.~00} %Still not sure if Issue Manuscript publication date
\setcounter{page}{126}\def\issueno{2}
\def\editors {\href{mailto:editors@mcfns.com} {Editor: Greg C Liknes}}
\def\submit {Mar.~20,~2013} %Submission date can be different than the issue year \issueyear
\def\accept {Jul.~8,~2013} %The works should be Accepted & Published in the year of the Current_Issue \issueyear
\def\lasterrata {Aug.~15,~2013} %Last Errata date can be different than the Issue-Year \issueyear
\def\citename {Van Deusen } %"Author"
\def\citeemail {pvandeus@gmail.com} % Use later: {\href{mailto://\citeemail}{\citename}}
\def\citeetal {and Roesch} % or {} %for a single author;
\author{ {\href{mailto:\citeemail}{Paul C~\citename}}$^1$, % Change only the first name of the first author
{\href{mailto:froesch@fs.fed.us}{Francis A~Roesch}}$^2$, }
\affiliation{
%---------------
\small \it {$^1$\href{http://www.ncasi.org}{National Council for Air and Stream Improvement, Mount Washington, MA, USA}}\\
\small \it{$^2${\href{http://www.srs.fs.usda.gov/}{USDA Forest Service, Southern Research Station, Asheville, NC, USA}}}\\ }
%%%comment the following and uncomment above section for names and addresses
%\def\citename {Author 1} %"Author"
%\def\citeemail {mail@mail.net}
%\def\citeetal {~et~al.} % or {} %for a single author; or
%\author{ {\href{mailto:\citeemail}{A.~\citename}}$^1$, % Change only the first name of the first author
% {\href{mailto:m@m}{B. ~Author}}$^2$, }
%\affiliation{
%---------------
%\small \it {$^1$\href{http://www}{address1}}\\
%\small \it{$^2${address2}}\\ }
%%%end of dummy author section
\def\yourtitle {{Trends and projections from annual forest inventory plots
and coarsened exact matching}} %need double {{ for \\ e.g.: {{Title \\ Subtitle}}
\def\yourkwords {Forest inventory, Moving average, Inventory projection, FIA data}
\def\yourabstract{
The coarsened exact matching (CEM) method is used to match annual
forest inventory plots awaiting remeasurement with plots that have already
been remeasured. This results in a model-free approach for short
term inventory projections. CEM has many desirable properties
relative to other matching methods and is easy to apply within a SQL
database. The combination of short term projections with a 3 or 5
year moving window is suggested for providing trend estimates that
include the current year and a few years into the future. The default
projection represents business as usual. A method to bias the plot
matching to generate desired scenarios is also developed. These ideas
and methods are demonstrated with several applications to forest
inventory data. Scenarios are generated where increasing future
harvest levels are stochastically controlled to demonstrate this
capability with operational data.
}%----------------------------------------------------------------------
% \usepackage[english]{babel}
\def\sqm{m$^2$}
\def\cum{m$^3$}
\def\sqmpha{m$^2$\, ha$^{-1}$}
\def\cumpha{m$^3$\,ha$^{-1}$}
\def\sqkm{km$^2$}
\def\dpcum{\$\,m$^{-3}$}
\def\dpha{\$\,ha$^{-1}$}
\def\cumpyr{m$^{-3}$\,y$^{-1}$}
% THE REST SHOULD BE AUTHOMATIC Go To the first Section
%\linenumbers
\title{\Large\bf\uppercase\yourtitle}
\begin{document}
\markright{\hfil{{{\href{mailto://\citeemail}{\citename}\citeetal}}~(\issueyear)/\mcfnshead}}
\twocolumn[
\begin{@twocolumnfalse}
\hypersetup{pdftitle={\mcfnshead},pdfauthor={\citename~(\issueyear)},pdfsubject={\yourtitle},pdfkeywords={\yourkwords}}
\maketitle
\hrule\begin{abstract}\yourabstract\\\\{\bf Keywords: }\yourkwords\end{abstract}\hrule\vspace{.3truein}
\end{@twocolumnfalse}
]
\section{Introduction}
The United States Department of Agriculture (USDA) Forest Service's
Forest Inventory and Analysis program (FIA) has implemented an annual
inventory system in the USA that involves measuring a proportion of
the plots every year in each state. The annual percentage of plots
measured is nominally 20\% in most eastern states and 10\%
elsewhere. The plots measured in the same year are said to be in the
same panel. The number of years required to measure all plots in a
state is called a cycle (Tab. \ref{tab:cycles}). The set of the most
recent measurements for all plots in a state is called the current
evaluation group (EVALGRP). Therefore, the first EVALGRP for a state
on a 5 year cycle is completed in year 5. Thereafter, a new EVALGRP is
created each year by updating the measurements for the current panel.
\begin{table}[tb]
\caption{FIA plots measured on 5 year cycles.
