P. 39: Should be "Sample size requirements for CRTs are discussed in more detail in Chapter 7."

 

Reference on p. 372: Should be "Greene, E. J. 2017. A SAS macro for covariate-constrained randomization of general cluster-randomized and unstratified designs. Journal of Statistical Software  77: Code Snippet 1."

   (mistake made by the publisher, when the reference          became available just before the book went to press) 

​

Example 12.2, pp. 287-289: The adjusted t-statistics should be compared to a Student t distribution with 4 d.f., not the 5 d.f. used near the end of Box 12.1 and in the paragraph just after it. One of the covariates, baseline HIV prevalence, is a cluster-level covariate, and to be conservative, it is best to subtract a d.f. when adjusting for it (as mentioned in Section 10.5.2.2. on p. 225). Thus, the two-sided p-value for the adjusted t-statistic of -4.1051 should be 0.015, instead of 0.009.

Note that when adjusting for cluster-level covariates, we should consider the general rule-of-thumb that says there should be 5-10 observations for each estimated parameter beyond the design-based covariates (i.e. dummy variables for strata). In the case of CRTs, we might consider this to apply to the number of clusters. In this example, it would be unwise to adjust for more than one cluster-level covariate.

​

Section 7.4.3 (Example 7.7), pp. 149-151: In the example, the estimated standard deviation of the response proportion across communities is seen to be 0.0786. The text then states "Note that the relative precision of the estimate 0.0786 is 0.0211/0.0786, or 27%..." However, Box 7.3 with the calculations does not have the value 0.0211. This number was left over from the penultimate version of the box, in which the Stata procedure xtmixed was used; that routine displayed, in its standard output, the estimated standard error of 0.0786, which was 0.0211. The mixed procedure has largely replaced xtmixed, but it is relatively easy to obtain an approximation of this SE. The "delta method" estimate of the variance of the square root of a random variable is σ /(4µ). Here the r.v. is var(_cons), so the variance of its square root is: 0.0033231 /(4*0.0061815), and the square root of this is 0.0211. 

 

In Box 11.8, p. 269, the output of the two calls to the xtgee routine is given as:

​

. xtgee debut i.arm if sex==1, family(binomial) link(logit) corr(exch) i(school

vce(robust > ) eform

. xtgee debut i.arm i.scpar if sex==1, family(binomial) link(logit) corr(exch)

i(school) vc > e(robust) eform

 

The ">" are what Stata uses to indicate the output has wrapped from the previous line, but in formatting for the box, the line break happened elsewhere. For both of these, what was input was "vce(robuste)" . Also, in the first call, the ") was cut off the end of i(school).