WGRWGGG
WGRWGGG
WGRWGGG
Computation
Gustavo H. A. Pereira
To cite this article: Gustavo H. A. Pereira (2019) On quantile residuals in beta regression,
Communications in Statistics - Simulation and Computation, 48:1, 302-316, DOI:
10.1080/03610918.2017.1381740
1. Introduction
Beta regression is often used to model the relationship between a dependent variable that
assumes values on the open interval (0, 1) and a set of predictor variables. It is useful to
model rates and proportions, quantities that are common in different areas. Beta regression
was introduced by Paolino (2001), Kieschnick and McCullough (2003), and Ferrari and
Cribari-Neto (2004). Recent contributions in this area were made by Bayer and Cribari-Neto
(2013), Figueroa-Zúñiga et al. (2013), Cribari-Neto and Queiroz (2014), Carrasco et al.
(2014), Espinheira et al. (2014), and Latif and Zafar Yab (2015).
Residuals play an important role in checking model adequacy and in the identification of
outliers and influential observations. An important challenge in regression models is to find
residuals whose distribution is well approximated by the standard normal distribution. When
the distribution of a residual is not well approximated by the standard normal distribution, it
is difficult to interpret residuals plots. This is especially true because, when this happens, in
general, residuals are not approximately identically distributed (see, e.g., Table 1 of Anholeto
et al., 2014).
Several residuals were proposed to perform diagnostic analysis in beta regression. Ferrari
and Cribari-Neto (2004) suggested the use of two residuals: standardized and deviance.
Espinheira et al. (2008) introduced two other residuals based on Fisher’s scoring iterative
algorithm for estimating the parameters of the beta regression model, and named them as
standardized weighted residual 1 and 2. Based on Monte Carlo simulation studies, these
authors concluded that the distribution of the two residuals proposed in their article is better
CONTACT Gustavo H. A. Pereira ghapereira@terra.com.br Department of Statistics, Federal University of São Carlos,
São Carlos - SP -, Brazil.
© Taylor & Francis Group, LLC
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 303
approximated by the standard normal distribution than that of the standardized residual.
They also revealed the shortcomings of the deviance residual and suggested it should not be
used in the beta regression model. In addition, they performed diagnostic analysis in three
applications and concluded that the standardized weighted residual 2 is a better residual than
the standardized weighted residual 1.
A different approach was used by Anholeto et al. (2014). They obtained approximations,
under suitable regularity conditions, for the means and variances of the standardized resid-
ual and of the standardized weighted residual 1 to order O(n−1 ), where n is the sample size.
Using these approximations, they introduced new adjusted residuals based on the original
residuals and their approximate moments. Employing Monte Carlo simulation studies, the
authors concluded that the distribution of the adjusted standardized weighted residual 1 is
better approximated by the standard normal distribution than that of the residuals proposed
by Espinheira et al. (2008).
The quantile residual (Dunn and Smyth, 1996) is a simple and general residual. It is usu-
ally used to perform diagnostic analysis in complex regression models, such as the generalized
additive models for location scale and shape (Rigby and Stasinopoulos, 2005). However, the
properties of the quantile residual in the beta regression model have never been studied. The
quantile residual is asymptotically standard normally distributed if the parameters are con-
sistently estimated (Dunn and Smyth, 1996), but it is important to study its properties when
sample size is not large. Ospina and Ferrari (2012) proposed using the randomized quantile
residual (randomization is necessary when response variable is discrete or a mixture of a dis-
crete and a continuous variable) to assess the overall adequacy of the zero-or-one inflated beta
regression model. Based on simulation studies, the authors concluded that the randomized
quantile residual perform well in detecting whether the distribution assumption is incorrect.
However, simulation results were not presented in the article and the authors did not compare
the randomized quantile residual with other residuals.
The main goal of this article is to study the properties of the quantile residual in the beta
regression model. Using Monte Carlo simulation studies and three applications, we compare
the behavior of the quantile residual, the standardized weighted residual 1 and 2, and the
adjusted standardized weighted residual 1 in several scenarios.
The remainder of the article is organized as follows. Section 2 defines the beta regression
model and the residuals considered in this article. In the following section, Monte Carlo sim-
ulation studies are performed to compare the properties of the defined residuals. Three appli-
cations are presented in Section 4. Concluding remarks are provided in Section 5.
