Errata for First Printing (Sept. 2015) of Regression Modeling Strategies, 2nd Edition

For replacements for R code that no longer works due to changes in the R system or R packages, to go Rms2AltR

Abbreviation	Meaning
Pg	Page
L	Line
Pr	Paragraph
b	From Bottom (ignoring equations)
Eq	Equation
	Corrected in master file

Pg i	Feinberg should be Fienberg
Pg 150	Changes in ggplot2 require code changes (4) below
Pg 174	Dotted blue line is missing in Figure 8.5. It should be at y=3958
Pg 179	Spline model in first large box - see (2)	*
Pg 184 Fig. 9.2	Caption should have log L not just L
Pg 240	Table 10.4 114.51 should be 114.41
Pg 256 R code	x=TRUE, y=TRUE should be x=TRUE, y=TRUE, data=sex.age.response
Pg 256 R code	which.influence(f, .4) should be round(resid(f, 'dfbetas'), 1)
Pg 306	five completed datasets should be twenty completed datasets plus 2 other places
Pg 360 2 lines under Eq 15.1	X should be X subscript i
Pg 362 Eq. 15.7	n should be k
Pg 373 Fig. 15.4	Was run with default family (logistic) instead of intended family='loglog'; agreement increases
Pg 378	family="loglog" can be family=loglog
Pg 474	g 0.033 gr 1.959 should be g=1.959 gr=7.095
Pg 508 Fig. 20.8	Omitted text (1) and index under figure
Pg 516 top	Code for Fig. 20.7: see Rms2AltR
Pg 516 bottom	Code for Fig. 20.10: see Rms2AltR
Pg 526	3 paragraphs (3) omitted before rcspline.plot(lvef, d.time, event=cdeath, nk=3) due to omission of @ in previous chunk

---

* Denotes places in the master text where changes are indicated with TODO and lines need to be uncommented.

(1) You can see that the ratio is not stable from 160 to 190 days, but is mostly flat thereafter. \index{hazard ratio!interval-specific} Interval-specific estimates of the group 2 : group 1 log hazard ratios are shown in Table~\ref{tab:cox-ratios-group} with intervals specified to yield equal numbers of deaths. The Efron likelihood was used. [had omitted @ to close knitr chunk above]

(2) To make the model converge, and to actually show a dotted blue line, the model had to have some splines reduced from 5 knots to 4 knots: f <- cph(S ~ rcs(sz,4) + rcs(sg,4) + rcs(log(ap),5) + rcs(sbp,4) + rcs(dbp,4) + rcs(age,3) + rcs(wt,4) + rcs(hg,4) + ekg + pf + bm + hx, tol=1e-14, data=imputed) New text interpreting the results: The model allowing all continuous predictors to be nonlinear is better than the linear full model but not b etter than the 5 component model.

(3) The standard errors of log survival probabilities computed by \fu{cph} and the confidence intervals derived by the associated functions \fu{survest} and \fu{survplot} are only proper when the model contains only stratification factors or when predictions are made at the means of all covariables, unless you store in the fit object the data used in fitting the model using the options \Co{x=TRUE, y=TRUE}. The \fu{survfit.cph} and \fu{survest} functions are modifications of Therneau's \fu{survfit.coxph} function that can obtain predicted survival probabilities and confidence intervals for any setting of predictors. \fu{survfit.cph} and \fu{survest} are called as needed when you request plots.
The \fu{survest} function is used to obtain predicted survival curves or survival at a given vector of times, for a desired set of predictor values and a fit produced by \fu{cph}. Confidence limits may be obtained also, as can the standard error of the log--log survival.
The \fu{survplot} function is similar in some ways to the \fu{Predict} and \Co{plot}, but it plots follow-up time on the $x$-axis. It plots step-function survival curves with confidence bands or confidence bars for Cox models.
Figure~\ref{fig:cox-ef-spline} was produced using ...

(4) require(data.table) cdystonia <- data.table(cdystonia) cdys <- cdystonia[, j=as.list(quantile(twstrs, (1 : 3)/4)), by = list(treat, week)] cdys <- upData(cdys, rename=c('25%'='Q1', '50%'='Q2', '75%'='Q3')) ggplot(cdys, aes(x=week, y=Q2)) + xl + yl + ylim(0, 70) + geom_line() + facet_wrap(~ treat, nrow=2) + geom_ribbon(aes(ymin=Q1, ymax=Q3), alpha=0.2)