Los Angeles Family and Neighborhood Survey (L.A.FANS): Sampling/Sample Weights

RESPONSE:

Whether you use “XWGT_RSCSIB_LAF” or “XWGT_RSCSIB_LAC” depends whether you want to use all children ages 12-19 in (a) just the 65 L.A.FANS tracts or (b) all of L.A. County. If it’s (a) then use the former weight, if it’s (b) then use the latter.

Yes, you should account for clustering of individuals using the original 1990 census tract indicator, which is the basis for sample selection.

Yes, you should still use the cluster option when you use the weights.

---

RESPONSE:

The user’s guide is a bit misleading it seems—something we didn’t catch. For RSAs and PCGs you use the XWGT_ADULT_LAC as their weight and for the aged up RSCs and SIBs, you use their XWGT_RSCSIB_LAC value as theirs. Note that an adult respondent may have a weight if they did not do the adult module but did do another module like the PCG or Health Measures.

---

RESPONSE:

The sample weights are explained in the LAFANS codebook (DRU-2400/2–it’s the pdf) starting on page 41. Please read the sample weights section for specific details on the LAFANS sampling weights.

The weights are designed for different analysis samples. If you are only working with RSCs, you use the WGTRSC weight. If you are using a sample that pools RSCs and SIBS, you use the WGTKID weight. If you are using a sample only of those who are PCGs (this includes PCGonlys and those RSAs who were also selected as the PCG), you use the WGTPCG. If you are using a sample of all adult respondents (this is all RSAs and all PCGs, remembering that some respondents are both an RSA and a PCG), then you use the WGTADLT sampling weight.

The sample weights for either the RSC or the PCG are what you need. It’s a classic issue in sampling or demography. If your analysis is based on a sample of mothers (or PCGs, to be more precise) and that is the group you want to make inferences to, you should start with the PCG sample, add in information on RSCs, and use the PCG weights. If, on the other hand, your analysis is based on children and that is the group you want to make inferences to, you should start with the RSC (or the RSC/SIB sample), add in information from the PCG for each child, and use the child weights. Before starting any analysis, I strongly recommend thinking about who your sample should be.

My guess is that you want to make inferences about children. In that case, as I said above, your analysis should be based on the child sample and you should use child weights.

For children, you have two choices in LAFANS, as you know. You can either use the RSCs, which are a random sample of one child per household. In that case, you should use the RSC weights. Or you should use the RSCs and SIBS, in which case we have provided a “child weight” that corrects for the sampling of children. If you use both RSCs and SIBS, however, you should use a statistical procedure that corrects for the fact that you have two children per household.

---

RESPONSE:

There is no other documentation about the L.A.FANS weighting procedures, other than what’s provided in the Codebook.

Just to elaborate on the write up there, the goal of the weighting procedure was to have the marginals of the two-way tables (that reflects the three variables–age, sex, and race/ethnicity) based on the survey match the same marginal distribution from the 2000 census (for Los Angeles County as a whole). Raking was necessary in order to get all the marginals to match. The weighting procedure thus incorporates an implicit adjustment for non-response.

The three variables (age, sex, and race/ethnicity) were obviously collected though interviews with L.A.FANS respondents–in other words, respondents reported on their individual characteristics and parents (typically mothers) reported on their children’s characteristics. The matching population figures were collected in the 2000 Census and reported in the summary files as cross-tabulations.

In constructing the weights, we used self-reported (or parent-reported, in the case of children) information on sex, age, and race/ethnicity from the adult (or parent) questionnaire. In cases where self-reported information was missing, we filled it in with information from the Roster. (Note that item non-response levels on race/ethnicity in the questionnaires were quite low.)

We also had interviewer observations of sampled respondents’ race/ethnicity–but these were not used for weighting. In fact, we have multiple different measures of race/ethnicity (roster, self-report, self-report “best” race if multiple races reported, interviewer observation, etc.).

Information from the screener was not used in constructing any of the weights. Recall that the screener respondent was not necessarily one of the sampled respondents. For non-response cases, we don’t have any further information on household composition to try to figure out the type of respondent and for many screener refusals and incompletes we didn’t have any interviewer observation information (because the interviewers did not complete it).

