\chapter{Sub-sampling a dataset}
\label{chap:sampling}

\section{Introduction}
\label{sample-intro}

Some subtle issues can arise here; this chapter attempts to explain
the issues.

A sub-sample may be defined in relation to a full dataset in two
different ways: we will refer to these as ``setting'' the sample and
``restricting'' the sample; these methods are discussed in
sections~\ref{sec:sample-set} and~\ref{sec:sample-restrict}
respectively. In addition section~\ref{sec:smpl-panel} discusses some
special issues relating to panel data, and
section~\ref{sec:resampling} covers resampling with replacement,
which is useful in the context of bootstrapping test statistics.

The following discussion focuses on the command-line approach. But you
can also invoke the methods outlined here via the items under the
\textsf{Sample} menu in the GUI program.


\section{Setting the sample}
\label{sec:sample-set}

By ``setting'' the sample we mean defining a sub-sample simply by
means of adjusting the starting and/or ending point of the current
sample range.  This is likely to be most relevant for time-series
data.  For example, one has quarterly data from 1960:1 to 2003:4, and
one wants to run a regression using only data from the 1970s.  A
suitable command is then

\begin{code}
smpl 1970:1 1979:4
\end{code}

Or one wishes to set aside a block of observations at the end of the
data period for out-of-sample forecasting.  In that case one might do

\begin{code}
smpl ; 2000:4
\end{code}

where the semicolon is shorthand for ``leave the starting observation
unchanged''.  (The semicolon may also be used in place of the second
parameter, to mean that the ending observation should be unchanged.)
By ``unchanged'' here, we mean unchanged relative to the last
\verb+smpl+ setting, or relative to the full dataset if no sub-sample
has been defined up to this point. For example, after

\begin{code}
smpl 1970:1 2003:4
smpl ; 2000:4
\end{code}

the sample range will be 1970:1 to 2000:4.  

An incremental or relative form of setting the sample range is also
supported.  In this case a relative offset should be given, in the
form of a signed integer (or a semicolon to indicate no change), for
both the starting and ending point. For example

\begin{code}
smpl +1 ;
\end{code}

will advance the starting observation by one while preserving the
ending observation, and

\begin{code}
smpl +2 -1
\end{code}

will both advance the starting observation by two and retard the
ending observation by one.

An important feature of ``setting'' the sample as described above is
that it necessarily results in the selection of a subset of
observations that are contiguous in the full dataset. The structure of
the dataset is therefore unaffected (for example, if it is a quarterly
time series before setting the sample, it remains a quarterly time
series afterwards).

\section{Restricting the sample}
\label{sec:sample-restrict}

By ``restricting'' the sample we mean selecting observations on the
basis of some Boolean (logical) criterion, or by means of a random
number generator.  This is likely to be most relevant for
cross-sectional or panel data.

Suppose we have data on a cross-section of individuals, recording
their gender, income and other characteristics.  We wish to select for
analysis only the women.  If we have a \verb+gender+ dummy variable
with value 1 for men and 0 for women we could do
%      
\begin{code}
smpl gender==0 --restrict
\end{code}
%
to this effect.  Or suppose we want to restrict the sample to
respondents with incomes over \$50,000.  Then we could use
%
\begin{code}
smpl income>50000 --restrict
\end{code}

A question arises: if we issue the two commands above in sequence,
what do we end up with in our sub-sample: all cases with income over
50000, or just women with income over 50000? By default, the answer is
the latter: women with income over 50000.  The second restriction
augments the first, or in other words the final restriction is the
logical product of the new restriction and any restriction that is
already in place.  If you want a new restriction to replace any
existing restrictions you can first recreate the full dataset using
%
\begin{code}
smpl --full
\end{code}
%
Alternatively, you can add the \verb+replace+ option to the
\verb+smpl+ command:
%
\begin{code}
smpl income>50000 --restrict --replace
\end{code}

This option has the effect of automatically re-establishing the full
dataset before applying the new restriction.

Unlike a simple ``setting'' of the sample, ``restricting'' the sample
may result in selection of non-contiguous observations from the full
data set.  It may therefore change the structure of the data set.

