Adaptive Design Experiment

Request for Approval of a Nonsubstantive Change:
National Health Interview Survey (NHIS)
ADAPTIVE DESIGN EXPERIMENT
OMB No. 0920-0214
(Expiration Date 12/31/2017)

Contact Information:
Marcie Cynamon
Director, Division of Health Interview Statistics
National Center for Health Statistics/CDC
3311 Toledo Road
Hyattsville, MD 20782
301.458.4174 (voice)
301.458.4035 (fax)
[email protected]

May 23, 2016

1

NATIONAL HEALTH INTERVIEW SURVEY (NHIS)
ADAPTIVE DESIGN EXPERIMENT
A1.
Circumstances Making the Collection of Information Necessary
This request is for a nonsubstantive change to an approved data collection (OMB No. 0920-0214; exp.
12/31/2017), the National Health Interview Survey (NHIS). With this nonsubstantive change request, the
Division of Health Interview Statistics seeks clearance to test the use of adaptive design in the NHIS. The
aim is to investigate the impact of adaptive case prioritization on sample representativeness and
nonresponse bias, while maintaining survey costs and minimizing any possible negative effect on the
overall response rate.
The National Health Interview Survey (NHIS), like many other large national sample surveys, has
experienced a steady decline in response rates over the past 15 to 20 years (see Figure 1). Low
participation rates in surveys matter to the extent that they introduce the potential for nonresponse
bias in survey estimates. Nonresponse bias assessments of 2013 and 2014 NHIS Early Release (ER)
Program estimates have been performed, with the results suggestive that bias may be present in some
key health estimates. In this section, we more fully describe the methods behind and the results from
these assessments.

Figure 1. Final Household, Family, Sample Child, and Sample Adult Response Rates: NHIS, 1997-2014
To assess if nonresponse bias may be present in ER measures, a logistic regression of family response
that included several covariates culled from the NOI and CPD was first estimated. The covariates in the
model were those found to be associated with several health outcomes on the NHIS. Examples include
the interviewer assessment of whether or not one or more residents may be disabled, handicapped, or
have a chronic health condition; interviewer assessment of whether or not the sample unit may include
smokers; average aggregated household income for the block group or tract in which the household is
2

located (ACS 2009-13); and the proportion of persons aged 25 and over with a college degree in the
block group or tract in which the household is located (ACS 2009-13). The two-level logistic regression
model included random interviewer effects and was estimated twice, once on 2013 data and once on
2014 data. The predicted values of response propensities from each run were used to group responding
families into response propensity quintiles. We then examined estimates for 20 ER measures in two
ways. First, we observed the health estimates for each of the response propensity quintiles. And
second, we observed the estimates cumulatively moving from the high response propensity quintile to
the low response propensity quintile. Note that the quintile-specific and cumulative estimates were
weighted using base weights (inverse of the probability of selection). Comparisons of estimates by
quintile, as well as systematic changes in the cumulative estimate as families with lower response
propensities are added to the sample, provide clues as to possible nonresponse bias with these key
health measures. The approach is akin to level-of-effort (LOE) analyses [1], where change in a statistic
over increased levels of effort, or in this case over response propensity quintiles, is indicative of the risk
of nonresponse bias. Conversely, little to no change in the statistic is suggestive of the absence of
nonresponse bias [2].
Table 1 presents person-level ER variables broken out by year. Starting with the percentage of persons
of all ages who had excellent or very good health, we can consider the persons in quintile 1 (high
response propensities) to be from families that were the easiest to recruit. In 2013, the estimate for this
group was 57.5%. This is roughly 16 percentage points lower than the estimate for quintile 5 (low
response propensities), or persons from families that were the least likely to participate. Table 1 also
presents the cumulative estimate moving from the high response propensity quintile to the low
response propensity quintile (left to right). As we recruit families with lower and lower response
propensities, the estimate of persons who had excellent or very good health increases from 57.5% to
65.9%. Furthermore, we can treat the persons from the low response propensity families as proxies for
nonresponders, and then compare the estimate for this group (73.2%) to the estimate for the remainder
of the sample (64.4%). The difference between the two estimates is statistically significant (two-tailed ttest conducted at the .05 level). Together, the quintile-specific and cumulative estimates suggest that
persons from nonresponding families may have a higher rate of reporting excellent or very good health.
Therefore, we may be underestimating the percentage of persons who have excellent or very good
health. The pattern identified for 2013 appears to hold for 2014.
For the three remaining person-level estimates, evidence of possible nonresponse bias is observed for
both years and the patterns are consistent. Generally speaking, we may be overestimating the
percentage of persons who failed to obtain need medical care due to cost in the past 12 months and the
percentage of persons who need help with personal care needs, while underestimating the percentage
of persons with health insurance coverage. The final column in Table 1 presents the final weighted
estimate for each measure. With the exception of health insurance coverage, applying the final person
weights appears to move the estimates in the anticipated direction, although the movements are small
and may be insufficient. For health insurance coverage, the final weighted estimate moved slightly in the
opposite direction from what the bias analysis suggests.

