 Research article
 Open Access
 Published:
A comparison of residual diagnosis tools for diagnosing regression models for count data
BMC Medical Research Methodology volume 20, Article number: 175 (2020)
Abstract
Background
Examining residuals is a crucial step in statistical analysis to identify the discrepancies between models and data, and assess the overall model goodnessoffit. In diagnosing normal linear regression models, both Pearson and deviance residuals are often used, which are equivalently and approximately standard normally distributed when the model fits the data adequately. However, when the response vari*able is discrete, these residuals are distributed far from normality and have nearly parallel curves according to the distinct discrete response values, imposing great challenges for visual inspection.
Methods
Randomized quantile residuals (RQRs) were proposed in the literature by Dunn and Smyth (1996) to circumvent the problems in traditional residuals. However, this approach has not gained popularity partly due to the lack of investigation of its performance for count regression including zeroinflated models through simulation studies. Therefore, we assessed the normality of the RQRs and compared their performance with traditional residuals for diagnosing count regression models through a series of simulation studies. A real data analysis in health care utilization study for modeling the number of repeated emergency department visits was also presented.
Results
Our results of the simulation studies demonstrated that RQRs have low type I error and great statistical power in comparisons to other residuals for detecting many forms of model misspecification for count regression models (nonlinearity in covariate effect, overdispersion, and zero inflation). Our real data analysis also showed that RQRs are effective in detecting misspecified distributional assumptions for count regression models.
Conclusions
Our results for evaluating RQRs in comparison with traditional residuals provide further evidence on its advantages for diagnosing count regression models.
Background
Count data consist of nonnegative integers that represent the number of times a discrete event is observed; for example, number of clinic visits, hospital admissions, adverse drug events, substance abuse and rates of cardiac arrest. Poisson regressions or negative binomial (NB) regressions are often used to model such data; however, this type of data may contain a large number of zero values that a Poisson or NB models can not adequately model. To overcome the issue with excessive zeros, zero inflated models [1] were proposed in literature to model the excessive number of zeros, which are a mixture of two components: a point mass at zero and a count regression model, such as a Poisson or NB model. Zero inflated models were widely used in many population and epidemiological studies [2–7]. Despite the advance in developing counts regression models, model diagnosis remains understudied and lacks of clarity on practical issues.
Examining residuals is a primary approach for identifying the overall discrepancies between models and data (e.g., nonlinear effects, overdispersion, zeroinflation), and observations that are not accommodated by the models (e.g., outliers). Residual analyses can also diagnose the overall goodnessoffit (GOF) and adequacies of a model. Pearson and deviance residuals have been often used for diagnosing generalized linear models (GLMs) [8, 9]. Pearson residuals are defined as the standardized distances between the observed and expected responses, and deviance residuals are defined as the signed square root of the individual contributions to the model deviance (i.e., the difference between the loglikelihoods of the saturated and fitted models). In normal regressions, Pearson and deviance residuals are the same and asymptotically follow a normal distribution under the true model [8, 9]. In order to assess the model fit, these residuals are commonly plotted against the fitted values and each covariate, as well as compared against the standard normal distribution. The chisquared (χ^{2}) test statistic is often used to measure the overall GOF of a normal regression model. The χ^{2} statistic was proven to have an asymptotic χ^{2} distribution with n−p degrees of freedom \(\left (\text {denoted by} \chi ^{2}_{np}\right)\) [8, 9]. However, in nonnormal regression, specifically in data with the response variable being distributed on a small number of distinct values, the Pearson and deviance residuals do not typically follow (marginally and conditionally) a normal distribution. The plots of their residuals contain nearly parallel curves, making their assessment difficult to interpret due to the lack of a unified reference distribution for comparison. Furthermore, although the χ^{2} tests with \(\chi ^{2}_{np}\) as the null distribution are widely used for quantitatively measuring the overall GOF in nonnormal regression, the \(\chi ^{2}_{np}\) distribution is often very poor for approximating the true null sampling distribution of χ^{2} statistic [8, 9] except for situations where the response variable is approximately normally distributed (e.g., Poisson with large means).
In a short communication paper, Dunn and Smyth [10] introduced the randomized quantile residual (RQR) method. The key idea of the RQRs is to introduce randomizations between the discontinuity gap of the cumulative distribution function (CDF) and then invert the fitted distribution function for each response value and finding the equivalent standard normal quantile. Dunn and Smyth [10] showed that the RQRs are approximately normally distributed under a correctly specified model. Klar and Meintanis (2012) [11] proposed the standardized RQRs for performing goodnessoffit tests in generalized linear models, which was shown to be approximately standard normally distributed. For modeling nonnormal and continuous outcome data, several studies have investigated the properties of quantile residuals for checking the model fit. For example, Pereira (2019) [12] investigated the properties of the quantile residual in the beta regression model and demonstrated that the distribution of the quantile residual is better approximated by the standard normal distribution than that of the other residuals in most scenarios. Lemonte and MorenoArenas (2019) [13] proposed the normalized quantile residual to check the adequacy of the generalized Johnson S_{B} (GJS) regression model, which were shown as a better choice to identify departures from the model assumptions and to assess the overall goodnessoffit than the deviance residual. Scudilio and Pereira (2020) [14] proposed an adjusted quantile residual for diagnosing inverse Gaussian or Gamma regression models, which was shown to be a better choice to perform diagnostic analysis compared to traditional residuals. Despite the increasing awareness of RQRs, the literature on using RQRs has primarily focused on diagnosing generalized linear models with the distribution of the response variable belonging to the exponential family, i.e., Poisson, Negative binomial, inverse Gaussian, Gamma, or being continuous nonnormal. A broader simulation study that evaluates the performance of RQRs for diagnosing count regression models, including zeroinflated count regressions would be useful for applied statisticians to understand the advantages of RQRs.
The purposes of this article are therefore to conduct simulation studies to (1) demonstrate that the RQRs approximately follow a normal distribution for count regression models when the model is correctly specified apart from the sampling variability in the estimated parameters, and (2) examine the power of the ShapiroWilk normality test for examining the performance of RQRs as an overall model diagnosis tool. More specifically, we show that the RQR method, in general, has high power and low type I error for detecting various model misspecification (i.e., having great statistical power in detecting many types of model inadequacy, such as nonlinearity, zeroinflation, and overdispersion).
For the remaining of this article, in “Methods” section, we review the commonly used count regression models and discuss the inadequacies of the Pearson and deviance residuals for detecting model misspecification followed by a brief review of RQRs. In “Results” section, the results of the simulation studies and a real data application are presented. More specifically, in “Simulation studies” section, we present the results of simulation studies to demonstrate the superior performance of RQRs for detecting various forms of model misspecification. In “Real data application” section, a real data application based on a health care utilization study is presented to illustrate the performance of RQRs in comparison with other traditional residuals. Implications and limitations of the study are presented in “Discussion” section. Concluding remarks are given in “Conclusion” sections.
