Using multi-year estimate data from American Community Survey in models?

Question

(EDITED FOR CLARITY) I have been looking around, and I can't seem the find the answer to my question. Apologies if it has been hiding in plain sight.

In order to obtain school district-level information, I have been using American Community Survey data obtained from American FactFinder. Variables I am pulling include demographic characteristics such as median income and percent of citizens over 65 years old. Five-year estimate American Community Survey data are provided for all geographic regions, whereas one-year estimate data are not provided for smaller regions. As such, there are estimates from 2005-2009, 2006-2010, etc. I understand that these data represent reliable estimates for geographies over a period and should not be used as estimates for a single year of data. (Also, note that I am not thinking yet about the standard errors in this question.)

How, then, can we use these data in data analyses where other data sources are available at the year level? Note that I do not want to use the one-year estimates because of reliability and availability reasons.

PART A: For example, if we want to use median income in a model for 2009, and we use the five-year estimates from ACS, which year would we use? 2005-2009? 2006-2010?

PART B: Let's say we use 2005-2009. I know from the documentation that we should not pretend as if any of the period data represent data from a single year. Would it be accurate to say, in this example model, that using 2005-2009 estimate median income data as a covariate would be "controlling for approximated median income data from the last five year period"?

PART C: AND, if that is the case, we are missing out on larger year-to-year changes, since the five-year estimates smooth the data trends. In this case, should we maybe then run limited models using one-year estimates as robustness checks?

(Does anyone have any resources with people using census data in this way?)

Note: I have thus far been downloading data from American Factfinder because I need it at the school district level.

Answer 1

Without knowing the goals and details of your model, I probably could not answer the question fully, so here are just some tentative suggestions.

There are cases to use, say, ACS 2006--10 income data to construct some income measure for 2010. See here for an example, and I quote:

The other substantial change in the methodology is the replacement of the decennial 2000 sample estimates of school-age poverty as one of the inputs to the school district estimation process with estimates derived from the latest five-year sample of the American Community Survey (ACS 2006-10). In the 2010 decennial census, no long-form survey questions were included. This decennial sample data has now been replaced by the continuous measurement ACS. So the SAIPE program also replaced the decennial sample estimates as an input to the school district process.

There are tradeoffs in using the two different sources as an input to the school district estimation process. The survey associated with the decennial 2000, in general, has lower errors associated with sampling than the ACS five-year survey. To the ACS advantage, the ACS five-year estimates are more timely relative to the decennial 2000 as an input to 2010 estimation. Combining the two effects, preliminary evaluation indicates that the ACS five-year estimate as an estimator of current-year poverty represents an improvement in relative error compared to using the decennial 2000 estimate. The impact on the final SAIPE school-district estimates will thus yield an improvement in precision, but a fuller evaluation is required to give more detail on how the uncertainty of these estimates have improved.

It depends on your empirical model to see whether the median income in 2009 is exactly what you needed, or is it sufficient to use a five-year average. Or if whatever you are trying to estimate takes time to affect your outcome variable, then previous-year averages may even be preferred to current-year estimates. If 2009 estimate is strictly preferred, have you considered ACS 3-year files instead of 5-year files. 3-year may strike some balance between data reliability and year-to-year changes.

Further more, is median income what you really needed? Maybe share of shcool-age children who are in poverty is as good a measure? Again, there is no way for me to know that. I only use it as an example, and the annual data is available here.

Also, for part C of your question, I would suggest using yearly data to check, despite lower precision of the variables. Or at least take a look at year-to-year changes in median income around 2009.