The plots are placed into 5 panels labeled a,b,c,d and e. The same
panel is measured during the same year within each cycle. All 5 year
intervals have an associated EVALGRP that consists of the
most current measurements for each panel. In practice, this idealized
plan is not exactly followed.}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{c c c c c c c c c c c c c c c}
\hline \hline
\multicolumn{5}{c}{cycle 1}&
\multicolumn{5}{c}{cycle 2}&
\multicolumn{5}{c}{cycle 3} \\
1&2&3&4&5&1&2&3&4&5&1&2&3&4&5\\
\hline
a& & & & &a& & & & &a& & & & \\
&b& & & & &b& & & & &b& & & \\
& &c& & & & &c& & & & &c& & \\
& & &d& & & & &d& & & & &d& \\
& & & &e& & & & &e& & & & &e \\
\hline
\end{tabular}
}
\label{tab:cycles}
\end{table}
A common way to obtain estimates for a state is to compute the mean
for the variable of interest over all plots in the most recent
EVALGRP. This estimator \citep{greenBook:2005,reams:1999,roesch:2003,vandeusen:1997} is often called a moving
average (MA),
since the average from one EVALGRP to the next changes as the current
panel is updated.
One concern with this approach is that the average of an EVALGRP does
not represent a particular year, rather it represents some time point
between when the first measurements and the last measurements were
taken for the EVALGRP. Clearly, a 5-year MA as defined by FIA,
applied to the center of the period rather than the end of the period,
is the 5-year moving window (MW). It follows that methods developed for variance
estimation with the MA apply directly to the MW, so the MW requires no
additional theoretical development. There is also no reason that
estimates should be constrained to plots within an EVALGRP, which is
an unnecessary limitation when looking at trends. For example, a
3-year MW is a very simple and useful way to look at trends in FIA
data.
Users of FIA data would like estimates that correspond to the current
year. They would also be interested in short term predictions for
5-10 years into the future. Short term projections of plots would be
one way to satisfy both interests. For example, taking the moving
window of a pseudo 5-year EVALGRP where years 4 and 5 are actually
predictions, would result in an estimate for the current year.
Likewise, the predicted years give an indication of what to expect in
the near future.
There are also time series methods that could provide current
estimates without using future predictions \citep{vandeusen:1999}, but
time series methods are more complex than a MW or MA
and don't necessarily work with categorical data. Creating a pseudo
EVALGRP where some of the panels are based on projections provides an
approach for obtaining current estimates that fits in with existing
procedures. If the projected data have all the features of real data,
it can be analyzed with existing software. The objective here is to
demonstrate how to use coarsened exact matching (CEM) to project
annual forest inventory data to provide current estimates and short
term projections.
\section{Plot matching for plot projection}
Plot matching methods have been suggested and used for projecting FIA
plots in previous studies
\citep{vandeusen:1997,vandeusen:2010a,wear:2011}. Those studies were
seeking multi-decade projections, whereas the interest here is in
short term projections that correspond to the number of years in an
FIA measurement cycle. A method is proposed that requires fewer
assumptions than the methods in these earlier studies, but it does
require availability of remeasured plots. In general, these
applications have been based on hot-deck methods \citep{sande:1983}
that seek matches from the large FIA data base \citep{FIADB:2010}.
The proposed method is based on a subject plot and a set of remeasured
plots. Suppose the remeasured plots have measurements for times 1 and
2. The idea is to match the subject plot with remeasured plots that
have similar time 1 characteristics. Then the time 2 measurements are
used as projections for the subject plot. Call this the match then
project (MTP) approach. Other applications of plot matching have used
a project then match (PTM) approach, where some of the variables on
the subject plot, like age and basal area, are projected and then
matched with variables on other plots.
The PTM approach can be applied in states or regions where no
remeasured plots are available. However, PTM requires that some
variables be projected before the matching takes place. The MTP
approach can only be used if remeasured plots are available. The
advantage of MTP is that no model based assumptions are required for
the projection. If the subject plot is similar to the time 1
measurement on the matched plot, then the time 2 measurement is likely
to be a good representation of a possible future state of the subject
plot.
The MTP approach will tend to give projections that correspond to the
cycle length, if the cycle length is relatively constant within the
set of remeasured plots. Additionally, the MTP method naturally
provides a business as usual (BAU) projection. For example, harvest
levels in the projected data will be consistent with harvest levels in
the matched plots. Levels of insect damage, storm damage and fire
will also reflect what influenced the years between time 1 and 2
measurements on the matched plots. A model based alternative for BAU
projections is possible \citep{fernandez:2012}, but requires specific
assumptions about various factors that have influenced recent trends.
A default BAU projection naturally results from the MTP method, but it
is also possible to manipulate the matching process to deviate from
BAU. After discussing the details of the matching process, a method to
bias the matching to simulate desired scenarios is developed. The
matching process and scenario generation are demonstrated with example
applications.
\subsection{Plot matching methods}
Matching procedures have been traditionally used to compare treated
and control groups in observational studies. Although the treated and
control analogy doesn't directly apply here, our treatment group would
be the plots that need to be projected and the remeasured plots would
be the control group. Regardless of terminology, it is clear that there are
2 sets of plots, subject and remeasured, and there is a need to find
remeasured matches for the subject plots. The main goal of matching
is to find remeasured plots that were similar to the subject plot at
time 1. Then the time 2 measurements can be used as projections for
the subject plot. A subject plot with earlier measurements can also be
placed in the pool of remeasured plots as a potential match for other
subject plots.