Methods
Regression models for count data
Poisson and negative binomial (NB) models
Let y_{i} denote a discrete random variable following a Poisson distribution with mean λ_{i}, i=1,⋯,n. The probability mass function (PMF) is given by \(dpois(y_{i};\lambda _{i})=\frac {e^{\lambda _{i}} \lambda _{i}^{y_{i}}}{y_{i} !}\), y_{i}=0,1,2⋯ and the cumulative probability function (CDF) is given by \(ppois(y_{i};\lambda _{i})=\sum _{y_{i}=0}^{y_{i}} dpois(y_{i};\lambda _{i})\). For a NB regression model with mean λ_{i} and shape parameter k, the PMF can be expressed as, \( dnb(y_{i};\lambda _{i}, k) = \frac {\Gamma (y_{i}+k)}{\Gamma (k) \Gamma (y_{i}+1)} \left (\frac {\lambda _{i}}{\lambda _{i}+k}\right)^{y_{i}} \left (\frac {k}{\lambda _{i}+k}\right)^{k}, y_{i}=0, 1, 2\cdots \) and the CDF can be expressed as \(pnb(y_{i};\lambda _{i}, k) = \sum _{y_{i}=0}^{y_{i}} dnb(y_{i};\lambda _{i}, k). \) Poisson or NB models with a log link function can be then written as \( \text {log}(\lambda _{i}) = \boldsymbol {X}^{T}_{i} \boldsymbol {\beta }\), where X_{i} denotes the design matrix of a set of covariates and β represents the corresponding regression coefficients.
Zeroinflated models
In practice, count data often contains excessive zeros that may not be accurately captured by a conventional Poisson or NB model; these data are commonly known as zeroinflated data. One popular approach to model such data is a mixture model of degenerate zeros from the nonrisk group (i.e., structural zeros) and responses with random zeros or positive values from the atrisk group [1, 2, 5, 15–17]. For example, one might be interested in the question: “How often do you revisit hospitals within the last 30 days after being released from the hospital?” Among the patients reported zero number of hospital revisits, some may be fully recovered from this disease leading to genuine nonusers (structural zeros); however, some might currently behave as nonusers (sampling zeros) and would have the potential to revisit the hospital depending on their health status.
The zeroinflated Poisson (ZIP) model with parameters λ_{i} and p_{i}, denoted by ZIP(λ_{i},p_{i}), is defined by
where λ_{i} is the mean of the Poisson component and p_{i} is the probability of belonging to the structural zeros component for the ith observation.
Let dzip(y_{i};λ_{i},p_{i}) denote the PMF of the ZIP distribution with the unconditional probability distribution written as
The CDF, denoted by pzip(y_{i};λ_{i},p_{i}), is then derived as
where ppois(J,λ_{i}) denotes the CDF of a Poisson distribution. The marginal mean and variance of a ZIP distribution can be derived as
As a result, zeroinflated model can accommodate overdispersion relative to a Poisson model, since V(y_{i})>E(y_{i}).
The ZIP model can include covariates for modeling both p_{i} and λ_{i}. Generally, p_{i} is modeled with a logistic regression and λ_{i} is modeled as a loglinear regression. ZIP model can be then written as,
where \(\boldsymbol {Z}^{T}_{i}\) and \(\boldsymbol {X}^{T}_{i}\) are vectors of covariates with corresponding vectors of regression parameters, γ and β, for p_{i} and λ_{i}, respectively. The zeroinflated negative binomial (ZINB) model can be defined similarly by replacing Poisson distribution with NB distribution in equation (1).
Pearson residuals
The Pearson residual, defined as the raw residual scaled by the estimated standard deviation of the response variable, is the most common measure for GOF, which can be expressed as,
where \(\hat {\mu }_{i}\) is the fitted value for y_{i} and \(\widehat {V}(y_{i})\) is the estimation of variance for y_{i}. Table 1 presents the Pearson residuals for some commonly used count regression models.
Deviance residuals
Deviance residuals are derived based on the deviance, which is defined as twice the difference between the loglikelihood for the saturated and fitted models and is given by \(D(y, \hat \mu)=2\sum _{i=1}^{n}\left \{{log} [p(y_{i}\hat \theta _{s})]{log} [p(y_{i}\hat \theta)]\right \}\), where \({log} [p(y_{i}\hat \theta _{s})]\) represents the loglikelihood function for the saturated model and θ_{s} denotes the set of parameter estimates for the saturated model, in which there are as many estimated parameters as data points [8, 18]. By definition, a saturated model leads to a perfect fit to the data and has the highest loglikelihood among all models. \(\text {log}[p(y_{i}\hat \theta)]\) represents the loglikelihood function of the fitted model and \(\hat \theta \) denotes the set of parameter estimates for the fitted model. Deviance residuals represent the contributions of individual samples to the deviance \(D(y, \hat \mu)\), which is defined as the signed square root of the corresponding component for \(D(y, \hat \mu)\) and can be written as
where \(d_{i}=2 \left \{\text {log} [p(y_{i}\hat \theta _{s})]\text {log} [p(y_{i}\hat \theta ])\right \}\). For a ZIP model, it can be shown that Poisson(y_{i}) is the saturated model [19]; hence, the deviance residual for a ZIP model is defined as the signed square root of the likelihood ratio between the fitted model (zeroinflated Poisson model) and the saturated model (Poisson model). The deviance residuals for a ZINB model can be defined similarly. Table 2 presents the deviance residuals for some commonly used count regression models.
Problems with pearson and deviance residuals
For a normal linear regression model, the Pearson and deviance residuals are identical and have an approximate normal distribution under the true model. However, their distributions are often skewed and nonnormally distributed for counts regression models [8, 20]. It is argued that the deviance residuals typically follow more closely a normal distribution than the Pearson residuals; nevertheless, as μ_{i}/ϕ→∞, both Pearson and deviance residuals from an exponential family model approach to the normal distribution due to the distribution for the response variable converging to normality. However, the asymptotic normal distribution only holds when the mean of the response variable is relatively large. Further, the residual plots often exhibit parallel curves according to distinct response values, imposing great challenges for visual inspection. Hence, Pearson and deviance residuals are difficult to use for graphically assessing the GOF of count regression models.
Further, the overall GOF of a regression model is often assessed based on the sum squares of the Pearson and deviance residuals, i.e., \( X^{2}=\sum _{i=1}^{n} r^{P^{2}}_{i}\) and \(D^{2}=\sum _{i=1}^{n} r^{D^{2}}_{i}\), respectively. Asymptotically, under a correctly specified normal regression model, we can expect X^{2} and D^{2} to have a chisquare distribution \(\chi ^{2}_{np}\), where n is the sample size, and p is the number of parameters. In practice, we often fail to achieve large samples, which renders the null distribution of this statistic invalid. It is also recognized that this approximation for diagnosing count regression models can be very poor even for large sample sizes [9, 21].