3. Simulation studies
Monte Carlo simulation studies were performed using a beta regression model, in which,
μi
log = β1 + β2 xi2 + β3 xi3 , i = 1, 2, . . . , n. (6)
1 − μi
Five scenarios were considered. The first four scenarios are those used by Anholeto et al.
(2014). The fifth scenario was inspired by real data. The median of the observed values of
the response variable is on the interval (0.15, 0.30) in certain databases used in some beta
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 305
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 10 and Scenario I.
a
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . . − . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . − . . . . . − . − . . . . . . . .
. . − . − . . . . − . − . . . . . . . .
. − . . . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. − . . . . . . − . − . . . . . . . .
. − . . − . . . . − . − . . . . . . . .
. . − . − . . . . − . − . . . . . . . .
. − . . − . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
Mean . . − . . . . − . − . . . . . . . .
SD . . . . . . . . . . . . . . .
regression papers, such as Prater’s gasoline data and food expenditure data (Ferrari and
Cribari-Neto, 2004). For this reason, in the fifth scenario the median of μ1 , μ2 , . . . , μn is on
the interval (0.15, 0.30). For the first three scenarios, the covariates values were generated
as independent draws from the standard uniform distribution. For the other scenarios, the
xi2 values were generated from the exponential distribution with mean equal to 2 and the
xi3 values from the standard normal distribution. The covariates values remained constant
throughout the simulations. In the first scenario (Scenario I), β1 = −2.3, β2 = −1.1, and
β3 = −0.7, which resulted in mean response values close to zero, μ ∈ (0.016, 0.091). In
the second scenario (Scenario II), β1 = −0.3, β2 = 0.3, and β3 = 0.7, which resulted in
mean response values close to 0.5, μ ∈ (0.425, 0.668). In the following scenario (Scenario
III), β1 = 4.0, β2 = −0.3, and β3 = −0.5, which resulted in mean response values close
to one, μ ∈ (0.960, 0.982). In the fourth scenario (Scenario IV), β1 = 1.0, β2 = 0.5, and
β3 = −0.5, which yielded μ ∈ (0.608, 0.924). In the last scenario (Scenario V), β1 = −2.5,
β2 = 2.0, and β3 = −0.5, which yielded μ ∈ (0.057, 0.687). For each scenario, two differ-
ent values of φ were considered: 10 and 100. All results are based on 5,000 Monte Carlo
replications and n = 16. Simulations were performed using the Ox language (Doornik,
2007).
Tables 1–4 present the sample mean, variance, skewness, and kurtosis coefficients for ri∗ ,
∗a q
ri , and ri out of 5,000 values for φ = 10 and Scenarios I, II, IV, and V. Table for Scenario III
was omitted because their results are similar to Scenario I. Results for ri∗∗ were also omitted
from the tables. As already noted by Anholeto et al. (2014) and confirmed in our simula-
tions, the distribution of ri∗∗ is worse approximated by the standard normal distribution than
that of ri∗ in all scenarios. Residual ri∗∗ has mean, skewness, and kurtosis coefficients similar
to ri∗ , but its variance is far from 1 than that of ri∗ . The tables also contain the value of the
Anderson–Darling statistic (Stephens, 1986) used to test whether each residual is standard
normally distributed. The critical value for this Anderson–Darling test is 2.492 at the 5% sig-
nificance level. However, as each test uses 5,000 replications, it is expected that most of them
reject the normality hypothesis. Therefore, the value of the Anderson–Darling statistic is used
as a closeness measure between each residual distribution and standard normal distribution.
306 G. H. A. PEREIRA
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 10 and Scenario II.
a
. . . . . . . . . . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . . . . . . . .
. − . − . − . . . . . . . . . . . . .
. . . . . . . − . − . . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. − . − . − . . . . . . . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . − . − . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
Mean . . . . . . . . − . . . . . . .
SD . . . . . . . . . . . . . . .