3

Table 1. Nonresponse Bias Analysis of Person-Level Early Release Estimates: NHIS, 2013-2014a
Quintile 1:
high response
propensities

Quintile 2

Quintile 3

Quintile 4

Quintile 5:
low response
propensities

2013
Percentage of persons of all ages who had excellent or very good health
Quintile estimate (s.e.)
57.5 (0.63)
64.1 (0.63)
67.9 (0.59)
69.3 (0.63)
73.2 (0.62)
Cumulative estimate (s.e.)
57.5 (0.63)
60.7 (0.44)
63.0 (0.37)
64.4 (0.32)
65.9 (0.30)
Percentage of persons of all ages who failed to obtain needed medical care due to cost in the past 12 months
Quintile estimate (s.e.)
7.6 (0.23)
6.5 (0.24)
5.1 (0.20)
5.1 (0.23)
4.9 (0.29)
Cumulative estimate (s.e.)
7.6 (0.23)
7.0 (0.17)
6.4 (0.14)
6.1 (0.12)
5.9 (0.11)
Percentage of persons of all ages who need help with personal care needs
Quintile estimate (s.e.)
3.2 (0.17)
1.9 (0.12)
1.8 (0.12)
1.5 (0.11)
1.2 (0.11)
Cumulative estimate (s.e.)
3.2 (0.17)
2.6 (0.10)
2.3 (0.08)
2.2 (0.07)
2.0 (0.06)
Percentage of persons of all ages with health insurance coverage
Quintile estimate (s.e.)
81.5 (0.48)
84.4 (0.46)
86.9 (0.41)
88.0 (0.42)
89.1 (0.47)
Cumulative estimate (s.e.)
81.5 (0.48)
82.9 (0.36)
84.2 (0.28)
85.1 (0.24)
85.8 (0.22)
2014
Percentage of persons of all ages who had excellent or very good health
Quintile estimate (s.e.)
59.5 (0.62)
66.1 (0.62)
67.5 (0.58)
68.9 (0.61)
71.0 (0.65)
Cumulative estimate (s.e.)
59.5 (0.62)
62.7 (0.46)
64.1 (0.37)
65.1 (0.32)
66.1 (0.28)
Percentage of persons of all ages who failed to obtain needed medical care due to cost in the past 12 months
Quintile estimate (s.e.)
5.9 (0.24)
5.2 (0.24)
5.3 (0.22)
5.0 (0.23)
4.7 (0.25)
Cumulative estimate (s.e.)
5.9 (0.24)
5.6 (0.19)
5.5 (0.15)
5.4 (0.13)
5.3 (0.12)
Percentage of persons of all ages who need help with personal care needs
Quintile estimate (s.e.)
3.0 (0.17)
1.9 (0.12)
1.5 (0.11)
1.5 (0.11)
1.5 (0.11)
Cumulative estimate (s.e.)
3.0 (0.17)
2.5 (0.11)
2.2 (0.08)
2.1 (0.07)
2.0 (0.06)
Percentage of persons of all ages with health insurance coverage
Quintile estimate (s.e.)
86.0 (0.43)
87.6 (0.47)
89.4 (0.35)
90.1 (0.36)
91.4 (0.39)
Cumulative estimate (s.e.)
86.0 (0.43)
86.8 (0.35)
87.6 (0.27)
88.1 (0.23)
88.6 (0.20)

Quintile 5 vs.
Top 4
Quintiles
(t-test)

Final
Weighted
Estimate

*

66.3 (0.29)

*

5.9 (0.11)

*

1.9 (0.06)

*

85.5 (0.22)

*

66.5 (0.28)

*

5.3 (0.12)

*

1.9 (0.06)

*

88.5 (0.19)

a

The quintile and cumulative estimates are base weighted.
* Indicates that a two-tailed t-test was significant at the .05 level.

