SiZer Analysis for the Comparison of Regression Curves

Cheolwoo Park1 and Kee-Hoon Kang2

Abstract

In this article we introduce a graphical method for the test of the equality of two regression
curves. Our method is based on SiZer (SIgnificant ZERo crossing of the differences) analysis,
which is a scale-space visualization tool for statistical inferences. The proposed method does
not require any specification of smoothing parameters, it offers a device to compare in a wide
range of resolutions, instead. This enables us to find the differences between two curves that
are present at each resolution level. The extension of the proposed method to the comparison of
more than two regression curves is also done using residual analysis. A broad simulation study
is conducted to demonstrate the sample performance of the proposed tool. Applications with
two real examples are also included.

Key words: Comparison of multiple curves, Kernel density estimation, Local linear smoothing,
SiZer.

1 Introduction
One of the most important problems in statistical inference is the comparison of two or more
populations, which is necessary in a variety of contexts. The comparison of several populations
can be done by looking at measures of location, measures of dispersion or some other variables,
measured in the sample of each population. The comparison of population curves deserves much
attention, including densities, regression curves, survival functions and some other characteristic
functions of the variables of interest. This comparison can be done either in a parametric or a
nonparametric way.
In this paper, we are interested in performing the comparison of several regression curves in
a nonparametric context. A statistical challenge in this problem is testing whether there is any
statistically significant differences of these curves. Suppose that we have k different samples and
P
n = ki=1 ni independent observations from the following regression models:

Yij = fi (Xij ) + σi (Xij )εij , j = 1, . . . , ni , i = 1, . . . , k, (1.1)

where Xij ’s are covariates, the εij ’s are independently distributed random errors with mean 0
and variance 1, fi (Xi ) = E(Yi |Xi ) is the unknown regression function of the ith sample and
σi2 (Xi ) = V ar(Yi |Xi ) is the conditional variance function of the ith sample (i = 1, . . . , k).
1
Department of Statistics, University of Georgia, Athens, GA 30602, USA. E-mail : [email protected]
2
Corresponding author. Department of Statistics, Hankuk University of Foreign Studies, Yongin 449-791, Korea.
E-mail : [email protected]
 For a motivation of our work we first introduce a real data set which contains monthly ex-
penditures in Dutch guilders of Dutch households on several commodity categories. This data set
has been analyzed by Adang and Melenberg (1995), Einmahl and Van Keilegom (2003), Pardo-
Fernández et al. (2004), and others. The data set is divided into three groups by the number of
members in the household. We model the data using (1.1), with the covariate ‘log of the total
monthly expenditure’ and the response variable ‘log of the expenditure on food’ for each different
family size.

Put Figure 1 around here.

Figure 1 is an example of local linear smooths with four different smoothing levels. The observed
data points are plotted at each panel of Figure 1, with crosses for the first group (two family
members), circles for the second group (three family members), and triangles for the third group
(four family members), respectively. The local linear kernel estimates are overlaid for each data
set, dotted lines for the first, dashed lines for the second, and solid lines for the third group,
respectively. The three estimated regression curves with the smallest bandwidth (h = 0.04) in
Figure 1 (a) seem to be quite different except for the central area. For h = 0.15 (Figure 1 (b)), the
extent of difference becomes smaller except for the beginning. Note that, however, for the second
and the third groups the few observations at the beginning make the local linear fits wavy. As the
bandwidth increases (Figures 1 (c) and (d)), most of these estimated curves seem to be similar in a
large part of the covariates. As one can see, the choice of smoothing parameter plays an important
role in this comparison. In addition, we need to take into account the number of data around the
point where the test is being done. Inference based on too small number of data often fails in
drawing a significant conclusion. This example will be revisited in Section 5.
Our main concern in this paper is to develop a graphical device for testing the following hy-
pothesis of the equality of mean regression functions

H0 : f1 = f2 = · · · = fk (1.2)

versus

H1 : fi 6= fj for some i 6= j ∈ {1, 2, · · · , k}.

The problem of testing the equality of nonparametric regression curves has been widely studied in
the literature. Härdle and Marron (1990) compared two regression curves estimated by using kernel
methods. A bootstrap method for testing the equality of two regression curves was proposed by Hall
and Hart (1990). Delgado (1993), Kulasekera (1995), Kulasekera and Wang (1997), and Neumeyer
and Dette (2003) used approaches related to the empirical process. Bowman and Young (1996)
introduced the idea of using a reference band for the comparison of two nonparametric curves.
Fan (1996), and Fan and Lin (1998) proposed tests based on the adaptive Neyman test and the
wavelet thresholding techniques. Some other relevant papers include Munk and Dette (1998), Dette
and Neumeyer (2001), and Pardo-Fernández et al. (2004).