Randomized quantile residual
Definition of randomized quantile residuals
Randomized quantile residual (RQR) proposed by Dunn and Smyth [10] gives a general tool for diagnosing regression models with independent responses, which is particularly ideal for diagnosing count regression models. RQR is an extension of quantile residuals, which is based on the idea of inverting the estimated distribution function for each observation to obtain approximately normally distributed residuals.
In the case of a normal regression model, quantile residuals are equivalent to Pearson and deviance residuals [14]. In the case of a nonnormal regression model for modeling a highly skewed and continuous outcome variable, Scudilio and Pereira (2020) [14] proposed an adjusted quantile residual to diagnose inverse Gaussian or Gamma regression models, which was shown to be a better choice to perform diagnostic analysis compared to traditional residuals. Pereira (2019) [12] and Arenas (2019) [13] investigated the properties of the quantile residual in the beta regression and generalized Johnson S_{B} regression models, respectively. Their results showed that the quantile residual is well approximated by a standard normal distribution.
In the case of the discrete outcome, such as Poisson or NB, the corresponding CDFs are discrete. To circumvent this issue, the RQR introduces randomization between two consecutive CDFs to produce continuous normal residuals [10]. More specifically, suppose we consider fitting a regression model with \(F(y_{i};\hat \mu _{i}, \hat \phi)\) denoting the CDF for a response variable y_{i} given a set of covariates x_{i}, where \(\hat \mu _{i}\) is typically a function of x_{i} (i.e., the conditional mean of y_{i}) and ϕ (i.e., the dispersion parameter) does not depend on x_{i}. Let \(p(y_{i}; \hat \mu _{i}, \hat \phi)\) be the corresponding PMF of \(F(y_{i}; \hat \mu _{i}, \hat \phi)\). The estimated CDF is defined as
where \(F(y_{i};\hat \mu _{i},\hat \phi)\) is the lower limit of F at y_{i} (i.e., \(\sup _{y < y_{i}} F(y;\hat \mu _{i})\)) and u_{i} is a random number from a uniform distribution on (0,1). An alternative definition for the randomized lower tail probability is a uniform random number between \(a=\sup _{y < y_{i}} F(y;\hat \mu _{i}, \hat \phi)\) and \(b = F(y_{i}; \hat \mu _{i}, \hat \phi)\) [10].
RQR for y_{i} is then calculated as the standard normal quantile of the estimated CDF, which is given by
where Φ^{−1} is the quantile function of the standard normal distribution. \(F^{\ast } (y_{i}, u_{i};\hat {\mu }_i,\hat {\phi })\) can be converted to any other previously mentioned standard distribution. The normal distribution is chosen due to its wellknown characteristic (i.e., the socalled “empirical rules”), and diagnostic and inspection methods for normal linear regression can be then applied to RQRs. Note that when F is continuous at y_{i}, in Eq. (10), PMF at y_{i} is \(p(y_{i};\hat \mu _{i},\hat \phi) = 0\); for clarification, \(p(y_{i};\hat \mu _{i},\hat \phi) \) is not the PDF of y_{i}. Therefore, there is no actual “randomness” in \(F^{\ast } (y_{i}, u_{i};\hat \mu _{i},\hat \phi)\) when response distribution is continuous. The formula given in (9) encompasses this situation. Therefore, in the case of continuous response variable, RQRs are equivalent to quantile residuals. Similarly, when y_{i} is discrete, the variability in F^{∗} tends to be smaller for narrower gaps at the discontinuity points of F. This scenario typically occurs when the mean of response variable is large.
Klar and Meintanis (2012) [11] suggested standardizing randomized quantile residuals, which are calculated as \(r^{Q}_{i}=\frac {z^{Q}_{i}\bar {z}^{Q}}{S_{z^{Q}}}\), where \(\bar {z}^{Q}=n^{1}\sum _{i=1}^{n} z^{Q}_{i}\) and \(S^{2}_{z^{Q}}=(n1)^{1}\sum _{i=1}^{n}\left (z^{Q}_{i}\bar {z}^{Q}\right)^{2}\) for performing goodnessoffit tests of common generalized linear models including Poisson, NB, Gamma and Inverse Gaussian. The standardized RQRs were shown to approximately follow a standard normal distribution. Following their suggestion, we standardized RQRs in our investigation. In this paper, we show that RQRs can also be applied for examining the model fits of count models in general, including zeroinflated models. Table 3 provides the derivation of the RQRs for the Poisson, NB, and ZIP and ZINB regression models, where ppois and dpois, pnb and dnb, pzip, dzip and pzinb and dzinb denote their respective CDFs and PMFs.
The idea of constructing F^{∗} in (9) is also closely related to the predictive pvalues proposed for diagnosing Bayesian models for discrete observations in which the u_{i} is fixed at 0.5 rather than being a uniform random number (See [22–25]). This approach creates nonrandom quantile residuals, to which we refer as “middlepoint quantile residuals (MQRs)”, which will also be compared with the RQRs in this study.
An illustrative example
To demonstrate the idea of the RQRs, we simulate a response variable of size n=1000 from a Poisson model with log(μ_{i})=−1+2sin(2x_{i}), where μ_{i} is the expected mean count for the ith subject and x_{i}∼Uniform(0,2π), i=1,…,n. Then, we fit a Poisson model with the same mean structure as well as a Poisson model with a misspecified mean structure log(μ_{i})=β_{0}+β_{1}x_{i} to illustrate the capability of RQRs for detecting the nonlinearity of the covariate effect.
The CDF of the response variable Y_{i} given x_{i} (under the considered model with estimated parameters) is denoted by F(kx_{i})=P(Y_{i}≤kx_{i}), for k=0,1,…. Figure 1 shows F(kx_{i}) as a function of x_{i}, with each coloured line representing a CDF curve associated with a value k. The distance between two curves, F(kx_{i}) and F(k−1x_{i}), is the “theoretical” (modelbased) probability of y_{i}=k given each x_{i}. Each observed y_{i} is then randomly scattered uniformly to a point between the CDF lines associated with k=y_{i}−1 and k=y_{i}. The pattern of randomized scattering of discrete y_{i} facilitates the comparison of the “observed” frequency (fraction of points) and “theoretical” frequency (distance of two lines). If the “observed” and “theoretical” frequencies agree well, the randomly scattered points of \(F^{\ast } (y_{i};\hat {\mu }_i,\hat {\phi }, u_{i})\) should be uniformly distributed on (0,1] in each neighbourhood of x_{i}. Figure 1 depicts that under the true model, the randomized CDFs are uniformly distributed on (0, 1] given each x_{i}. By contrast, under the misspecified model, the randomized CDFs are not uniformly distributed, exhibiting a nonlinear trend. The results also show that MQRs have the same difficulties as the Pearson and deviance residuals for model assessment with residuals clustering along curved lines for both true and wrong models.
