很可能是变量间的共线性导致的

https://bbs.pinggu.org/thread-1311860-1-1.html
可以查看这里，有很详细的解释

Resolving the problem


AMOS is functioning as it was designed to do when it encounters a nonpositive variance during bootstrapping and you are estimating standardized estimates. The program can't calculate some standardized estimates because it would require either taking the square root of a negative variance, or dividing by a zero standard deviation.

Admittedly, there is something alarming about the message "An error occurred..." because it sounds like a program error occurred, but that is not the case.

It is true that the probability of this error message occurring can be reduced by reducing the number of bootstrap samples. There are a couple of other solutions that we think are better:

1. Don't ask for standardized estimates if you don't really need them. Of course this is not a useful solution of the main purpose of the bootstrapping is to estimate standard errors or confidence intervals for standardized estimates.

2. Click View->Analysis Properties->Bootstrap, and put an integer larger than 1 in the Bootfactor box. Say you put "4" in the Bootfactor box. Then the program will quadruple the sample size by using each observation 4 times. If you have a sample of 200, then you end up with a sample of 800, where each observation in the original sample of 200 appears 4 times in the sample of 800. Assuming that standard errors are inversely proportional to the square root of sample size, using a Bootfactor of 4 will have the effect of cutting the standard errors in half, so the program corrects for this by doubling each standard error before displaying it. With a larger sample, the estimates in the bootstrap samples will tend to be closer to the estimates in the original sample. If the original sample does not have nonpositive variance estimates it is less likely that the bootstrap samples will have nonpositive variance estimates. Of course if the original sample does have nonpositive variance estimates, a large Bootfactor values makes it more likely that the bootstrap samples will have nonpositive variance estimates. The online help gives a little more information on Bootfactor. (Look for "Bootfactor Method" in the online help index.)

If you don't request standardized estimates, Amos does not check for negative variance estimates during bootstrapping. Even though negative variance estimates don't have any meaningful interpretation that we know of, they don't create any numeric problems as long as you don't try to compute standardized estimates

Thank for your advice，it is so useful for me！

You really did me a favor!