2
 Recently, Pardo-Fernández et al. (2004) proposed two types of test statistics for comparing two
regression curves. Their approach is based on the estimation of the distribution of the residuals
in each population. The calculation is not straightforward and the procedure requires bootstrap
approximation to obtain the critical values of the test. As in other nonparametric curve estima-
tion methods, their approach also requires bandwidth selection. This motivates us to consider a
graphical device to visualize the differences of two or more regression curves for a wide range of
bandwidths.
In this paper we present a graphical device, called SiZer (SIgnificant ZERO crossing of the
differences), for comparing multiple regression curves. SiZer, which originally stands for SIgnificant
ZERo crossing of the derivatives, was proposed by Chaudhuri and Marron (1999) as a graphical
goodness-of-fit test. It combines the scale-space idea of simultaneously considering a family of
smooths (e.g. local linear fits) with the statistical inference that is needed for exploratory data
analysis in the presence of noise. It brings an immediate insight into a central scientific issue in
exploratory data analysis: which features observed in a smooth of data are “really” there? By
studying the derivatives of smooths, one can compare sample data to white noise. Also, SiZer
avoids the classical problem of bandwidth selection by considering a wide range of bandwidths.
Subsequently, several SiZer tools have been developed and they have been proven to be very
powerful in many applications (see, e.g. Park et al., 2006). Hannig and Lee (2005) developed a
robust version of SiZer in a regression setting, which can be used for identifying outliers. Hannig
and Marron (2006) proposed a new method to reduce unexpected features in the SiZer map. The
existing SiZer tools, however, have limitations. For example, they are applicable only to one data
set, i.e. one curve. They compare the observed curve with one generated from an assumed (true)
model, and test the difference between them. Therefore, it is not possible to directly compare two
or more observed curves.
In this paper, a SiZer tool which is capable of comparing multiple curves is developed based on
the differences of smooths. It gives insightful information about the differences of the curves by
combining statistical inference with visualization. The method not only keeps the advantages of
the original SiZer tools but also extends their usefulness to a broader range of scientific problems;
for example seismic recordings of earthquakes and nuclear explosions, gait analysis, temperature-
precipitation patterns, brain potentials evoked by flashes of light, packet/byte counts in Internet
traffic, and so on.
This paper is organized as follows. Section 2 describes a SiZer for the comparison of two
regression curves. The extension to the comparison of multiple regression curves is done in Section
3. As a byproduct, we develop a SiZer for the comparison of two density functions. Section 4
investigates the finite sample performances of the proposed method via several simulated examples.
Applications to real data are illustrated in Section 5. Future work is discussed in Section 6. The
quantile for constructing confidence intervals in Section 2 is derived in Section 7.

3
2 SiZer for the comparison of two regression curves
The original SiZer (Chaudhuri and Marron, 1999) is a visualization method based on nonparametric
curve estimates. SiZer analysis enables statistical inference for the discovery of meaningful structure
within a data set, while doing exploratory analysis. SiZer addresses the question of which features
observed in a smooth are really there, or represent an important underlying structure, and not
simply artifacts of the sampling noise.
SiZer is based on scale-space ideas from computer vision, see Lindeberg (1994). Scale-space is
a family of kernel smooths indexed by the scale, which is the smoothing parameter or bandwidth
h. SiZer considers a wide range of bandwidths which avoids the classical problem of bandwidth
selection. Furthermore, the target of a SiZer analysis is shifted from finding features in the “true
underlying curve” to inferences about the “smoothed version of the underlying curve”, i.e. the
“curve at the given level of resolution”. The idea is that this approach uses all the information that
is available in the data at each given scale. The details underlying the statistical inference using
SiZer can be found in Chaudhuri and Marron (1999). The essential idea is that SiZer investigates
the derivative of smooths and reports the results of a large number of simultaneous hypothesis
tests. Chaudhuri and Marron (2000) studied weak convergence of the empirical scale space surface
and some related asymptotic results under appropriate regularity conditions.
While the conventional SiZer compares one sample data set to a theoretical model (white noise),
our method compares two or more data sets and tests whether they are significantly different. Thus,
we do not investigate the derivatives of curves as the original SiZer does, but study the differences
of curves. For this reason, in this paper, SiZer stands for SIgnificance of ZERo crossing of the
differences.
Let us start with the two-sample problem and extend it to multiple samples in Section 3. SiZer
applies the local linear fitting method, see e.g. Fan and Gijbels (1996), for obtaining a family of
kernel estimates in a regression setting. Precisely, at a particular point x0 , fˆi,h (x0 ) (i = 1, 2) are
obtained by fitting lines

βi0 + βi1 (x0 − Xij )

to the (Xij , Yij ), with kernel weighted least squares. Then, fˆi,h (x0 ) = β̂i0 (i = 1, 2) where β̂ i =
(β̂i0 , β̂i1 )0 minimizes
ni
X
{Yij − (βi0 + βi1 (x0 − Xij ))}2 Kh (x0 − Xij ), (2.1)
j=1

where Kh (·) = K(·/h)/h. K is a kernel function, usually a symmetric probability density function.
In this paper, we use a Gaussian kernel. Since the solution of (2.1) provides estimates of a regression
function for different bandwidths, we can construct the family of smooths parameterized by h and
the confidence intervals of the difference of two curves.
In SiZer analysis, we seek confidence intervals for the scale-space version f1,h (x) − f2,h (x) ≡
E fˆ1,h (x) − E fˆ2,h (x) instead of seeking confidence intervals for f1 (x) − f2 (x) (see Chaudhuri and

4
Marron, 1999, for discussion on this subject). From this point of view, significance of any difference
depends on the scale of resolution, h. Thus, the hypotheses we are testing are

H0 : f1,h (x0 ) = f2,h (x0 ) vs. H1 : f1,h (x0 ) 6= f2,h (x0 ) (2.2)

for a fixed point x0 .

SiZer visually displays the significance of differences between two regression functions in families
of smooths {fˆi,h (x), i = 1, 2} over both location x and scale h, using a color map. It is based on
confidence intervals for fˆ1,h (x) − fˆ2,h (x), which will be defined soon, and uses multiple comparison
level adjustment. Each pixel shows a color that gives the result of a hypothesis test in (2.2) at the
point indexed by the horizontal location x, and by the bandwidth corresponding to the row h. At
each (x, h), if the confidence interval is above (below) 0, meaning that the curves are significantly
different, i.e., f1,h (x) > f2,h (x) (f1,h (x) < f2,h (x)), then that particular map location is colored
black (white, respectively). On the other hand, if the confidence interval contains 0, meaning that
the curves are not significantly different, then that map location is given the intermediate gray.
Finally, if there are not enough data points to carry out the test, then no decision can be made and
the location is colored darker gray. To determine the darker gray areas, based on the definition of
Chaudhuri and Marron (1999), we define the estimated effective sample size (ESS), for each (x, h)
as
Pni
j=1 Kh (x − Xij )
ESSi (x, h) = i = 1, 2,
Kh (0)
ESS(x, h) = min(ESS1 (x, h), ESS2 (x, h)).