Propensity score (PS) matching is perhaps the most popular strategy
for causal analysis in observational
studies \citep{rosenbaum:1983,rubin:2001,dehejia:2002}. The PS
provides a scalar summary of the covariates. The main result
from \cite{rosenbaum:1983} is that if 2 groups of observations have
the same propensity score then they have the same multivariate
distribution of $\textbf{X}$, no matter how big the dimension of
$\textbf{X}$
\citep{rubin:2001}. The importance of this result cannot be
overstated. It means that PS matching can result in comparisons
between control and treated groups where bias due to differences in
the covariates has been eliminated. Unmatched comparisons with
observational data could otherwise be dominated by differences in
$\textbf{X}$ between the 2 groups.
In practice, there is no guarantee that PS matching will reduce the
bias in comparisons of the 2 groups. Proper application of the method
requires careful checking of the results and iterative re-application.
PS methods and Mahalanobis matching are members of the equal percent
bias reducing (EPBR) class \citep{iacus:2011a,iacus:2011b} which works
on average, but does not guarantee bias reduction for any particular
data set. In fact, \cite{king:2011} found that PS matching often
approximates random matching, and can degrade inferences relative to
not matching at all. \cite{iacus:2011a} present a new monotonic
imbalance (MIB) class that offers improved operational performance.
\subsection{Coarsened exact matching}
CEM is a member of the MIB class that is used here, in part, because
it is easily applied to large databases with SQL code. More
significantly, CEM-based estimates have many desirable statistical
properties \citep{iacus:2011a,iacus:2011b}. It has been shown that
CEM dominates commonly used matching methods (EPBR and others) as
measured by its ability to reduce imbalance, model dependence,
estimation error, bias, variance, mean square error, and other
criteria for a range of data sets \citep{iacus:2011a,iacus:2011b}.
CEM is surprisingly easy to implement. Matching on continuous
variables may involve rounding or creating meaningful
categories. Categorical variables can be used directly or after
reducing the number of categories. CEM allows the user to operate in
the natural k-dimensional data space. PS matching or Mahalanobis
distance matching create a transformed measure that has less intuitive
appeal. As such, CEM maintains congruence between the data space and
analysis space and does not violate \emph{The Congruence
Principle} \citep{iacus:2011b}.
CEM might be best explained by considering how to apply it to the
problem at hand, i.e. matching subject FIA plots with the time 1 state
of a remeasured plot. This is a special instance of what are generally
called hot-deck methods \citep{sande:1983}. The first decision is
whether to match on plots or conditions. Each mapped plot condition
can represent a unique situation, so it was decided to match
conditions. Hence, the FIA variable that corresponds to condition
proportion (CONDPROP) became the first matching criterion. Since
CONDPROP ranges between 0 and 1 it needs to be coarsened. CONDPROP
was put into 5 categories ranging from 0 to 4 by applying the
following function, $C=round(round(CONDPROP*100)/25)$. The coarsened
value of CONDPROP, C, ensures that conditions get matched with other
conditions of similar size. This eliminates the need to make
adjustments to the post-matched magnitude of the
measurements. Post-match condition size adjustment would involve a
number of decisions about how to reduce or increase the condition
size. In particular, increasing the size could lead to problems if a
small condition contained a few unusual trees, like giant redwoods.
In addition to CONDPROP, the following variables were used in the matching
process:
\begin{description}
\item[OWNGRPCD] The owner group code was not initially coarsened.
This indicates if the owner of the FIA plot is in one of four
groups:
\begin{enumerate}
\item National Forest
\item Other Federal
\item State and local government
\item Private
\end{enumerate}
OWNGRPCD is a coarsened version of OWNCD provided by FIA.
\item[FORTYPCD] The FIA forest type code was used without any
coarsening. Hence, plot conditions had to be matched with
conditions having the same FORTYPCD.
\item[STDORGCD] Matching was also done on uncoarsened stand origin
code, which indicates if the condition originated
naturally or from planting.
\item [SITECLCD] The other matching variable that was not initially
coarsened was site productivity class code which
indicates site growth potential and is explained in more detail
below.
\item [STDAGE] Stand age was coarsened into a 10, 20, 30, and 40
year age class for younger stands. Stands greater than 40 and less
than 60 went into class 50. Stands greater than 60 and less than
80 went into class 70. Stands 80 or older went into class 80.
\item[BA] Basal area was coarsened into increments of 20 (square feet
per acre) with the following function, $B=round(BA/20)$, which sets
basal areas of 11-29 to 1, and basal areas of 190-210 to 10.
\end{description}
Decisions on what variables to use for matching and how to coarsen
them are based largely on knowledge about important descriptors of
the inventory plots and what level of detail is needed for matching.
Too much detail results in many plots remaining unmatched and not
enough detail leads to poor matches. In any event, the decisions made
here are easily modified, and allow us to proceed with an example
application.