In all scenarios, the mean of the three residuals is close to zero. In scenarios II, IV, and
a
V, the variance of ri∗ is closer to one than that of the other two residuals, but none of the
residuals have variance very close to one for all 16 observations in scenarios I and III. The
skewness coefficient is close to zero in all scenarios for ri and only in scenario II for ri∗ and
q
a a
ri∗ . In scenarios IV and V, the skewness coefficient of ri∗ and ri∗ is not far from zero, but in
scenarios I and III these residuals are considerably skewed. The kurtosis coefficient is not very
close to 3 in any of the scenarios for the three residuals. However, only in scenarios I and III
a
and residuals ri∗ and ri∗ , the kurtosis coefficient is extremely far from 3 for some observations.
In scenarios I and III, the Anderson–Darling statistic’s value (ADS) is considerably smaller for
a
ri than that for ri∗ to all 16 observations. On the other hand, in scenario II, the ADS is smaller
q
a
for ri∗ than that for ri to 15 out of 16 observations. In scenarios IV and V, the sample mean
q
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 10 and Scenario IV.
a
i μi ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗
a
ri∗ riq
. . . − . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. . . . . . . . . . . . . . . .
. − . − . . . . . . . − . . . . . . .
. − . − . − . . . . . . . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. . − . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
Mean . . . . . . . . − . . . . . . .
SD . . . . . . . . . . . . . . .
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 307
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 10 and Scenario V.
a
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . . . . . . − . − . . . . . . . .
. − . . . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . − . − . . . . . . . . . . . . .
. . − . . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . . . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . − . − . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
Mean . . . . . . − . − . . . . . . . .
SD . . . . . . . . . . . . . . .
a
and the sample standard deviation of the ADS is smaller for ri than that for ri∗ , but the latter
q
a
has a lower ADS for several observations. Residual ri∗ has a larger ADS than ri and ri∗ in
q
a
scenarios IV and V but a similar ADS to ri∗ in scenarios I and III and a similar ADS to ri in
q
scenario II.
Tables 5–8 present the simulation results for φ = 100. The main differences are observed
in the skewness coefficient. In all scenarios, its absolute value reduces considerably when
a
φ = 100, especially for residuals ri∗ and ri∗ whose skewness coefficients are shown higher for
a
φ = 10. Even in scenarios I and III, the skewness coefficient of ri∗ and of ri∗ are not far from
zero when φ = 100. As a consequence, except for scenario II, ADS decreases substantially for
a
residuals ri∗ and ri∗ and it reduces slightly for ri . When φ = 100, in scenarios IV and V, the
q
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 100 and
a
Scenario I.
Mean Variance Skewness Kurtosis A-D Statistic
i μi ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗ ri∗
a
riq
. . . . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . − . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . . − . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . . . . . . − . − . − . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. − . . − . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
Mean . . . . . . − . − . . . . . . . .
SD . . . . . . . . . . . . . . .
308 G. H. A. PEREIRA
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 100 and
a
Scenario II.
Mean Variance Skewness Kurtosis A-D Statistic
. . . . . . . − . − . − . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. − . − . − . . . . . . . . . . . . .
. − . − . − . . . . . . . . . . . . .
. − . − . − . . . . . . . . . . . . .
. − . − . − . . . . . . . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . . . . . . . .
. − . − . − . . . . − . − . − . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
Mean . . . . . . . . . . . . . . .
SD . . . . . . . . . . . . . . .
a
sample mean and the sample standard deviation of the ADS are smaller for ri∗ than that for
q
ri , but the latter is still better in scenarios I and III.
Simulations were also carried out for n = 40. Table 9 summarizes the results of the ADS
for all scenarios with n = 16 and n = 40. In all scenarios, the estimates of the mean, standard
q
deviation, and quartiles of the ADS reduce considerably for ri when sample size increases
∗ ∗a
from n = 16 to n = 40. The reduction is smaller for ri and ri in some scenarios and it does
not exist in others. As a consequence, except for scenario II, the sample mean and the sample
a
standard deviation of the ADS are lower for ri than that for ri∗ and ri∗ when n = 40. In scenario
q
a
II, ri and ri∗ have similar ADS’s sample mean and ri∗ has the smallest values.
q
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 100 and Scenario
a
IV.
Mean Variance Skewness Kurtosis A-D Statistic
. . . . . . . . . . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. − . − . − . . . . . . − . . . . . . .
. − . − . − . . . . . . . . . . . . .
. . − . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. . . . . . . . . − . . . . . . .
. − . − . − . . . . . . . . . . . . .
