This page shows an example of negative binomial regression analysis with footnotes explaining the output. The data collected were academic information on 316 students. The response variable is days absent during the school year (daysabs), from which we explore its relationship with math standardized tests score (mathnce), language standardized tests score (langnce) and gender (female).
As assumed for a negative binomial model our response variable is a count variable, and each subject has the same length of observation time. Had the observation time for subjects varied, the model would need to be adjusted to account for the varying length of observation time per subject. This point is discussed later in the page. Also, the negative binomial model, as compared to other count models (i.e., Poisson or zero-inflated models), is assumed to be the appropriate model. In other words, we assume that the dependent variable is ill-dispersed (either under- or over- dispersed) and does not have an excessive number of zeros.
The dataset can be downloaded here.
data preg; set "C:\temp\lahigh"; female = (gender = 1); run; proc genmod data = preg; model daysabs = mathnce langnce female / link=log dist=negbin; run; The GENMOD Procedure Model Information Data Set WORK.PREG Distribution Negative Binomial Link Function Log Dependent Variable DAYSABS number days absent Number of Observations Read 316 Number of Observations Used 316 Criteria For Assessing Goodness Of Fit Criterion DF Value Value/DF Deviance 312 356.9348 1.1440 Scaled Deviance 312 356.9348 1.1440 Pearson Chi-Square 312 337.0888 1.0804 Scaled Pearson X2 312 337.0888 1.0804 Log Likelihood 2149.3649 Algorithm converged. Analysis Of Parameter Estimates Standard Wald 95% Confidence Chi- Parameter DF Estimate Error Limits Square Pr > ChiSq Intercept 1 2.2849 0.2099 1.8735 2.6962 118.52
Model Information Data Seta WORK.PREG Distributionb Negative Binomial Link Functionc Log Dependent Variabled DAYSABS number days absent Number of Observations Reade 316 Number of Observations Usede 316
a. Data Set – This is the SAS dataset on which the negative binomial regression was performed.
b. Distribution – This is the assumed distribution of the dependent variable. Negative binomial regression is a type of generalized linear model. As such, we need to specify the distribution of the dependent variable, dist = negbin, as well as the link function, superscript c.
c. Link Function – This is the link function used for the negative binomial regression. By default, when we specify dist = negbin, the log link function is assumed (and does not need to be specified); however, for pedagogical purposes, we include link = log. When we write our model out, log( μ ) = β0 + β1x1 + … + βpxp, where μ is the count we are modeling, log( ) defines the link function (i.e., how we transform μ to write it as a linear combination of the predictor variables).
d. Dependent Variable – This is the dependent variable used in the negative binomial regression.
e. Number of Observations Read and Number of Observations Used – This is the number of observations read and the number of observation used in the poisson regression. The Number of Observations Used may be less than the Number of Observations Read if there are missing values for any variables in the equation. By default, SAS does a listwise deletion of incomplete cases.
Criteria For Assessing Goodness of Fit
Criteria For Assessing Goodness Of Fit Criterionf DFg Valueg Value/DFh Deviance 312 356.9348 1.1440 Scaled Deviance 312 356.9348 1.1440 Pearson Chi-Square 312 337.0888 1.0804 Scaled Pearson X2 312 337.0888 1.0804 Log Likelihood 2149.3649
Prior to discussing the Criterion, DF, Value and Value/DF, we need to discuss the logic of this section. Attention is placed on Deviance and Scaled Deviance; the argument naturally extends to Pearson Chi-Square. Also, this section is more applicable to Poisson regression where issues of dispersion are relevant. Negative binomial regression handles dispersion issues by modeling the dispersion parameter of the response variable.
First, note the Deviance has an approximate chi-square distribution with n-p degrees of freedom, where n is the number of observations, p is the number of predictor variables (including the intercept), and the expected value of a chi-square random variable is equal to the degrees of freedom. Then, if our model fits the data well, the ratio of the Deviance to DF, Value/DF, should be about one. Large ratio values may indicate model misspecification or an over-dispersed response variable; ratios less than one may also indicate model misspecification or an under-dispersed response variable. A consequence of such dispersion issues is that standard errors are incorrectly estimated, implying an invalid chi-square test statistic, superscript p. Importantly, however, assuming our model is correctly specified, the regression estimates remain unbiased in the presence of ill-dispersion. A “fix” is to adjust the standard error of the estimates. The standard error correction corresponds to the approach for the scaled criterion. A naive explanation is that when the scale option is specified (scale = dscale), the Scaled Deviance is forced to equal one. By forcing Value/DF to one (dividing Value/DF by itself), our model becomes “optimally” dispersed; however, what actually happens is that the standard errors are adjusted ad hoc. The standard errors are adjusted by a specific factor, namely the square root of Value/DF.
f. Criterion – Below are various measurements used to assess the model fit.
Deviance – This is the deviance for the model. The deviance is defined as two times the difference of the log-likelihood for the maximum achievable model (i.e., each subject’s response serves as a unique estimate of the negative binomial parameter), and the log likelihood under the fitted model. The difference in the Deviance and degrees of freedom of two nested models can be used in likelihood ratio chi-square tests.
Scaled Deviance – This is the scaled deviance. The scaled deviance is equal to the deviance since we did not specify the scale=dscale option on the model statement.
