Download:
pdf |
pdfAPPENDIX B
FIELD TEST RESULTS
This page has been left blank for double-sided copying.
MEMORANDUM
TO:
FoodAPS TWG Members
FROM:
Mathematica Policy Research
955 Massachusetts Avenue, Suite 801
Cambridge, MA 02139
Telephone (617) 491-7900
Fax (617) 491-8044
www.mathematica-mpr.com
DATE: 7/11/2011
Rev. 9/29/2011
SUBJECT:
FoodAPS Field Test -- Sampling Design and Response Rates
Mathematica Policy Research is currently conducting the field test of the National
Household Food Acquisition and Purchase Survey (FoodAPS), referred to in the field as the
National Food Study. The results of this field test will be used to inform the design of the fullscale survey scheduled for a six-month field period from March 2012 through September 2012.
This memo provides a summary of the field test sampling design, sampling procedures,
weighting, and response rates. We also provide an overview of operational challenges that may
have affected response rates, and recommended changes in procedures for the full-scale survey.
Some of the first section of this memorandum was provided to you in the “Overview of the Field
Test Design” dated May 2.
SAMPLING DESIGN
The design for the full-scale survey calls for a multistage design with 50 primary sampling
units (PSUs) selected at the first stage, and 8 secondary sampling units (SSUs) selected within
each PSU at the second stage for a total of 400 SSUs. Addresses will be selected at the third
stage of sampling within each SSU.
For the field test, our target sample size was 400 completed cases, divided among the groups
of interest:
200 SNAP households;
120 non-SNAP households with income between poverty and 185 percent of poverty
(low income); and
80 non-SNAP households with income less than poverty (very low income).
These sample sizes were chosen to allow adequate precision for estimates of response rates,
household burden, and data quality, both overall and for the SNAP and non-SNAP groups. For
estimates of response rates, we expected 95 percent confidence intervals of between ± 5.7 and ±
6.2 percentage points for the sample as a whole (depending on the extent of the design effect)
and of no more than ± 8.1 percentage points for each of the groups of SNAP and non-SNAP
households. This design would also enable us to evaluate sampling and data collection
procedures for the two groups. The sample of 80 very low-income non-SNAP households was
An Affirmative Action/Equal Opportunity Employer
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
2
not intended to` provide precise estimates for this group but can inform us of large differences
between this group and others.
First Stage of Selection
The field test was conducted in two purposively selected PSUs (counties) in New Jersey.
The TWG had recommended purposive selection of PSUs that provide a mix of observed food
acquisitions so as to test our survey protocols and gauge the quality of the data obtained from
FoodAPS. We measured the expected diversity of food acquisitions by the racial/ethnic mix of
the population and by the retail food environment (distribution of SNAP redemptions by type of
retailer). Atlantic and Essex counties were selected for the field test. These counties have the
characteristics shown in Table 1.
TABLE 1. FIELD TEST PSUS
SNAP retailers
County
Estimated SNAP/non-SNAP
distribution of households
# SNAP %SNAP redemptions % SNAP % non-SNAP % non-SNAP
retailers
at supermarkets
HHs
<100% FPL <185% FPL
Population distribution by race /
ethnicity
White
Hisp
Black Asian
Atlantic
219
78.5%
8.1%
2.8%
15.7%
65.6% 14.3% 17.8% 6.9%
Essex
763
57.1%
11.5%
3.8%
14.1%
42.6% 18.7% 42.0% 4.6%
Second Stage of Selection
Eight SSUs (Census Block Groups) were selected within each PSU, with a goal of obtaining
low-income areas with a variety of retail food environments. SSUs were selected using
probability proportional to size (PPS) sampling. The measure of size (MOS) was a composite of
three measures: the number of SNAP households in the SSU (as calculated from SNAP
administrative data), estimates of the numbers of low-income non-SNAP households in the SSU,
and estimates of the numbers of very low-income non-SNAP households in the SSU. The
composite measure, of numbers within each SSU of households in each of the sampling strata,
reflects the relative overall sampling rate of households within the SSU. The composite MOS is
intended to enable us to obtain samples of households that have nearly equal probabilities of
selection within each study population group.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
3
Potential inaccuracies in the data used to create the MOS1 and the fact that the SNAP frame
contained a lower than expected proportion of SNAP households (discussed below) could have
led to a higher than hoped for increase in sampling error (design effect) due to unequal
weighting.
We set the sampling rates for the field test assuming that the SNAP frame would yield
almost all of the SNAP households. The sampling rates for the two frames would have been a bit
different if we had better anticipated the respective yields of the two frames.
Some of these issues will be alleviated in the full study. We plan to use five-year ACS files2
to create the MOS. We will also try to make sure the SNAP frame used for sampling within
SSUs is more current. If needed we can adjust the sampling rates within frames to reflect the
actual yields. Either or both should reduce the design effects of weighting within Quota groups,
and the first approach should reduce field costs somewhat since we will able to more efficiently
identify SNAP households. However, there are “costs” to these approaches: ensuring the
currency of the SNAP data may entail getting multiple SNAP frames over the course of the field
period; additional fine-tuning of sampling rates during the field period will increase the costs of
sampling, and will not reduce field costs.
Third Stage of Selection
Sampling of addresses at the third stage of selection followed the same procedures planned
for the full-scale survey. Within sampled SSUs, we sampled addresses for screening from two or
three sources (sampling frames):
SNAP frame. Address list of SNAP participant households as of November 2010,
obtained from the State SNAP Agency. These addresses comprised the frame for
selecting households expected to be receiving SNAP.
Non-SNAP frame. An Address-Based Sampling (ABS) frame, a commercial list of
addresses compiled from the United States Postal Service Delivery Sequence File.
Addresses in the SNAP frame were eliminated from the non-SNAP frame.
1
Because current Block Group level data were not available from the American Community Survey (ACS) or
the Decennial Census, we used estimates provided by our sample vendor, Marketing Systems Group. Their
estimates are derived from Claritas, now Nielsen Claritas.
2
The ACS Block Group level files are tabular rather than microdata, so constructing more accurate MOS may
be more challenging than we anticipated.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
4
Listing frame. Some multi-unit buildings appeared as a single record in the ABS
frame; such records contained an indication of the number of units, but no unit
numbers.3 If these buildings were “hit” in the sampling process, they were field listed
to obtain unit numbers and units were sampled from the listings.
Addresses in the SNAP and non-SNAP frame were selected directly. Buildings in the
Listing frame were selected with PPS and after listing, addresses were selected randomly from
each listed building.
Initial samples of 400 SNAP and 4,000 non-SNAP addresses were selected for screening.4
We used two methods for sampling addresses from the non-SNAP frame. In 12 of the 16 SSUs
we selected an equal probability sample of non-SNAP addresses. In the other 4 SSUs (2 per
PSU), we oversampled addresses on the non-SNAP frame that are adjacent to addresses on the
SNAP frame. The theory underlying this test is that nonparticipants who are eligible for SNAP,
or close to the SNAP’s income cutoff, are more likely to live in close proximity to SNAP
households. We will evaluate this procedure to see if it can reduce data collection costs, and also
estimate the extent to which it may increase the design effect.
The addresses in the initial samples were randomly sorted into 20 replicate subsamples and
assigned to 3 release waves. Assignment to replicates provided flexibility. After the target
number of interviews for any of the three groups (SNAP, low-income non-SNAP, very lowincome non-SNAP) was attained, that group could be made ineligible for interviewing in
subsequent releases. Further, if the SNAP target was met before all replicates had been released,
subsequent releases would not include any addresses from the SNAP frame.
As a check on the completeness of the sampling frame, we identified an adjacent address for
each sampled address. Field interviewers were instructed to confirm that the adjacent address
was in fact adjacent to the sampled address and to report any building or unit that was between
3
4
These buildings are called “drop points” on the ABS frame.
