Several explanations have been identified. Although the factors listed below are sufficient to explain the observed differences, it is possible that other factors may also play an important role.

(1) Different approaches to counting seasonal migrants.

Recent decennial censuses have been conducted primarily in March and April, and the 2000 Census did not include any special instructions to ensure that seasonal migrants were counted at their "usual residence." It is therefore likely that some usual residents of Michigan were counted by the census at seasonal homes in the Sunbelt. Counting retiree households elsewhere would tend to result in higher median and average incomes for Michigan.

The ACS counts people where they are found, provided that they are staying there for at least two months. This means that seasonal residents are generally counted in Michigan if they are surveyed during a month when they are here, and they are counted in other states if they are surveyed during a month when they are elsewhere. Counting more seasonal households in Michigan would tend to depress average earnings and median income. Counting more individual household members in Michigan, on the other hand-such as students who are home from out-of-state colleges in the summer or individuals who spend the summer with family members in Michigan-would tend to increase average and median household incomes.

(2) Higher rates of missing data in the 2000 Census.

Wage data had to be imputed for 16.4% of the individuals in the 2000 Census, but imputations were made for only 5.1% of the individuals in the 2004 ACS. (These statistics are based on analysis of the nation-wide Public Use Microdata Samples from the 2000 Census and the 2004 ACS). This factor would tend to make ACS data more accurate relative to the 2000 Census.

3) A tendency for respondents to report higher wage income when the question is asked on a calendar-year basis.

The 2000 Census collected income data for the 1999 calendar year at a time when respondents had access to W-2 forms, 1099 forms, and other documents that contained information about their annual income. ACS data, on the other hand, are collected year-round with respect to the twelve months prior to the survey. Respondents do not necessarily even have documentation for their income over that time period, and they do not necessarily base their responses on careful calculations even if they do have the necessary documentation. This factor would tend to decrease the accuracy of the ACS relative to past censuses.

(4) A tendency for the "optical character recognition" software used in the 2000 Census to produce higher income figures than human beings reading the same entries.

The Census Bureau found a tendency for the OCR software used in the 2000 Census to produce higher income levels than human coders interpreting the same entries. This may reflect the fact that most potential misinterpretations of digits involve small numbers being interpreted as larger numbers, such as a carelessly drawn zero being mistaken for a six.

(5) Different time periods covered by the two data sources.

The Census collected income data for previous calendar year. That is the same time span covered by January ACS interviews, but each subsequent month's interviews cover a progressively later period of time. On average, the ACS interviews conducted in 2000 covered the period from mid-June of 1999 to mid-June of 2000. Any real decreases in income which occurred during 2000 would be reflected in the ACS, but not in the Census.

Click here for a Census Bureau report on differences between income data from Census 2000 and the ACS.