Mean . . . . . . . . − . . . . . . .
SD . . . . . . . . . . . . . . .
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 309
Table . Simulation results of ri∗ , ri∗ , and riq , using logit link function and n = 16 for φ = 100 and Scenario
a
V.
Mean Variance Skewness Kurtosis A-D Statistic
i μi ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗ ∗a
ri riq ri∗
a
ri∗ riq
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . . − . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . − . . . . . − . − . − . . . . . . .
. . . . . . . − . − . − . . . . . . .
. − . . − . . . . . . . . . . . . .
. − . − . − . . . . − . − . − . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . . . . . . − . − . − . . . . . . .
. . . . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . − . − . . . . − . − . . . . . . . .
. . . . . . . − . − . . . . . . . .
. − . . . . . . − . − . . . . . . . .
Mean . . . . . . − . − . . . . . . . .
SD . . . . . . . . . . . . . . .
The logit is the most used link function in beta regression because it facilitates interpre-
tation of the parameters. However, it is a symmetric link function and this may facilitate the
normal approximation of the distribution of the residuals considered in this article. In order
to check whether the results change considerably when an asymmetrical link function is used,
the same scenarios were considered in a beta regression model with complementary log-log
link function (McCullagh and Nelder, 1989). The value of the parameters in each scenario
was changed in such a way that the mean, standard deviation, minimum, and maximum of
the vector (μ1 , μ2 , . . . , μn ) keep almost the same of the logit link function case. Table 10
summarizes the results of the ADS. The adjusted weighted standardized residual 1 was not
considered because it was derived by Anholeto et al. (2014) only for the logit link function.
Except in scenario II where ri and ri∗ have similar behavior, the sample mean of the ADS
q
is considerably smaller for ri than that for ri∗ , especially when n = 40. For n = 16, the dis-
q
tribution of ri and ri∗ show worse approximation to the standard normal distribution when
q
the complementary log-log link function is used instead of the logit, notably in the scenar-
ios where the predictor variables are not generated from the uniform distribution. Moreover,
q
ri has similar behavior using these two link functions when n = 40. This suggests that the
distribution of these residuals approximate to standard normal ones when the sample size is
q
larger if the complementary log-log link function is used, being n = 40 large enough for ri .
4. Applications
In this section, we present three applications, two of them based on real data and one that
employs simulated data. Applications were performed using Ox and the betareg package for
the R software.
In Section 3, we studied whether the distribution of the residuals considered in this article
are well approximated by the standard normal distribution. Other essential property is the
ability to identify model misspecification. The first application uses data on the 2006 Peruvian
a
general election to investigate whether ri∗ , ri∗∗ , ri∗ , and ri have this property. The response
q
variable is the proportion of blank votes of an electoral district and the single predictor
310 G. H. A. PEREIRA
Table . Anderson–Darling statistic’s values for the simulations of ri∗ , ri∗ , and riq , considering the logit link
a
function and Scenarios = (I, II, IV, V), φ = (10, 100), and n = (16, 40).
Quartiles
Standard
Scenario φ n Residual Mean deviation Minimum Q1 Q2 Q3 Maximum
variable is the Human Development Index. It is known that beta regression does not provide
a reasonable fit to these data, see, for example, Bayes, et al. (2012) and Lemonte and Bazán
(2016), who proposed alternative models because of the lack of fit. We fitted beta regres-
sion model with logit link function for Peruvian election data and obtained the four set of
residuals. Figure 1 presents a plot of residuals against linear predictor and a half-normal
a
residual plot with simulated envelope (Atkinson, 1985, sec. 4.2) for ri∗ , ri∗∗ , ri∗ , and ri . Clearly,
q
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 311
Table . Anderson–Darling statistic’s values for the simulations of ri∗ , ri∗ , and riq , considering the comple-
a
mentary log-log link function and Scenarios = (I, II, IV, V), φ = (10, 100), and n = (16, 40).
Quartiles
Standard
Scenario φ n Residual Mean deviation Minimum Q1 Q2 Q3 Maximum
the four residuals suggest lack of fit of the beta regression model, in agreement with previous
findings.