Mathematica expected to need 300 SNAP addresses, with plans to keep a random sub-sample of 100 in
reserve. For SNAP households, we assumed 95 percent of addresses contacted would be eligible. (The other 5
percent would either be invalid addresses, non-household housing units or would, at the time of the contact, be
occupied by a household that is not eligible for the survey.) In addition, we assumed that 87 percent of the addresses
provided will result in a contact. For other households, we expected the screening eligibility rate for the eligible nonSNAP group to be 25 percent in the SSUs selected. We expected to make contact with 90 percent of households and
obtain a screener completion rate of 75 percent and complete data collection among 81 percent of eligibles. We
expected 80 percent of the addresses on the ABS frame to be deliverable household addresses. If these assumptions
were correct, we would need a sample of 1828 addresses after unduplicating with SNAP records, and a total of 2504
before unduplicating. To allow for inaccuracies in our assumptions we selected an initial sample of 4000 addresses.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
5
the sampled address and adjacent address. Adjacent addresses were assigned according to the
type of sampled address:
Single unit building – adjacent address was identified by sorting street addresses that
are on the same side of the street.5 The adjacent address was the address before the
sampled address in numerical order.
Multiple unit buildings – If the sampled unit was not the highest numbered unit in
the building, field interviews confirmed the number of units in the building. If the
sampled unit was the highest numbered unit, field interviews confirmed the adjacent
address which was identified the same way as for single unit buildings.
We will report on the results of the checking procedures in a follow-up memo.
Sample Release
The sample was released in stages as shown below.
TABLE 2. TIMING OF SAMPLE RELEASES
Release
Replicates
Number of
sampled
addresses
Mailing date for
advance letters
Date sent to
field
Comments
1
1
1
2
1-9
1-9
1-9
10-15
1,223
423
151
1,104
January 14
January 23
February 8
February 8
January 31
January 31
February 14
February 14
12 of 16 SSUs
3 SSU
1 SSU
SNAP & ABS frames released.
On March 13, we pulled back all
non-SNAP cases leaving 118
SNAP addresses from release 2.
1
3
From listing
16-20
55
102
March 23
April 9
March 30
April 14
1-18
2,017
Total in field
SNAP only
As noted above, each sampled address was randomly assigned to a survey protocol and
incentive level. Field interviewers confirmed the presence of a housing unit at each sampled
address and administered a screener to determine the household’s eligibility for the survey.
Eligibility was determined by membership in a quota group:
5
In practice, all street addresses were assigned the postal service delivery point code and sorted by that code.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
6
1. Quota group A – Non-SNAP, household income ≤ 100% of poverty
2. Quota group B – Non-SNAP, household income between 100-185% of poverty
3. Quota group C – Non-SNAP, household income 185% of poverty
4. Quota group D – SNAP participant household
Households in Quota group C were not eligible for the field test. Households screened into
quota group B from releases 2 and 3 were also not eligible for the field test. (Quota group B was
closed to cases in release 2 when the SNAP frame of release 2 was re-released to the field.)
SNAP participant households were identified in both the SNAP and ABS frame. Among all
households completing a screener, 46 percent of those reporting SNAP participation were from
the ABS frame. Among completed interviews (defined as households completing a data
collection week), 48 percent of SNAP households were from the ABS frame.
Table 3 shows the final status, at the time that we ended field operations, of the 2017
sampled addresses released to the field. A small percentage (2.5) of released addresses were
outside the sampling area.6 We determined the dwelling unit status for 97.5 percent of addresses,
of which 84.3 percent were occupied dwelling units; 83.1 percent of occupied dwelling units
were contacted for screening (12 percent of cases were retired after a maximum number of
attempts7). Of those contacted and eligible for screening, 72.2 percent completed the screener
and 27.8 percent refused.
Among households completing the screener, 76.1 percent were eligible for the study and
62.9 percent completed the initial household interview. Overall, half of refusals occurred during
screening, with the other half occurring during Household Interview #1 or training. Among
households who refused during the screener, most (90 percent) agreed to complete additional
questions (“Refusal Form”) which provide demographic information for non-response analysis.
6
7
50 of the addresses were outside the sampled SSUs and resolving their status did not require field contact.
Starting on April 11, we retired cases after a maximum number of attempts (18 attempts for replicates 1-6 and
10-20; 12 attempts for replicates 6-9). The maximum number was based on a review of the distribution of completed
cases which showed that 99 percent of completes had 18 or fewer attempts.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
7
TABLE 3. FINAL STATUS OF RELEASED SAMPLE
(Each line shown in bold is the denominator for calculating the unweighted percentages in the next section.)
Total
Number
Atlantic County
Percent
Number
Distribution of Released Sample by Screener Status
Sample Released
2017
100.00
1085
Outside sampling area
50
2.5
48
Within sampling area
1967
97.5
1037
DWELLING UNIT DETERMINATION RATE (DDR)
Dwelling unit not determineda
58
2.9
49
Dwelling units determined
1909
97.1
988
DWELLING UNITS
Dwelling unit = No (vacant)
299
15.7
156
Percent
Number
Percent
100.00
4.4
95.6
932
2
930
100.00
0.2
99.8
4.7
95.3
9
921
1.0
99.0
15.8
143
15.5
1610
84.3
832
84.2
SCREENING ELIGIBILITY DETERMINATION RATE (EDR)
272
16.9
123
14.8
Not eligible for screeningb
Eligible for screening
1338
83.1
709
85.2
SCREENER COOPERATION RATE (SCR)
Refusal to complete screener
371
27.8
199
28.2
778
84.5
149
629
19.2
80.8
Dwelling unit = Yes
Screener complete
Ineligible for study (Income >185%
FPL)
Eligible for study
172
27.3
963
72.2
ELIGIBILITY FOR STUDY
506
71.8
457
72.7
230
128
25.3
102
22.3
355
77.7
247
69.6
15
6
35
4.2
1.7
9.9
50
14.1
23.9
733
76.1
378
74.7
HH #1 COMPLETION RATE (HH1CR)
Completed HH #1
461
62.9
214
56.6
DISTRIBUTION OF ELIGIBLE HOUSEHOLDS THAT DID NOT COMPLETE HH1 #1
1. Refused HH#11 at screening
Completed refusal form
89
12.0
73
19.3
No refusal form
10
1.4
4
1.1
2. Refused at HH#1 or training
94
13.0
60
15.9
3. Appointment for HH#1 pending and
not completed
76
10.4
26
6.9
a
Essex County
The dwelling unit could not be determined for some addresses in locked buildings and gated communities.
Not eligible for screening if (a) case expired due to maximum attempts (N=199); language was not English or Spanish (N=52);
respondent had physical impairment (N=5); respondent unavailable during field period (N=16); group quarters (N=4).
b
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
8
WEIGHTING
For analysis of the field test data, it is appropriate that each PSU be given equal weight at
least at the screener level. We note that if the 2 field test PSUs had been selected with PPS, as the
PSUs for the main study have been, the sums of the analysis weights for the households in each
would be approximately equal (the composite MOS and non-response adjustments would
introduce some differences). Within PSUs the field test analysis weights adjust for differences in
probability of selection and propensities to respond.
The field test screener weights have 3 components:
1. The first component of a household’s weight is the inverse of its probability of
selection. We calculated the probabilities of selection for each household as the
product of the selection probabilities of the SSU to which it belongs and the
address at which it resides. The probability of selection for the address had only
one component if the building was not listed. For buildings that were listed the
address probability of selection is the product of the probability of the building
being selected and the probability of the listed address unit being selected within
the building.
2. The next component isa non-response adjustment calculated separately within PSU
for each frame: SNAP addresses and the ABS frame. Before constructing the
weights we examined the distribution of responses by SSU to determine if SSUs or
their characteristics should be used in adjusting for non-response, but decided not
to. Insufficient data about non-responding units precluded using other
characteristics as weighting factors.
3. The screener weights were scaled so that the sums of the weights are the same for
each PSU. The non response adjusted weights were examined to determine if
trimming is appropriate and it was decided not to trim them.
Two additional weights were constructed: HH1 weights are for respondents to Household
Interview #1 (462 households that started the data collection week); HH3 weights are for
households that completed Household Interview #3 (411 households that started and completed
the data collection week). For the HH1 and HH3 data the weights are the screener weights
adjusted for further sampling8 and for response to these particular data collection efforts. The
non-response adjustment cells were the same as used for the screener.
8
Depending on the quota group, a completed case might not be selected to continue. Quota Group B
households were included in the Household Interview 1 sample only if they were part of Release 1.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
9
RESPONSE RATES
Response rates were calculated in a way the is equivalent to AAPOR Response Rate 4
(AAPOR 2011). The screener response rate (SRR) is the product of
The dwelling unit determination rate (DRR) or the percent of attempted addresses
where it was determined whether the unit was a dwelling unit.
The screening eligibility determination rate (EDR) or the percent of dwelling units
where we determined whether or not household lived there.
The screener cooperation rate (SCRR) or the percent of known households that
completed the screener.
Thus:
SRR
=
=
=
DRR*EDR*SCR
(97.05%)(83.11%)(72.19%)
58.2%
For Household Interviews 1 and 3 (HH1 and HH3), the response rates (RRH1 and RRH3)
were calculated as the product of the SRR and the completion rates for HH1 and HH3 (CRH1
AND CRH3).
RRH1
RRH3
=
=
58.2% (62.9%) = 36.6%
58.2% (56.3%) = 32.7%
RRH3 can be considered the response rate for the entire survey.
Response Rates Compared to Expectations
Our proposed sampling and data collection budget was based on four assumptions about the
behavior of the sample. These assumptions are shown in Table 4 along with field test results.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
10
TABLE 4. SAMPLING ASSUMPTIONS AND PRELIMINARY INDICATORS
FROM THE FIELD TEST
Proposal
SNAP
Percent of addresses ineligible for survey (invalid address; nonhousing unit; group home)
Screener completion rate (completed screeners as a percent of eligible
addresses)
Percent of screened households eligible for survey
Percent of eligible households agreeing to participate
Field Test
ABS
SNAP
ABS
5.0%
10.0%
12.3%
16.5%
87.0%
67.5%
76.9%
70.89%
95.0%
90.0%
39.0%
90.0%
88.0%
74.8%
72.3%
72.3%
These finding from the field test have implications for the full-scale survey:
1. Percent of ineligible addresses is higher than expected in both sampling frames –
implies that we must release and work a larger sample
2. Screener completion rate is lower than expected for the SNAP frame – implies
that we must release and work a larger sample from the SNAP frame
3. Eligibility rate among screened households is higher than expected in the ABS
frame – implies a smaller release of households from the ABS frame
4. Participation rate among eligible households is lower than expected in both
frames9 – implies that we must release and screen more households
The implications of these findings for sample release are offsetting because the eligibility
rate in the ABS frame is higher than expected. However, there are two factors that make
extrapolating from our experience a bit difficult:
1. The high ABS eligibility rate is because both field test PSUs were low income
urban areas, whereas the full study will have a broader mix of areas (the 39 percent
estimate was based on national data).
2. The overall eligibility of the ABS frame was increased by the inaccuracies (or
aging) of the SNAP frame. (A large proportion of the SNAP households came
from the ABS.) If the SNAP frame is more current (through change #4), the
eligibility rate for ABS will be lower.
9
The low and high incentives provided a small difference in participation rates (74 versus 78 percent).
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
11
Therefore the main implication from these findings is the lower than expected screener
completion rate for SNAP and lower than expected participation rate in both groups. These
findings imply a larger than anticipated release of sample for the full-scale survey. If we assume
that rate #3 (participation of eligible households) is 80% and not 90%, then we must release and
screen 16% more addresses than anticipated.
VARIANCE ESTIMATION
The field test sample was designed to provide adequate precision to evaluate data collection
procedures and data quality and to provide adequate precision and adequate power to detect
differences between the experimental treatments on cooperation rates.
Expected design effects for the full-scale survey are shown in Table 5, and estimated design
effects from the field test are shown in Table 6. For the full-scale survey, we estimated design
effects (Deffs) to range from 1.38 to 2.38 for measures with low intracluster correlations (ICCs).
These estimates are based on Deffs reported in Cohen et al. 199910 and further analysis of the
same data. We expect values of the ICCs to be between 0.01 and 0.05, and Table 5 shows the
values of Deff at the ICC values of both 0.01 (first two rows) and 0.05 (last two rows). The value
of Deff_w of 1.07 was derived from the same study.
TABLE 5. EXPECTED DESIGN EFFECTS FOR THE FULL-SCALE
NATIONAL FOOD STUDY, FOR THE WHOLE SAMPLE AND SUBGROUPS
Group
Completed
Households
PSUs
b-1
ICC
Deff_c
Deff_w
Deff
Effective n
All
3,000
50
59
0.01
1.59
1.5
2.38
1,257
SNAP
1,500
50
29
0.01
1.29
1.07
1.38
1,086
All
3,000
50
59
0.05
3.95
1.5
5.92
506
SNAP
1,500
50
29
Notes: Deff = deff_c * deff_w
deff_c is the design effect due to clustering
deff_w is the design effect due to unequal weights
deff_c = 1 + ICC(b-1)
ICC is the intracluster correlation
b is the number of cases per PSU
0.05
2.45
1.07
2.62
572.
10
Cohen, Barbara, James Ohls, Margaret Andrews, Michael Ponza, Lorenzo Moreno, Amy Zambrowski, and
Rhoda Cohen. “Food Stamp Participants’ Food Security and Nutrient Availability: Final Report.” Princeton, NJ:
Mathematica Policy Research, July 1999.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
12
Estimated design effects and ICC from the field test are means of DEFFs and ICC for 19
outcome measures (Attach A).11 The design effects are computed treating the PSUs as strata and
the SSUs as PSUs. If these estimates of the ICC are indicative of the SSU level clustering for the
main study, the overall design effect of clustering will be somewhat higher, but not inconsistent
with the higher level in Table 5. For the main study, the number of households per SSU will be
smaller than in the field test so the impact of the SSU level clustering will be less. Since we
expect the level of clustering (ICC) at the PSU level to be much less than at the SSU level, the
total design effects of clustering for the main study should be less than the higher level presented
in Table 5.
The design effect of weighting was higher than anticipated for subgroups, but this is
probably due to two factors: sampling frame issues (discussed above) and the strategy of over
sampling households next to SNAP addresses.
TABLE 6. DESIGN EFFECTS12 FROM THE FIELD TEST
Quota Group
Deffw
Deff
Effective n
All
390.4
23.4
.035
1.82
1.48
A. Non-SNAP, income
64.0
4.0
.020
1.05
1.32
<100% FPL
B. Non-SNAP, income
134.4
8.4
.034
1.31
1.47
100-185% FPL
D. SNAP
193.6
11.1
.019
1.21
1.31
Notes: Deff = deff_c * deff_w
deff_c is the design effect due to clustering
deff_w is the design effect due to unequal weights
deff_c = 1 + ICC(b-1)
ICC is the intracluster correlation
b is the average number of cases per PSU (for the field test, these are the SSUs)
2.69
1.39
144.9
46.0
1.93
69.6
1.59
121.8
11
Average
Completed
Households
b-1
ICC
Deff-c
These include five measures normalized per adult male equivalent (number of food-at-home (FAH) and foodaway-from-home (FAFH) acquisitions, FAH and FAFH expenditures, total food expenditures); number of school
meals per school-age children; percent of FAH acquisitions that were (a) free, (b) at a supercenter), (c) respondent
reported a saved receipt, (d) scanned data matched a receipt; percent of FAFH acquisitions that were (a) free, (b) at a
top 30 fast food or full service restaurant, (c) respondent reported a saved receipt; number of stores where household
shopped last month; whether primary store is a supercenter; whether household reported no change in food shopping
in the data collection week; household food insecurity; awareness of the MyPyramid; whether respondent uses
nutrient facts panel when shopping.
12
Deff-w was estimated as 1 plus the relvariance of the weight variable. Deff was computed by SUDAAN, the
software used to analyze the field test data. Deff-c was estimated as Deff-c=Deff/Deff-w. The ICC was derived as
ICC=(Deff-c-1)/(b-1). Nominal n is (average completed households/Deff).
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
13
NON-RESPONSE BIAS ANALYSIS
We will report on non-response bias analysis in a subsequent memo.
OVERVIEW OF OPERATIONAL CHALLENGES
The FoodAPS Field Test was an ambitious undertaking. We learned that households were
willing to participate in the survey (information about the quality of collected data is presented in
a separate memo). However, the collection of food acquisition data over a 7-day period had not
previously been tested and this field test revealed some challenges which led to a longer than
expected field period and possible impact on response.
Field activities were initially scheduled to run eight weeks from January 23 through March
17. The field period lasted 15 weeks from January 31 through May 22. The beginning of field
operation was postponed to begin with the SNAP benefit distribution schedule, and a one-week
break in field activities was scheduled to conduct preliminary analyses.13 The length of the field
period was affected by the following:
1. Delay in obtaining printed field materials – OMB clearance was obtained Dec 27,
2020. Printing the large number of materials required to field two different protocols
took longer than expected. During the first week of the field period, interviewers
were able to work only half of the sample assigned to the single book protocol.
2. New Jersey experienced a very harsh winter – Winter weather slowed field activities
even when interviewers were in the field. Interviewers experienced navigation
problems, limited parking, and respondents’ reluctance to converse in their doorway.
3. Greater need for training and monitoring for the screening effort – Our planning for
the field test focused primarily on procedures for collecting and processing food
acquisition data. We designed an efficient screening instrument, but the three-day
field interviewer training provided little room for training interviewers on efficiently
organizing and implementing the screening effort.14
13
The one-week break was taken from March 6-13 while we conducted preliminary analyses requested by
OMB.
14
The three-day field interviewer training focused on training households to use the food reporting instruments
under two separate survey protocols (day 1), administering the screener and CAPI interviews (day 2), practicing
mock interviews and reporting time and expenses (day 3).
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
14
Two corrections were made while in the field:
o In mid-March we trained “screener-only” interviewers. These interviewers
worked in teams with other interviewers and only administered the screener.
After screening an eligible household, they handed the case off to an
interviewer who was trained and equipped to administer the household
interviews and train the household to track food acquisitions. This strategy
allowed us to increase field staff without the full training, however, it
necessarily imposed a lag between screening and the start of the data
collection week for the household and may have contributed to the number of
refusals that occurred after screening.
o In April we provided specific guidelines to interviewers regarding days of the
week and time slots that must be filled when making attempts on each case
before a case could be retired. At the same time we set a maximum number
of attempts after which the case was retired.
4. Computer systems were not well suited for managing the large screening effort –
Mathematica’s standard sample management system proved to be a cumbersome tool
for logging the large number of attempts that interviewers made on days when they
spent all of their time knocking on doors and screening households. As a result, the
ratio of administrative time to field time was much higher than expected, resulting in
less field time.
Our mid-field correction included deployment of “screener-only” interviewers to
increase field time; and implementation of an alternate logging system whereby field
interviewers reported their activity to the survey operations center by telephone each
night.
5. Too few scanners in the field – During the field test it became apparent that the
project budget had planned for a number of scanners to provide for average demand
(average number of households active in a week), rather than peak demand. One
hundred scanners were in the field, allocated across about 35 interviewers. If an
interviewer identified an eligible household and did not have an available scanner,
they made an appointment to begin data collection at a future date.
During our debriefings with field interviewers, we learned that: (1) Some
interviewers did not initially understand that they could continue screening when
they had no scanner available. This led to a delay in screening activities. (2) Some
interviewers hoarded their scanners rather than share them with team members
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
15
because they firmly believed that they would “lose a household” if they had to make
an appointment. (3) Only a minority of interviewers thought they had a high success
rate when appointments were made.15
RECOMMENDED CHANGES FOR THE FULL-SCALE SURVEY
The field test revealed several opportunities for improving field operations. These are:
1. Acquire additional SNAP data – One of the findings from the field test is that the SNAP
caseload is fluid, thus reducing the efficiency of the SNAP sampling frame. Half of all SNAP
households were identified from the ABS sampling frame. Presumably SNAP households
from the ABS frame enrolled in SNAP after the time that we obtained the SNAP caseload
data for sampling (SNAP participation will be confirmed with administrative data at the end
of the field period).
One option for increasing the efficiency of the sampling design is to refresh the SNAP frame.
This would be achieved by randomly splitting the ABS master list of addresses into two parts
to work in the first and second halves of the field period. SNAP data obtained in January
2012 will be matched to the first half of the ABS frame, providing a SNAP frame to sample
cases for the first half of the field period (March-May). SNAP data obtained in April will be
matched to the second half of the ABS frame and provide sample for the second half of the
period (June-August).
Use of additional SNAP data will involve:
Coordinating with 27 States for an additional data extract
Geocoding additional SNAP extract from each state and matching it with the ABS frame
Managing the split sample and the added complexity of developing sampling weights
2. Revise the field interviewer training schedule – Training for the full-scale survey will not
include training on two survey protocols and two incentive levels. Therefore, the schedule
may be revised to include more structured training on screening activities, including finding
sampled addresses, mapping out efficient routes, and making attempts within a structured
schedule of time slots and days of the week.
15
Information about the lag between screening and the start of the data collection week is presented in a
second memo.
MEMO TO: FoodAPS TWG Members
FROM:
Mathematica Policy Research
DATE:
7/11/2011
PAGE:
16
3. Implement a new Sample Management System – Mathematica suggests implementation of
a web-based system to manage the screening effort for the full-scale survey. A web-based
system will be more responsive and intuitive for field interviewers, compared with the legacy
system used for the field test. A web-based system will also provide real-time reports with
drill-down capability for supervisory staff who monitor the screening effort.
4.
Acquire additional scanners – The number of scanners acquired for the field test was
based on an average number of households participating per week. It became clear in the
field test that scanners should be fielded to serve peak rather than average demand.
5. Provide replicate weights – ERS has asked about the feasibility of providing replicate
weights for FoodAPS. While such weights were not budgeted, they may be useful if ERS
wishes to produce a Public (or Restricted) Use File, Mathematica has the experience and
capabilities to create replicates and replicate weights for such a file.
REFERENCE
Standard Definitions Final Dispositions of Case Codes and Outcome Rates for Surveys.
http://www.aapor.org/Content/aapor/AdvocacyandInitiatives/StandardsandEthics/StandardD
efinitions/StandardDefinitions2011.pdf
ATTACHMENT A
DESIGN EFFECTS USED TO ESTIMATED
INTRA-CLUSTER CORRELATION
TABLE A.1. DESIGN EFFECTS USED TO ESTIMATED INTRA-CLUSTER CORRELATION
MEMORANDUM
TO:
FoodAPS TWG Members
FROM:
Mathematica Policy Research
SUBJECT:
FoodAPS Field Test Findings
955 Massachusetts Avenue, Suite 801
Cambridge, MA 02139
Telephone (617) 491-7900
Fax (617) 491-8044
www.mathematica-mpr.com
DATE: 7/11/2011
This memo summarizes findings from National Household Food Acquisition and Purchase
Survey (FoodAPS) Field Test, conducted by Mathematica Policy Research from February
through May 2011.
FIELD TEST OBJECTIVES
The primary objectives of the FoodAPS field test were:
1. Assess survey response, including burden and item non-response;
2. Assess data quality, including the consistency of information provided in multiple
survey instruments, the completeness of data provided, and the potential to match
data to external databases;
3. Assess the impact of alternative incentive levels and survey protocols (single book
and multiple book) on participation, burden, and data quality.
This memo addresses each objective. Within the memo, we make frequent reference to
findings presented in a set of appendix tables provided as a separate PDF document. Tabulations
and were prepared to assess the quality and completeness of data obtained from each survey
instrument. All of the appendix tables are weighted by sampling weights and estimated with
Sudaan to account for the complex sampling design. The tables show tests of statistical
significance for differences between the following subgroups of interest:
Single Binder versus Multiple Book survey protocol
Low versus High incentive level
SNAP versus Non-SNAP households with income <100% FPL
SNAP versus Non-SNAP households with income between 100-185%FPL
Appendix tables include asterisks to indicate statistically significant between-group
differences, with the asterisks appearing on estimates for the second group listed in each bullet
above.
An Affirmative Action/Equal Opportunity Employer
PAGE:
1.
2
SUMMARY OF THE FIELD TEST
The objective of FoodAPS is to obtain a comprehensive picture of household food
acquisitions over a 7-day period, including food-at-home (FAH) and food-away-from-home
(FAFH) from all sources including purchases and food obtained for free. Addresses were
sampled from an address-based sampling frame and a SNAP sampling frame. Field interviewers
screened households at sampled addresses to determine eligibility for the study. By design, the 7day data collection week for the household was intended to begin immediately after screening,
when possible. Participating households were in contact with survey staff over a 9-day period
with contact at multiple points in the data collection week, as shown in Figure 1.