If ESS(x, h) < 5, then the corresponding pixel is colored darker gray. In order to achieve rea-
sonable computational speed, fast binned implementation of the smoothers and the corresponding
hypothesis tests are used, as discussed in Chaudhuri and Marron (1999).
Confidence intervals for f1,h (x) − f2,h (x) are of the form

fˆ1,h (x) − fˆ2,h (x) ± q · SD(

d fˆ1,h (x) − fˆ2,h (x)), (2.3)

d is the estimated standard deviation, which will be

where q is an appropriate quantile, and SD
discussed soon. Chaudhuri and Marron (2000) showed that the asymptotic simultaneous level of
the test for the entire family of hypotheses is α as long as the true q can be approximated properly.
For the approximation of the quantile, Chaudhuri and Marron (1999) suggested several methods
including pointwise Gaussian quantiles, number of independent blocks, and bootstrap. Recently,
Hannig and Marron (2006) improved the multiple comparison tests using advanced distribution
theory. A similar calculation can be done for the comparison of two regression curves. As a result,
the quantile for significance level α is defined as
µ³ ¶
−1 α ´1/(θg)
q=Φ 1−
2

5
where Φ is the standard normal distribution function and g is the number of bins. The “cluster
index” θ is given by
à !
p ∆˜
θ = 2Φ log g − 1,
2h

˜ is the distance between the pixels of the SiZer map. We use this quantile in our imple-
where ∆
mentation, and the brief derivation of the cluster index θ is provided in Section 7.
For the estimation of the standard deviation, note that fˆi,h (x) obtained from (2.1) can be
written as
ni
1 X
fˆi,h (x) = Wi,h (x, Xij )Yij
ni
j=1

where
{ŝ2 (x; h) − ŝ1 (x; h)(x − Xij )}Kh (x − Xij )
Wi,h (x, Xij ) = (2.4)
ŝ2 (x; h)ŝ0 (x; h) − ŝ1 (x; h)2
and
ni
1 X
ŝr (x; h) = (x − Xij )r Kh (x − Xij ).
ni
j=1

Then, by independence

V ar(fˆ1,h (x) − fˆ2,h (x)) = V ar(fˆ1,h (x)) + V ar(fˆ2,h (x)),

and
ni
1 X
V ar(fˆi,h (x)) = (Wi,h (x, Xij ))2 V ar(Yij )
n2i j=1
ni
1 X
= σ 2 (Xij )(Wi,h (x, Xij ))2 .
n2i j=1 i

The estimation of σi2 (Xij ) can be found in Chaudhuri and Marron (1999).

3 SiZer for the comparison of more than two regression curves

This section is devoted to testing the equality of several regression curves. We consider testing the
following scale-space version of the hypotheses in (1.2),

H0 : f1,h (x0 ) = f2,h (x0 ) = · · · = fk,h (x0 ) vs. H1 : not H0 (3.1)

However, the extension of the approach in Section 2 is not straightforward for this testing prob-
lem. We borrow an idea from Pardo-Fernández et al. (2004), and compare the residual distributions
under the null and alternative hypotheses, respectively. In other words, first we obtain two residual

6
sets by fitting local linear estimates under the null and alternative hypotheses in (3.1), and then
compare their density functions by fitting kernel density estimates. In this way, we convert the
comparison of several regression curves into the comparison of two density functions.
For obtaining the residuals, we use a pilot bandwidth hp , that is different from the bandwidth
h used for constructing the SiZer map. If one takes hp = h, then it is much more difficult to draw
a SiZer plot due to the various ranges of the residuals. Hence, we treat h and hp separately, which
means the addition of another dimension to the SiZer plot. Our visualization approach for this
problem is to draw a series of SiZer plots indexed by the pilot bandwidth hp .
Let
ni
X
σ̂i2 (x) = n−1
i Wi,hp (x, Xij )Yij2 − fˆi,h
2
p
(x)
j=1

be the estimators of the conditional variance function, where the weights Wi,hp (x, Xij )’s are given
in equation (2.4). Let fˆhp (·) be the local linear estimator of the common scale-spaced regression
function fhp (·) under H0 , which has the following form:
ni
k X
X
1
fˆhp (x) = Pk Wij,hp (x, Xij )Yij ,
i=1 ni i=1 j=1

where the weights Wij,hp (x, Xij )’s are similar to those in the equation (2.4) with
ni
k X
X
1
ŝr (x; h) = Pk (x − Xij )r Kh (x − Xij ).
i=1 ni i=1 j=1

Let (Yij − fˆi,hp (Xij ))/σ̂i (Xij ) be the estimate of the error εij from the ith population in model (1.1)
and let (Yij − fˆh (Xij ))/σ̂i (Xij ) be the estimate of the same quantity under the null hypothesis in
p

(3.1). The idea is that if H0 is true, (Yij − fˆi,hp (Xij ))/σ̂i (Xij ) and (Yij − fˆhp (Xij ))/σ̂i (Xij ) would
be quite similar and have the same distribution. Hence, we may check the equality of the regression
curves with these two types of residual distributions. Then, we need a SiZer tool for comparing the
two densities.
The construction of SiZer for the comparison of two densities is also based on the difference of
two functions. Using (Yij − fˆi,hp (Xij ))/σ̂i (Xij )’s and (Yij − fˆhp (Xij ))/σ̂i (Xij )’s, we estimate the
difference of the two residual density functions, say g1 − g0 , at a point t:
k ni
( Ã ! Ã !)
1 XX Yij − fˆi,hp (Xij ) Yij − fˆhp (Xij )
ĝ1,hp (t) − ĝ0,hp (t) = Khp t − − Khp t − (3.2)
n σ̂i (Xij ) σ̂i (Xij )
i=1 j=1

where g1 (g0 ) is a probability density function of the residuals under the H1 (H0 ). The idea for
constructing the confidence intervals for the difference of the two residual density functions is similar
to that of the regression case shown in (2.3).
In our data analysis, we try a wide range of pilot bandwidths hp to get the residuals. However,
if we show all the SiZer plots with the full selection of pilot bandwidths, the complete series of SiZer