The second application uses data on men professional tennis. The aim is to study the pro-
portion of service games won as a function of the number of aces per 5 games of service
(AP5GS) and the first service percentage (FSP), that is, the percentage of services that the
player got in on his first try. Data were obtained from ATP Website (ATP Tour Inc, 1994) and
refer to the performance of the best 50 tennis players of the world in the 10 first months of
2015. Figure 2 presents the residuals plots of the beta regression model with logit link func-
tion. None of the residuals suggest model misspecification.
Table 11 presents parameter estimates and standard errors. When the logit link function
is used in beta regression, the exponential of the parameters are odds ratios (Ferrari and
Table . Parameter estimates and standard errors for tennis data.
Estimate Standard error Exp(estimate)
Figure . Residuals against linear predictor (left) and half-normal residual plots with simulated envelope
(right) for Peruvian election data.
Cribari-Neto, 2004). It is estimated that for every ace per 5 games increase, FSP held constant,
the odds of the proportion of service games won increase by 20.6%. It is also estimated that
for every percentage point increase in the first service percentage, AP5GS held constant, the
odds of the proportion of service games won increase by 4.2%. The pseudo R2 of the fitted
regression model is 0.719, suggesting that a substantial proportion of the variability of the
proportion of service games won can be explained by a beta regression model with AP5GS
and FSP as covariates.
In the second application, none of the residuals suggest model misspecification. However,
a
the simulation studies presented in Section 3 indicated that ri∗ , ri∗∗ , and ri∗ could suggest lack
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 313
Figure . Residuals against linear predictor (left) and half-normal residual plots with simulated envelope
(right) for tennis data.
of fit in some scenarios when the beta regression model is correct. To illustrate this fact, we
simulated data with 160 observations, which is 10 times greater than n = 16 used in Section
3, to reduce sample variability. Covariates and response were generated as in Scenario 1 of the
simulation studies.
Figure 3 presents the residuals plots of the beta regression model with logit link function
a
for this third application. Residuals ri∗ , ri∗∗ , and ri∗ incorrectly suggest model misspecification
because there are several residuals lower than −3 and four out of the five higher residuals (in
absolute value) are outside the boundaries of the simulated envelope. On the other hand, plots
q
of ri do not suggest lack of fit.
314 G. H. A. PEREIRA
Figure . Residuals against linear predictor (left) and half-normal residual plots with simulated envelope
(right) for simulated data.
5. Concluding remarks
In this work, we studied the properties of four residuals in beta regression using three appli-
cations and Monte Carlo simulation techniques. Five scenarios were simulated for two values
of the precision parameter, two sample sizes, and two link functions.
It is very desirable that a residual has the following two properties. First, it should be able
to detect lack of fit of the model. Second, its distribution should be well approximated by the
standard normal distribution. The first application suggested that the four residuals consid-
ered in this article can detect lack of fit of the beta regression model. However, the simulation
studies indicated that they are not similar regarding the second property.
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 315
In all scenarios, the distribution of the weighted standardized residual 2 is worse approx-
imated by the standard normal distribution than that of the weighted standardized residual
1. In addition, in none of the scenarios, the weighted standardized residual 1 has the closest
distribution to the standard normal distribution. Therefore, these residuals do not seem to
be the best choice to perform diagnostic analysis in beta regression. The best residual regard-
ing normal approximation depends on the scenario. For n = 40, except in the case where all
mean responses are close to 0.5, the distribution of the quantile residual is better approxi-
mated by the standard normal distribution than that of the adjusted weighted standardized
residual 1. This result suggests that the quantile residual is better than the adjusted weighted
standardized residual 1 for moderate to large sample sizes. When n = 16, for most scenarios,
the quantile residual is the best for φ = 10 and the adjusted weighted standardized residual
1 performs better for φ = 100. These findings seem to indicate that, for small sample sizes,
the quantile residual is better than the adjusted weighted standardized residual 1 when the
variance of the response variable is high, and worse in the low variance case. In addition, the
distribution of the quantile residual satisfactorily approximates to the standard normal dis-
tribution in all scenarios, but the distribution of the adjusted weighted standardized residual
1 displays considerable asymmetry when φ is small and the mean responses are close to one
of the limits of the standard unit interval. This feature of the adjusted weighted standardized
residual 1 may lead to incorrect conclusions as shown in the third application.
