Hello IPUMS Community,
I’m analyzing ACS 5-year dataset for Florida and have come across households with
NFAMS > 1. I want to split these households into individual families using
FAMUNIT, but how should I adjust the
HHWT? Currently, I’m dividing
NFAMS to allocate weights to each family. Is this method appropriate, or is there a recommended approach to recalculating
HHWT in this scenario?
Thanks for your guidance!
It sounds like you are interested in conducting a family-level analysis using the ACS public use microdata sample (PUMS) data. The Census Bureau does not provide a family weight in their release of the PUMS data. However, the Census Bureau’s Weighting and Estimation Chapter for the 2010 ACS documentation notes that estimates of families are based on the household weight (HHWT in IPUMS USA). Following this approach seems reasonable as it is in line with how the Census calculates estimates and should allow you to benchmark your analytical sample against published estimates, but I will share a few things to consider.
First, you should determine if the family-level variables you are working with are actually for families or if they are for households; the two concepts can be a bit obscured in some variables (if you are using household-level variables, then household-level weights are perfectly appropriate!). Additionally, NFAMS is an IPUMS-constructed variable based on our family interrelationship algorithm, and it may not map directly onto Census Bureau definitions of families; depending on your application (e.g., trying to match official estimates) you may be interested in reviewing alternate variables about the household composition. How you are defining families is relevant because if you are applying household weights to family-level analyses, households with more than one family will be counted more than once. In contrast, under the approach you propose, multi-family households will only be counted once and their weight will be evenly divided amongst all families in the household.
I am also trying to do an analysis at the family level using the IPUMS definition of family (variable FAMUNIT). Specifically, I am trying to compute the average, number of families represented in the selected population, and the median (and other quantiles) of family income using FAMUNIT and FTOTINC variables but I am facing similar challenges as @Shane_Dabney because I am not sure how to use weights in this case. In my analysis, I do not intend to restrict my sample to PERNUM = 1 because I do not want to drop any families that are denoted as a second, third,… family within a given household. Would create a new family weights variable that divides HHWT by NFAMS work in my situation? What are some alternative methods for applying weights at the family level (using FAMUNIT definition of family) without restricting the sample to PERNUM = 1?
For reference, I am using the ACS 5-year dataset (2017-2021).
I hope my questions make sense. I would appreciate any guidance on this.
For family-level analysis, instead of using PERNUM=1, you should identify a single case from each family, and use only those in your analysis. It is appropriate to use HHWT for analyses at the family level because within a given stratum, households are sampled with equal probability. However, there are differences in response rates by demographics, mode of response, and other factors. The household weights are designed to correct for these imbalances. Once they are corrected, the households included in the survey are as good as randomly sampled (when using HHWT). Families are nested within households, meaning the sample of families observed in the survey is also as good as random. We would not expect to see systematic nonresponse that is correlated with the number of families in a household.
If you restrict your analytical sample to PERNUM=1, you would drop out second (and third, etc.) families from each household. The same caveats Kari noted about ensuring you are using family variables to measure family characteristics, and household variables to measure family characteristics, would still apply if you restrict your analytical sample with PERNUM=1.
Thank you so much, @Isabel_Pastoor for the quick response; this is very useful.
In my analysis, I am using the family definition of FAMUNIT. As you suggested I will be identifying a single case from each family and then using HHWT as my weights for calculations, including means, shares, and quantiles (if I understood that correctly).
Would HHWT also work for obtaining a count of families represented in the population(using the FAMUNIT definition)? I am considering what @KariWilliams mentioned previously, as I wouldn’t want to count households with more than one family more than once: "How you are defining families is relevant because if you are applying household weights to family-level analyses, households with more than one family will be counted more than once. "
If using HHWT is not appropriate for computing counts, would dividing HHWT by NFAMS work here or what are some alternative options?
If you narrow your analysis to just one person in each family from each household (i.e., unique observations of YEAR, SERIAL, and FAMUNIT), you can use HHWT to estimate the number of families fitting some parameters without double counting. In her post, Kari meant that households with more than one family would be counted multiple times, but families would not be double counted. There is no need to divide HHWT by NFAMS.