7
plots would be too long. The simultaneous view of all these SiZer plots is hard to comprehend and
the information contained in several such plots is often redundant. This motivates us to choose only
a subset of SiZer plots. We found three plots are usually enough to convey the needed information.
Our choice among the several plots is intended to reflect ‘a wide array of trade-offs’ between
undersmoothing and oversmoothing. We first get the range of the covariates in the given data, and
divide it into 11 equally spaced values in a logarithmic scale. Then, we obtain residuals and draw
SiZer plots for each set of residuals. Finally, we choose three plots that make good representatives.
We recommend not choosing the smallest pilot bandwidth since the degree of overfitting may be too
high. We always include the second one in our analysis to see the effect of a small pilot bandwidth.

4 Simulation
Section 4.1 shows the simulated examples for comparing two regression curves. In Section 4.2, the
comparisons of two density functions and multiple regression curves are provided.

4.1 Comparison of two regression curves

Figure 2 about here.

This section provides six simulated examples. Each example has sample sizes n1 = 1000 and
n2 = 2000. X1 and X2 are generated independently from a U (0, 1) distribution. The first example
has the same constant mean 0 with independent N (0, 1) errors:

(i) Yij = εij , j = 1, . . . , ni , i = 1, 2,

where ²ij ∼ N (0, 1) for i = 1, 2. The correct SiZer plot would show no significant difference. Figure
2 (a) displays its SiZer plot. In the top panel, the thin curves display the family of smooths, which
are the differences of two local linear smooths, fˆ1,h (x) − fˆ2,h (x). These differences are located
around 0 because both samples have zero constant functions and the same noise distribution. The
SiZer map in the lower panel reports the equality test of the two samples by investigating the
confidence intervals in (2.3) at each (x, h). The horizontal locations in the SiZer map are the same
as in the top panel, and the vertical locations correspond to the logarithm of bandwidths of the
family of smooths shown as thin curves in the top panel. The white dotted curves show effective
window widths for each bandwidth, as intervals representing ±2h. Each pixel shows a color that
gives the result of a hypothesis test for the sign of the thin curve, at the point indexed by the
horizontal location, and at the bandwidth corresponding to that row. The result shows only gray
color, meaning no significant difference, as expected.
For the second example, one sample has mean 2 and the other has mean 0 with both having
error distribution N (0, 1):

(ii) Y1j = 2 + ε1j , and Y2j = ε2j

8
where ²ij ∼ N (0, 1) for i = 1, 2. Figure 2 (b) displays its SiZer plot. The upper panel shows the
difference of two smooths amounts to approximately 2, which corresponds to the difference of two
true regression functions. The SiZer map shows positive differences (black) across all scales since
the mean of the first sample is greater than that of the second sample by 2. This shows that SiZer
for two samples can successfully detect differences when the data have different mean levels.
The third example studies two different regression functions and has the following regression
models:

(iii) Y1j = sin(6πX1j ) + ε1j , and Y2j = ε2j

where ²ij ∼ N (0, 0.25) for i = 1, 2. Figure 2 (c) displays its SiZer plot. The difference of the two
smooths clearly reveals the sine curves in the top panel, and the SiZer map shows positive (black)
and negative (white) differences along the sine curve. From these three simulations, we show that
SiZer for the comparison of two regression curves performs well in various settings.

Figure 3 about here.

We make the examples more challenging in the next three simulations by adding errors with
different variances. The first error ε1j is drawn from N (0, 0.5) and the second error ε2j from
N (0, 0.25). The following curves are added to the errors:

(iv) f1 (x) = f2 (x) = sin(2πx),

(v) f1 (x) = sin(2πx), f2 (x) = sin(2πx) + x,
(vi) f1 (x) = exp(x), f2 (x) = exp(x) + sin(2πx).

Figures 3 (a)-(c) show the SiZer plots of these three examples. No significant features are found in
the SiZer map of Figure 3 (a) since they have the same regression curve. When the two regression
curves are different, the SiZer maps capture the differences and flag the linear trend (Figure 3
(b)) and the sine wave (Figure 3 (c)) as significant. These three examples show that SiZer for the
comparison of two regression curves performs well for differing error variances.
We check the behavior of the proposed method in multiple replications. We repeat the six
examples presented above 100 times and average their SiZer maps. Each averaged SiZer map is
very close to its corresponding one in Figures 2 or 3. To save the space figures are not included
here but available from the authors.

4.2 Comparison of multiple regression curves

Figure 4 about here.

As explained in Section 3, to compare several regression curves we compare density functions of

two residual sets, one under H0 and one under H1 in (3.1). To get an intuitive idea for comparing

9
two densities, we simulate three different examples. Each example has sample sizes n1 = 1000
and n2 = 2000. The top panels of Figure 4 show the difference between the two kernel density
estimates and the lower panels report the results of testing the equality of the two density functions.
In the first example, X1 and X2 are separately generated from a N (0, 1) distribution. The family
of smooths in the top panel of Figure 4 (a) is located around 0, and the SiZer map in the lower
panel shows no features since the two samples are drawn from the same density. The darker gray
in the bottom corners of the plots is due to the lack of data points, which can easily happen near
the boundaries and at small smoothing levels.
In the second example, X1 and X2 are drawn from N (2, 1) and N (0, 1) distributions, respec-
tively. Since they have different centers, the difference of the densities tends to be positive at one
side and negative at the other side. Also, note that the range of x axis is wider compared to the
first example.
In the third example, X1 and X2 are drawn from N (0, 0.25) and N (0, 1), respectively. Since
the density of X1 is more concentrated around the mean in comparison to X2 , the difference of the
two densities tends to be negative at both sides and positive in the middle.

