A method for the
anova function, for use on
svycoxph objects. With a single model argument it produces a sequential anova table, with two arguments it compares the two models.
# S3 method for svyglm anova(object, object2 = NULL, test = c("F", "Chisq"), method = c("LRT", "Wald"), tolerance = 1e-05, ..., force = FALSE) # S3 method for svycoxph anova(object, object2=NULL,test=c("F","Chisq"), method=c("LRT","Wald"),tolerance=1e-5,...,force=FALSE) # S3 method for svyglm AIC(object,...,k=2, null_has_intercept=TRUE) # S3 method for svyglm BIC(object,...,maximal) # S3 method for svyglm extractAIC(fit,scale,k=2,..., null_has_intercept=TRUE) # S3 method for svrepglm extractAIC(fit,scale,k=2,..., null_has_intercept=TRUE)
Use (linear combination of) F or chi-squared distributions for p-values. F is usually preferable.
Use weighted deviance difference (LRT) or Wald tests to compare models
For models that are not symbolically nested, the tolerance for deciding that a term is common to the models.
BIC, optionally more
Does the null model for AIC have an intercept or not?
Force the tests to be done by explicit projection even if the models are symbolically nested (eg, for debugging)
svyglm model that
object (and ... if supplied) are nested in.
Multiplier for effective df in AIC. Usually 2. There is no choice of
k that will give BIC
The reference distribution for the LRT depends on the misspecification effects for the parameters being tested (Rao and Scott, 1984). If the models are symbolically nested, so that the relevant parameters can be identified just by manipulating the model formulas,
anova is equivalent to
regTermTest.If the models are nested but not symbolically nested, more computation using the design matrices is needed to determine the projection matrix on to the parameters being tested. In the examples below,
model2 are symbolically nested in
model0 can be obtained just by deleting terms from the formulas. On the other hand,
model2 is nested in
model1 but not symbolically nested: knowing that the model is nested requires knowing what design matrix columns are produced by
as.numeric(stype). Other typical examples of models that are nested but not symbolically nested are linear and spline models for a continuous covariate, or models with categorical versions of a variable at different resolutions (eg, smoking yes/no or smoking never/former/current).
A saddlepoint approximation is used for the LRT with numerator df greater than 1.
AIC is defined using the Rao-Scott approximation to the weighted
loglikelihood (Lumley and Scott, 2015). It replaces the usual penalty term p, which is the null expectation of the log likelihood ratio, by the trace of the generalised design effect matrix, which is the expectation under complex sampling. For computational reasons everything is scaled so the weights sum to the sample size.
BIC is a BIC for the (approximate) multivariate Gaussian models
on regression coefficients from the maximal model implied by each
submodel (ie, the models that say some coefficients in the maximal model
are zero) (Lumley and Scott, 2015). It corresponds to comparing the models with a Wald test and replacing the sample size in the penalty by an effective sample size.
For computational reasons, the models must not only be nested, the names of the coefficients must match.
extractAIC for a model with a Gaussian link uses the actual AIC based on maximum likelihood estimation of the variance parameter as well as the regression parameters.
Object of class
seqanova.svyglm if one model is given, otherwise of class
At the moment,
AIC works only for models including an intercept.
Rao, JNK, Scott, AJ (1984) "On Chi-squared Tests For Multiway Contingency Tables with Proportions Estimated From Survey Data" Annals of Statistics 12:46-60.
Lumley, T., & Scott, A. (2014). "Tests for Regression Models Fitted to Survey Data". Australian and New Zealand Journal of Statistics, 56 (1), 1-14.
Lumley T, Scott AJ (2015) "AIC and BIC for modelling with complex survey data" J Surv Stat Methodol 3 (1): 1-18.
data(api) dclus2<-svydesign(id=~dnum+snum, weights=~pw, data=apiclus2) model0<-svyglm(I(sch.wide=="Yes")~ell+meals+mobility, design=dclus2, family=quasibinomial()) model1<-svyglm(I(sch.wide=="Yes")~ell+meals+mobility+as.numeric(stype), design=dclus2, family=quasibinomial()) model2<-svyglm(I(sch.wide=="Yes")~ell+meals+mobility+stype, design=dclus2, family=quasibinomial()) anova(model2) #> Error in .svycheck(design): object 'dclus2' not found anova(model0,model2) #> Working (Rao-Scott+F) LRT for stype #> in svyglm(formula = I(sch.wide == "Yes") ~ ell + meals + mobility + #> stype, design = dclus2, family = quasibinomial()) #> Working 2logLR = 21.52228 p= 0.0013407 #> (scale factors: 1.7 0.3 ); denominator df= 34 anova(model1, model2) #> Working (Rao-Scott+F) LRT for stype - as.numeric(stype) #> in svyglm(formula = I(sch.wide == "Yes") ~ ell + meals + mobility + #> stype, design = dclus2, family = quasibinomial()) #> Working 2logLR = 25.10744 p= 1.816e-05 #> df=1; denominator df= 34 anova(model1, model2, method="Wald") #> Wald test for stype - as.numeric(stype) #> in svyglm(formula = I(sch.wide == "Yes") ~ ell + meals + mobility + #> stype, design = dclus2, family = quasibinomial()) #> F = 17.73552 on 1 and 34 df: p= 0.00017598 AIC(model0,model1, model2) #> eff.p AIC deltabar #> [1,] 3.538287 142.11041 1.179429 #> [2,] 6.248895 119.31208 1.562224 #> [3,] 7.202561 97.03351 1.440512 BIC(model0, model2,maximal=model2) #> p BIC neff #> [1,] 4 120.8183 72.43313 #> [2,] 6 111.6461 NaN ## from ?twophase data(nwtco) dcchs<-twophase(id=list(~seqno,~seqno), strata=list(NULL,~rel), subset=~I(in.subcohort | rel), data=nwtco) a<-svycoxph(Surv(edrel,rel)~factor(stage)+factor(histol)+I(age/12), design=dcchs) b<-update(a, .~.-I(age/12)+poly(age,3)) ## not symbolically nested models anova(a,b) #> Working (Rao-Scott+F) LRT for poly(age, 3) - I(age/12) #> in svycoxph(formula = Surv(edrel, rel) ~ factor(stage) + factor(histol) + #> poly(age, 3), design = dcchs) #> Working 2logLR = 9.743054 p= 0.0086592 #> (scale factors: 1.2 0.82 ); denominator df= 1146