Figure 2 displays the scatter plots of various types of residuals against the covariate x under the true and wrong models, which show that the Pearson, deviance, and MQRs are clustered as curves according to the distinct and discrete values of the response variable. By contrast, under the true model, RQRs are randomly scattered between 3 and 3; under the wrong model, the plot of RQRs against x exhibits a sinusoidal trend.
To examine whether standardized RQR is well approximated by a standard normal distribution under the true model as compared with other types of residauls, Table 4 reports the mean, variance, skewness, excess kurtosis for the Pearson residual, deviance residual, MQR and RQR, under the true and wrong models, respectively. The skewness is a measure of symmetry. Negative skewness indicates that the data is leftskewed and positive skewness indicates that the data is rightskewed. The excess kurtosis describes the tail shape of the data distribution, so a normal distribution has zero excess kurtosis. Negative excess kurtosis would indicate a thintailed data distribution, and positive excess kurtosis indicates a fattailed distribution. We also used ShapiroWilk (SW) test [26] to evaluate the normality of the residuals with the null hypothesis H_{0}: The residuals are normally distributed versus the alternative hypothesis H_{a}: The residuals are not normally distributed. Previous research showed that SW test is more powerful compared to other normality tests, such as Kolmogorov–Smirnov test, the Lilliefors test, the Cramer–von Mises test, the Anderson–Darling test, the D’Agostino–Pearson test, etc., over a wide range of asymmetric distributions [27].
The results presented in Table 4 indicate that the means of the Pearson residuals, MQRs and RQRs are all close to zero, but the means of the deviance residuals are below zero under both the true and wrong models. This finding is consistent with the previous investigations [13, 14], which showed the deviance residuals do not follow a standard normal distribution regardless of the true or wrong models. The variances of Pearson residuals under both the true and wrong models are above one. Under the true model, the variances of the deviance residuals and MQRs are all below one, but above one under the wrong model. By contrast, the variances of RQR are equal to one for all models, since the RQRs in the present study are all standardized to have unit variance. In terms of skewness and kurtosis, the Pearson residual, deviance residual, and MQR are rightskewed and heavytailed relative to a normal distribution, and the extent of skewness and heavytailed is more pronounced for Pearson residual. RQRs are approximately symmetric with the tail shape close to a normal distribution under the true model and are rightskewed under the wrong model. For testing the normality of the residuals based on the SW normality test, the results clearly showed that both true and wrong models were rejected according to the Pearson residual, deviance residual, and MQR. By contrast, RQRs confirm the adequacy of the true model with the pvalue of the SW normality test equal to 0.684 and inadequacy of the wrong model with the SW pvalue almost equal to zero.
Further, normal quantilequantile (QQ) plots can be produced by plotting the ordered values of the residuals versus the expected order statistics of a normal distribution, approximated as \(\Phi ^{1}(\frac {i3/8}{n+1/4})\), where i is the ith order statistic, 1≤i≤n and n is the sample size [28]. If the residuals are normally distributed, the points in the QQ plot should follow a straight diagonal line. However, for the count regression models, the asymptotic distributions of the residuals are unknown; therefore, the normal QQ plot is not informative for diagnosing model fits of count regression models. To overcome this challenge, a simulated envelope proposed by Atkinson (1981) [29] can be added to the QQ plot to detect departures from the distributional assumptions as well as outlying observations of the fitted model. The simulated envelope can be obtained as follows [30]: (1) fitting a model, (2) extracting and sorting the residuals, (3) simulating 100 response variable using the same model matrix, error distribution and parameter estimates, (4) fitting the same model to each simulated response variable and extracting and sorting the residuals, and (5) computing the 2.5% and 97.5% percentiles of the simulated residuals at each ordered residuals to form the envelope. For a wellfitted model, most of the residuals are expected to fall within the simulated envelope. As displayed in Fig. 3, all types of residuals under the true model fall within the simulated envelopes. By contrast, under the wrong model, a bulk of residuals fall outside of the simulated envelopes.
The results indicate that the normality probability plot with a simulated envelope can help distinguish the true and wrong models; however, visualizing these plots could not provide information on the nature of deficiencies with respect to the fitted model. In contrast, the scatter plot of RQRs against covariates or fitted values can provide information on the inadequacy of different aspects of the models, such as the misspecified functional form of the covariate effect, overdispersion, and zeroinflation, etc. Moreover, a numerical measure of overall model goodness of fit is desirable to summarize the discrepancy between fitted and observed values under the fitted model rather than only relying on visualizing the QQ plots with simulated envelope to determine the model fit, which can be subjective.
Results
Simulation studies
In this Section, we investigate the performance of the RQRs in comparison with MQRs, deviance, and Pearson residuals via simulations. The simulations consist of testing nonlinearity of the covariate effect, overdispersion, and zeroinflation. A numerical measure of overall model goodness of fit is desirable to summarize the discrepancy between expected values under the fitted model and observed. It is also desired to have a unifying framework for all types of models using residuals that are approximately standard normally distributed in line with a normal regression model. We, therefore, use the SW test evaluating the normality of residuals as a goodnessoffit test with the null hypothesis H_{0}: The model fits the data well versus the alternative hypothesis H_{a}: The model does not fit the data well.
Under each simulation scenario, we randomly generated 5000 datasets from the true model to examine the type I error rate and statistical power. The type I error rate is defined as the probability of rejecting the true model (i.e., the proportion of times that the SW test pvalue <5% for the residuals under the true model) and statistical power is the probability of rejecting the wrong model (i.e., the proportion of times that the SW test pvalue <5% for the residuals under the wrong model). Ideally, a desirable GOF test should have a probability of type I error close to 5% with high statistical power.
For each simulation, we first present the performance of RQR for detecting model misspecification under one simulation scenario. Further, to gain more insight into the finitesample performance of RQRs in comparison with other residuals, we perform power analysis by setting the sample sizes to n=20,50,100,200,400,600,800 and 1000 with varying degrees of model misspecification.
Detection of nonlinearity
In this subsection, we first investigate the performance of the RQRs for detecting nonlinearity of the covariate effect based on a single simulation setting. We first simulate the covariate x∼Uniform(−1.5,1.5) of size n=1000. The response variable is then simulated from a NB regression model \(\text {log}(\mu _{i})=\beta _{0}+\beta _{1} x_{i}^{2}\), where μ_{i} is the expected mean count for the ith subject. We then consider fitting the model assuming x_{i} having a linear effect, i.e., log(μ_{i})=β_{0}+β_{1}x_{i}. The regression parameters were set as β_{0}=0 and β_{1}=1 while the reciprocal for the dispersion parameter associated with the NB distribution was set as 2.