This can be seen in the case of panel data.  Say we have a panel of
five firms (indexed by the variable \verb+firm+) observed in each of
several years (identified by the variable \verb+year+).  Then the
restriction
%
\begin{code}
smpl year==1995 --restrict
\end{code}
%
produces a dataset that is not a panel, but a cross-section for the
year 1995.  Similarly
%
\begin{code}
smpl firm==3 --restrict
\end{code}
%
produces a time-series dataset for firm number 3.

For these reasons (possible non-contiguity in the observations,
possible change in the structure of the data), gretl acts differently
when you ``restrict'' the sample as opposed to simply ``setting'' it.
In the case of setting, the program merely records the starting and
ending observations and uses these as parameters to the various
commands calling for the estimation of models, the computation of
statistics, and so on. In the case of restriction, the program makes a
reduced copy of the dataset and by default treats this reduced copy as
a simple, undated cross-section---but see the further discussion of
panel data in section~\ref{sec:smpl-panel}.

If you wish to re-impose a time-series interpretation of the reduced
dataset you can do so using the \cmd{setobs} command, or the GUI menu
item ``Data, Dataset structure''.

The fact that ``restricting'' the sample results in the creation of a
reduced copy of the original dataset may raise an issue when the
dataset is very large.  With such a dataset in memory, the creation of
a copy may lead to a situation where the computer runs low on memory
for calculating regression results.  You can work around this as
follows:

\begin{enumerate}
\item Open the full data set, and impose the sample restriction.
\item Save a copy of the reduced data set to disk.
\item Close the full dataset and open the reduced one.
\item Proceed with your analysis.
\end{enumerate}

\subsection{Random sub-sampling}
\label{sample-random}

Besides restricting the sample on some deterministic criterion, it may
sometimes be useful (when working with very large datasets, or perhaps
to study the properties of an estimator) to draw a random sub-sample
from the full dataset.  This can be done using, for example,
%
\begin{code}
smpl 100 --random
\end{code}
%
to select 100 cases.  If you want the sample to be reproducible, you
should set the seed for the random number generator first, using the
\cmd{set} command.  This sort of sampling falls under the
``restriction'' category: a reduced copy of the dataset is made.

\section{Panel data}
\label{sec:smpl-panel}

Consider for concreteness the Arellano--Bond dataset supplied with
gretl (\texttt{abdata.gdt}). This comprises data on 140 firms
$(n=140$) observed over the years 1976--1984 $(T=9)$. The dataset is
``nominally balanced'' in the sense that that the time-series length
is the same for all countries (this being a requirement for a dataset
to count as a panel in gretl), but in fact there are many missing
values (\texttt{NA}s).

You may want to sub-sample such a dataset in either the
cross-sectional dimension (limit the sample to a subset of firms) or
the time dimension (e.g.\ use data from the 1980s only). The simplest
(but limited) way to sub-sample on firms keys off the notation used by
gretl for panel observations. The full data range is printed as
\texttt{1:1} (firm 1, period 1) to \texttt{140:9} (firm 140, period
9). The effect of
%
\begin{code}
smpl 1:1 80:9
\end{code}
%
is to limit the sample to the first 80 firms. Note that if you instead
tried \texttt{smpl 1:1 80:4}, gretl would insist on preserving the
balance of the panel and would truncate the range to ``complete''
firms, as if you had typed \texttt{smpl 1:1 79:9}.

The firms in the Arellano--Bond dataset are anonymous, but suppose you
had a panel with five named countries. With such a panel you can
inform gretl of the names of the groups using the \cmd{setobs}
command. For example, given
%
\begin{code}
string cstr = "Portugal Italy Ireland Greece Spain"
setobs country cstr --panel-groups 
\end{code}
%
gretl creates a string-valued series named \texttt{country} with group
names taken from the variable \texttt{cstr}. Then, to include only
Italy and Spain you could do
%
\begin{code}
smpl country=="Italy" || country=="Spain" --restrict
\end{code}
%
or to exclude one country,
%
\begin{code}
smpl country!="Ireland" --restrict
\end{code}

To sub-sample in the time dimension, use of \option{restrict} is
required. For example, the Arellano--Bond dataset contains a variable
named \texttt{YEAR} that records the year of the observations and if
one wanted to omit the first two years of data one could do
%
\begin{code}
smpl YEAR >= 1978 --restrict
\end{code}
%
If a dataset does not already incude a suitable variable for this
purpose one can use the command \texttt{genr time} to create a simple
1-based time index.