4

Table 2. Nonresponse Bias Analysis of Sample Child Early Release Estimates: NHIS, 2013-2014a
Quintile 1:
high response
propensities

Quintile 2

Quintile 3

Quintile 4

Quintile 5:
low response
propensities

2013
Percentage of children aged 17 and under with a usual place to go for medical care
Quintile estimate (s.e.)
94.0 (0.65)
94.9 (0.55)
96.3 (0.41)
96.6 (0.48)
97.5 (0.39)
Cumulative estimate (s.e.)
94.0 (0.65)
94.5 (0.42)
95.1 (0.31)
95.4 (0.27)
95.8 (0.22)
Percentage of children aged 17 and under who had received an influenza vaccination in the past 12 months
Quintile estimate (s.e.)
43.5 (1.26)
45.6 (1.25)
47.0 (1.32)
47.0 (1.36)
48.2 (1.42)
Cumulative estimate (s.e.)
43.5 (1.26)
44.6 (0.88)
45.4 (0.77)
45.7 (0.68)
46.2 (0.61)
Percentage of children aged 17 and under who had an asthma attack in the past 12 months
Quintile estimate (s.e.)
5.6 (0.60)
5.3 (0.55)
4.3 (0.46)
5.5 (0.63)
4.3 (0.47)
Cumulative estimate (s.e.)
5.6 (0.60)
5.4 (0.41)
5.1 (0.33)
5.2 (0.28)
5.0 (0.25)
Percentage of current asthma among children aged 17 and under
Quintile estimate (s.e.)
9.6 (0.69)
8.6 (0.69)
7.5 (0.61)
9.9 (0.79)
7.3 (0.61)
Cumulative estimate (s.e.)
9.6 (0.69)
9.1 (0.48)
8.6 (0.40)
8.9 (0.35)
8.6 (0.31)
2014

Quintile 5 vs.
Top 4
Quintiles
(t-test)

Final
Weighted
Estimate

95.9 (0.20)

46.1 (0.62)

*

4.9 (0.25)

*

8.5 (0.31)

Percentage of children aged 17 and under with a usual place to go for medical care
Quintile estimate (s.e.)
Cumulative estimate (s.e.)

95.4 (0.45)
95.4 (0.45)

96.1 (0.51)
95.7 (0.33)

96.5 (0.51)
96.0 (0.29)

97.0 (0.45)
96.2 (0.25)

97.0 (0.44)
96.3 (0.22)

96.4 (0.22)

Percentage of children aged 17 and under who had received an influenza vaccination in the past 12 months
Quintile estimate (s.e.)
Cumulative estimate (s.e.)

46.4 (1.35)
46.4 (1.35)

48.6 (1.24)
47.5 (0.93)

48.7 (1.34)
47.9 (0.78)

48.9 (1.39)
48.1 (0.69)

50.0 (1.55)
48.4 (0.63)

48.8 (0.64)

4.8 (0.54)
4.4 (0.30)

3.9 (0.49)
4.3 (0.26)

4.3 (0.27)

9.2 (0.74)
8.8 (0.38)

8.9 (0.72)
8.8 (0.34)

8.8 (0.35)

Percentage of children aged 17 and under who had an asthma attack in the past 12 months
Quintile estimate (s.e.)
Cumulative estimate (s.e.)

3.7 (0.47)
3.7 (0.47)

4.3 (0.65)
4.0 (0.40)

4.8 (0.59)
4.3 (0.35)

Percentage of current asthma among children aged 17 and under
Quintile estimate (s.e.)
Cumulative estimate (s.e.)

8.1 (0.68)
8.1 (0.68)

9.0 (0.79)
8.5 (0.52)

8.9 (0.74)
8.7 (0.44)

a

The quintile and cumulative estimates are base weighted.
* Indicates that a two-tailed t-test was significant at the .05 level.

5

Table 3. Nonresponse Bias Analysis of Sample Adult Early Release Estimates: NHIS, 2013-2014a
Quintile 1:
high response
propensities