The panels in the first column of Fig. 4 display the scatter plots of RQRs against the covariate under the true and wrong models. Under the true model, RQRs are randomly scattered without exhibiting any pattern and being mostly distributed between 3 and 3 as standard normal variates; conversely, under the wrong model, the RQRs clearly show a quadratic pattern. The panels of the second column of Fig. 4 present the quantilequantile (QQ) plots of the RQRs under the true and wrong models. Under the true model, the points almost perfectly align along the diagonal line, whereas under the wrong model, the points deviate from the diagonal line in both upper and lower ends with some of the points on the ends falling outside of the simulated envelope.
To examine the power of the SW test for RQRs as an overall GOF, we repeatedly simulated 5000 datasets from the true model. The panels in the third column of Fig. 4 present the histograms of 5000 SW pvalues under the true and wrong models. The SW pvalues under the true model are nearly uniform, indicating the wellcalibration of this overall GOF test. In contrast, under the wrong model, the SW pvalues are highly distributed near 0. Thus, the SW test for the RQRs as an overall GOF test seems to perform well with type I error close to 5% and great power in detecting the nonlinear relationship in this simulation setting.
To further investigate the performance of the SW test for RQRs as an overall GOF test, we conducted the power analysis by setting β_{1}=0.5,1,2 to increase the degree of nonlinearity of the covariate effect. As shown in the top panels of Fig. 5, type I errors of the SW test for RQRs are consistently centered around the 5% nominal level for all scenarios. In contrast, type I errors for the SW tests for MQRs, deviance, and Pearson residuals approach to 1 as sample size increases. The bottom panels of Fig. 5 indicate that the power of this GOF test based on RQRs increases as the sample size increases, especially when the misspecified model deviates from the true model at a greater degree. Despite the high power of the SW tests based on MQRs, deviance, and Pearson residuals, they suffer from substantially high type I errors. As a result, overall GOF tests based on SW tests of MQRs, deviance, or Pearson residuals are undesirable in this scenario.
Detection of overdispersion
As in the previous section, the same approach is implemented to investigate the performance of the RQRs in detecting overdispersion [31] in the data. We first simulate a covariate x∼Uniform(−1,2) of size n=1000. Then, the response variable is simulated from a NB model log(μ_{i})=β_{0}+β_{1}x_{i}, where μ_{i} is the expected mean count for the ith study subject. We set the regression parameters as β_{0}=1 and β_{1}=2 and reciprocal for the dispersion parameter as 2. To examine the performance of the various types of residuals in diagnosing overdispersion, we considered fitting a Poisson model as the misspecified model, which has the same mean structure as the NB model.
The panels in the first column of Fig. 6 present the scatter plots of the RQRs against the covariate under the true and wrong models. Under the true model, the residuals are mostly scattered between 3 and 3 without any discerning pattern. In contrast, under the wrong model, the residuals “fan out” from left to right, suggesting the presence of overdispersion at increasing values of x_{i}. The panels in the second column of Fig. 6 present the QQ plots of RQR residuals under the true and wrong models. Under the true model, the points align along the diagonal line well; whereas, under the wrong model, the points significantly deviate from the diagonal line with a substantial amount of the points falling outside of the simulated envelope, indicating the RQRs are approximately normally distributed under the true model, but not under the wrong model. The panels in the third column of Fig. 6 present the histograms of 5000 SW pvalues for testing the normality of the RQRs under the true and wrong models. As shown, the SW pvalues under the true model are nearly uniform, while the SW pvalues under the wrong model are clustered around 0. These results demonstrate that the SW test for the RQRs as an overall GOF test can help detect overdispersion.
In the power analysis, we increased the level of overdispersion in the data by setting the dispersion parameter as 1,2, and 10. Figure 7 shows that the type I errors of the SW test for the RQRs remain at the nominal level 0.05 for all scenarios. In contrast, the type I errors of the SW tests for the MQRs, deviance and Pearson residuals exceed 5% as sample size increases. Further, SW tests of the RQRs are able to maintain high statistical power at all scenarios when the sample size is sufficiently large (i.e., n>100).
Detection of zeroInflation
Finally, we conduct simulations to investigate the performance of the RQRs in detecting zeroinflation. We first simulate a covariate x∼Uniform(−1,2) of size n=1000. Then, the response variable is derived from a ZIP model, where the expected mean for the Poisson component is λ_{i}= exp(β_{0}+β_{1}x_{i}). Under the true model, we set the regression parameters as β_{0}=1 and β_{1}=2, and the percentage of excessive zeros as 30%. A Poisson model with the same expected mean λ_{i}, but ignoring zeroinflation is considered as the misspecified model.
The panels in the top row of Fig. 8 display the scatter plot of RQRs against the covariate, QQ plot of RQRs, and histogram of 5000 SW pvalues of the RQRs under the true model. The results indicate that RQRs are mostly distributed between 3 and 3 as standard normal variates without any unusual patterns. The points in the QQ plot align along the diagonal line, and the histogram of the SW pvalues are nearly uniform. By contrast, under the misspecified model, a clear separation of the RQRs is observed from the residuals associated with the zero responses, as shown in the bottom left panel of Fig. 8. The points in the QQ plot under the misspecified model deviate from the diagonal line, with a substantial amount of the points falling outside of the simulated envelope. This histogram of SW pvalues is highly distributed near 0, indicating that the wrong model was rejected most of the time. Hence, the SW tests for the RQRs as the overall GOF test can effectively detect zeroinflation.
In the power analysis, we set the probability of generating excessive zeros as 0.1, 0.3, and 0.5. Figure 9 shows that the type I error rates of the SW test for RQRs remain at the nominal level 5% for all scenarios. In contrast, the type I error rates of the MQRs, deviance, and Pearson yield extremely large type I error rates almost across all scenarios. In all scenarios, RQRs demonstrate descent statistical power even at small sample sizes, and the power increases as the sample size increases. Overall, RQRs outperform other types of residuals having low type I error and high power.
Real data application
In this section, we examine the performance of RQRs in comparison with other types of residuals in a real data application based on the National Medical Expenditure Survey (NMES) (see further descriptions of this data in [32, 33]). In this data set, 4406 elderly in the United States were surveyed about their demands of health care. The response variable considered in this study is the number of emergency department (ED) visits. The covariates considered include demographic characteristics (e.g., age, race, sex, marital status, education, and region), socioeconomic variables (e.g., family income, employment status, supplementary private insurance status, and public insurance status) and health measures (e.g., selfperceived health, the number of chronic conditions and a measure of disability status).