\subsubsection{What to do with unmatched plots}
After the initial application of CEM, there will usually be some plot
conditions that aren't matched. As the name suggests, a subject plot
condition can only be matched if there are time 1 values for
remeasured plot condtions that exactly match all of its coarsened
values. Either these unmatched conditions must be discarded from the
analysis, or further coarsening must be applied. The approach used
here involved 3 iterations. Iteration 1 was to apply the CEM method
described above. All successfully matched subject plots are then set
aside.
Iteration 2 seeks matches for the remaining subject plot conditions by further
coarsening owner group, stand age, basal area and site productivity,
\begin{description}
\item [OWNGRPCD]
The National Forest and Other Federal categories are combined. This results in
the following 3 categories
\begin{enumerate}
\item Federal
\item State and local government
\item Private
\end{enumerate}
\item [STDAGE]
Stand age is coarsened into 3 categories.
\begin{enumerate}
\item $STDAGE < 30$
\item $30 \leq STDAGE < 80$
\item $STDAGE \geq 80$
\end{enumerate}
\item[BA]
Basal area is coarsened into 3 categories
\begin{enumerate}
\item $BA < 30$
\item $30 \leq BA < 120$
\item $BA \geq 120$
\end{enumerate}
\item[SITECLCD]
Site productivity class code is coarsened into 3 categories
\begin{enumerate}
\item $SITECLCD \leq 2$
\item $3 \leq SITECLCD < 5$
\item $SITECLCD \geq 5$
\end{enumerate}
\end{description}
The original SITECLCD consists of 7 levels of productivity
\citep{FIADB:2010}. It is a classification of forest land in terms of
inherent capacity to grow crops of industrial wood. It is an estimate
of the potential growth and is based on the
culmination of mean annual increment of fully stocked natural stands,
\begin{enumerate}
\item 225+ ft$^3$/ac/yr
\item 165-224 ft$^3$/ac/yr
\item 120-164 ft$^3$/ac/yr
\item 85-119 ft$^3$/ac/yr
\item 50-84 ft$^3$/ac/yr
\item 20-49 ft$^3$/ac/yr
\item 0-19 ft$^3$/ac/yr
\end{enumerate}
A third and final CEM iteration is applied to subject plot conditions
that were not matched in the first 2 iterations. This involved
further coarsening of OWNGRPCD, STDAGE, BA and SITECLCD,
\begin{description}
\item [OWNGRPCD]
Owner group is eliminated entirely for this final attempt to match plots.
\item [STDAGE]
Stand age is coarsened into 2 categories.
\begin{enumerate}
\item $STDAGE < 40$
\item $STDAGE \geq 40$
\end{enumerate}
\item[BA]
Basal area is coarsened into 2 categories
\begin{enumerate}
\item $BA < 50$
\item $BA \geq 50$
\end{enumerate}
\item[SITECLCD]
Site productivity class code is coarsened into 2 categories
\begin{enumerate}
\item $SITECLCD \leq 3$
\item $SITECLCD > 3$
\end{enumerate}
\end{description}
The few plot conditions that have no matches after the 3 CEM iterations are
excluded from further analysis. Increased coarsening for CEM will increase
the potential for bias. However, bias due to coarsening one variable does
not effect bias in other variables. This is an important property of the
CEM method not enjoyed by other non-MIB methods such as propensity score
matching \citep{iacus:2011a,iacus:2011b}.
\section{Biasing the matching process to generate scenarios}
The CEM process results in nearly all plot conditions having at least
one match, and many will have multiple matches. The multiple matches
could be used to implement multiple
imputation \citep{rubin:1987,vandeusen:1997,reams:2000,mcroberts:2001}
if obtaining better variance estimates for BAU predictions is a
primary interest. However, an important objective here is to select
from the multiple matches to control how the results will deviate from
a BAU scenario.
The BAU scenario arises naturally when each subject condition is
randomly assigned a single match from its CEM results. The selected
matches will reflect whatever has been occurring during the recent
remeasurement intervals. Suppose the BAU selection is accomplished
by assigning a uniform (0,1) random variate ($rv$) to each matched
condition and then selecting the matched condition with the largest
$rv$. Now we want to modify this approach to bias the selections toward
matches that were subjected to certain events. Specifically, the event
must be something that happened after the time 1 measurement and is
recorded in the database.
The event could be storm damage, insect attack, harvesting or any well
defined occurrence. Sort the matches for subject condition $i$ so the
matches where the event occurred come first, $M_i =
\{e_1,e_2,...,O_1,O_2,...\}$, where $e_j$ denotes a matched condition
that experienced the event and $O_j$ is a match where the event did
not occur. Let $u$ and $u'$ represent draws from a uniform (0,1)
distribution and $r$ be some small positive value. Suppose the
following $rv$ vector corresponds to $M_i$,
\begin{equation} \label{eq:RV}
RV_i~=~\left \{ u_1+r,u_2+r,...,u'_1,u'_2,... \right \}
\end{equation}
where the random variables for the event impacted matches have $r$
added to them. Now select the condition from the matched set $M_i$
that corresponds to the largest value in $RV_i$. This approach
increases the probability of selecting a match that experienced the
event. The value of $r$ directly controls how likely it is that an
event impacted match is selected.