Quintile 2

Quintile 3

Quintile 4

Quintile 5:
low response
propensities

2013
Percentage of adults aged 18 and over who had an asthma attack in the past 12 months
Quintile estimate (s.e.)
3.6 (0.27)
3.3 (0.26)
3.7 (0.29)
3.3 (0.27)
3.7 (0.34)
Cumulative estimate (s.e.)
3.6 (0.27)
3.5 (0.19)
3.6 (0.16)
3.5 (0.14)
3.5 (0.13)
Percentage of current asthma among adults aged 18 and over
Quintile estimate (s.e.)
7.4 (0.38)
6.8 (0.44)
7.7 (0.38)
7.2 (0.39)
7.3 (0.47)
Cumulative estimate (s.e.)
7.4 (0.38)
7.1 (0.29)
7.3 (0.23)
7.3 (0.20)
7.3 (0.19)
Percentage of adults 18 and over with diagnosed diabetes
Quintile estimate (s.e.)
12.4 (0.46)
9.9 (0.46)
10.0 (0.43)
8.4 (0.43)
8.3 (0.45)
Cumulative estimate (s.e.)
12.4 (0.46)
11.2 (0.32)
10.8 (0.25)
10.3 (0.22)
10.0 (0.20)
Percentage of adults 18 and over who had 5 or more drinks in 1 day at least once in the past year
Quintile estimate (s.e.)
18.1 (0.57)
21.4 (0.66)
22.7 (0.70)
22.7 (0.74)
23.9 (0.84)
Cumulative estimate (s.e.)
18.1 (0.57)
19.6 (0.46)
20.6 (0.39)
21.1 (0.35)
21.5 (0.33)
2014
Percentage of adults aged 18 and over who had an asthma attack in the past 12 months
Quintile estimate (s.e.)
3.6 (0.25)
3.1 (0.24)
3.9 (0.34)
3.3 (0.27)
3.3 (0.33)
Cumulative estimate (s.e.)
3.6 (0.25)
3.4 (0.17)
3.5 (0.17)
3.5 (0.15)
3.4 (0.13)
Percentage of current asthma among adults aged 18 and over
Quintile estimate (s.e.)
7.7 (0.36)
7.1 (0.36)
8.7 (0.50)
7.5 (0.40)
7.5 (0.51)
Cumulative estimate (s.e.)
7.7 (0.36)
7.4 (0.26)
7.8 (0.25)
7.7 (0.22)
7.7 (0.19)
Percentage of adults 18 and over with diagnosed diabetes
Quintile estimate (s.e.)
12.7 (0.47)
10.1 (0.45)
9.5 (0.47)
7.7 (0.42)
8.6 (0.50)
Cumulative estimate (s.e.)
12.7 (0.47)
11.5 (0.33)
10.9 (0.27)
10.2 (0.24)
10.0 (0.21)
Percentage of adults 18 and over who had (male=5/female=4) or more drinks in 1 day at least once in the past year
Quintile estimate (s.e.)
19.2 (0.67)
24.4 (0.73)
25.0 (0.82)
26.2 (0.84)
26.9 (0.82)
Cumulative estimate (s.e.)
19.2 (0.67)
21.7 (0.50)
22.6 (0.46)
23.4 (0.42)
23.9 (0.37)

Quintile 5 vs.
Top 4
Quintiles
(t-test)

Final
Weighted
Estimate

3.4 (0.13)

7.1 (0.18)

*

9.3 (0.20)

*

22.6 (0.35)

3.4 (0.14)

7.6 (0.20)

*

9.1 (0.20)

*

25.3 (0.40)

6

Table 3. (continued)
Quintile 1:
high response
propensities

Quintile 2

Quintile 3

Quintile 4

Quintile 5:
low response
propensities

2013
Prevalence of current cigarette smoking among adults aged 18 and over
Quintile estimate (s.e.)
23.7 (0.68)
19.2 (0.63)
15.4 (0.57)
15.1 (0.60)
13.5 (0.63)
Cumulative estimate (s.e.)
23.7 (0.68)
21.6 (0.47)
19.7 (0.38)
18.7 (0.33)
17.8 (0.29)
Percentage of adults aged 18 and over who had received an influenza vaccination in the past 12 months
Quintile estimate (s.e.)
41.0 (0.79)
40.9 (0.79)
42.7 (0.76)
43.2 (0.86)
45.2 (0.93)
Cumulative estimate (s.e.)
41.0 (0.79)
40.9 (0.59)
41.5 (0.47)
41.9 (0.43)
42.4 (0.39)
Percentage of adults aged 18 and over who met the federal physical activity guidelines
Quintile estimate (s.e.)
38.8 (1.10)
46.9 (0.87)
51.1 (0.87)
51.6 (0.83)
54.8 (0.98)
Cumulative estimate (s.e.)
38.8 (1.10)
42.6 (0.73)
45.2 (0.60)
46.6 (0.50)
48.0 (0.45)
Prevalence of obesity among adults aged 18 and over
Quintile estimate (s.e.)
30.9 (0.81)
29.1 (0.71)
29.6 (0.70)
28.5 (0.75)
26.5 (0.78)
Cumulative estimate (s.e.)
30.9 (0.81)
30.1 (0.58)
29.9 (0.45)
29.6 (0.39)
29.1 (0.35)
2014
Prevalence of current cigarette smoking among adults aged 18 and over
Quintile estimate (s.e.)
20.3 (0.68)
16.7 (0.59)
16.7 (0.75)
14.1 (0.60)
13.9 (0.73)
Cumulative estimate (s.e.)
20.3 (0.68)
18.6 (0.46)
18.0 (0.39)
17.2 (0.34)
16.7 (0.31)
Percentage of adults aged 18 and over who had received an influenza vaccination in the past 12 months
Quintile estimate (s.e.)
45.1 (0.79)
42.4 (0.77)
45.2 (0.85)
42.6 (0.84)
43.2 (0.97)
Cumulative estimate (s.e.)
45.1 (0.79)
43.8 (0.58)
44.2 (0.49)
43.9 (0.43)
43.8 (0.40)
Percentage of adults aged 18 and over who met the federal physical activity guidelines
Quintile estimate (s.e.)
41.2 (0.90)
47.9 (0.93)
50.5 (0.92)
51.8 (0.94)
53.0 (1.04)
Cumulative estimate (s.e.)
41.2 (0.90)
44.3 (0.68)
46.2 (0.56)
47.3 (0.49)
48.2 (0.45)
Prevalence of obesity among adults aged 18 and over
Quintile estimate (s.e.)
30.9 (0.65)
30.5 (0.78)
29.8 (0.70)
29.0 (0.75)
27.6 (0.88)
Cumulative estimate (s.e.)
30.9 (0.65)
30.7 (0.53)
30.4 (0.43)
30.1 (0.38)
29.7 (0.35)