Note that if you apply a sample restriction that just selects certain
units (firms, countries or whatever), or selects certain contiguous
time-periods---such that $n>1$, $T>1$ and the time-series length is
still the same across all included units---your sub-sample will still
be interpreted by gretl as a panel.


\subsection{Unbalancing restrictions}

In some cases one wants to sub-sample according to a criterion that
``cuts across the grain'' of a panel dataset. For instance, suppose you
have a micro dataset with thousands of individuals observed over
several years and you want to restrict the sample to observations on
employed women.  

If we simply extracted from the total $nT$ rows of the dataset those
that pertain to women who were employed at time $t$ $(t = 1,\dots,T)$
we would likely end up with a dataset that doesn't count as a panel in
gretl (because the specific time-series length, $T_i$, would differ
across individuals). In some contexts it might be OK that gretl
doesn't take your sub-sample to be a panel, but if you want to apply
panel-specific methods this is a problem. You can solve it by giving
the \option{balanced} option with \cmd{smpl}. For example, supposing
your dataset contained dummy variables \texttt{gender} (with the value
1 coding for women) and \texttt{employed}, you could do
%
\begin{code}
smpl gender==1 && employed==1 --restrict --balanced
\end{code}
%
What exactly does this do? Well, let's say the years of your data are
2000, 2005 and 2010, and that some women were employed in all of those
years, giving a maximum $T_i$ value of 3. But individual 526 is a
women who was employed only in the year 2000 ($T_i = 1$). The effect
of the \option{balanced} option is then to insert ``padding rows'' of
\texttt{NA}s for the years 2005 and 2010 for individual 526, and
similarly for all individuals with $0 < T_i < 3$. Your sub-sample
then qualifies as a panel.


\section{Resampling and bootstrapping}
\label{sec:resampling}

Given an original data series \varname{x}, the command
%
\begin{code}
series xr = resample(x)
\end{code}
%
creates a new series each of whose elements is drawn at random from
the elements of \varname{x}.  If the original series has 100
observations, each element of \varname{x} is selected with probability
$1/100$ at each drawing.  Thus the effect is to ``shuffle'' the
elements of \varname{x}, with the twist that each element of
\varname{x} may appear more than once, or not at all, in \varname{xr}.

The primary use of this function is in the construction of bootstrap
confidence intervals or p-values.  Here is a simple example.  Suppose
we estimate a simple regression of $y$ on $x$ via OLS and find that
the slope coefficient has a reported $t$-ratio of 2.5 with 40 degrees
of freedom.  The two-tailed p-value for the null hypothesis that the
slope parameter equals zero is then 0.0166, using the $t(40)$
distribution.  Depending on the context, however, we may doubt whether
the ratio of coefficient to standard error truly follows the $t(40)$
distribution.  In that case we could derive a bootstrap p-value as
shown in Listing~\ref{resampling-loop}.  

Under the null hypothesis that the slope with respect to $x$ is zero,
$y$ is simply equal to its mean plus an error term.  We simulate $y$
by resampling the residuals from the initial OLS and re-estimate the
model.  We repeat this procedure a large number of times, and record
the number of cases where the absolute value of the $t$-ratio is
greater than 2.5: the proportion of such cases is our bootstrap
p-value.  For a good discussion of simulation-based tests and
bootstrapping, see Davidson and MacKinnon
(\citeyear{davidson-mackinnon04}, chapter 4); Davidson and Flachaire
(\citeyear{davidson-flachaire01}) is also instructive.

\begin{script}[htbp]
  \caption{Calculation of bootstrap p-value}
  \label{resampling-loop}
\begin{scode}
ols y 0 x
# save the residuals
genr ui = $uhat
scalar ybar = mean(y)
# number of replications for bootstrap
scalar replics = 10000
scalar tcount = 0
series ysim
loop replics --quiet
  # generate simulated y by resampling
  ysim = ybar + resample(ui)
  ols ysim 0 x
  scalar tsim = abs($coeff(x) / $stderr(x))
  tcount += (tsim > 2.5)
endloop      
printf "proportion of cases with |t| > 2.5 = %g\n", tcount / replics
\end{scode}
%$
\end{script}

    
%%% Local Variables: 
%%% mode: latex
%%% TeX-master: "gretl-guide"
%%% End: