2019 NSCH Sample Frame and Sampling Flags Creation

Appendix B – 2019 NSCH Sample Frame and Sampling Flags Creation

2019 National Survey of Children’s Health sample frame
John Voorheis
Center for Economic Studies
Research and Applications
US Census Bureau
john.l.voorheis@census.gov
301-763-5326
April 16, 2018
This document describes using administrative records to build a sample frame for the National
Survey of Children’s Health (NSCH) for 2019. We include tables and figures for the 2018 sample
frame for reference.

Population of interest
The population of interest is all children residing in housing units in the US on the date of the
survey.

A sample frame for all households with children
The sample frame identifies three mutually exclusive strata:
• [1] Households with explicit links to children in administrative data.
• [2a] Households without explicit links to children in administrative data, but predicted to
be likely to have children conditional on administrative data.
• [2b] Households without explicit links to children in administrative data, but predicted to
be unlikely to have children conditional on administrative data.
This document first explains the construction of the Stratum 1 flag, and then documents the
separation of Strata 2a and 2b.

1
2019 National Survey of Children’s Health sample frame

1

Stratum 1: identifying explicit links from children to addresses
The Stratum 1 flag for all households with explicit links to children comes from three data sources:
the Numident, a list of Social Security Number applicants with data updated from various
administrative records; and the CARRA kidlink file, a prototype linkage between children and
parents based on Census and administrative records. Household addresses are updated with the
Master Address Auxiliary Reference File, a file that links person identifiers with the latest location
updates from a variety of administrative data.

Using the Numident to identify children
The Numident is based on off the all individuals who have been assigned Social Security
Numbers. Demographic data from the Numident is updated from federal tax data and various
administrative records. There are about 83 million children in the 2017 Numident who will be
aged 0–17 years on June 1, 2018. Figure 1 shows the distribution of date of birth for these
children.
Figure 1: Distribution of date of birth, aged 0–17 years as of June 1, 2018 (2017 Numident)

1/2000 1/2002 1/2004 1/2006 1/2008 1/2010 1/2012 1/2014 1/2016 1/2018
Date of birth

Identifying the households containing the children in the Numident
To sample households with children, we must connect the children in the Numident to the
households in which they live. We do this with the CARRA kidlink file.
CARRA kidlink
2019 National Survey of Children’s Health sample frame

2

The CARRA kidlink file uses data from Census survey and federal administrative records to link
children PIKs to parent PIKs. We can use this file to identify the parents of children in the
Numident.
The source data for the CARRA kidlink file are: the Census Numident, the 2010 Census Unedited
File, the IRS 1040 and 1099 files, the Medicare Enrollment Database (MEDB), Indian Health
Service database (IHS), Selective Service System (SSS), and Public and Indian Housing (PIC) and
Tenant Rental Assistance Certification System (TRACS) data from the Department of Housing and
Urban Development. Of these, the IRS 1040 provides the most significant information.
In the CARRA kidlink file generated March 2018, there are about 66 million unique records for
children who will be aged 0–17 years on June 1, 2018.
Let us consider how many children from the Numident have been linked to a parent in the
CARRA kidlink file. Table 1 shows the number of children linked with both a mother and a father,
linked with a mother only, linked with a father only, or not linked with any parent.
Table 1: Child-parent links in the CARRA kidlink file relative to the Numident population, aged 0–
17 years as of 2018, March 2018 CARRA kidlink file
Type of link

Frequency Percent

Mother and father
57,920,000
70%
Mother only
15,380,000
19%
Father only
2,836,000
3.4%
No link
6,821,000
8.2%
All children in Numident 82,956,000
100%
Figure 2 compares the distributions of date of birth for these children against the distribution
shown in Figure 1.
Figure 2: Frequency distributions of date of birth, Numident vs. kidlink entries, aged 0–17 years
as of June 1, 2018

2019 National Survey of Children’s Health sample frame

3

The CARRA kidlink file was updated in March 2018 for NSCH sample frame production. We will
use the same CARRA kidlink file for production in 2019. We will, however, supplement this file
with additional parent-child linkages identified in sources which are not used to build the
CARRA kidlink file, including ACS and CPS-ASEC data.

Updating household location using the MAF-ARF
In order to update household location, we use a Census dataset called the Master Address
Auxiliary Reference File (MAF-ARF). The MAF-ARF links person identifiers to address identifiers
using Census survey data and federal administrative data. The source data for the MAF-ARF file
are: the Census Numident, the 2010 Census Unedited File, the IRS 1040 and 1099 files, the
Medicare Enrollment Database (MEDB), Indian Health Service database (IHS), Selective Service
System (SSS), and Public and Indian Housing (PIC) and Tenant Rental Assistance Certification
System (TRACS) data from the Department of Housing and Urban Development, and National
Change of Address data from the US Postal Service. Of these, the IRS 1040 provides the most
significant information.
Out of about 83 million children in the Numident, about 68 million, are matched directly to a
MAFID. Out of about 73 million kidlink-matched mothers, about 67 million are matched to a
MAFID. Out of about 60 million kidlink-matched fathers, about 56 million are matched to a
MAFID.
For each child observation from the Numident, we now have three possible MAFIDs: the kid
to MAF-ARF MAFID, the child-to-kidlink-to-mother-to-MAF-ARF MAFID, and the child-tokidlinkto-father-to-MAF-ARF MAFID. I allocate the single MAFID using that order. First, I assign
the directly identified child MAFID (about 65 million cases). If the MAFID is missing, I assign the
mother MAFID (about 6 million cases). Finally, if the MAFID is still missing, I assign the father
MAFID (about 3 million cases). That leaves about 9 million children from the Numident not
assigned MAFIDs (a MAFID match rate of 87.2%).
There are some MAFIDs associated with a great number of children. As an example, out of
74 million children associated with a MAFID, about 7 million children are associated with a MAFID
with more than 20 child-MAFID links.
The 74 million children associated with a MAFID are then collapsed down to 38 million unique
MAFIDS. This implies 1.94 children per household for households assigned a flag.
For 2019, one additional step will be conducted in the construction of stratum 1. We will use
administrative HUD PIC and TRACS data, which contain flags for the number of children present
at the household level for all public housing and voucher households, to enhance the existing
stratum 1 process. We will merge all MAFIDs not assigned a stratum 1 flag using the above kidlinkMAF-ARF process, with the most recent data on all public housing and voucher households in the
PIC-TRACS data. We will then assign a stratum 1 flag to all households which have a child present
flag in the HUD data.
We then need to scale up the MAFID list to the universe of MAFIDs to allow sampling of
unflagged households. A merge of the 38 million unique child-flagged MAFIDS with the January
2018 ACS MAF-X file matches 38 million MAFIDS with child flags, removes 164 million MAFIDS
2019 National Survey of Children’s Health sample frame
4

with child flags, and adds 289 MAFIDs without child flags. The sample frame file now has about
203 million valid MAFIDS, of which 38 million MAFIDS include child flags. Compare this
with the 2011 ACS, in which about 37 million out of 115 million households included related
children. 1

Stratum 1 construction visualization
Figure 3 shows a visualization of the sample frame construction.
Figure 3: Stratum 1 construction

1 http://www.census.gov/prod/2013pubs/p20-570.pdf

2019 National Survey of Children’s Health sample frame

5

Strata 2a and 2b: identifying probabilistic links from children to
addresses
In 2016, the Stratum 1 flag performed well. That is, it contained approximately the same rate of
children after as sampling as had been predicted before the survey. The survey team would like
to further increase the sampling efficiency of the survey by adding more information to the
second stratum. By definition, Stratum 2 does not have explicit links from children to
households in the administrative data. In 2018 as in 2017, we will further bifurcate Stratum 2
into those households more likely to have children and those households less likely to have
children.
Households will be assigned to Stratum 2a based on a model of child presence as a function
of variables available in administrative data for all households in the MAF. The model is
estimated with data from the most recent year of the ACS, in which child presence can be
observed. Then parameter estimates from that model can be used to predict the likelihood of
child presence for all households. These models are estimated separately for each state, and the
threshold for bifurcation is based on an objective of minimizing the size of Stratum 2a while also
maintaining 95% coverage of children in Strata 1 and 2a.