Quintile 5 vs.
Top 4
Quintiles
(t-test)

Final
Weighted
Estimate

*

17.8 (0.30)

*

41.0 (0.39)

*

49.4 (0.45)

*

28.6 (0.36)

*

16.8 (0.33)

42.2 (0.42)

*

49.3 (0.45)

*

29.3 (0.37)

7

Table 3. (continued)
Quintile 1:
high response
propensities

Quintile 2

Quintile 3

Quintile 4

Quintile 5:
low response
propensities

2013
Percentage of adults aged 18 and over who had ever been tested for HIV
Quintile estimate (s.e.)
36.0 (0.85)
36.4 (0.79)
35.2 (0.79)
37.0 (0.84)
38.6 (0.93)
Cumulative estimate (s.e.)
36.0 (0.85)
36.1 (0.59)
35.8 (0.48)
36.1 (0.43)
36.5 (0.40)
Percentage of adults aged 18 and older who experienced serious psychological distress during the past 30 days
Quintile estimate (s.e.)
5.5 (0.33)
4.4 (0.33)
3.4 (0.27)
3.2 (0.28)
2.4 (0.29)
Cumulative estimate (s.e.)
5.5 (0.33)
5.0 (0.24)
4.5 (0.19)
4.2 (0.16)
3.9 (0.15)
Percentage of adults aged 18 and over with a usual place to go for medical care
Quintile estimate (s.e.)
81.4 (0.65)
83.2 (0.57)
84.7 (0.57)
86.3 (0.56)
87.7 (0.62)
Cumulative estimate (s.e.)
81.4 (0.65)
82.3 (0.44)
83.0 (0.36)
83.7 (0.31)
84.4 (0.28)
Percentage of adults 18 and over who received the pneumococcal vaccination in the past 12 months
Quintile estimate (s.e.)
27.2 (0.71)
23.1 (0.70)
22.0 (0.68)
21.3 (0.68)
18.7 (0.66)
Cumulative estimate (s.e.)
27.2 (0.71)
25.3 (0.53)
24.3 (0.43)
23.6 (0.37)
22.8 (0.33)
2014
Percentage of adults aged 18 and over who had ever been tested for HIV
Quintile estimate (s.e.)
35.5 (0.84)
36.7 (0.77)
36.7 (0.87)
36.6 (0.89)
37.9 (0.96)
Cumulative estimate (s.e.)
35.5 (0.84)
36.1 (0.58)
36.3 (0.49)
36.3 (0.44)
36.6 (0.39)
Percentage of adults aged 18 and older who experienced serious psychological distress during the past 30 days
Quintile estimate (s.e.)
4.1 (0.27)
3.1 (0.25)
3.0 (0.25)
3.1 (0.29)
2.4 (0.26)
Cumulative estimate (s.e.)
4.1 (0.27)
3.6 (0.18)
3.5 (0.15)
3.4 (0.13)
3.2 (0.12)
Percentage of adults aged 18 and over with a usual place to go for medical care
Quintile estimate (s.e.)
85.3 (0.58)
85.3 (0.56)
86.7 (0.52)
86.3 (0.58)
87.7 (0.55)
Cumulative estimate (s.e.)
85.3 (0.58)
85.3 (0.43)
85.7 (0.33)
85.8 (0.30)
86.1 (0.27)
Percentage of adults 18 and over who received the pneumococcal vaccination in the past 12 months
Quintile estimate (s.e.)
29.8 (0.75)
23.8 (0.68)
22.6 (0.69)
20.2 (0.72)
20.0 (0.76)
Cumulative estimate (s.e.)
29.8 (0.75)
26.9 (0.53)
25.7 (0.44)
24.5 (0.38)
23.8 (0.35)

Quintile 5 vs.
Top 4
Quintiles
(t-test)

Final
Weighted
Estimate

*

37.3 (0.41)

*

3.8 (0.15)

*

83.7 (0.30)

*

21.0 (0.30)

37.5 (0.41)

*

3.1 (0.12)

*

85.3 (0.29)

*

21.8 (0.34)

a

The quintile and cumulative estimates are base weighted.
* Indicates that a two-tailed t-test was significant at the .05 level.

8

Table 2 presents the same bias analysis for four sample child ER estimates. Overall, we tend to see less
evidence for potential nonresponse bias in these measures, especially for 2014. In 2014, there were no
significant differences across these four measures when comparing the estimate for sample children in
the low response propensity quintile to the estimate for the remainder of the sample. In 2013, we
observe possible nonresponse bias in current asthma and having an asthma attack in the past 12
months. The data suggest we may be underestimating the prevalence of both of these measures.
Looking at the final weighted estimate (person weights that include nonresponse adjustments), it
appears that we may be mitigating the potential bias, but only slightly.
Finally, Table 3 presents the nonresponse bias analysis results for 12 sample adult ER estimates, again
broken out by year. Across the two years, evidence of potential bias is observed for all but two of the 12
measures: current asthma and having an asthma attack in the past 12 months. The quintile-specific and
cumulative estimates suggest that relatively large amounts of nonresponse bias could be present in
some health measures. Current cigarette smoking and meeting the federal physical activity guidelines
provide good examples. In 2013, the current smoking estimate for adults in quintile 1 (high response
propensities) was 23.7%. The estimate for adults in quintile 5 (low response propensities) was 13.5%,
roughly 15 percentage points lower. The corresponding quintile 1 and quintile 5 estimates for 2014 were
20.3% and 13.9% respectively. The cumulative estimate of current cigarette smoking in 2013, moving
from the high response propensity quintile to the low response propensity quintile, drops from 23.7% to
17.8% (20.3% to 16.7% in 2014). Finally, the estimate for adults in quintile 5 (proxies for nonresponders)
is significantly lower than the estimate for the other four quintiles combined (2013: 13.5% vs. 18.7%;
2014: 13.9% vs. 17.2%). Together, this information suggests that the estimate of current cigarette
smoking could be an overestimate. For both years, applying final sample adult weights (that include
nonresponse adjustments) has little impact on the estimates.
With regard to meeting federal physical activity guidelines, the analysis suggests a possible
underestimate. The quintile 1 (high response propensities) estimate for 2013 was 38.8%, while the
quintile 5 estimate was 54.8%. The corresponding figures for 2014 were 41.2% and 53.0%. Hence,
interviewers encountered more difficulties in securing the participation of physically active sample
adults. Not surprisingly, a steady increase in the cumulative estimates of exercise is observed for both
years when moving from the high to low propensity quintile (2013: 38.8% to 48.0%; 2014: 41.2% to
48.2%). The estimate for adults in quintile 5 (again, proxies for nonresponders) is significantly higher
than the estimate for the other four quintiles combined (2013: 54.8% vs. 46.6%; 2014: 53.0% vs. 47.3%).
Unlike current cigarette smoking, however, applying the final sample adult weights has the effect of
increasing the estimate for both years, as suggested by the bias analysis.
In sum, we found evidence for possible nonresponse bias in 16 of 20 key health outcomes in 2013 and
12 of 20 outcomes in 2014. The pattern of possible bias appears to be consistent across the two years of
data, and may be substantial for estimates of needing help with personal care needs, diagnosed
diabetes, serious psychological distress, binge drinking, current smoking, and exercise. Final weighted
estimates suggest that the current nonresponse adjustment procedures used with the NHIS moves most
estimates in directions suggested by the bias analysis. However, for some estimates where bias may be
more substantial, the movement in estimates was small and may be insufficient. Hence, there appears
to be room for improvements in minimizing nonresponse bias in NHIS health estimates. Additionally, by
addressing nonresponse bias during data collection, via adaptive design, we may minimize the size of
needed post-survey nonresponse weighting adjustments, producing reductions in variance and
therefore gains in precision.

9

Declining response rates and increasing data collection costs—like those faced by the NHIS—have
forced survey organizations to consider new data collection strategies. Adaptive design has emerged as
a tool for tailoring contact to cases, based on data monitoring both between and during data collection
periods. There is a range of data collection features that can be tailored, including mode of data
collection, incentives, number of contacts, and case prioritization. The feature used for tailoring should
be chosen to achieve specific survey goals. For example, one may want to increase response rate or
sample representativeness for a given cost, or maintain the same response rate, while reducing cost.
The planned case prioritization experiment is a test of the former; prioritization entails altering the
amount of effort an interviewer expends on a case based on data monitoring metrics.

A2.