Pearson Chi-Square – This is the Pearson chi-square statistic. The Pearson chi-square is defined as the squared difference between the observed and predicted values divided by the variance of the predicted value summed over all observations in the model.
Scaled Pearson X2 – This is the scaled Pearson chi-square statistic. The scaled Pearson X2 is equal to the Pearson chi-square since we did not specify the scale=pscale option on the model statement.
Log Likelihood – This is the log likelihood of the model. Instead of using the Deviance, we can take two times the difference between the log likelihood for nested models to perform a chi-square test.
g. DF and Value – These are the degrees of freedom DF and the respective Value for the Criterion measures. The DF equals n-p, where n is the Number of Observation Used and p is the number of parameters estimated.
h. Value/DF – This is the ratio of Value to DF given in superscript g. Refer to the discussion at the beginning of this section for an interpretation/use of this value.
i. Algorithm Converged – This is a note indicating that the algorithm for parameter estimation has converged, implying that a solution has been found.
Analysis of Parameter Estimates
Analysis Of Parameter Estimates Standard Wald 95% Confidence Chi- Parameterj DFk Estimatel Errorm Limitsn Squareo Pr > ChiSqo Intercept 1 2.2849 0.2099 1.8735 2.6962 118.52
j. Parameter – Underneath are the intercept, the predictor variables and the dispersion parameter.
k. DF – These are the degrees of freedom DF spent on each of the respective parameter estimates. The DF define the distribution used to test Chi-Square, superscript o.
l. Estimate -These are the estimated negative binomial regression coefficients for the model. Recall that the dependent variable is a count variable, and the regression models the log of the expected count as a linear function of the predictor variables. We can interpret each regression coefficient as follows: for a one unit change in the predictor variable, the difference in the logs of expected counts of the response variable is expected to change by the respective regression coefficient, given the other predictor variables in the model are held constant.
Also, each subject in our sample was followed for one academic year. If this was not the case (i.e., some subjects were followed for half a year, some for a year and the rest for two years) and we neglect exposure time, the regression estimates would be biased since our model assumes all subjects had the same observation time. If this is an issue, we could use the offset option in the model statement, offset=logvarname, where logvarname corresponds to the log of a variable specifying length of time an individual was followed to adjust the regression estimates. The log of the time-followed variable must be calculated in an earlier data step.
Intercept – This is the negative binomial regression estimate when all variables in the model are evaluated at zero. For males (the variable female evaluated at zero) with zero mathnce and langnce test scores, the log of the expected count for daysabs is 2.2849 units. Note that evaluating mathnce and langnce at zero is out of the range of plausible test scores. Had the test scores been mean-centered, the intercept would have a natural interpretation: the log of the expected count for males with average mathnce and langnce test scores.
mathnce – This is the negative binomial regression estimate for a one unit increase in math standardized test score, given the other variables are held constant in the model. If a student were to increase her mathnce test score by one point, the difference in the logs of expected counts would be expected to decrease by 0.0016 unit, while holding the other variables in the model constant.
langnce – This is the negative binomial regression estimate for a one unit increase in language standardized test score, given the other variables are held constant in the model. If a student were to increase her langnce test score by one point, the difference in the logs of expected counts would be expected to decrease by 0.0143 unit while holding the other variables in the model constant.
female – This is the estimated negative binomial regression coefficient comparing females to males, given the other variables are held constant in the model. The difference in the logs of expected counts is expected to be 0.4312 unit higher for females compared to males, while holding the other variables constant in the model.
Dispersion – This is the estimate of the dispersion parameter. If the dispersion parameter equals zero, the model reduces to the simpler Poisson model; if dispersion is greater than zero the response variable is over-dispersed; and if dispersion is less than zero the response variable is under-dispersed.
m. Standard Error – These are the standard errors of the individual regression coefficients. They are used in both the Wald 95% Confidence Limits, superscript n, and the Chi-Square test statistic, superscript o.
n. Wald 95% Confidence Limits – This is the Wald Confidence Interval (CI) of an individual regression coefficient, given the other predictors are in the model. For a given predictor variable with a level of 95% confidence, we’d say that we are 95% confident that upon repeated trials, 95% of the CI’s would include the “true” population negative binomial regression coefficient. It is calculated as Estimate (zα/2)*(Standard Error), where zα/2 is a critical value on the standard normal distribution. The CI is equivalent to the Chi-Square test statistic: if the CI includes zero, we’d fail to reject the null hypothesis that a particular regression coefficient is zero, given the other predictors are in the model. An advantage of a CI is that it is illustrative; it provides information on the precision of the point estimate.
o. Chi-Square and Pr > ChiSq – These are the test statistics and p-values, respectively, testing the null hypothesis that an individual predictor’s regression coefficient is zero, given the rest of the predictors are in the model. The Chi-Square test statistic is the squared ratio of the Estimate to the Standard Error of the respective predictor. The Chi-Square value follows a central chi-square distribution with degrees of freedom given by DF, which is used to test against the alternative hypothesis that the Estimate is not equal to zero. The probability that a particular Chi-Square test statistic is as extreme as, or more so, than what has been observed under the null hypothesis is defined by Pr>ChiSq.