It is important to understand the influence of $r$ to control the
scenario generation process. There are 2 ways that an event impacted
condition, $e_i$, will be selected, so the overall probability of an
event winning the selection can be broken into 2 parts,
\begin{description}
\item [Part 1] At least one of the random variates corresponding to $u_j+r \geq 1$ in
$RV_i$.
\item [Part 2] All $u_j + r \le 1$ but an event impacted match still wins.
\end{description}
Part 1 of the overall probability is equivalent to
\begin{equation} \label{eq:part1}
P \left ( \text{At least one } u_j+r \geq 1 \right ) ~=~
1~-~p_1^{n_e}
\end{equation}
where there are $n_e$ event impacted matches and
$p_1~=~P(u+r\le 1)~=~1/(1+r)$. This is 1 minus the probability
that none of the $u_j+r\geq 1$
Part 2 is
\begin{equation} \label{eq:part2}
P \left ( u_j +r \leq 1 ~\forall j ~\text{and an event match wins} \right ) ~=~p_1^{n_e} \frac{n_e}{N}
\end{equation}
With this information it is possible to solve for $r$ to control the
probability of selecting an event impacted match. This is done
relative to the probability of an event occurring in the BAU scenario,
which is $P_B=n_e/N$, where N is the total number of matches for a
particular set of subjects. The overall probability of selecting an
event impacted match as a function of r is the sum of the 2 parts,
\begin{equation} \label{eq:parts}
P_r~=~1~-~p_1^{n_e}~+~~p_1^{n_e} \frac{n_e}{N}
\end{equation}
Now define a value, $Q=P_r/P_B$, which defines a scenario. For
example, if r is set so that $Q=1.5$ then the selected matches will
have an expected 50\% increase in events relative to the BAU scenario.
Finally, solve for r as a function of Q,
\begin{equation} \label{eq:Q}
r(Q)~=~\left [
\left ( 1-\frac{n_e}{N} Q \right )\frac{N}{N-n_e}
\right ] ^{-\frac{1}{n_e}} -1
\end{equation}
Setting $Q$=1 results in $r$=0, which yields the BAU scenario. Values
of $Q>1$ will cause the event impacted matches to be selected at a
higher rate than under the BAU scenario. However, equation
(\ref{eq:Q}) must be carefully applied. It is not sufficient to
compute a single $r$-value for an entire state. The most accurate method
is to compute a unique $r$-value for each plot and condition based on
its specific set of matches. A plot condition that has 10 possible
matches where only 1 gets harvested needs a different $r$-value than if
8 of the matches get harvested. This is controlled by the $n_e$ and
$N$ parameters (eq \ref{eq:Q}).
\section{Example applications}
The CEM matching and projection method is demonstrated with two
applications. The scenarios that are considered involve selecting more
plots that will experience harvesting than what would occur in a
BAU scenario. In addition to BAU where $Q$=1, we evaluate
$Q$=0.5 (reducing harvest by 50\%), $Q$=1.5 (increasing harvest by 50\%),
and $Q$=2.0 (increasing harvest by 100\%).
\subsection{Application to Maine data}
The first application is to Maine FIA data for the 2006-2010 EVALGRP.
The CEM algorithm is applied in 3 stages as described above. This
results in projections for 2011-2015. There is actual data for 2011,
which results in the 2011 moving window consisting of actual data for
2010 and 2011 and projected data for 2012. The gross-growth over
harvest-removals ratio (G/R) is displayed (Tab. \ref{tab:ME1}) for
increasing harvest rate scenarios. G/R$>1$ for all years from
2005-2010 and for the BAU and +50\% projections. This suggests that
Maine forests are being sustainably managed, overall. However, a 50\%
increase in harvest levels would barely be sustainable (Table
\ref{tab:ME1}) according to this analysis, but the +100\% G/R falls
below 1. Sample sizes (Tab. \ref{tab:ME2}) are adequate for each of
the scenarios.
The CEM plot matching and projection process is stochastic.
Therefore, the desired change in harvest level for the scenarios
(-50\%,+50\% and +100\%) is not exactly attained. The removals
(Tab. \ref{tab:ME3}) are displayed for each scenario.
They increase or decrease as expected depending on the harvest
scenarios, but reflect the stochastic nature of the process.