RESPONSE:

Please read the section of the L.A.FANS codebook on sampling weights. The sampling weights take care of both the oversample by neighborhood poverty level and the oversample of households with children. It may also help you to read about The Design of a Multilevel Survey of Children, Families, and Communities.

---

RESPONSE:

If you look at the L.A.FANS Codebook produced by RAND on pp. 41-46 you will find a description of the sample weights that account for the sampling design of the survey. If you are using special estimation procedures, such as “svy” commands in Stata, you can use the weights in conjunction with the stratifying and clustering variables (also described in the manual) to obtain corrected standard errors.

---

RESPONSE:

The POVCAT variable in the MODSTAT1 and ROSTHH1 files contains the sampling strata in the sample design. The POVCAT variable is also described in Chapter 4 of the main LAFANS codebook under the “Community Characteristics” subsection.

The sample design is described in detail in:

Sastry, Narayan, Bonnie Ghosh-Dastidar, John Adams, and Anne R. Pebley
The Design of a Multilevel Survey of Children, Families, and Communities: The Los Angeles Family and Neighborhood Survey — 2003

The short answer to your question is that there were two oversamples: (1) we oversampled in poor and very poor strata, and (2) we oversampled households with children. But please see the publication listed above for a more complete description.

If you’re using Stata for your analysis, you might want to use the “svy” set of routines, which would involve specifying the following sample design parameters:

One thing to note is that if you are using all children (i.e., both RSCs and SIBs) in your analysis, then you will have an additional level of clustering. A similar situation would occur if you are using all adults (i.e., both RSAs and PCGs). There are a number of ways to account for this additional level of clustering–you should probably consult with a statistician to get some specific guidance that best applies to your situation.

Note that to use TRACT, you must have either restricted version 1 data which has the pseudo tract variable TRACTX or the restricted version 2 data that has the actual census tract in TRACTH90.

---

RESPONSE:

The LAFANS excludes all non-English and non-Spanish speaking households. So using weighted tabulations of native vs immigrant from the LAFANS will not give one an estimate of the native vs immigrant distribution in LA county. One would be missing Asian immigrants, Russian immigrants, etc.

If you were focusing just on natives and immigrants from English- or Spanish-speaking households, you probably could use weighted tabs from the LAFANS and apply them to census numbers based on English- and Spanish-speaking households only.

---

RESPONSE:

I’d suggest rereading the section of the codebook on weights. As the codebook says, you use the wgtrsa weights if you are using the RSA only sample and wgtadlt if you are using all adults. The weights all account for the strata oversampling, the household selection probabilities by tract, and tract-specific rates of oversampling of households with children and of household nonresponse. To apply the sample weights in STATA, there are at least two ways to do it. You can use the svy procedures or the weight= subprocedure for each procedure (e.g., logit). In the weight= procedure you do not need to specify the strata or PSUs, since they are already incorporated into the weights.

If you want to use the svy procedures and specify both the PSUs and the strata, you have to use restricted version 1 data and not the public use data. Remember that the public use data does not specify which respondents are in the same census tract (PSU). That’s the reason the data are not restricted. The public use data do, however, provide information on which stratum a household is in. This variable is povcat.

So basically you have two options: (1) use the public use data with the weights and correcting the standard errors for stratum only (using povcat) and the weight= option or (2) use the version 1 restricted data with the svy commands and full correction of the standard errors for both PSU and stratum.

---

RESPONSE:

Providing household weights significantly increases the probability of deductive identification of respondents. Therefore, we are not able to release this information in the public use data.

We believe that the household weight in the restricted data is a good weight and you should feel comfortable using it. Given the risk of deductive disclosure, and the fact that LAFANS data were designed to use as individual and not household data, we originally decided not to release a household weight at all. But subsequently, we found that users of the restricted data were interest in having it, so we have released it in the restricted data. Like other variables in the restricted data, you have to comply with the security restrictions you agreed to while using this variable.