Definitions
Population or sample concepts
• 2016 ACS sample, edited and swapped
– unit of observation is the household, unless noted otherwise
– sample includes sampled vacant dwellings, unless noted otherwise
• MAF
– population but restricted to MAFIDs marked as valid for ACS
Sample frame notation
•
•
•
•

h indexes household
s indexes states
C equals 1 if a household has any children, 0 otherwise
Strata:
– S 1 : household with children
– S 2a : household likely to have children – S 2b : household unlikely to have children

• Strata sizes:

2019 National Survey of Children’s Health sample frame

6

– p(S 1 )
– p(S 2a )
– p(S 2b )

• Strata child rates:
– p(C|S 1 )
– p(C|S 2a )
– p(C|S 2b )

• Coverage with unsampled S 2b :
– p(S 1 ∪ S 2a |C)

Model
Our goal is a scalar measure of the likelihood of a child being associated with a MAFID. This
measure must be available for all ACS-valid MAFIDs in the MAF. Using a sample in which the
presence of children is observable, we will estimate a model of child presence. The regressors
used to make the index prediction must be observable for all MAFIDs (i.e., to predict outside of
the estimation sample to the entire MAF).
The general model is:
C h = f(X h ;θ),

where C is equal to one if a household includes any children and zero otherwise, X is a vector of
characteristics available for all households, and θ is an unknown vector of parameters.
We estimate the model using the most recent ACS 1-year sample:
E[C h |X h ] = f(X h ;βˆ ACS ) for households h in the ACS.

With parameter estimates from the ACS, we make predictions for the entire MAF:
Cˆ h = f(X h ;βˆ ACS ) for households h in the MAF.

In practice, we estimate models separately for each state. We do this to account for systematic
differences in administrative records coverage and MAF quality across states. The model can now
be specified as:
E[C hs |X hs ] = f(X hs ;βˆ s,ACS ) for households h in state s in the ACS,

where s is the MAFID’s state and the parameters βˆ s,ACS now vary across states. The state-specific
predictions become:
2019 National Survey of Children’s Health sample frame

7

Estimation

Cˆ hs = f(X hs ;βˆ s,ACS ) for households h in state s in the MAF.

The model above is estimated as a linear probability model separately for each state using the
edited and swapped 2015 ACS sample. The outcome is child_present, a flag for whether a child
is present at the sampled MAFID.
The following covariates are included (with associated data sources) and are available for each
MAFID (except where a missingness flag is used):
• 2016 ACS 5-year published aggregate data
– acs_blkgrp_childrate_lvout: proportion of residents of block group who are children,
excluding the own-observation child counts from the numerator and denominator
• MAF-ARF
– female2050: flag for female between ages 20 and 50 at MAFID
– adult2050: flag for adults between ages 20 and 50 at MAFID
– coresid_sexdiff: flag for coresidence of men and women between ages 20 and 50 at
MAFID
– miss_adult2050: flag for missingness from MAF-ARF
• IRS 1040 filings, tax year 2015
– any_kid_deduct_max: does any tax form associated with this MAFID have any
deduction related to children?2
– itemized_max: does any tax form associated with this MAFID use itemized
deductions?
– miss_any_kid_deduct_max: flag for MAFIDs without associated tax forms
• VSGI NAR commercial data
– vsgi_nar_homeowner_max: does any observation associated with this MAFID record
it as homeowener-occupied?
– miss_vsgi_nar_homeowner_max: flag for MAFIDs without associated VSGI data
• Targus commercial data
–
–
–
–

targus_homeowner_0: various flags for homeowner-occupied MAFID
targus_homeowner_A: various flags for homeowner-occupied MAFID
targus_homeowner_B: various flags for homeowner-occupied MAFID
targus_homeowner_C: various flags for homeowner-occupied MAFID

2 The following IRS variable were used to make this variable: child exemptions and EITC qualifying children.

2019 National Survey of Children’s Health sample frame

8

– targus_homeowner_D: various flags for homeowner-occupied MAFID
– targus_homeowner_E: various flags for homeowner-occupied MAFID
– targus_homeowner_F: various flags for homeowner-occupied MAFID
miss_targus_homeowner: flag for MAFIDs without associated Targus data

–

Parameter estimates are stored in the file frame2018_child_present_bystate.csv.

Sample frame objective function
In order to choose an optimal Strata 2a, we use the following objective function:
• Minimize the size of Strata 2a while maintaining coverage of at least 95%
Strata 2a is defined as:
S 2a = {households in the MAF with Cˆ h > C¯ but not in S 1 }.

Strata 2b is defined as

S 2b = {households in the MAF but not in S 1 or S 2a }.

With state-specific modeling, the objective function and coverage constraint also becomes
state specific:
• Minimize the size of Strata 2a in each state while maintaining coverage of at least 95% in
each state
State-specific Strata 2a is defined as:
S 2a = {households in the MAF with Cˆ hs > C¯ s but not in S 1 }.

Strata 2b is defined as before.

Optimization algorithm
The optimization parameter is a threshold on the child-present prediction probability, such that
MAFIDs with values above the threshold are assigned to Stratum 2a. Starting at a low threshold
(C¯) 3, follow this algorithm:
3 The most conservative starting threshold would be at p(S

1),

where p(S2b) = 0.

2019 National Survey of Children’s Health sample frame

9

1. Under the current threshold C¯, calculate the proportion of MAFIDs in Stratum 2a, p(S 2a ),
and the coverage of Strata 1 and 2a under no sampling of Strata 2b, (p(S 1 ∪ S 2a |C)).

2. If p(S 2a ) > 0 and p(S 1 ∪ S 2a |C) ≥ 0.95, then increase the child prediction threshold C¯ one
step (e.g., 0.01) and return to (1). If p(S 1 ∪ S 2a |C) < 0.95, then the previous threshold C¯ is
the optimal cutoff for S 2a .

Under state-specific modeling, this algorithm is applied separately to each state.

Optimal strata
Table 2 shows the optimal strata under a 95% coverage constraint for Strata 1 and 2a. The
coverage constraint assumes non-sampling of Stratum 2b. The notation is as defined above. The
strata were optimized separately for each state using parameter estimates from separate state
regressions of child presence in the 2016 ACS microdata.

Auditing the sample frame against the ACS
To examine the performance of the administrative records used to build the sample frame, we
merge the list of MAFIDs constructed above with the American Community Survey housing-unit
sample from 2016. Currently, this audit uses unedited ACS data (i.e., item nonresponse are left
as missing and are not imputed including children’s age). If item nonresponse is random with
respect to the presence of children in the household, this should not cause any systematic bias
in the audit.
All estimates are weighted with the housing-unit-level weights, which include weight for
vacant units (209,556 vacant housing units in the 2016 ACS). In vacant housing units, we assign
zero children. These estimates should reflect the NSCH survey production process.

State-specific performance
In 2018, the smallest oversample strata were in Hawaii, Maine, Vermont, and West Virginia. The
largest oversample strata are in California, Texas, and Utah. The highest rates of Type 1 error are
in DC, Florida, Louisiana, Mississippi, Nevada, and South Carolina. The highest rates of Type 2 error
were in Alaska, Hawaii, New Mexico, Texas, and Utah. For 2019, we will perform similar audits of
the frame against the 2017 ACS , and will additionally audit the frame against an early release file
of 2018 ACS microdata.

2019 National Survey of Children’s Health sample frame

10

Local-area Internet-accessibility
Here we describe the construction of a tract-varying Internet-accessible household flag.
Since 2012, ACS respondents have been able to submit survey forms over the Internet. ACS
paradata record whether a respondent chose the online option. The ACS paradata has been
summarized at the tract level. Our Internet-accessible household measure is equal to a
weighted proportion of the respondents that chose to submit the ACS survey over the Internet if
given the option to do so. Figure 4 shows the kernel-smoothed distribution of tract-level
Internet response for the 2013–2014 ACS survey years.
Figure 4: Kernel-smoothed probability distribution function of tract-level ACS Internet response
rate, ACS paradata, 2013–2014 survey years

0

.2
.4
.6
.8
ACS Internet response rate, weighted, by tract

1

To construct an Internet-access flag, we use the first tritile for a cut-off. A block is considered
to have low Internet access if the Internet accessibility index is below the first tritile of the blocklevel distribution. For low-population blocks, we replace missing values of the block-varying lowInternet flag with the modal value from the corresponding block group. For very new housing
units without assigned Census blocks, we assign a value of zero for this binary variable (i.e., the
default for these new households is high Internet accessibility.)

Local-area household income relative to the poverty rate
The frame has a set of poverty variables from the 2016 5-year American Community Survey file.
These variables measure the proportion of households with household income in an interval
defined by the poverty rate. Figure 5 shows the kernel-smoothed probability distribution function
2019 National Survey of Children’s Health sample frame

11

of the proportion of households in the block group that have household income less than 150%
of the poverty rate.

2019 National Survey of Children’s Health sample frame

12

Figure 5: Kernel-smoothed probability distribution function of block-group-level 150% poverty
rate, ACS, 2016 5-year file

0
.2
.4
.6
.8
1
Proportion of individuals below 150% of poverty line, weighted, by block group

Final sample frame data layout
The component data files are merged together based on MAFID. The data layout for this
combined file is given in Table 2.
Table 2: NSCH population data file layout
Level of

Any

Variable name

Label

variation

Type

Domain

missing?

mafid
maf_curstate

Master Address File ID
State

MAFID
State

long
str2

9 digits

no
no

maf_curcounty

County

County

str3

no

maf_curblktract

Tract

Tract

str6

yes

2018 National Survey of Children’s Health sample frame

13

maf_curblkgrp

Block group

Block group

str1

yes

maf_curblk

Block

Block

str4

yes

stratum1
Stratum 1 identifier
MAFID
stratum2a
Stratum 2a identifier
MAFID
stratum2b
Stratum 2b identifier
MAFID
acs_tract_net_response ACS Internet response
Tract
web_low
Low web use (lowest tritile)
Tract
blkgrp_lt_100_povrate Pr. HH w/ inc. < 100% poverty rate Block group
blkgrp_100_150_povratePr. HH w/ inc. 100–150% poverty rateBlock group
blkgrp_150_185_povratePr. HH w/ inc. 150–185% poverty rateBlock group
blkgrp_185_200_povratePr. HH w/ inc. 185–200% poverty rateBlock group
blkgrp_gt_200_povrate Pr. HH w/ inc. > 200% poverty rate
Block group
blkgrp_lt_150_povrate Pr. HH w/ inc. < 150% poverty rate Block group
valdf18
Valid mailing address
MAFID
Filename: nsch_pop_file.sas7bdat
Population: all MAFIDs in 2017 MAF-X
Unit of observation: household (MAFID)
Number of observations: 202,800,000
Filesize: 20GB

2019 National Survey of Children’s Health sample frame

byte
byte
byte
float
byte
float
float
float
float
float
float
byte

{0,1}
{0,1}
{0,1}
[0,1]
0,1
[0,1]
[0,1]
[0,1]
[0,1]
[0,1]
[0,1]
{0,1}

no
no
no
yes
no
yes
yes
yes
yes
yes
yes
yes

14