\begin{table*}[!t]
\caption{Gross growth over harvest removals (G/R) projections for
private timberland in Maine using a 3-year moving window. Values
for 2006-2010 are from FIA data. Values for 2011-2015 are based on
CEM matching and show the business as usual (BAU) projection along
with stochastic changes in harvest levels of -50, +50 and +100\%}
\centering
\begin{tabular}{ rc|rcccc }
\hline \hline
Year & Actual G/R & Year & G/R BAU & BAU -50\% &BAU +50\% &BAU +100\% \\
\hline
2006 &1.626& 2011& 1.718& 2.007& 1.466& 1.330 \\
2007 &1.536& 2012& 1.871& 2.721& 1.322& 1.059 \\
2008 &1.382& 2013& 1.635& 2.941& 1.092& 0.811 \\
2009 &1.406& 2014& 1.678& 2.917& 1.128& 0.813 \\
2010 &1.582& 2015& 1.618& 2.965& 1.148& 0.839 \\
\hline
\end{tabular}
\label{tab:ME1}
\end{table*}
\begin{table*}[!t]
\caption{Sample sizes associated with the estimates in Table (\ref{tab:ME1})}
\centering
\begin{tabular}{ rc|rcccc }
\hline \hline
Year & Actual & Year & BAU & BAU -50\% &BAU +50\% &BAU +100\% \\
\hline
2006 &1739& 2011& 1739& 1729& 1730& 1733 \\
2007 &1741& 2012& 1726& 1714& 1727& 1730 \\
2008 &1750& 2013& 1716& 1698& 1715& 1716 \\
2009 &1760& 2014& 1712& 1699& 1717& 1721 \\
2010 &1767& 2015& 1702& 1688& 1702& 1710 \\
\hline
\end{tabular}
\label{tab:ME2}
\end{table*}
\begin{table*}[!htbp]
\caption{Annual per acre $ft^3$ harvest removals projections for private
timberland in Maine using a 3 year moving window. Values for
2006-2010 are from FIA data. Values for 2011-2015 are based on CEM
matching and show the business as usual (BAU) projection along with
stochastic increases in harvest levels of 50 and 100\% and a reduction
of 50\%. Estimates are based on 3 year moving windows.}
\centering
\begin{tabular}{ rc|rcccc }
\hline \hline
Year & Actual R & Year & R BAU & R -50\% & R +50\% & R +100\% \\
\hline
2006 &30.0& 2011& 31.1& 26.7& 36.1& 39.6 \\
2007 &31.7& 2012& 27.7& 19.2& 38.5& 47.5 \\
2008 &35.1& 2013& 30.6& 17.1& 44.5& 59.1 \\
2009 &35.9& 2014& 30.4& 17.5& 44.2& 60.4 \\
2010 &33.4& 2015& 31.6& 17.3& 43.8& 59.1 \\
\hline
\end{tabular}
\label{tab:ME3}
\end{table*}
\subsection{Application to Wisconsin data}
Wisconsin has enough national forest and state and local government
land to make a comparison with privately owned land possible.
The G/R values (Tab. \ref{tab:WI1}) for national forest land are quite
variable. This is due to the relatively small number of FIA plots on
national forest land in Wisconsin, and also because there is less
harvesting on national forest land in general. Often, G/R ratios are shown
as net-growth/removals. The values here are (gross growth)/(harvest removals),
so mortality is not subtracted from growth.
\begin{table}[!b]
\caption{Gross-growth over harvest-removals ratios (G/R) for Wisconsin
for National Forest (NF), state and local government (State), and
private owners on timberland computed with a 3 year moving
window. Sample size is shown in parentheses.} \centering
\begin{tabular}{ rccc }
\hline \hline
Year & NF & State & Private \\
\hline
2007 & 8.957 (316) & 2.629 (768) & 2.764 (2576)\\
2008 & 7.183 (325) & 2.762 (789) & 2.610 (2605)\\
2009 & 9.635 (337) & 2.623 (771) & 2.584 (2587)\\
2010 & 5.377 (334) & 2.429 (766) & 2.695 (2643)\\
2011 & 5.572 (315) & 2.431 (797) & 2.768 (2696)\\
\hline
\end{tabular}
\label{tab:WI1}
\end{table}
Now we can look at the G/R projected trend for the 3 ownership types
for years 2012-2016. The results (Tab. \ref{tab:WI2}) suggest that the
national forest could easily sustain a doubling of harvest and still
maintain a large G/R. The state and local government land could
barely sustain a doubling of harvest, but it would border on being
unsustainable. The private land could also sustain a harvest
doubling, according to the G/R results, since the G/R values (Table
\ref{tab:WI2}) are all larger than 1.0 for the projected years. Of course,
other measures of sustainability should also be considered before such a
drastic increase in harvest levels is implemented.
\begin{table}[!h]
\caption{Projected gross-growth over harvest-removals ratios (G/R) for
Wisconsin for National Forest (NF), state and local government
(State), and private owners under scenarios that represent business
as usual (BAU) harvesting and a doubling of harvest (+100\%). These
are 3 year moving windows applied to timberland.}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{ rcccccc }
\hline \hline
\multicolumn{1}{c}{} &
\multicolumn{2}{c}{NF}&
\multicolumn{2}{c}{State}&
\multicolumn{2}{c}{Private}\\
%\hline
Year & BAU & +100\% & BAU & +100\%& BAU & +100\%\\
\hline
2012 & 5.067 & 3.365 & 2.437 & 1.739 & 2.960 & 2.210\\
2013 & 8.470 & 4.030 & 2.438 & 1.396 & 2.783 & 1.762\\
2014 & 9.036 & 4.437 & 2.427 & 1.039 & 1.643 & 1.390\\
2015 & 6.803 & 3.907 & 2.494 & 1.050 & 2.671 & 1.402\\
2016 & 5.462 & 3.450 & 2.634 & 1.055 & 2.755 & 1.462\\
\hline
\end{tabular}
}
\label{tab:WI2}
\end{table}
\section{Discussion}
FIA data analysis has depended very heavily on the 5-year MA since the
beginning of the annual inventory system around 1998. The MA has
become linked to the concept of an EVALGRP, which is basically the
most recent set of measurements for all FIA plots in a state. This
linkage between an EVALGRP and the MA has become a standard analysis
concept with FIA online tools, but it should be clear that the FIA
sample design is not that restrictive. For example, the 3-year MW is
a legitimate option, even though it does not align with any EVALGRP.