Purpose and Use of the Information Collection

The Division of Health Interview Statistics, in collaboration with the Center for Adaptive Design (CAD) at
the U.S. Census Bureau U.S. Census Bureau, is planning to test the use of adaptive design to increase
sample representativeness and, therefore, reduce nonresponse bias in the National Health Interview
Survey (NHIS). The planned experiment would be carried out beginning in July 2016, to encompass the
third quarter of data collection. Data monitoring metrics and quality indicators will be used to assign a
priority level to cases in each interviewer’s workload, with the aim of increasing sample
representativeness and reducing nonresponse bias within a fixed cost.
This experiment requires no additional data collection from respondents, or changes to data collection
procedures experienced by survey participants. No personally-identifiable information is used in building
the response propensity models. Instead, response propensity and balancing propensity models will be
constructed using Contact History Instrument (CHI) variables, Neighborhood Observation Instrument
(NOI) variables, and Census 2010 or ACS 2009-13 variables (block group level) available on the 2015
Census Planning Database (CPD). Mid-month, each incomplete case will be assigned a response
propensity score and a balancing propensity score based on these models and, using these scores,
assigned a priority (low, medium, high) for contact by field representatives. Thus, this prioritization
involves altering the amount of effort an interviewer expends on a case based on data monitoring
metrics.
Experimental Design
Figure 2 provides a high-level visual overview of the proposed experiment. The experiment will be
randomized at the interviewer level, meaning an entire interviewer’s workload will either be a control
workload or an experimental workload. Each arm of the experiment will include 50% of interviewers, so
the survey will be split evenly between experiment and control groups. For information on the
minimum detectable differences that we expect to find, given conservative assumptions, see
Appendix A.
While cases will be assigned to either the experimental or control groups before data collection starts,
all cases will be treated the same for the first 15 days of data collection in each month. During this time,
interviewers (referred to as “FR” in the figure, short for “field representative”) must ensure that at least
one personal visit attempt is made for each case, and that contact history data and interviewer
observations are documented for each case in the Contact History Instrument (CHI) and Neighborhood

10

Observation Instrument (NOI). The NOI data1 will be combined with covariate data that is available
before the start of data collection (CPD variables) and contact history instrument (CHI) data2 that is
collected during every contact attempt for each case to estimate two models: a response propensity
(likelihood) model and a balancing propensity (or bias likelihood) model.

Figure 2. Conceptual Overview of NHIS Adaptive Design Experiment
Response Propensity (Likelihood) Model. A response propensity or likelihood model (logistic regression)
will be run on the 15th to 16th of each month in quarter three. Response status will be the dependent
variable, and variables from CHI, NOI, and the CPD will serve as predictors. Variable selection relied
1

These data include information the field representative observes about the housing unit that may be predictive of key
survey outcomes (as well as response).
2
These data include information about the date, time, and outcomes of contact attempts, contact strategies used, and
respondent reluctance.

11

heavily on survey theories of response and past assessments of CHI and NOI variables related to NHIS
response [10, 11]. Unlike propensity models used in weighting adjustments [12], where we are
concerned about finding variables that predict both response and the outcomes of interest, here we are
concerned solely with predicting likelihood of response. After the model is run, each case will be
assigned a response propensity or likelihood score. Table 4 presents the set of variables to be used in
the response propensity model.
Table 4. Variables Selected for Inclusion in the Response Propensity Model
CHI Variables
Access barriers
Prior peak contact
Prior hard refusals
Prior privacy concerns
Total prior personal visits
Total prior telephone attempts
Total prior incoming contacts
Day of month
NOI Variables
Interviewer assessment of the condition of the sample unit
Interviewer assessment of whether or not the sample unit has 3+ door locks
Interviewer assessment of whether or not all household occupants are over the age of 65
Interviewer assessment of whether or not a language other than English is spoken by residents
Interviewer assessment of whether or not one more residents is disabled, handicapped, or has a
chronic health condition
Interviewer assessment of whether or not residents may be smokers
Interviewer assessment of whether or not one or more adults of the household are employed
CPD Variables
% of 2010 Census housing units with no registered occupants on Census day (vacant)

Balancing Propensity (Bias Likelihood) Model. A balancing propensity model (logistic regression) will also
be estimated on the 15th and 16th of the month. Like the response propensity model, response status will
be the dependent variable and NOI and CPD variables will be used as predictors. (Table 5 presents the
set of variables to be used in the balancing propensity model.) What separates the balancing propensity
model from a standard response propensity model is the inclusion of predictors or covariates related to
key survey outcomes, in this case health variables. (The selection of variables for inclusion in the
balancing propensity model is discussed in Appendix B.) Based on model output, each case will be
assigned a balancing propensity score that will be used to identify nonrespondents most likely to reduce
bias in key survey variables if converted to respondents. Essentially, the balancing propensity model will
be used to identify cases that are most unlike the set of sample members that have already responded.
In essence, we are attempting to balance response across groups defined by variables, available for both
responding and nonresponding households, related to health outcomes on the NHIS.