Figure 5 about here.

We saw how SiZer compares two density functions, and move to the simulation for comparing
multiple regression curves. We simulate three examples to compare three regression curves. Each
example has sample sizes n1 = 500, n2 = 1000, and n3 = 1500. X1 , X2 and X3 are generated from
U (0, 1) independently. In the first example, the first sample is drawn from N (0, 0.25), the second
from N (0, 0.5), and the third from N (0, 0.75), and the mean regression functions are all zero:

(a) Y1j = ε1j , Y2j = ε2j , and Y3j = ε3j

where ²1j ∼ N (0, 0.25), ²2j ∼ N (0, 0.5), and ²3j ∼ N (0, 0.75).
To draw the SiZer plots for comparing the three regression curves, we first obtain two residual
sets, one under H0 and the other under H1 in (3.1) with 11 different pilot bandwidth hp as explained
in Section 3. Then, with the two residual sets, we construct a SiZer plot based on the difference
between two density estimates in (3.2) for each hp . Figures 5 (a)-(c) show the SiZer plots with the
second, the third, and the fourth pilot bandwidths. As expected, the three families of smooths look
similar to Figure 4 (a), and the three SiZer maps show no significant features. Different sample
sizes and variances do not appear to make any wrong decisions in our method.

Figure 6 about here.

In the second example the error structures remain the same as in the first, but we increase the
third mean by 2, that is

(b) Y1j = ε1j , Y2j = ε2j , and Y3j = 2 + ε3j

10
where ²1j ∼ N (0, 0.25), ²2j ∼ N (0, 0.5), and ²3j ∼ N (0, 0.75). Figures 6 (a)-(c) show the SiZer
plots of the difference between two residual sets with the second, the fifth, and the eighth pilot
bandwidths. Even with the large hp , the SiZer map clearly flags the significant difference. The
SiZer plots look similar to Figure 4 (c). The third sample with the higher mean increases the
variance (and possibly shifts the mean as well) of the residuals obtained under H0 , and this makes
the difference between the two densities.

Figure 7 about here.

In the third example, the error structures remain the same as in the first, but we add the sine
curve f3 (x) = 0.4 sin(6πx) to the third sample, that is

(c) Y1j = ε1j , Y2j = ε2j , and Y3j = 0.4 sin(6πX3j ) + ε3j

where ²1j ∼ N (0, 0.25), ²2j ∼ N (0, 0.5), and ²3j ∼ N (0, 0.75). . Compared to the second example,
the differences of the three curves are not trivial. Figures 7 (a)-(c) show the SiZer plots of the
difference between two residual sets with the second, the fourth, and the sixth pilot bandwidths.
While the SiZer maps with the second and the fourth hp ’s correctly rejecting H0 and revealing a
few significant features, the map with the sixth hp shows no significant differences. This happens
because the differences between the three curves diminish as hp increases, as we observed this
phenomenon in Figure 1. A bandwidth selection approach might conclude no difference between
these three curves, since it selects one particular bandwidth, but our method can correctly detect
the difference because it considers a wide range of bandwidths.
These three examples show that the comparison of several regression curves can be successfully
done by comparing one residual set under H0 and the other under H1 .
We repeated the three examples presented above 100 times and confirmed that each averaged
SiZer map was very close to its corresponding one in Figures 5, 6, or 7. We do not report the result
to save the space but it is available from the authors.

5 Real examples
This section is devoted to illustrating our procedure applied to real data.

Example 1. The first example revisits the example which has been discussed previously by Hall and
Hart (1990). This example was reanalyzed in Munk and Dette (1998). They compared the towns
of Coweeta and Lewiston, North Carolina, for the concentration of sulfate in rain as a function
of time in a 261 week period between 1979 and 1983. The measurement of these data were taken
weekly and the sample sizes are n1 = 220 and n2 = 215. The data actually used in comparing two
locations were the natural logarithms of the acid concentration. For a scatterplot of the adjusted
data together with kernel regression estimates, see Figure 1 of Hall and Hart (1990). In the analysis,

11
they found no indication that the error terms were correlated across time using a residual analysis.
Here, we compare the sulfate concentration as a function of time in the two towns.

Figure 8 about here.

Since this example compares two towns, a SiZer plot, which is depicted in Figure 8, is constructed
based on (2.3). Many little spikes are found in the top panel, but they are not flagged as significant
by the SiZer map. Significant features are found at low resolutions (large bandwidth) which suggest
that their grand mean is different. Since this SiZer map indicates negative values at those levels
the concentration of sulfate in rain at Coweeta is less than that at Lewiston. These results coincide
with those of Hall and Hart (1990)[Section 3.4] and Munk and Dette (1998) [Section 4.3]. Our
approach offers a more effective visual understanding in a wide range of resolution levels.

Example 2. The second real data example, as introduced in Section 1, was obtained from Data
Archive of the Journal of Applied Econometrics, and consists of monthly expenditures in Dutch
guilders of Dutch households on several commodity categories, as well as on a number of background
variables. We will compare the regression curves for three groups of households: households con-
sisting of 2 members (1575 in total), 3 members (377 in total) and 4 members (292 in total). The
data were collected from April 1984 to September 1987, and the average was taken over the 42
months for each household. The autocorrelation plots of the data (not reported), do not indicate
any evidence of correlation.

Figure 9 about here.

