Tim Riffe's profile photoCarl Schmertmann's profile photoGeorge Hough's profile photoEddie Hunsinger's profile photo
Small ACS sample sizes are a big headache, but the data on response rates also appears to be misleading.

Can anyone explain the unusual definition?

For example, in 2010 the data at the original link above show 1.918 mil interviews out of 2.900 mil household addresses selected. That's 66.1%. Yet the "weighted" response rate data show "97.5%" for 2010.

What does 97.5% mean in this context?
Carl - of the 66.1% that actually got contacted, 97.5% responded, the other 2.5 Percent were comprised of refusals, no on home, insufficient data, etc. I have been looking at these rates too and have a spreadsheet that shows that not only do they NOT survey 3 million households, but that the Census 200 Long Form reaches almost ALL the Households from their original Sample Frame (not 66.1%), and hence reaches the stated goals of over-sampling small geogs. ACS does NOT. Check you email as I do not know how to link google share docs here.
Sorry if I'm asking dumb questions, but I'm still uncertain about the difference between (1) a household address that is chosen for the ACS sample, and (2) a household address that is contacted.

Apparently the response rate is calculated with (2) in the denominator rather than (1). Why?
That seems like a reasonable question to me. Guessing it doesn't matter technically once we know the definitions, but it does seem strange. Maybe they don't want to include vacant units in the rate. Or maybe it's appearance?? I don't know. I bet they'd follow up-- they're really good (though ACS is a mess in a lot of ways I think (just using consistent residency requirements would be a huge help))-- would be interested to hear.
So is the high response rate an estimate of "What % of non-vacant housing units mailed back the form (or were eventually interviewed) with at least some responses?".

Because of time lags between the sample month and followup, there are some technical questions about vacancy (namely, vacant WHEN?), but is that the main idea?

As George says, the main question is about the unweighted sample sizes, not the 'response rate'. But I was curious.

Edit after posting: Added (or were eventually interviewed)
I think I'd need to know more about this-- that (low) 66% might also have to do with how Census does their mailings and follow up (and what proportion they follow up on). Any Census folks out there to respond?
Eddie and Carl - Here is the skinny. Census 2000 LF sent out 18.3 million Forms to Housing Unit Addresses (15.8% of all HUs); then followed up with 16.4 million HHold samples (15.6% of all HHolds) = 89.7% of Total LFs resulted in Sample HHold Data. ACS 2006-2010 sent out 14.5 million forms to Housing Unit Addresses (11.1% of all HUs); then followed up by receiving 9.6 million HHold Samples (8.5% of all HHolds) = 66.9% of Total ACS Forms returned HHold data. Census 2000 LF at 15.6% of Households did not miss the 1/6 mark by much. ACS 2006-2010 is about a 1/12 sample, hence the missing data and high standard errors in the ACS data. Think if how much better the data would be with 7 million more households of data.
Yes I wish they could have invested in a huge (maybe 15-20%) sample to be done every five years rather than a small every year ACS, but oh well. Now I just wish they would get their residency definitions consistent (2 month residence for ACS controlled to primary residence for Estimates just doesn't work). Well I know they work very hard (I see it all the time, and a couple points at which I've been impressed stick in my mind), and this project is of huge importance in any case.
Add a comment...