Multicollinearity among all the covariates, was evaluated using the generalized variance inflation factor (GVIF) [34], which is a generalization of the variance inflation factor (VIF). GVIF is applicable to measure the collinearity among covariates, such as dummy regressors from a polytomous categorical variable, by considering the size of the joint confidence region for the related coefficients. Literature suggests reporting GVIF^{1/(2·df)}, where df is the number of dummy variables in a categorical variable, which is analogous to reporting the square root of the VIF for a single coefficient [34]. As a rule of thumb, VIF of 5 (i.e., \(\sqrt {{VIF}}\approx 2.236\)) or greater indicates multicollinearity is high. Table 7 in Appendix reported no values of GVIF^{1/(2·df)} are greater than 2.236; therefore, all the covariates considered are not highly correlated in this analysis.
Over 81% of the patientyear records were zero, implying that the majority of patients did not make any ED visits during the year of the study. The number of nonzero visits ranged from 1 to 12, with only 5% having more than one visit in the study year. Given the high percentage of zeros and skewness of the response variable, we considered fitting Poisson, NB, ZIP, and ZINB regression models. After backward elimination at the 5% significance level, the final models included the following covariates: The number of chronic conditions, selfperceived health (excellent vs. poor; average vs. poor), limited activities of daily living (yes vs. no) and the number of years of education. In addition to those covariates, the black race was significantly associated with increased ED use for the Poisson and ZIP models, but not for the NB and ZINB models. This discrepancy highlighted the importance of examining the model GOF; that is, fitting a model with unsatisfactory model fit may lead to biased estimates, incorrect standard errors, and erroneous inferences.
For the binary components of the ZIP and ZINB models, no covariates were statistically significant at the 5% level by including one variable at a time or using backward selection. Table 5 presented the results of the analyses, which showed that the standard errors of the estimated regression coefficients for the NB and ZINB models are all larger relative to their counterpart Poisson models (i.e., Poisson and ZIP models), indicating that the choice of model distribution has a significant impact on covariate effects. NB and ZINB models yield almost identical estimated regression coefficients, with ZINB model giving slightly less efficient estimates compared to the NB model; that is, the standard errors for the estimated regression coefficients of chronic conditions and selfperceived health (Average vs. poor) were slightly larger under the ZINB model as compared to the corresponding standard errors under the NB model.
For comparing the competing models, the Akaike Information Criterion (AIC) is used with smaller values indicating better and more parsimonious model fit. The AIC scores for the Poisson, NB, ZIP, and ZINB are 5648, 5352, 5418, and 5354, respectively; this suggests that the NB and ZINB models are superior to their counterpart Poisson and ZIP models, respectively. NB model performs slightly better than the ZINB model. The intercept term for the binary component of the ZINB model is estimated as −8.912(SE: 41.106, pvalue =0.828); therefore, the estimated probability of excessive zeros is about exp(−8.912)/[1+exp(−8.912)]≈0, which further confirms that the data does not exhibit significant more zeros than are expected under an NB distribution.
Although AIC can be used to compare the GOF of competing models, it cannot measure the adequacy of the model fit, assess the need for additional complexity, and validate the distribution assumption of the response variable. Diagnosing the residuals is, therefore, imperative to address these concerns. Figure 10 presented the scatter plots of Pearson, deviance, MQR, and RQRs versus fitted values under the Poisson, NB, ZIP, and ZINB models for modeling the number of ER visits. It is evident that Pearson, deviance, and MQR are clustered as curved lines. RQRs achieved continuity correction via randomization of the cumulative distribution function. The NB and ZINB models fitted the data fairly well with residuals ranging mostly between 3 and 3, and no discernible pattern. By contrast, Poisson and ZIP models do not appear to accommodate larger values of the response variable. The QQ plots of all types of residuals are presented in Fig. 11, which indicates that under the NB and ZINB models, the residuals fall within the simulated envelope. By comparison, under Poisson and ZIP models, a portion of residuals fall outside of the simulated envelopes. This indicates that NB and ZINB models fit the data adequately well, and Poisson and ZIP do not fit the data well, which further supports the need to model overdispersion in this data.
One concern of using RQRs is the fluctuation in the residuals introduced by randomization for producing continuously distributed residuals. As suggested by Dunn and Smyth [10], multiple realizations of the RQRs should be produced to ensure that the discrepancies are not made by the randomization in producing the residuals. To assess the uncertainty in the GOF test due to randomization in the real data application, we generated 1000 realizations of the RQRs. Figure 12 presents the histograms of 1000 replicated pvalues of the SW tests, which indicates that randomization introduced little variation to the SW pvalues of the RQRs in this application. More specifically, the SW pvalues of RQRs under the fitted Poisson and ZIP models were close to 0, as depicted on the left panels of Fig. 12 with the histogram concentrated at 0, confirming the inadequacies of both models. Conversely, the SW pvalues of the RQRs for the fitted NB and ZINB models varied between 0 and 1 with about 96% of the pvalues being above 0.05, as depicted on the right panels of Fig. 12, demonstrating the adequacies of both NB and ZINB models, although about 5% of the SW pvalues are below 0.05.
To further illustrate RQR can be well approximately by a standard normal distribution in comparison with other types of residuals in the real data application, we calculated the mean, variance, skewness, kurtosis and the pvalue of the SW normality test for the Pearson residual, deviance residual, MQR and RQR in the real data analysis. Note that for the SW normality test of the RQRs, the average of SW pvalues based on 1000 replicated RQRs was presented to account for the fluctuation in the residuals introduced by randomization. As shown in Table 6, the results indicate that the means of the Pearson residuals, MQRs and RQRs are close to zero, but the means of the deviance residuals are consistently lower than zero under all models. The variances of Pearson residuals under all models are above one, and the variances of the deviance residuals and MQRs are all below one. By contrast, the variances of RQR are equal to one for all models, since the RQRs in the present study are all standardized. In terms of skewness and kurtosis, the Pearson residual, deviance residual, and MQR are rightskewed and heavytailed relative to a normal distribution, but RQRs are approximately symmetric with the tail shape close to a normal distribution under the NB and ZINB models. For testing the normality of the residuals based on the SW normality test, the results clearly showed that all models were rejected under the Pearson residual, deviance residual, and MQR. By contrast, RQRs confirm the adequacy of NB and ZINB models, with an average of SW pvalues close to 0.5.
Discussion
Diagnosing regression models is crucial to ensure the validity of the results and implications that heavily rely on the tenability of the model assumptions. Pearson and deviance residuals and their corresponding χ^{2} tests are commonly used in practice. However, the use of these tools in count regression models are often not valid, since Pearson and deviance residuals do not have an asymptotic normal distribution under the correctly specified model. This paper reminisces that RQRs as a diagnostic tool to show that it has great advantages over other types of considered residuals for diagnosing count regression models, including zeroinflated count regression models.