First, we compare two groups of three pairs and then the three groups altogether. Figure 9
shows SiZer plots for comparing (a) the two and three, (b) the two and four, and (c) the three and
four member groups. Since two curves are being compared, SiZer plots are constructed based on
(2.3). According to the SiZer maps in the lower panels no difference is found because the other
SiZer maps show either gray (not significant) or darker gray (lack of data). A spike and a valley
are observed at the beginning in all the families of smooths (with small bandwidths), but they are
not flagged as significant since there are not sufficient data around the points where the tests are
being performed and the map shows darker gray. As mentioned in Figure 1, there are very few
observations at the beginning for the second and the third groups.

Figure 10 about here.

Figure 10 shows SiZer plots for comparing the three groups simultaneously using their residual
distributions. The result shows the same conclusion as the comparison of two groups, that is no
difference among three groups is found at the second, the fourth, and the sixth pilot bandwidths.

12
6 Future work
Another approach using ANOVA type statistics can be developed for the comparison of multiple
curves. Instead of converting the problem into comparing two densities, this approach compares
multiple curves directly at each point. The test statistics for comparing the curves at x can be
roughly written as
Pk ˆ
i=1 (fi,h (x) − fˆh (x))2
(Constant) × Pk Pni
i=1 j=1 (Yij − fˆi,h (Xij ))2 Kh (x − Xij )

where fˆi,h is a local linear fit using ith sample and fˆh using the combined sample under the null
hypothesis. This statistic mimics the ratio of variations from the model and the error in ANOVA.
To conduct a test, one needs to find the approximate distribution of this statistic and its degree of
freedom. Also, an appropriate multiple adjustment needs to be designed for SiZer. The advantage
of this approach over the previous one is that one can compare several curves directly and get
the information of their local differences, which reflects the original intention of SiZer. If some
differences are found among the curves, multiple pairwise comparisons can be performed as done
in ANOVA analysis. We are currently developing this approach.
SiZer is a useful tool to find meaningful structures in the given data, but its usefulness can
be limited in the case of dependent data because we assume independent errors. For dependent
data, significant features appear in SiZer, which are due to the presence of dependence. The great
challenge in time series for applications of SiZer in the trend (of difference between two curves)
estimation context is that “trend” and“serial dependence artifacts” cannot be distinguished. For
the one curve case, Rondonotti, Marron, and Park (2007) extended the original SiZer to SiZer for
time series. This tool finds features of the underlying trend function, while taking into account
the dependence structure using the estimated autocovariance function. Another version of SiZer
for dependent data is so-called ‘Dependent SiZer’ proposed by Park, Marron, and Rondonotti
(2004). Dependent SiZer has a slightly different goal from SiZer for time series. It uses a true
autocovariance function of an assumed model instead of estimating it from the observed data. By
doing so, a goodness of fit test can be conducted and we can see how different the behavior of the
data is from that of the assumed model. For two curves, confidence intervals for f1,h (x) − f2,h (x)
have the same form as the independent case in (2.3), but the estimation of the autocovariance
function and the choice of the quantile q should be adjusted. Some work was done in Rondonotti,
Marron, and Park (2007), but we plan to improve their method and extend it to the comparison
of several time series in the future.

13
7 Appendix
To color the pixels SiZer checks whether the difference of the estimates of the two regression
functions
  
Xni ni
X
β̂i0 = c−1  Kh (x − Xij )Yij   (x − Xij )2 Kh (x − Xij )
i
j=1 j=1
  
Xni ni
X
− c−1  (x − Xij ) Kh (x − Xij )  (x − Xij )Kh (x − Xij ) Yij  , (7.1)
i
j=1 j=1
    2
Xni ni
X ni
X
ci =  Kh (x − Xij )  (x − Xij )2 Kh (x − Xij ) −  (x − Xij ) Kh (x − Xij ) ,
j=1 j=1 j=1

for i = 1, 2, is significantly different from 0. By appropriate binning procedure, the points Xij
can be converted into the fixed design points Xl = l∆, l = 1, . . . , m, where ∆ > 0 is the distance
between design points and m is the number of grid points. If x is away from the boundary, it
follows from symmetry of the kernel that
m
X
(x − Xl ) Kh (x − Xl ) ≈ 0,
l=1

which means that the second term in (7.1) disappears.

˜ denote the distance between the pixels of the SiZer map, and p = ∆/∆,
Let ∆ ˜ which denotes
the number of data points per SiZer column. For simplicity of notation we can assume that p is a
positive integer.
Let g be the number of pixels on each row, and T1 , . . . , Tg denote test statistics of a row in the
SiZer map. Then Tk is proportional to β̂10 − β̂20 calculated for x = k ∆ ˜ = kp∆. In particular
m
X
h
Tk ≈ Wkp−q (Y1,q − Y2,q ).
q=1

h
The form of the Wkp−q is given in the first term of (7.1) with appropriate binning adjustment.
h
Note that Wkp−q is proportional to Kh/∆ (kp − q) and thus the weights Wqh are proportional to the
Gaussian kernel with standard deviation h/∆.
If the null hypothesis of two mean curves being equal is true, then Y1,q − Y2,q are independently
distributed Gaussian random variables with mean zero.
The full joint distribution of T1 , ..., Tg depends on the correlation between them and this corre-

