数据表：

临床分期 有骨转移 无骨转移 合计 转移率

Ⅰ期 233 0(4.33) 233 100%

IIA期 413 3(7.7) 416 99.28%

IIB期 356 4(6.6) 360 98.89%

IIIA期 121 11(2.45) 132 91.67%

IIIB期 26 1(0.5) 27 97.30%

IIIB期 13 3(0.29) 16 81.25%

合计 1162 22(21.87) 1184 98.14%

上表是我的数据，分析目的是不同分期样本率的比较。表中有好几个单元格中理论频数T<1，这个数据是否可以直接用行*列表资料的卡方检验？

如果不能，该用什么方法或做怎样的处理。请指教！谢谢。

回复 第1楼 的 liuxingbei：T<5需要做修正，如果你样本总体数目大于40，就做一个校正即可，如果小于40，用Fisher精确概率（这也是最为常用的）

回复 第2楼 的 Tracy X：我的样本个数大于40，其中有T<1的单元格。我的数据是多组样本的，不是两组样本，请问fisher精确概率可以用于多组样本率的比较吗？

回复 第1楼 的 liuxingbei：观测值无论多小都不影响，真正有影响的是期望值，是每一个格子的期望值不能太小。这种误解似乎传播得极为广泛。

回复 第5楼 的 remember, discover, invent：现在我的数据就是期望值太小，才不确定用哪种方法。请问，fisher精确检验可以用于多组样本率的比较吗？

回复 第6楼 的 liuxingbei：Sorry for my careless reading. Fisher's exact test can surely be used. Its less often application in larger contingency tables is mainly due to computational burden. But this does not seem to be a big problem for your data.

As Tracy X suggested, a continuity correction is also possible.

As for jamestony's comments, the question of how to do a reasonable test still remains after fitting a logistic regression. This is not to say I disagree with the use of logistic regression. It may actually be more powerful against certain alternatives, like the trend test (the trend test is equivalent to the socre test in logistic regression).

My point is that probably the difference in your data is so obvious that any reasonable tests should reject the null. Thus, probably there is not much difference in choosing any one of them.

PS: Just for curiosity about your criterion: why do more advanced tumors have lower metastatis rates? Because most of those with metastatsis died before they enter more advanced stages? Moreover, isn't the criterion you used to classify the stages considering metastatsis status? If so, they will surely be dependent, without needing to see any data.

回复 第9楼 的 remember, discover, invent：OK