CEM was applied to FIA plots in the current EVALGRP, since those are
the plots that are next in line to be remeasured. CEM identifies, for
each subject plot in the current EVALGRP, already remeasured plots
where the time 1 measurement matches the most recent measurement of
the subject plot. Then the time 2 measurements from the matching
plots can be used as predictions of what the subject plot will look
like at its next measurement time.
CEM predictions combined with a 3-year or 5-year MW provide a simple
way to estimate trends from FIA data that include the current year
and a few projected years. Variance estimators that have been used
for the 5-year MA can also be applied to the MW estimates. However,
the same variance estimators applied to projected years would
understate the true uncertainty in the projections. It's not clear
that an unbiased estimator exists for the variance of projected means.
One option would be to use methods developed for multiple
imputation \citep{rubin:1987}, which would involve applying CEM
several times and combining the resulting variance estimates.
We did not compare CEM with other plot matching methods, such as those
based on the propensity score or Mahalanobis distance. We deferred to
the extensive work done elsewhere \citep{iacus:2011a,iacus:2011b} that
demonstrates the improved performance of CEM over other methods.
On the basis of that work, users can be confident that the CEM process
will result in good matches with respect to the coarsened variables
that are selected by the user to describe the forest inventory plot data.
\section{Conclusions}
This study focused on FIA annual inventory plot data, but the methods
should be applicable to any data base consisting of remeasured forest
inventory plots. Methods were presented and demonstrated for
estimating current trends and making short term projections. Few
assumptions are required to implement these methods. Annual estimates
can be based on simple 3-year or 5-year moving windows. Projections
can be based on plot matching techniques that eliminate the need for
growth and yield models.
CEM was suggested as the plot matching method, because it has
desirable statistical properties and is easy to implement. In fact,
CEM can be operationally implemented with standard SQL database
programming. The focus here was on short term projections, which
minimizes required assumptions and leads to a default BAU projection.
The short term projections were demonstrated with real data that was
collected in the recent past. Therefore, the projections reflect what
has occurred recently. This includes recent levels of disturbance,
harvesting and land use change. Also, a method was developed to bias
the projections to have increased frequency of an event type of
interest. This provides an opportunity to implement stochastic
scenario development. Scenarios involving increasing levels of
harvesting were demonstrated in the example applications.
It would be inadvisable to evaluate rare events or conditions with
these methods, because few matches will be found in the
database. Therefore, rare conditions would not be reliably
projected. On the other hand, rare conditions are not likely to have
reliable estimates for current values either, since there is typically
only 1 FIA plot per 6000 acres. The methods here are recommended for
short term projections, which should minimize the possibility that
users would wrongly conclude that frequency of rare types is
changing.
The methods presented here require fewer assumptions than when
projections are model derived. The biasing method provides a
stochastic alternative to modeling changes in probability of event
occurrence, and it fits in well with the CEM hotdeck approach. Candidate
matches are found for a plot based on time 1 characteristics, then
candidates that experience the event of interest are selected with a
greater probability than candidates that don't have the event. The
event (harvesting in the examples) happens after time 1, so there is
no deterministic way to predict it. We think the stochastic biasing
method is a good way to mimic reality in a model independent manner.
The example applications were implemented with a combination of SQL
and R \citep{R:2010}. These methods provide simple yet effective
procedures for extracting trend information from forest inventory
data. The short term projections give a look ahead that requires few
assumptions. Since the projections are short term, they could be
assessed for bias and perhaps improved by fine tuning the matching
process every few years.
\section*{Acknowledgements}
We thank the guest editors and the four anonymous reviewers for their
helpful comments.
\begin{thebibliography}{}
% \bibitem[Names(Year)]{label}
% Text of bibliographic item
\bibitem[Bechtold and Patterson (2005)]{greenBook:2005}
Bechtold, W. A., and P.L. Patterson. Editors. 2005. The enhanced forest
inventory and analysis program - national sampling design and
estimation procedures. Gen. Tech. Rep. SRS-80. Asheville, NC:
U.S. Department of Agriculture, Forest Service, Southern Research
Station. 85 p.
\bibitem[Dehejia and Wahba (2002)]{dehejia:2002}
Dehejia, R.H., and S. Wahba. 2002. Propensity score-matching methods
for nonexperimental causal studies. Rev. Econ. Stat. 84(1): 151-161.
\bibitem[Fernandez and Astrup (2012)]{fernandez:2012} Fernandez, C.A.,
and R. Astrup. 2012. Empirical harvest models and their use in
regional business-as-usual scenarios of timber supply and carbon
stock development. Scand. J. For. Res. 27:379-392.