Although using the QQplots for the Pearson residual, the deviance residual or MQR with simulated envelope could be used to check the model fit, visual inspection of the points falling outside of the simulated envelope can be subjective, and no single numerical measure of the overall model fit could be easily summarized based on such plots. A key strength of RQRs is that the plots of RQRs against covariates or fitted values contain important information concerning the inadequacy of different aspects of the model, which could not be easily reflected in the QQ plots with simulated envelopes. Further, the distribution of RQRs under the correctly specified count model is well approximated by a standard normal distribution. As a result, a numerical measure of overall model fit, i.e., SW normality test, can be readily derived, which was shown in our study to perform well for testing the overall model lack of fit.
Note that in our investigations, RQRs are standardized to have unit variance, as suggested by Klar and Meintanis (2012) [11]. Therefore, testing the normality of RQRs is essentially the same as testing standard normality in our investigation. Our study showed that Pearson, deviance, and MQR residuals for count regressions are not well approximated by a normal distribution. As such, testing standard normality of those residuals would be more stringent than testing normality, which might give even stronger evidence of the inadequacy of those types of residuals for diagnosing count regression models.
RQRs depend on a random uniformly distributed number u_{i} to convert the discrete cumulative distribution function into continuous values. Although the randomness in u_{i} may produce special patterns in the RQRs, we note that the chance that the pure random numbers will make any observable pattern decreases as the sample size n increases. Additionally, the influence from the randomness in u_{i} decreases as the number of possible values of y_{i} increases due to the shrinkage in the continuity gaps in the PMF. Nevertheless, as suggested by [10], multiple realizations of the RQRs are needed to ensure that any pattern shown in the RQRs is not caused by the randomness in u_{i}. Although this offers a solution to alleviate the impact of the randomness in the RQRs, it is still desired to have a “nonrandom” overall GOF test pvalue for the RQRs. Using the mean of the normality test pvalues from multiple sets of the RQRs is a natural choice. Further research is needed to investigate the null distribution of the mean or other summary for replicated normality test pvalues under the true model.
Further, although deviance residuals may not be appropriate for diagnosing counts regressions, they may be constructed for each systematic component (for example, location and dispersion) to assess specific goodnessoffit. For example, Paula (2013) [35] used a normal probability plot with a simulated envelope for the deviance component residual for the mean and precision models in double generalized linear models [36]. Extension of randomized quantile residuals for each systematic component could be developed to give local information for assessing the specific goodness of fit. In addition, in generalized additive models for location, scale and shape proposed by Rigby and Stasinopoulos (2005) [37], the worm plot (a detrended normal QQplot of the normalized quantile residuals) [38], a diagnostic tool for checking the residuals within different ranges of the explanatory variable(s), was proposed to identify regions (intervals) of the explanatory variable within which the model does not adequately fit the data. Such a plot could be constructed based on randomized quantile residuals for checking model inadequacies for count regression models.
In many applications, counts data are often clustered (i.e., longitudinal, spatial, or multilevel data) due to unmeasured clusterlevel confounders. For modeling the complex dependence structure in these types of data, mixedeffects models are widely used (e.g., [39]). Programs to extend normal and nonnormal regression models to clustered or longitudinal data are widely available (e.g., lme4 and mgcv packages in the R software, and glimmix and nlmixed procedures in the SAS software); however, model diagnosis for the mixedeffects models for counts data are still underdeveloped. Further development of extending the RQR method to examine the GOF of the mixedeffects models for counts data is underway by our research group, where the aforementioned data features are often encountered.
RQRs can also be extended as an alternative of the widely used posterior predictive diagnostics [40] for validating hierarchical Bayesian models in the Bayesian framework. However, extending RQRs to diagnose these complex models is nontrivial due to the optimistic bias of the posterior predictive diagnostics, where the actual observations are used twice for sequentially estimating parameters and testing the predictive distribution. The optimal bias could lead to posterior predictive pvalues concentrating around 0.5 rather than being truly uniformly distributed (even after the randomness is applied). Leaveoneout crossvalidation (LOOCV) is an alternative to the posterior predictive diagnostics. However, the actual LOOCV approach is timeconsuming because several Markov chains are required in order to sample from each posterior distribution in which an observation is excluded as a test case. There have been numerous computational methods proposed to apply model diagnostics with the Markov chain samples from the posterior distribution based on the full dataset without the actual LOOCV being implemented [22–25, 41, 42]. It is possible to apply these methods to compute LOOCV RQRs in the Bayesian framework.
Conclusion
In this paper, we demonstrated that the RQRs is approximately standard normally distributed under the correctly specified model, and their overall goodnessoffit (GOF) test (i.e., ShapiroWilk (SW) normality test) is wellcalibrated for diagnosing count regression models including zeroinflated models. Our simulation results indicate that the RQRs perform reasonably well to detect many forms of model misspecification: Nonlinearity, zeroinflation, and overdispersion. As expected, the statistical power of the RQRs for detecting model misspecification tends to be low for small sample sizes with minor deviations from the fitted models to the true models. Nevertheless, RQRs have substantive appeal in diagnosing counts regression models with moderate or large sample sizes and deviations between the fitted and true models.
Appendix A: Supplementary materials
Availability of data and materials
The dataset analysed during the current study is available in the Journal of Statistical Software repository www.jstatsoft.org/article/view/v027i08, named as DebTrivedi.rda.zip.
Abbreviations
 RQR:

Randomized quantile residuals
 NB:

Negative binomial
 GOF:

Goodnessoffit
 GLMs:

Generalized linear models
 ZIP:

Zeroinflated Poisson
 ZINB:

Zeroinflated negative binomial
 PMF:

Probability mass function
 CDF:

Cumulative distributions function
 MQR:

Middlepoint quantile residuals
 QQ plot:

Normal quantilequantile plot
 SW:

ShapiroWilk
 ED:

Emergency department
 LOOCV:

Leaveoneout crossvalidation VIF: Variance inflation factor
References
Lambert D. Zeroinflated Poisson regression, with an application to defects in manufacturing. Technometrics. 1992; 34(1):1–14.
Hall DB. Zeroinflated Poisson and binomial regression with random effects: a case study. Biometrics. 2000; 56(4):1030–9.
Lee AH, Wang K, Yau KKW. Analysis of zeroinflated Poisson data incorporating extent of exposure. Biom J. 2001; 43(8):963–75.
Baetschmann G, Winkelmann R. Modeling zeroinflated count data when exposure varies: With an application to tumor counts. Biom J. 2013; 55(5):679–86.
