Margin of Error for a zero cell count

The way they handle MOE's except in the random replicates is startistically wrong.  They report negative counts as part of the MOE.  See this correspence with the Bureau that I had for more  www.dropbox.com/.../ABu7stdM8PGZGmcgQnuqKLU

I think the formulae for calculating the variance for zero count cells is included in the ACS Design and Methodology report but I find it easier to just use variance replicate table documentation:

https://www2.census.gov/programs-surveys/acs/replicate_estimates/2022/documentation/5-year/2018-2022_Variance_Replicate_Table_Documentation.pdf

https://www2.census.gov/programs-surveys/acs/replicate_estimates/2022/documentation/5-year/2018-2022%20Appendix%20A%20Average%20Weights%20and%20k%20Values.pdf 

Basically the MOE(0) = 1.645 * sqrt( Average weight * k-value ), both of which are geography specific.

I don't need to know on how I should compute the MoE for a zero cell -- I want to know how they compute the MoE in published tables based on the ACS raw data.  I think that they make use some type of jackknife style procedure.

Best,

Dave

David--

They use the normal approximation and that is why it goes negative with zero counts.  This is for there basic tables not the Random Reeplicates, which were done after I started complaining.  Generally the doctrine is that you are not supposed to use the normal approximation except when the % is in the middle of the distribution, something like 30 to 70 percent.  Whne you get towards one of the tails of the distribution, one can expect to have non-asymetric standard errors.  The Census still publishes the normal approximation, which leads them to publish a number that is a margin of error that can go negative.  Plainly for a zero in a sample one cannot go below zero, but there could some that were missed.  Sudaan has a variety of methods, they are also buiilt into SAS and Sata and they can work with complex samples.  

That is why the suggestion of using the replicates make sense.  Take a a look at my memo and their response, the only change is that they said that you should not take the negative part of the interval seriously.  Their original notion was that any estimate was better than none.  Similarly, their approximation for the median is based upon the distribution being symetric.  Have you ever seen a Symetric income distribution?

Andy Beveridge