Besides the advantages shown in the simulation results, the quantile residual has two other
advantages compared to the adjusted weighted standardized residual 1: it is simpler and it
can be used for any link function. Overall, the quantile residual is therefore a better choice to
perform diagnostic analysis in beta regression than the competing residuals.
Acknowledgments
The author thanks professor Rafael Izbicki for helpful comments and suggestions on an earlier draft
and Fapesp for financial support.
ORCID
Gustavo H. A. Pereira http://orcid.org/0000-0002-5491-0937
References
Anholeto, T., M. C. Sandoval, and D. A. Botter. 2014. Adjusted Pearson residuals in beta regression
models. Journal of Statistical Computation and Simulation 84:999–1014.
Atkinson, A. C. 1985. Plots, Transformations and Regression. Oxford: Clarendon Press.
ATP Tour Inc. 1994. ATP Scores and Stats. St. Johns County: ATP Tour Inc.
Bayer, F. M., and F. Cribari-Neto. 2013. Bartlett corrections in beta regression models. Journal of Statis-
tical Planning and Inference 143:531–547.
Bayes, C. L., J. L. Bazán, and C. García. 2012. A new robust regression model for proportions. Bayesian
Analysis 7 (4):841–866.
Carrasco, J. M. F., S. L. P. Ferrari, and R. B. Arellano-Valle. 2014. Errors-in-variables beta regression
models. Journal of Applied Statistics 41 (7):1530–1547.
Cribari-Neto, F., and M. P. F. Queiroz. 2014. On testing inference in beta regressions. Journal of Statistical
Computation and Simulation 84 (1):186–203.
Cribari-Neto, F., and A. Zeileis. 2010. Beta regression in r. Journal of Statistical Software 34 (2):1–24.
Doornik, J. A. 2007. An Object-Oriented Matrix Language: Ox 5. London: Timberlake Consultants Press.
316 G. H. A. PEREIRA
Dunn, P. K., and G. K. Smyth. 1996. Randomized quantile residuals. Journal of Computational and
Graphical Statistics 5 (3):236–244.
Espinheira, P. L., S. L. P. Ferrari, and F. Cribari-Neto. 2008. On beta regression residuals. Journal of
Applied Statistics 35 (4):407–419.
Espinheira, P. L., S. L. P. Ferrari, and F. Cribari-Neto. 2014. Bootstrap prediction intervals in beta regres-
sions. Computational Statistics 29:1263–1277.
Ferrari, S. L. P., and F. Cribari-Neto. 2004. Beta regression for modelling rates and proportions. Journal
of Applied Statistics 31 (7):799–815.
Figueroa-Zúñiga, J. I., R. B. Arellano-Valle, and S. L. P. Ferrari. 2013. Mixed beta regression: A Bayesian
perspective. Computational Statistics and Data Analysis 61:137–147.
Kieschnick, R., and B. D. McCullough. 2003. Regression analysis of variates observed on (0,1): Percent-
ages, proportions and fractions. Statistical Modelling 3:193–213.
Latif, S., and M. Zafar Yab. 2015. D-optimal designs for beta regression models with single predictor.
Journal of Statistical Computation and Simulation 85 (9):1709–1724.
Lemonte, A. J., and J. L. Bazán. 2016. New class of Johnson SB distributions and its associated regression
model for rates and proportions. Biometrical Journal 58 (4):727–746.
McCullagh, P., and J. A. Nelder. 1989. Generalized Linear Models. 2nd ed. London: CRC press.
Nocedal, J., and S. J. Wright. 2006. Numerical Optimization. 2nd ed. New York: Springer.
Ospina, R., and S. L. Ferrari. 2012. A general class of zero-or-one inflated beta regression models. Com-
putational Statistics & Data Analysis 56 (6):1609–1623.
Paolino, P. 2001. Maximum likelihood estimation of models with beta-distributed dependent variables.
Political Analysis 9 (4):325–346.
Rigby, R. A., and D. M. Stasinopoulos. 2005. Generalized additive models for location, scale and shape.
Journal of the Royal Statistics Society: Series C 54 (3):507–554.
Stephens, M. A. 1986. Tests based on EDF statistics. In: D’Agostino, R. B., Stephens, M. A., eds. Goodness-
of-Fit Techniques. New York: Marcel Dekker, pp. 97–193.