\bibitem[Iacus et al. (2011a)]{iacus:2011a} Iacus, S.M., G. King, and
G. Porro. 2011a. Multivariate matching methods that are Monotonic
Imbalance Bounding. J. Am. Stat. Assoc. 106(493): 345-361.
\bibitem[Iacus et al. (2011b)]{iacus:2011b}
Iacus, S.M., G. King, and G. Porro. 2011b. Causal Inference without
Balance Checking: Coarsened Exact Matching. Polit. Anal. doi:
10.1093.
\bibitem[King et al. (2011)]{king:2011}
King, G., R. Nielsen, C. Coberley, J.E. Pope, and A. Wells. 2011.
Comparative Effectiveness of Matching Methods for Causal
Inference. Working Paper, 2011. Available online at http://j.mp/jCpWmk;
last accessed Feb. 25, 2013.
\bibitem[McRoberts (2001)]{mcroberts:2001}
McRoberts, R.E. 2001. Imputation and model-based updating tech-
niques for annual forest inventories. For. Sci. 47: 322-330.
\bibitem[R Development Core Team (2010)]{R:2010}
R Development Core Team 2010. R: A language and environment for
statistical computing. R Foundation for Statistical Computing,
Vienna, Austria. ISBN 3-900051-07-0, Available online at
http://www.R-project.org.
\bibitem[Reams and McCollum (2000)]{reams:2000}
Reams, G.A., and J.M. McCollum. 2000. The use of multiple imputation
in the southern annual forest inventory system. In Integrated Tools
for Natural Resource Inventories, Proceedings of IUFRO
Conference. Edited by M.H. Hansen and T. Burk. USDA
For. Serv. Gen. Tech. Rep. NC-212. pp. 228-233.
\bibitem[Reams and Van Deusen (1999)]{reams:1999}
Reams, G.A. and P.C. Van Deusen. 1999. The Southern Annual Forest
Inventory System. J. Agric. Biol. Environ. Stat. 4(4): 346-360.
\bibitem[Roesch et al. (2003)]{roesch:2003}
Roesch, F.A., J.R. Steinman, and M.T. Thompson. 2003. Annual forest
inventory estimates based on the moving average. P. 21-30 in
Proc. third annual forest inventory and analysis symp. October 17-19
2001, Traverse City, Michigan. McRoberts, R.E., G.A. Reams, P.C. Van
Deusen, and J.W. Moser (eds.). Gen. Tech. Rep. NC-230, USDA
For. Serv., North Central Res. Stn., St. Paul, MN. 208 p.
\bibitem[Rosenbaum and Rubin (1983)]{rosenbaum:1983}
Rosenbaum, P.R., and D.B. Rubin. 1983. The central role of the
propensity score in observational studies for causal
effects. Biometrika 70(1): 41-55.
\bibitem[Rubin (1987)]{rubin:1987}
Rubin, D.B. 1987. Multiple imputation for nonresponse in surveys. John Wiley \& Sons, Inc., New York, New York, USA. 258 p.
\bibitem[Rubin (2001)]{rubin:2001} Rubin, D.B. 2001. Using Propensity
Scores to Help Design Observational Studies: Application to the
Tobacco Litigation. Health Serv. Outcomes Res. Methodol. 2:169-188.
\bibitem[Sande (1983)]{sande:1983}
Sande, I.G. 1983. Hot-deck imputation procedures. In: Madow WG,
Olkin I, editors. Incomplete data in Sample Surveys, vol. 3.
New York: Academic Press: 334-50.
\addtolength{\textheight}{-7.75truein}
\bibitem[Van Deusen (1997)]{vandeusen:1997}
Van Deusen, P.C. 1997. Annual forest inventory statistical concepts
with emphasis on multiple imputation. Can. J. For. Res. 27:
379-384.
\bibitem[Van Deusen (1999)]{vandeusen:1999}
Van Deusen, P.C. 1999. Modeling trends with annual survey data.
Can. J. For. Res. 29: 1824-1828.
\bibitem[Van Deusen (2010)]{vandeusen:2010a} Van Deusen, P.C.
2010. Carbon sequestration potential of forest land: Management
for products and bioenergy versus preservation. Biomass Bioenergy
34: 1687-1694.
\vspace{.15in}
\bibitem[Wear (2011)]{wear:2011}
Wear, D.N. 2011. Forecasts of county-level land uses under three
future scenarios: a technical document supporting the Forest Service
2010 RPA Assessment. Gen. Tech. Rep. SRS-141. Asheville, NC:
U.S. Department of Agriculture Forest Service, Southern Research
Station. 41 p.
\vspace{.02in}
\bibitem[Woudenberg et al. (2010)]{FIADB:2010}
Woudenberg, S.W., B.L. Conkling, B.M. O'Conell, E.B. LaPoint, J.A. Turner, and K.L. Waddell. 2010. The Forest Inventory and Analysis Database: Database description and users manual version 4.0 for Phase 2. USDA For. Serv. Gen. Tech. Rep. RMRS-245. 336 p.
\label{docend}
\end{thebibliography}
\end{document}