Yau KK, Lee AH. Zeroinflated poisson regression with random effects to evaluate an occupational injury prevention programme. Stat Med. 2001; 20:2907–20.
Buu A, Li R, Tan X, Zucker RA. Statistical models for longitudinal zeroinflated count data with applications to the substance abuse field. Stat Med. 2012; 31:4074–86.
Neelon B, O’Malley AJ, Smith VA. Modeling zeromodified count and semicontinuous data in health services research Part 1: background and overview. Stat Med. 2016; 35(27):5070–93.
McCullagh P, Nelder JA. Generalized Linear Models, Second Edition. Chapman and Hall/CRC Monographs on Statistics and Applied Probability Series: Chapman & Hall; 1989.
Agresti A. Categorical Data Analysis. New York: Wiley; 2012.
Dunn PK, Smyth GK. Randomized quantile residuals. J Comput Graph Stat. 1996; 5(3):236–44.
Klar B, Meintanis SG. Specification tests for the response distribution in generalized linear models. Comput Stat. 2012; 27:251–67.
Pereira GHA. On quantile residuals in beta regression. Commun Stat Simul Comput. 2019; 48(1):302–16.
Lemonte AJ, Moreno–Arenas G. On residuals in generalized johnson sb regressions. Appl Math Modell. 2019; 67:62–73.
Scudilio J, Pereira GHA. Adjusted quantile residual for generalized linear models. Comput Stat. 2020; 35:399–421.
Welsh A, Cunningham RB, Donnelly CF, Lindenmayer DB. Modeling the abundance of rare species: statisticalmodels for counts with extra zeros. Ecol Model. 1996; 88:297–308.
Min Y, Agresti A. Random effect models for repeated measures of zeroinflated count data. Stat Modell. 2005; 5:1–19.
Yu Q, Chen R, Tang W, He H, Gallop R, CritsChristoph P, Hu J, Tu XM. Distributionfree models for longitudinal count responses with overdispersion and structural zeros. Stat Med. 2013; 32(14):2390–405.
Agresti A. Foundations of Linear and Generalized Linear Models. New York: Wiley; 2015.
Lee AH, Wang K, Yau KKW. Analysis of zeroinflated Poisson data incorporating extent of exposure. Biom J. 2001; 43(8):963.
Pierce DA, Schafer DW. Residuals in generalized linear models. J Am Stat Assoc. 1986; 81(396):977–86.
Davison AC. Statistical Modeling. Cambridge: Cambridge University Press; 2003.
Stern HS, Cressie N. Posterior predictive model checks for disease mapping models. Stat Med. 2000; 19(1718):2377–97.
Marshall EC, Spiegelhalter DJ. Approximate crossvalidatory predictive checks in disease mapping models. Stat Med. 2003; 22:1649–1660.
Marshall EC, Spiegelhalter DJ. Identifying outliers in Bayesian hierarchical models: a simulationbased approach. Bayesian Anal. 2007; 2(2):409–44.
Li L, Feng CX, Qiu S. Estimating crossvalidatory predictive pvalues with integrated importance sampling for disease mapping models. Stat Med. 2017. https://doi.org/10.1002/sim.7278.
Shapiro SS, Wilk MB. An analysis of variance test for normality (complete samples). Biometrika. 1965; 52(3/4):591–611.
Yap BW, Sim CH. Comparisons of various types of normality tests. J Stat Comput Simul. 2011; 81(12):2141–55.
McCullagh P. On the asymptotic distribution of Pearson’s statistic in linear exponentialfamily models. Int Stat Rev. 1985; 53(1):61–67.
Atkinson AC. Two graphical displays for outlying and influential observations in regression. Biometrika. 1981; 68(1):13–20.
Moral R, Hinde J, Demétrio C. Halfnormal plots and overdispersed models in R: The hnp package. J Stat Softw Artic. 2017; 81(10):1–23.
Dean C. Testing for overdispersion in poisson and binomial regression models. J Am Stat Assoc. 1992; 87:451–7.
Deb P, Trivedi PK. Demand for medical care by the elderly: A finite mixture approach. J Appl Econ. 1997; 12:313–36.
Zeileis A, Kleiber C, Jackman S. Regression models for count data in R. J Stat Softw Artic. 2008; 27(8):1–25.
Fox J, Monette G. Generalized collinearity diagnostics. Journal of the American Statistical Association. 1992; 87(417):178–183.
Paula GA. On diagnostics in double generalized linear models. Comput Stat Data Anal. 2013; 68:44–51.
Smyth GK. Generalized linear models with varying dispersion. J R Stat Soc Ser B (Methodol). 1989; 51(1):47–60.
Stasinopoulos D, Rigby R. Generalized additive models for location scale and shape (gamlss) in R. J Stat Softw Artic. 2007; 23(7):1–46.
van Buuren S, Fredriks M. Worm plot: a simple diagnostic device for modelling growth reference curves. Stat Med. 2001; 20 8:1259–77.
Breslow NE, Clayton DG. Approximate inference in generalized linear mixed models. J Am Stat Assoc. 1993; 88:9–25.
Gelman A, Meng X, Stern H. Posterior predictive assessment of model fitness via realized discrepancies. Stat Sin. 1996; 6:733–60.
Gelfand AE, Dey DK, Chang H. Model determination using predictive distributions with implementation via samplingbased methods (with discussion) In: Bernardo JM, Berger JO, Dawid AP, Smith AFM, editors. Bayesian Statistics 4. Oxford: Oxford University Press: 1992. p. 147–67.
Vehtari A, Gelman A. Pareto smoothed importance sampling. 2015. arXiv:1507.02646 [stat].
Acknowledgements
The authors would like to thank the suggestions and comments from the two reviewers, which significantly helped to improve the quality of this manuscript. The authors would also like to acknowledge financial support from the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grants and the Canadian Foundation for Innovation.
Funding
Feng’s research was partially supported by Natural Sciences and Engineering Research Council of Canada (NSERC). Li’s research was partially supported by the NSERC and the Canadian Foundation for Innovation (CFI). The views expressed are those of the author(s) and not necessarily those of the NSERC and CFI. The funders played no role in any of the design of the study, analysis, interpretation of data, and writing the manuscript.
Author information
Authors and Affiliations
Contributions
LL and CF developed the conceptual design of the study. CF, LL, and AS contributed to the simulation study, real data analysis, and drafted the manuscript. All authors have read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Feng, C., Li, L. & Sadeghpour, A. A comparison of residual diagnosis tools for diagnosing regression models for count data. BMC Med Res Methodol 20, 175 (2020). https://doi.org/10.1186/s12874020010552
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12874020010552
Keywords
 Randomized quantile residual
 Model checking
 Model diagnostics
 Goodnessoffit test