How should I weight IPUMS-CPS data for representative HHINCOME statistics at the state-year level?

I would like to get the 90th percentile of the income distribution for each state and year combination from 1989 to 1998. I extracted a pooled file from these years to do so.

Is it possible to summarize the 90th percentile by state and year applying the household weights, or would I need to extract year samples and state samples separately to get the accurate weights?

The weight variables in IPUMS-CPS are representative of the entire nation in that year. This means that if you analyse multiple years in a single file, and do not include the YEAR variable, the weighted frequencies will seem extremely large. However, if you do include YEAR in your analysis the correct estimates will be generated. Since you are interested in an household-level variable, you will need to limit the sample to one person per household. Because IPUMS-CPS rectangularizes the data, if you do not limit the sample to one person per household, your frequencies will be too large since each household will be multiplied by the number of persons living within the household. You can use the variable PERNUM to select a single person per household. As long as you include YEAR in your analysis and select only one person per household, it is appropriate to apply HHWTSUPP in the way you have described.

I hope this helps.

Thank you for the help. Just to clarify, I extracted a pooled 1989-1998 file with year, region, and state variables in addition to household income.

Let’s say I want a weighted mean for 1990 household income in Alabama. To get the accurate weighted mean for the household variable, I would need to first only keep observations in the data where PERNUM = 1. Then, I could take the mean of observations in YEAR = 1990 and STATEFIPS = 01 (Alabama), applying HWTSUPP weights, and this would be representative?

You are correct. You will want to be sure to account for the NIU code of 9999999, as group quarters were not included in the universe for this question.

Rather than using PERNUM, another way of subsetting the data by household is to keep if RELATE == 0101. The out-of-universe households such as group quarters will probably look like single-person households with RELATE > 1000.