14
lation is approximated by

ρk = corr(Tl , Tl+k )
P h h
q Wq−kp Wq
= P h 2
q (Wq )
R
Kh/∆ (x − kp) Kh/∆ (x) dx
≈ R
Kh/∆ (x)2 dx
˜ 2 /(4h2 )
= e−(k∆)

where the third line follows by replacing the sums by integral approximations and the last step
˜
follows by observing that p∆ = ∆.
By the Theorem 1 of Hannig and Marron (2006), for each fixed g we can get k step correlation
ρk,g such as
2 C 2 /(4 log g)
ρk,g = e−k ,
√
˜
by setting ∆/h = C/ log g and

k2 C 2
lim log(g)(1 − ρk,g ) = .
g→∞ 4
Following the similar arguments in paragraphs after Theorem 1 of Hannig and Marron (2006), we
conclude that in the case of SiZer for comparing two curves,
· ¸
P max Ti ≤ x ≈ Φ(x)θg ,
i=1,...,g

where the cluster index θ

à !
p ∆˜
θ = 2Φ log g − 1.
2h

Acknowledgement

We would like to thank Jan Hannig for helpful discussion. The first author was supported by
UGA Faculty Research Grants Program. The second author was supported by Korea Research
Foundation Grant funded by Korea Government (MOEHRD, Basic Research Promotion Fund)
(KRF-2005-015-C00069)

References
[1] Adang, P. J. M. and Melenberg, B. (1995). Nonnegativity constraints and intratemporal un-
certainty in multi-good life-cycle models. Journal of Applied Econometrics, 10, 1–15.

15
 [2] Bowman, A. and Young, S. (1996). Graphical Comparison of Nonparametric Curves. Applied
Statistics, 45, 83–98.

[3] Chaudhuri, P. and Marron, J. S. (1999). SiZer for exploration of structures in curves. Journal
of the American Statistical Association, 94, 807–823.

[4] Chaudhuri, P. and Marron, J. S. (2000). Scale space view of curve estimation. The Annals of
Statistics, 28, 408–428.

[5] Delgado, M. A. (1993). Testing the equality of nonparametric regression curve. Statistics &
Probability Letters, 17, 199–204

[6] Dette, H. and Neumeyer, N. (2001). Noparametric analysis of covariance. The Annals of Sta-
tistics, 29, 1361–1400.

[7] Einmahl, J. H. J. and Van Keilegom, I. (2003). Goodness-of-fit tests in nonparametric regres-
sion. Discussion Paper.

[8] Fan, J. (1996). Test of significance based on wavelet thresholding and Neyman’s truncation.
Journal of the American Statistical Association, 91, 674–688.

[9] Fan, J., and Gijbels, I. (1996). Local Polynomial Modelling and Its Applications. Chapman &
Hall, London.

[10] Fan, J. and Lin, S. (1998). Test of significance when data are curves. Journal of the American
Statistical Association, 93, 1007–1021.

[11] Hall, P. and Hart, J. D. (1990). Bootstrap test for difference between means in nonparametric
regreeeion. Journal of the American Statistical Association, 85, 1039–1049.

[12] Hannig, J. and Lee, T. (2005). Robust SiZer for exploration of regression structures and outlier
detection. Journal of Computational & Graphical Statistics, 15, 101–117 .

[13] Hannig, J. and Marron, J. S. (2006). Advanced Distribution Theory for SiZer. Journal of the
American Statistical Association, 101, 484–499 .

[14] Härdle W. and Marron, J. S. (1990). Semiparametric comparison of regression curves. The
Annals of Statistics, 13, 63–89.

[15] Kulasekera, K. B. (1995). Comparison of regression curves using quasi-residuals. Journal of

the American Statistical Association, 90, 1085–1093.

[16] Kulasekera, K. B. and Wang, J. (1995). Smoothing parameter selection for power optimality
in testing of regression curves. Journal of the American Statistical Association, 92, 500–511.

[17] Lindeberg, T. (1994). Scale-Space Theory in Computer Vision. Kluwer, Boston.

16
[18] Munk, A. and Dette, H. (1998). Nonparametric comparison of several regression functions:
exact and asymptotic theory. The Annals of Statistics, 26, 2339-2368.

[19] Neumeyer, N. and Dette, H. (2003). Noparametric comparison of regression curves: an empir-
ical process approach. The Annals of Statistics, 31, 880–920.

[20] Pardo-Fernández, J. C., Van Keilegom, I. and González-Manteiga, W. (2004). Comparison of

regression curves based on the estimation of the error distribution. Discussion Paper DP0416.
Available at http://www.stat.ucl.ac.be/ISpub/ISdp.html

[21] Park, C., Marron, J. S. and Rondonotti, V. (2004). Dependent SiZer: goodness of fit tests for
time series models. Journal of Applied Statistics, 31, 999–1017.

[22] Park, C., Hernandez-Campos, F., Le, L., Marron, J. S., Park, J., Pipiras, V., Smith, F. D.,
Smith, R. L., Trovero, M., and Zhu, Z. (2006). Long range dependence analysis of Internet
traffic. Under revision, Technometrics.

[23] Rondonotti, V., Marron, J. S., and Park, C. (2007). SiZer for time series: a new approach to
the analysis of trends. Electronic Journal of Statistics, 1, 268–289.

17
 13 13
2 member
3 member
12.5 4 member 12.5
est2
est3
12 12
est4

11.5 11.5

log(expenditure on food)
log(expenditure on food)

11 11

10.5 10.5

10 10

9.5 9.5

9 9

8.5 8.5

8 8
9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14
log(total monthly expenditure) log(total monthly expenditure)

(a) h = 0.04 (b) h = 0.15

13 13

12.5 12.5

12 12

11.5 11.5
log(expenditure on food)

log(expenditure on food)

11 11

10.5 10.5

10 10

9.5 9.5

9 9

8.5 8.5

8 8
9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14
log(total monthly expenditure) log(total monthly expenditure)

(c) h = 0.30 (d) h = 1.00

Figure 1: Local linear estimates for each group with different smoothing levels.

18
 2.5 1
0.5 0.5

2 0
0
−0.5
−0.5 1.5 −1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

0 0 0

−0.5 −0.5 −0.5

log10(h)

log10(h)

log10(h)

−1 −1 −1

−1.5 −1.5 −1.5

−2 −2 −2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Example (i) (b) Example (ii) (c) Example (iii)

Figure 2: SiZer plots for comparing two regression curves. The two samples are drawn from (a)
normal errors with the same mean, (b) normal errors with different means, and (c) normal errors
with a sine curve versus a constant mean.

