can't replicate median household income in ACS

Hi,

I was wondering if I should NOT expect to exactly match the median household income figures published by the Census Bureau in FactFinder when I run my own analysis of the ACS PUMS data in Stata. I calculated the median household income in California from 2007 to 2015 and I get figures that are close to what’s published in FactFinder, but they don’t exact match, and in some years, the difference between my values and those in FactFinder is off by close to or more than 2%. Is this to be expected or could I be doing something wrong? If this is to be expected, is there any way I can confirm that my analysis is correct? The PUMS estimates for user verification on the ACS website do not include the median household income or any other income values. (Also, I should note that my figures are based on the same number of households as published in FactFinder so that’s not the problem.)

Here are my values vs FactFinder’s:

|

Median Household Income, CA, 2016 Dollars

| |
| | | |
| |

Stata

|

FactFinder

|

% diff

|
|

2007

|

68,297

|

69,395

|

1.6%

|
|

2008

|

66,907

|

68,046

|

1.7%

|
|

2009

|

66,672

|

65,923

|

-1.1%

|
|

2010

|

62,949

|

63,508

|

0.9%

|
|

2011

|

59,763

|

61,136

|

2.3%

|
|

2012

|

60,631

|

60,974

|

0.6%

|
|

2013

|

61,823

|

62,019

|

0.3%

|
|

2014

|

61,842

|

62,788

|

1.5%

|
|

2015

|

65,416

|

65,315

|

-0.2%

|

Thanks in advance for any help!

-Alissa

Estimates derived from the ACS public use microdata are expected to differ slightly from the summary tables available through American FactFinder. This is because the summary tables use the full (and restricted access) ACS samples while the public use microdata files include only a subset of the sample. The microdata sample is then re-weighted to match total population figures. Therefore, while we do expect estimates to differ slightly, we do expect them to fall within the published margins of error.

I am having a similar issue, where the median income I have calculated is off by more than the margin of error. How should I interpret this discrepancy? For reference, I am using the 2019 5-Year sample and the difference between my number ($47,998) and the American FactFinder number ($48,851) is $853, but the margin of error is only +/-146.

Is there something significant wrong with my analysis, or is this still within the acceptable range of discrepancy between the two samples?

This seems to me that it is within the expected range of discrepancy. These are two separate estimates based on different (though overlapping) samples, so they will each have an associated confidence interval. You can read more about the reasons these estimates can differ at this Census FAQ. You should note that the Census Bureau’s margin of error denotes a 90% confidence interval, so there is a 10% chance that the actual population number will be outside that range. Since the Bureau gives millions of estimates, plenty of them will not contain the actual population number.