P是常数？
如果因变量0不是很多的话，普通的计数模型就可以了，没有必要使用ZI模型

这里面的P貌似是常数，想知道怎么编程的？

第156页里For ease of interpretation in current setting and to avoid possible identifiability problem in parameter estimation, the proportion of zeros is set as a constant
and covariates are accommodated in the Poisson part only. With the setting of a constant in the logit part, a direct
estimate of the proportion of structural zeros (φ) can be obtained;
这应该是说P为常数也可以吧？
我也郁闷，一般P都不是常数。这有什么区别？代码怎么改呢？

把所有的在zeromodel的自变量去掉即可。我的看法，应用同样的自变量组合，无论在model还是zeromodel，是比较难理解的。不过很多网上的例子都是这么做的。题外话，那个p本就是符合constant分布的。京剧

http://blog.sina.com.cn/s/blog_a3a92636010153r9.html
http://blog.sina.com.cn/s/blog_a3a92636010153ra.html

jingju，P本身符合常数分布？这句如何理解。为什么网上的例子都是用相同的协变量，对于model和zeromodel可以引入不同的协变量吗？如果这样，关于各自协变量的选择又改如何？望指教！

both questions are yes

Jingju，想请教你一段代码。我看你博客写的，然后按你写的编了下。
但proc countreg和proc nlmixed求解参数估计结果不一样，不知道是哪里出错了？
更奇怪的是当P和协变量不相关时，P为常数时，两个过程得到的结果是一样的。

/*-- Zero-Inflated Negative Binomial Model */
proc countreg data=clmfreq  covest=hessian  method=nra;
model CLM_FREQ=new_CAR_USE new_MARRIED new_DENSITY new_INCOME  new_GENDER/ dist=ZINB;
zeromodel CLM_FREQ ~new_CAR_USE new_MARRIED new_DENSITY new_INCOME  new_GENDER /link=logistic;
run;

/*--Proc Nlmixed--*/
proc nlmixed data=clmfreq;
parms b0=0 b1=0 b2=0 b3=0 b4=0 b5=0 a0=0 a1=0 a2=0 a3=0 a4=0 a5=0 infprinfl=0 k=1;
infprinfl= a0+a1*new_CAR_USE+        a2*new_MARRIED+a3*new_DENSITY+a4*new_INCOME+a5* new_GENDER;
infprob=1/(1+exp(-infprinfl));
linp=b0+b1*new_CAR_USE+        b2*new_MARRIED+b3*new_DENSITY+b4*new_INCOME+b5* new_GENDER;
mu=exp(linp);
p=1/(1+mu*k);
ll=(CLM_FREQ=0)*(log(infprob+(1-infprob)*pdf('negbinomial',CLM_FREQ,p,1/k)))+
     (CLM_FREQ^=0)*(log(1-infprob)+log(pdf( 'negbinomial',CLM_FREQ,p,1/k)));
model  CLM_FREQ~general(ll);
predict exp(ll) out=nb0Out(keep=CLM_FREQ pred);
run;

what i can see is, initiating  infprinfl =0 in PARMS statement was not correct.Remove it.
as my understanding, countreg uses no distinct nonlinear optimaztion methods as proc nlmixed does. so, you can expect same results, at least similiar.
i don't have sas/ets so i cannot check countreg. years ago, my sas indeed contained ETS but now it is not accessible to me.
hopefully it is helpful. jingju

我把 infprinfl =0去掉了，但出现
WARNING: The final Hessian matrix is full rank but has at least one negative eigenvalue.
                Second-order optimality condition violated
而且结果和countreg结果还是不一样。（ZIP的countreg和nlmixed结果却是一样的）

yeah. 我的建议是不要过分简单化的给定初始值.比如全是0等等.靠近真实值或许可以避免一些计算上的问题.JINGJU

二楼正解

谢谢你的回复！我将初始值给定为countreg结果，算出的结果很接近。但是仍然有WARNING: The final Hessian matrix is full rank but has at least one negative eigenvalue.Second-order optimality condition violated。这样的做法可靠吗？我设置别的初始值算的结果相差就会大一点，请问我该如何知道初始值？

initiating logk =0.001;
set k =exp(logk);
set qpoints =1 at nlmixed model.

that is what i can think of, just have a try. jingju

我用上面的程序，操作不成功呀，哪位高手帮助写一下零堆积负二项回归sas程序呀，谢谢呀