19
 0.5
0 1

0 −0.5 0

−1 −1
−0.5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

0 0 0

−0.5 −0.5 −0.5

log10(h)

log10(h)

log10(h)

−1 −1 −1

−1.5 −1.5 −1.5

−2 −2 −2

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Example (iv) (b) Example (v) (c) Example (vi)

Figure 3: SiZer plots for comparing two regression curves. The two samples have different variances
and the two true regression curves are (a) sine curves, (b) sine curves with a linear function versus
a zero mean, and (c) exponential functions with a sine curve versus a zero mean.

20
 0.4 0.6
0.1
0.05 0.2 0.4

0 0 0.2
−0.05
−0.2 0
−0.1
−0.4
−0.2
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 4 5 −2 −1 0 1 2 3

1
0.5 0.5 0.5

0
log10(h)

log10(h)

0
(h)

0
10

−0.5
log

−0.5 −0.5
−1 −1
−1
−1.5 −1.5
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 4 5 −2 −1 0 1 2 3
(a) Same density (b) Different means (c) Different variances

Figure 4: SiZer plots for the comparison of two densities. The two samples are drawn from (a)
normal errors with the same mean, (b) normal errors with different means, and (c) normal errors
with different variances.

21
 0.08
0.08
0.06
0.04
0.06
0.04

0.02 0.04 0.02

0 0.02

−0.02 0
0

−0.04
−0.02
−0.02
−0.06
−0.04
−0.08
−0.06 −0.04
−0.1
−0.08
−0.12 −0.06
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

0.5 0.5 0.5

0 0 0
log10(h)

log10(h)

log10(h)

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

(a) Example (a) hp (2) (b) Example (a) hp (3) (c) Example (a) hp (4)

Figure 5: SiZer plots for comparing the densities of two sets of residuals. The three samples are
from normal errors with the same regression function and different variances.

22
 0.1 0.1 0.1

0.05 0.05
0.05

0 0
0
−0.05 −0.05
−0.05
−0.1 −0.1
−0.1
−0.15
−0.15
−0.15
−0.2
−0.2
−0.2 −0.25
−0.25

−0.25 −0.3
−0.3

−0.3 −0.35
−0.35
−20 −15 −10 −5 0 5 −6 −4 −2 0 2 4 −6 −4 −2 0 2 4

1.5
1 1

1
0.5 0.5

0.5
log10(h)

log10(h)

log10(h)

0 0

0
−0.5 −0.5

−0.5
−1 −1

−20 −15 −10 −5 0 5 −6 −4 −2 0 2 4 −6 −4 −2 0 2 4

(a) Example (b) hp (2) (b) Example (b) hp (5) (c) Example (b) hp (8)

Figure 6: SiZer plots for comparing the densities of two sets of residuals. The three samples are
from normal errors with the different mean and different variances.

23
 0.06 0.08 0.04

0.04
0.06
0.02
0.02
0.04
0 0
0.02
−0.02
0 −0.02
−0.04

−0.06 −0.02
−0.04

−0.08 −0.04
−0.06
−0.1 −0.06

−0.12 −0.08
−0.08
−0.14
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 4

0.5
0.5 0.5

0
0 0
log10(h)

log10(h)

log10(h)

−0.5
−0.5 −0.5

−1
−1 −1

−1.5
−1.5 −1.5
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 4 −3 −2 −1 0 1 2 3 4

(a) Example (c) hp (2) (b) Example (c) hp (4) (c) Example (c) hp (6)

Figure 7: SiZer plots for comparing the densities of two sets of residuals. The three samples are
from normal errors with the different regression function and different variances.

24
 2

1.5

0.5

−0.5

−1

−1.5

−2

50 100 150 200 250

2.5

1.5
log10(h)

0.5

0
50 100 150 200 250

Figure 8: SiZer plots of North Carolina rain data in Example 1.

25
 4
1.5 1

3
1 0.5
2

0.5 0
1

0 −0.5 0

−1
−0.5 −1

−2
−1 −1.5
−3

−1.5 −2
−4
10 10.5 11 11.5 12 12.5 13 13.5 14 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0
0
log10(h)

log10(h)

log10(h)

−0.2 −0.2
−0.2
−0.4 −0.4
−0.4
−0.6 −0.6

−0.6 −0.8 −0.8

−0.8 −1 −1

−1.2 −1.2
−1
−1.4 −1.4
10 10.5 11 11.5 12 12.5 13 13.5 14 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14

(a) 2 and 3 (b) 2 and 4 (c) 3 and 4

Figure 9: SiZer plots of Dutch households data in Example 2. (a) The two and three, (b) the two
and four, and (c) the three and four members groups are compared each other, respectively.

26
 0.05
0.06 0.1
0.04

0.03 0.04

0.05
0.02
0.02
0.01

0 0 0

−0.01
−0.02
−0.02
−0.05
−0.03 −0.04

−0.04
−0.06
−0.05 −0.1
−6 −5 −4 −3 −2 −1 0 1 2 −5 −4 −3 −2 −1 0 1 2 −5 −4 −3 −2 −1 0 1 2 3

1 1 1

0.5 0.5 0.5

0 0 0
log10(h)

log10(h)

log10(h)

−0.5 −0.5 −0.5

−1 −1 −1

−6 −5 −4 −3 −2 −1 0 1 2 −5 −4 −3 −2 −1 0 1 2 −5 −4 −3 −2 −1 0 1 2 3

(a) Second hp (b) Fourth hp (c) Sixth hp

Figure 10: SiZer plots of residuals obtained from Dutch households data.

27