Figure 1. Data Collection Week for Households
The goal of the field test was to obtain 400 completed interviews with households, with a
complete defined by starting and finishing the data collection week with completion of
Household Interviews #1 and #3.
The survey collected food acquisition data through multiple instruments and three “food
reporting calls.” The instruments are listed in Table 1. Instruments for FAH provided redundant
reporting; instruments for FAFH provided some redundancy for respondents (multiple booklet
sections and receipt) with the goal of resolving the separate sources of information in a single
complete record during the food reporting calls.
As shown in Figure 1, households were asked to complete three household interviews in
addition to the initial screener and food reporting. The topics included in each interview are
listed in Table 2.
PAGE:
3
Table 1. Instruments for Reporting Food Acquisitions
Instrument
Method of reporting
Information reported
Data processing
Food-at-home (FAH) Reporting
Daily List
(booklet)
Record in booklet; report
by telephone
Place, date, total amount paid
Entered in Food Reporting
System (FRS) during
telephone interview
Blue page
(booklet)
Record in booklet
Place, day, total amount paid, payment
methods, use of coupons/loyalty card, “who
got the food,” items that could not be scanned
Data entry at Mathematica
Scanner
Scan items
Date/time stamp, barcode for place, barcodes
from items or barcode book
Processed to delimit
transactions and match
UPCs to item descriptions
Receipts
Attach receipt to Blue
page
Item descriptions, prices, store savings,
coupon savings, weights (if applicable),
quantities (if applicable) payment type, total
paid
Data entry at Mathematica
in system pre-loaded with
scanner data
Daily List
(booklet)
Record in booklet; report
by telephone
Place, date, total amount paid
Entered in FRS during
telephone interview
Red page
(booklet)
Record in booklet; report
by telephone
Place, day, total amount paid, payment
methods, “who got the food”, items that are
not on receipt
Entered in FRS during
telephone interview
Receipts
Attach receipt to Red page
and report details by
telephone
Item descriptions, prices, payment type, total
paid
Entered in FRS during
telephone interview
Food-away-from-home (FAFH) Reporting
Table 2. Household Interviews
Instrument
Expected Burden
Topics
Confirm address
Identify additional housing units, if applicable
Usual frequency of food shopping
Household size and income
Identification of food shopper and meal planner
For eligible refusals: (a) primary food store; (b) food bank/pantry in last 30
days?; (c) types of stores where household shopped for food in last 30 days; (d)
number of household members by age group,
Screener
8 minutes
Household
Interview #1
(HH1)
19 minutes
Household roster
For each household member: age, race, education, marital status, employment
Nutrition program participation (SNAP, WIC, school meals)
Child care and community meal programs
Usual food shopping location and means of travel
Household
Interview #2
(HH2)
25 minutes
Number of jobs and earned income last month
Unearned income by source, last month
Non-food expenditures last month for: housing, utilities, out-of-pocket medical
expenses, education, recreation
Vehicle ownership/leasing and expenses
Identification of household assets (no amounts)
Life events: change in household membership or job, of major illness in past 12
months
PAGE:
4
Table 2. Household Interviews - continued
Instrument
Household
Interview #3
(HH3)
Expected Burden
18 minutes
Topics
Meals prepared in past 7 days / any guests for meals
General health status of respondent and family
Height and weight of each household member
Nutrition knowledge and attitudes
Special dietary needs
Food security
Previous residence / citizenship
Household Interview #1 (HH1) was to be conducted immediately after screening, however,
field interviewers made appointments when respondents were not available immediately or when
the interviewer did not have a scanner available for the household to begin the data collection
week. Household Interview #2 (HH2) was attempted by telephone in the middle of the data
collection week, with attempts continuing until the end of the field period if necessary.
Household Interview #3 (HH3) was conducted at the end of the data collection week when field
interviewers returned to the household to pick-up survey booklets and scanner, and to distribute
the incentive.
Two additional data collections are not shown in Figure 1. First, households were asked to
complete a “Meals and Snacks Form” during the data collection week to indicate, for each
household member, the meals (breakfast, lunch, dinner) and snacks (morning, afternoon,
evening) consumed each day during the week. Second, primary respondents were asked to
complete a “Respondent Feedback Form,” self-administered on paper, at the end of the data
collection week to report on the ease or difficulty of the data collection and whether they
changed behavior because of the survey.
2.
SURVEY RESPONSE AND BURDEN
Screening was completed for 1,665 addresses, with 961 households responding to the
screener. From the 961 screeners completed, 731 households were determined eligible and 537
agreed to participate. Household Interview #1 was completed with 462 households; 411
households completed the data collection week.
Survey response rates are shown in Table 3 for the overall sample and for subgroups of
interest (unweighted). We do not expect differences in the dwelling determination rate (DRR) or
screening eligibility determination rate (EDR) for survey groups or incentive levels because
addresses were randomly assigned to these groups. The SNAP frame has a higher eligibility rate,
as expected.
Screener completion and response rates do not differ significantly by incentive group,
suggesting that the incentive has little impact on households’ initial willingness to discuss the
survey and determine their eligibility. The SNAP frame had higher screener completion and
response rates than the ABS frame.
PAGE:
5
Response rates for the household interviews do not differ by survey protocol; thus if one
protocol was significantly more difficult, the difficulty was not reflected in a significantly greater
drop off in participation through the survey week. The higher incentive was associated with
higher rates of household interview response, with the difference in response increasing from
HH1 to the HH2 and HH3.
Table 3. Weighted Response Rates
Response rate
Survey Protocol
Overall
Dwelling unit
determination rate (DRR)
Screening eligibility
determination rate (EDR)2
Screening completion rate
(SCR)1
Single
Book
Multiple
Books
Incentive Level
Low
High
Sampling Frame
SNAP
ABS
96.79
96.68
96.90
96.80
96.78
98.16
96.68
84.93
84.73
85.12
85.15
84.72
89.12
84.57
70.25
69.11
71.35
67.59
72.73
76.56
69.68*
Screening response rate
(SRR = DRR*EDR*SCR)
57.75
56.61
58.85
55.71
59.63
66.97
56.97
Household interview
completion rates (CR)
HH #1
HH #2
HH #3
60.98
53.26
55.19
60.80
53.01
54.62
61.16
53.51
55.75
56.02
47.03
49.00
65.25*
58.61*
60.50*
67.02
57.40
59.79
60.18
52.71
54.57
Household interview
response rates
(RR = SRR*CR)
HH #1
HH #2
HH #3
35.21
30.76
31.87
34.42
30.01
30.92
35.99
31.49
32.81
31.21
26.20
27.30
38.91*
34.95*
36.08*
44.89
38.44
40.04
34.29
30.03
31.09
1
2
Completed screeners as a percentage of eligible addresses.
Percentage of screened households eligible for the survey.
Reporting of Food Acquisitions
In addition to the screener and household interviews, respondents were asked to participate
in three food reporting telephone calls, track FAH and FAFH acquisitions, and provide scanner
data (if applicable). Table A-1 shows the percentages of households responding to each
component.
Only half of all households completed at least three food reporting calls, with no statistically
significant differences across groups. Ninety percent of households, however, reported
PAGE:
6
information for all seven days of the data collection week during the food reporting calls. Thus, a
significant percentage of households missed a call and recalled data for a longer period during a
subsequent call. If households did not complete a call as scheduled, the telephone center
attempted to reach them the next day. If the final food reporting call for the week was not
complete when the field interviewer picked up materials, the field interviewer initiated the call
for the respondent. Sixteen percent of households completed only one call, with most of these
households reporting the entire 7-day week during one phone call.
FAH and FAFH acquisitions were reported by 90 and 82 percent of households,
respectively, with 71 percent of households providing scanner data (FAH acquisitions may be
reported without scanner data if items are written on blue pages). Overall, 55 percent of
households reported all components of the data collection: 7 days reported by phone, at least one
FAH and FAFH acquisition, and scanner data. This low rate of “complete” reporting partly
reflects underreporting during telephone calls. As we discuss later, FAH data was collected using
several methods and accounting for all reports indicates that 90 percent of households had at
least one FAH acquisition and only 5 households had no food acquisitions.
Burden of Survey Components
Burden estimates are provided in Table A-2. Response burden for the screener was not
determined because the screener was administered as a paper instrument. The overall estimate
for HH1 was slightly above expectations (24 versus an expected 19 minutes); HH2 was
somewhat below expectations (22 versus an expected 25 minutes); and HH3 matched
expectations (18 minutes). There were no significant differences by survey type or incentive
level. Non-SNAP households had greater burden in responding to HH2 regarding income and
expenditures.
Total household burden for food reporting calls was 41 minutes, which is slightly above
with our estimate of 13 minutes per call for three calls. The estimate is higher for multiple book
households but not statistically significant.
The burden of scanning was measured by the date/time stamps recorded on the scanner at
the start and end of each transaction. This estimate overstates actual scanning time if a
respondent filled out the booklet as he scanned items; it understates the full burden of FAH
reporting if the scanning was done entirely before or after completing the booklet. Thus it is not
surprising that the mean time of 20 minutes is below our estimate of 25 minutes for reporting
FAH. Multiple booklet households had higher scanning time, possibly because they had to
complete forms (Daily List and Blue page) that were located in separate booklets.
Timing of Survey Components
FoodAPS conducted in-person screening. After determination of eligibility, households were
asked to complete HH1 and training on the food reporting booklets, which could take up to 1.5
hours. Two-thirds of the households that ultimately completed the survey did not complete HH1
and training immediately after screening (Table A-3). Early in the field period these delays were
at the request of the household (with some due to the lack of a scanner); in the last month of the
PAGE:
7
field period, appointments were also necessary when “screener only” interviewers identified
eligible households and referred them to field staff trained for the full survey. Non-SNAP higher
income households were more likely to delay HH1 than other groups.
Households were compliant in starting the data collection week on the day of HH1 or the
day after (Table A-3). HH2 was completed on the third day of the data collection week, as
scheduled, for half of the sample, and completed after the data collection week for only 5 percent
of households. HH3 was completed on the day after the data collection week, as scheduled, for
61 percent of households and more than one-week later than scheduled for only 6 percent of
households. Compared with the low incentive group, the high incentive group was somewhat
less compliant with the study schedule for HH2 and HH3, but differences are not statistically
significant.
Item Non-response
Most of the data items from the household interviews and food booklets were tabulated to
assess item non-response. Table 4 shows the correspondence between survey instrument and
appendix tables. (The consistency and quality of data reported for food acquisitions is discussed
in Section 3.)
Item non-response on the household interviews is indicated by a response of “don’t know”
or “refused.” Non-response to the Meals and Snacks Form is indicated when no meal or snack is
checked off for a person on a given day. Non-response to the Respondent Feedback Form is
indicated when no response category is checked for a question.
Overall, item non-response was more likely when the primary respondent was reporting
information for other household members than when reporting own information. Item nonresponse was minimal, except for income measures reported on HH2. Item non-response did not
vary significantly by subgroups of interest.
HH1 and HH3 exhibited minimal item non-response. On HH1, age, employment, and
citizenship status of household members other than the main respondent were missing for at least
one member in 0.7, 4.3, and 3.4 percent of households, respectively (Table A-4). HH3 had item
non-response only for the height and weight questions. This missing data resulted in missing
weight status (normal, overweight obese) for 2 percent of primary respondents and for at least
one household member in 11 percent of households (Table A-14).1
The paper Meals and Snacks Form was returned by 93.1 percent of households (Table A16). The form was considered complete if there were no missing days for any household
1
In an effort to reduce non-response to the question about body weight, we gave respondents the option to
enter weight directly into interviewers’ laptops rather than providing the information verbally. Sixteen percent of
households used this option to enter weight for one or more household members.
PAGE:
8
member: 79.7 percent of households returned a complete form and 13.3 percent returned an
incomplete form. The survey protocol did not affect the likelihood of returning a complete form
(80 percent), but single book households were more likely to return an incomplete form whereas
multiple book households were more likely to return no form (10 versus 3 percent).
Among households that returned the Meals and Snacks Form, 5 percent of household
members failed to complete the form on one day, and 4.2 percent of household members were
missing more than one day of data. The percent of households returning any form was not
significantly different by incentive level, but the low incentive group was significantly more
likely to have any missing data (13.8 versus 6 percent).
Table 4. Correspondence between Survey Instrument and Appendix Tables
Instrument
Appendix Table
HH1
Table A-4–Characteristics of Households
Table A-5–Nutrition Assistance Program Participation
Table A-6–Characteristics of Primary Respondents
Table A-7–Usual Food Acquisitions
HH2
Table A-8–Non-Wage Sources of Income
Table A-9–Non-Wage Income Amounts
Table A-10–Employment Data: Persons Age 16 and Over
Table A-11–Percent of Households with Any Employed Persons and Missing
Earnings
Table A-12–Household Income as Percent of the Poverty Level
HH3
Table A-13–Nutrition Knowledge and Attitudes
Table A-14–Household Member Health and Weight Status
Table A-15–Food Security and Diet Assessment
Meals & Snacks
Form
Table A-16–Consumption During the Week As Reported on the Meals and
Snacks Form
Respondent
Feedback Form
Table A-17–Respondent Feedback
The Respondent Feedback Form was completed on paper and returned to us by 96.8 percent
of households. A very small additional item non-response occurred on 3 of the 4 questions, for a
total non-response never exceeding 5 percent (Table A-16). There were no differences in
response by subgroups of interest.
PAGE:
9
HH2 Item Non-response and Impact on Household Income
Household income and non-food expenditures from HH2 present the most significant
challenge in terms of non-response. The measures of interest are total household income and
total monthly non-food expenditures; thus a single missing component of these measures renders
the measure of interest missing, or potentially largely understated. Due to time constraints, the
non-food expenditure data from HH2 have not been examined.
Sixty percent of households reported at least one source of non-wage income (Table A-8).
The most common sources are unemployment compensation, retirement income, disability or
welfare payments, and fuel assistance (Table A-8). Of the 26 sources of non-wage income
reported by any household, 15 sources exhibit some item non-response on dollar amounts.
Overall, 16 percent of households reporting any non-wage income (9.7 percent of all households)
had missing data for dollar amounts of non-wage income (not shown in table).2
Seventy-one percent of households reported at least one employed household member
(Table A-4), and 43.5 percent of all household members over age 16 were reported to be
employed (Table A-10). Survey questions about employment were structured as follows:
How many jobs (do you/does NAME) work for pay?
How many hours (do you/does NAME) usually work per week or per month at (your / his
/ her) (first/second/third) job?
How often (are you/is [NAME]) paid from (your / his / her) (first/second/third) job?
What is the amount of pay that (you/NAME) get per check from your (first/second/third)
job before taxes and any deductions? IF NEEDED: If you were paid in cash not in a
check, that is fine. We are just looking for the amount paid in this item.
Primary respondents were less likely to be employed than spouse/partners (50 versus 57
percent).3 Primary respondents were also less likely to have missing data (13.3 versus 23.5
percent were employed and missing data). Substantially more household members were missing
data on amounts of pay than on hours (13.0 versus 5.9 percent for primary respondents; 23.5
versus 3.9 for spouse/partners) (Tables A-10). About one-third of children age 16 and over were
employed, but only 8 percent were employed with no missing earnings data.
Overall, 4.3 percent of households did not respond to HH2; 30.4 percent have no employed
persons; 38.3 percent have employed persons and no missing earnings data; and 27.1 percent
have employed persons and missing earnings data (Table A-11).
2
The percentages of households reporting each non-wage source of income in Tables A-8 and the percentages
of all households with missing non-wage amounts in Table A-9 are measured among all households, including the
4.3 percent who did not respond to HH2.
3
Seventy-five percent of primary respondents were female (Table A-6).
PAGE:
10
As a result of missing data for earnings and non-wage income, 35 percent of households are
missing total household income as a percentage of the poverty level. However, it is important to
assess the accuracy of income reported at screener to determine if the screener properly
identified households eligible for the survey and allocated them to the correct strata. Table A-12
provides an indication of the accuracy of income reported on the screener. The top panel shows
the mean and median household income as a percentage of the poverty level for: (a) households
with no missing data; (b) households with no missing data or missing earnings, with earnings
imputed; and (c) all households, with missing earnings and non-wage income imputed. All of
these measures indicate that households underreported income on the screener. The bottom panel
compares only earnings to the poverty level to assess whether households primarily underreport
(or forget to report) non-wage income. These measures provide a closer correspondence of
household income with screener category.
3.
QUALITY OF FOOD ACQUISITION DATA
The quality of food acquisition data was assessed by examining the following:
1. Consistency of FAH reporting across instruments: Daily List, Blue Page, receipt,
scanner data
2. Percent of scanned items matched to item descriptions and prices
3. Completeness of FAFH telephone reporting (examined by comparing food
booklets to the telephone data for a sample of households)
4. Item non-response on blue and red pages
One of the main questions for our assessment of data quality is whether or not households
report all acquisitions. We will be able to validate SNAP purchases with EBT data, but have not
completed that analysis yet. For non-SNAP purchases, the best indicators of whether households
report acquisitions may be the degree to which they follow survey protocols (i.e., consistency
across instruments) and whether we observe differences in reported acquisitions by survey
protocol or incentive level.
Between-group differences discussed below should be viewed with caution because they are
not based on multivariate analysis that controls for household characteristics. For example, the
Single Book group has a somewhat higher percentage of SNAP households than the Multiple
Book group (41 versus 32 percent; Table A-5), and likewise for the low versus high incentive
group. In addition, SNAP households are significantly less likely to be single-person households
than non-SNAP households (Table A-4), thus influencing results by survey strata.
Consistency of FAH Reporting Across Instruments
The greatest challenge for data processing was the match of FAH acquisitions across the
following instruments:
PAGE:
11
Daily List – place name and total paid was reported by telephone;
Blue page – place name, total paid, and other details were recorded on paper;
Scanner data – electronic file contains a date/time stamp and place category (if
respondent scanned the place code);
Receipt – contains place name, total paid and item details.
Three matching processes were used to match information across these sources:
1. Booklet information was compiled by matching Blue pages with the Daily List
2. Item information was compiled by matching receipts to scanner data during the
price entry process
3. A single matched file of FAH places was compiled by matching item information
with booklet information
Households reported more acquisitions on Blue pages than on the Daily List (overall, 1,133
versus 1,047). Booklet information was compiled by matching blue pages to the daily list based
on household ID, place name, date, and total amount. This match was not exact due to variations
in spelling of the place name, and missing total amount on the blue page (respondents were
instructed to fill total amount only if they did not save a receipt). Two-thirds of matches were
made in the first round of matching by exact match on household ID, date and first three letters
of the place name.4 After completing multiple rounds of matching, we identified 1314 unique
FAH events; 66 percent were reported on both the Daily List and Blue page, 12 percent only on
the Daily List, and 20 percent only on a Blue page (Table A-18). Compared with Multiple Book
households, Single Book households were more likely to report on both the Daily List and Blue
page (71 versus 65 percent of acquisitions) but the difference is not statistically significant.
There was no difference in reporting compliance by incentive level.
Households were more likely to scan transactions than save receipts (937 scanned
transactions and 763 receipts). The match of receipts with scanner data was a completely manual
operation. Scanner files received from the field were processed in batches to: (a) drop items
scanned in training; (b) identify separate transactions based on scanned delimiters or date/time
stamps; and (c) match barcodes with item descriptions from an external UPC database or the list
of barcodes printed in the barcode book. The “cleaned” scanner data was loaded in a web-based
entry system, shown in Figure 2.
4
We implemented a series of sequential matches wherein each record from the blue page file was joined to
each record in the daily list file. Matches were identified if there was a unique match on our match criteria;
nonmatches or non-unique matches were passed to the next round of matching.
PAGE:
12
Figure 2. Interface for the Price Entry System
Screen #1: Home page – Instructions and list of cases
Screen #2: List of transactions (left) after selecting a case
Screen #3: Entry screen for item details
PAGE:
13
Using the price entry system, data entry staff could select a case, view the list of scanned
transactions identified by date and place type (supermarket, convenience store, liquor store, etc.
or not specified)5; select a transaction and view the items scanned. Items were listed with an item
description, if available, or a blank field for filling the item description. Blank lines were also
provided in a separate section for items that may have been missed by the respondent.
Information entered for each item, if applicable, included: description, price, store savings,
coupon amount, weight, quantity, and indicators that the item is a nonfood or should be dropped
because it was not on the receipt.
At the completion of the price entry process, 1,146 unique transactions were identified for
the 411 households completing a data collection week: 52 percent of transactions had scanner
data and a receipt, 16 percent had only scanner data, and 30 percent had only a receipt (Table A18). Adherence to survey protocols for scanning and saving receipts did not differ by survey
protocol or incentive level. Non-SNAP households with income between 100 and 185% of
poverty were more likely to provide both scanner data and receipts, compared with SNAP
households and non-SNAP households below poverty (56 percent versus 47 and 49 percent).
The match of booklet information with item information used an algorithm similar to the
match of blue pages with the daily list. This match required more sequential rounds of matching
because of three complications: the place name was missing on the item file if the respondent
scanned items but did not save a receipt; the “price total” from the receipt often did not exactly
match the “total paid” because we entered prices only for food items; the SNAP total paid, if
available, could be used as an alternative to the “total paid.” This match was manageable
primarily because the average number of transactions per household is small and most
households have one FAH transaction per date. The number of unique transactions identified in
booklets or by item information is 1,616: 55 percent have both booklet and item information, 18
percent have only booklet information, and 27 percent have only item information.
The resulting correspondence of information across sources reflects both non-response to
survey protocols and errors in the matching processes. The large percentage of transactions with
no booklet information causes us to question the validity of the data: were households playing
with the scanner? Did they provide receipts that were not their own or outside the data collection
week?6
At the same time, collection of redundant information, each with a different level of burden,
provides a potential indicator of true acquisition behavior. The end result, however, is that we
have a significant amount of missing data when a household reports a FAH acquisition on the
5
After receipt information was entered into the system, a transaction was identified by the place name entered
from the receipt.
6
We will verify the date ranges of receipts within the data collection week prior to the TWG meeting.
PAGE:
14
Daily List but provides no detail on a Blue page (12 percent of booklet entries); provides scanner
data but no receipt with prices (16 percent of transaction with either scanner or receipt), or
provides only booklet or item information but not both.
Percent of Scanned Items Matched With Prices and Item Descriptions
Table A-19 describes the scanner and receipt information at the transaction and item level.
As noted above, 52 percent of transactions had scanner data and receipt; 30 percent had scanner
data only; 18 percent had a receipt only. Table A-19 shows the source of price or expenditure
data for transactions. While 70 percent of transactions had a receipt, 65 percent have item-level
price data because some receipts did not contain item prices or the item descriptions could not be
matched with scanned items (for example, the receipt might say “Grocery” or “Misc item.”).
Twelve percent of transactions have a “total paid” amount from the booklets and 22 percent of
transactions have no expenditure data. The large percentage of transactions with no “total paid”
amount is partly due to the fact that the Blue page instructs respondents to fill “total paid” only if
they do not have a receipt, and respondents apparently skipped this item even when they did not
attach a receipt or did not know the total if they did not have a receipt.
The bottom portion of Table A-19 presents statistics at the item-level, including all items
scanned or obtained from receipts: 50 percent of all observed items were scanned and validated
by a receipt; 7 percent were scanned and dropped (a receipt was provided for the transaction but
the item was not on the receipt); 14 percent of items were scanned but not validated by a receipt;
and 29 percent of items were added to the database from receipts. (The last category overstates
the underreporting by scanner because respondents had the option of writing items on the blue
page if they could not scan the item and we were unable to supplement the scanner data with the
“blue page items” prior to price entry.)
The percent of scanned items matched with the UPC database was lower than we expected:
50 percent of scanned items matched the Gladson database; 10 percent matched the barcode
book; 28 percent did not match Gladson but we obtained an item description from the receipt;
and 11 percent of scanned items have missing item description. An early review indicated that
store brands (private labels) did not match the Gladson database. Gladson confirmed that there is
wide variation in the percentage of each retailer’s private label captured in the database, and
some retailers do not provide the private label data.
Overall, only 7 percent of items observed in the scanner file or on receipts are missing item
descriptions (Table A-19). However, item descriptions entered from receipts are abbreviated, not
standardized across retailers, and not linked with information such as package size,
manufacturer, and nutrients. In addition, entry of item descriptions from receipts is significantly
more costly than obtaining those data from a database.
Completeness of FAFH Telephone Reporting
Respondents were asked to call the survey center three times during the data collection week
to report information recorded in their booklets (Daily Lists and red pages). These calls provided
an opportunity to reconcile information from the Daily Lists and red pages, and to probe when
PAGE:
15
information was not complete. Telephone interviewers entered reconciled information into our
Food Reporting System (FRS). Thus, unlike FAH data, the FRS data did not include multiple
records in need of reconciliation.
Our main concern with FAFH reporting was whether or not the telephone calls captured all
of the information recorded in booklets. To address this concern, we selected a random 20
percent of households (from among the 411 completes). Reviewers examined booklets and
obtained counts of the number of places on the Daily Lists, the number of red “places” (in some
instances respondents used a single red page to record multiple places), and the number of red
page receipts. We then compared counts from the booklets with counts from the FRS database.
Table A-20 presents the results of this quality control review of booklets. We found a
surprising amount of information in booklets was not reported during telephone calls: 30 percent
of households had information on the Daily List that was not reported by phone and 28 percent
of households had red page information not reported by phone (Table A-20). This non-reporting
by phone does not represent a true loss of data because booklet information may be recovered,
however, information not reported by telephone may be less complete.7
While some booklets contained more information than was reported by phone, the reverse
was also true (and often for the same households): 27 percent of households reported Daily List
information by phone that was not in their book, and 22 percent of households reported red
places by phone that were not in their book. Overall about 60 percent of households had an exact
match of booklet information and phone data and 70 percent of households had phone data that
was not missing booklet information.
An exact match of booklets and telephone data was significantly less likely for Single Book
households than for Multiple Book households (46 versus 73 percent for Daily Lists; 46 versus
79 percent for red pages). However, this is primarily due to the fact that Single Book households
reported more information by phone than they recorded in their book. Thus, much of the
information from Single Book households was based on recall.
We examined underreporting of booklet information by a number of household
characteristics, such as language or household size, but did not find a significant relationship.
Underreporting did increase with the total length of telephone calls, which suggest that
respondent fatigue played a role.
7
This nonreporting by phone partially explains why the count of FAH acquisitions from Daily Lists was less
than the count from blue pages.
PAGE:
16
Item Non-response on Blue and Red Pages
The data items from blue pages (entered from booklets) and red pages (reported by phone)
were tabulated to assess item non-response. These data are presented in Tables A-21 and A-22.
Rates of item nonresponse for FAFH are for acquisitions reported by phone and do not account
for the underreporting of information in booklets.
Ninety percent of households reported FAH acquisitions, with an average of 3.3 acquisitions
per household. Rates of item nonresponse are 4 to 7 percent for information about payment type,
use of frequent shopper card, and coupon use. Three percent of bue pages did not have a day of
week checked off (this hindered out matching process). Seventeen percent of blue pages did not
have a box checked for whether the household scanned all, some, or none of their items. Just
over half (52.5 percent) of blue pages had food items written on the page, which were not
accounted for in our match of scanned items with receipts.
Only 46 percent of FAH acquisitions were from a supercenter or supermarket (Table A-21).
This is an important characteristic of the field test sample, as it is significantly below the national
average of 64 percent of SNAP purchases at supercenters and supermarkets.8 Nearly all
supercenter purchases were from two chains: Pathmark and ShopRite. A large number of FAH
acquisitions were from markets or grocery stores that are grouped in the “Other” category in
Table A-21.
Eighty-nine percent of households reported FAFH acquisitions, with an average of 8.3
acquisitions per household. Item nonresponse was under 2 percent for payment type and
purchase amount (Table A-22); there was no missing data for date, and almost no missing data
for place name. Less than half of acquisitions were associated with a saved receipt, but this
should be examined conditional on whether the acquisition was free and/or obtained from a
source that typically does not provide receipts (school or workplace). For 7 percent of
acquisitions, respondents did not report whether they purchased food for non-household
members; however, these purchases were rare, when reported.
Nearly half of FAFH acquisitions were obtained for free (mostly school meals). Less than 30
percent were obtained from a “top 30” fast food or “top 30” full service restaurant. (Our Food
Reporting System was preloaded with menu items from these 60 establishments to standardize
food item data upon entry.) The precise location of FAFH establishments was obtained for all
acquisitions that were not school meals, workplace meals, or meals in a private home.9 After the
completion of the field period, the establishment locations were used together with household
8
We did not tabulate the distribution of FAH purchase amounts by store type due to missing data on purchase
amount. The PSUs were chosen on the basis of this characteristic so that we could test data collection procedures in
a diverse retail setting.
9
When a precise location was not obtained, the town name was recorded.
PAGE:
17
addresses to measure driving distances via a Google map API. The distribution of driving
distances is shown in Table A-22.10
4.
WEEKLY FOOD ACQUISITIONS AND EXPENDITURES
Table A-23 provides a summary of food acquisitions and expenditures. As noted above, 90
and 89 percent of households reported FAH and FAFH acquisitions, respectively. The 10 percent
with no FAH and FAFH acquisitions were mostly mutually exclusive with only five households
reporting no acquisitions during the data collection week. Average weekly spending per
household was $38.30 for FAFH and $93.20 for FAH. The only statistically significant betweengroup difference is in the percentage of households with no reported FAH acquisition, which is
significantly greater for households in the low incentive group. However, among households
reporting acquisitions, there were no significant between-group differences in number of
acquisitions or dollar amounts. Therefore, we cannot conclude that survey protocol or incentive
level significantly affected reporting.
Household acquisitions and expenditures, normalized for household size (per adult male
equivalent or AME), are shown in the middle section of Table A-23. The only statistically
significant differences in the normalized data are for income groups, with non-SNAP, higher
income households having more food acquisitions and expenditures than SNAP and lowerincome non-SNAP households.
Total weekly expenditures on FAH and FAFH, per AME, is $76.30 for higher income
households, $49.60 for SNAP households, and $43.90 for non-SNAP, lower income households.
For comparison, the 2008 Consumer Expenditure Survey (CEX) reported average annual food
expenditures of $6,443 per household (= $124/week), and average household size of 2.5 persons,
which yields an estimate of $50 per person per week without accounting for male equivalence.11
Thus, it appears that the FoodAPS field test captured more food spending per household, on
average, than the CEX.
5. SUMMARY
The FoodAPS field test revealed several challenges for a 7-day collection of food
acquisition data. The greatest challenge may not be an underreporting of acquisitions, as feared,
but our ability to manage the several redundant sources of information so as to assemble a
10
Precise locations were obtained during the telephone interviews via an automated Google search for place
name within the vicinity of a household’s address. We experiences problems with this feature of our Food Reporting
System during the first weeks of the field period. Similar distances to FAH locations could not be obtained due to an
error in the CAPI programming for the questions about usual “food stores” on HH1.
11
The Consumer Price Index has risen just over 2 percent from the 2008 annual average to March 2011.
PAGE:
18
complete record for each household. Retooling our processing of FAH data may yield
significantly more consistent data across reporting instruments. Adding a formal review of
booklets will eliminate underestimation of FAFH.
An important question for the field test was “Would respondents be willing to participate?”
We will complete our analysis of nonresponse bias in time for presentation at the TWG meeting.
Other questions were “Could respondents follow the survey protocols?” and “Would
participation in the survey change household acquisition behavior?” Table A-17 presents data
collected from the self-administered respondent feedback form at the end of the data collection
week. Seventy percent of respondents reported that it was easy or very easy to keep track of
foods during the data collection week, with no statistically significant differences between
groups. Two-thirds of respondents reported that it was easy or very easy to get other household
members to participate, with a significant difference between Single Book and Multiple Book
households (72 versus 60 percent). Finally, nearly 80 percent of households reported that they
did not change the way they got food because they were participating in the survey, however,
those in the higher incentive group were less likely to report no change (75 versus 85 percent).
| File Type | application/pdf |
| Author | ECurley |
| File Modified | 2011-11-14 |
| File Created | 2011-11-11 |