12

Table 5. Variables Selected for Inclusion in the Balancing Propensity Model
NOI Variables
Interviewer assessment of the household’s income relative to the general population
Interviewer assessment of whether or not all household occupants are over the age of 65
Interviewer assessment of whether or not a language other than English is spoken by residents
Interviewer assessment of whether or not one or more adults of the household are employed
Interviewer assessment of whether or not one more residents is disabled, handicapped, or has a
chronic health condition
Interviewer assessment of whether or not residents may be smokers
Interviewer assessment of whether or not the sample unit has a well-tended yard or garden
Interviewer assessment of the condition of the sample unit
Interviewer assessment of whether or not the walls of the sample unit are damaged
CPD Variables
% of ACS pop. 25+ that have a college degree or higher
% of ACS pop. 25+ that are not high school grads
% of 2010 Census occ. housing units with female householder(s) and no husband
% of 2010 Census occ. housing units where householder and spouse in same household
Average aggregated household income of ACS occupied housing units
Average aggregated value for ACS occupied housing units
% of ACS population that is uninsured
% of ACS civilians 16+ that are unemployed
% of ACS population classified as below the poverty line
% of 2010 Census housing units with no registered occupants on Census day (vacant)
As to how sample balancing helps to improve representativeness, take the following example. We know
from past analyses of NHIS data that interviewer observations of the household’s income relative to the
general population (bottom third, middle third, top third) is related to current smoking status among
adults (polychoric correlation from 2013: 0.2745). Assume that the true relationship between
interviewer assessment of household income and current smoking status is as follows:
x1: observation of
household income

y1: current smoking status

Bottom Third

30%

Middle Third

20%

Top Third

10%

Finally, assume that the survey achieves a 75% response rate. To get the most accurate estimate of
current smoking status, how would we want the response rate to be distributed across the household
income groups? In Figure 4, three response rate options are presented. In option A, all of the
nonresponse is coming from the top third income group. Based on our knowledge of the true
13

relationship between income and smoking status (above), this would likely lead to an overestimate of
current smoking among adults. In option B, while nonresponse is more equitably distributed, both the
middle and top third income groups are under-represented. As a result, we are likely still overestimating
current smoking status. Ultimately, option C is our goal. In this option, all three income groups have a
75% response rate. By ensuring that households from these groups respond at the same rate, our
estimate of current smoking status should be a very close reflection of the true current smoking rate.
This is what is meant by sample balancing.

Figure 3. Balancing Sample to Ensure Representativeness

For a more thorough discussion of sample balancing and how it may be optimal for some estimators and
can protect against nonresponse bias, see: Royall and Herson 1973[13]; Royall and Herson 1973a[14];
Royall 1992[15]; Sarndal and Lundquist 2015[16]; Shouten et al. 2015[17]; and Valliant et al. 2000[18].
Also, for a discussion of how the balancing propensity model covariates were selected for this
experiment and response rates for groups defined by these covariates, see Appendix B.
Using the combination of response propensity and balancing propensity scores, the priority (H: high, M:
medium, L: low) for each incomplete case will be determined. Together, the respective scores allow us
to calculate how strategic a case is to achieving the experimental goal of improving sample
representativeness and managing costs without negatively affecting overall response. For example, we
would want to assign high priority to a case that, if completed, would improve representativeness.
However, this decision would have to be balanced by how likely the case is to be completed. This is
conceptually visualized in Figure 4.

14

HIGH
LOW

y
r it
r io
)P

MED

ar
ul
eg
(R

Likelihood of Response

No

Figure 4. Conceptual Visualization of Prioritization

If the case would contribute moderately to representativeness but has a very low likelihood of
responding based on the response propensity model, then assigning the case to receive greater priority
may simply lead to increased costs with no gains in sample representativeness. Similarly, if a case had a
very high likelihood of response, but would not improve representativeness because we already had an
overabundance of respondents with the same characteristics, we would not want to assign a high
priority to the case. All cases (both experimental and control) will be assigned a priority status, which
may be useful later for propensity matching or direct comparisons among subgroups in the
experimental vs. control groups. However, only experimental cases will have the priority statuses
displayed in field management reports and on interviewer laptops. Interviewers will then be asked to
work their cases according to the assigned priorities.
It is important to note that the response and balancing propensity models, as shown in Figure 2, will be
re-estimated on the 22nd to 23rd of each month. Updated priorities on remaining incomplete cases will
be pushed out to FR’s laptops at that time. The experimental cases with low priority status may be
pulled from the interviewer’s workload in the final week of data collection.
At each prioritization point in data collection (the 15th-16th and 22nd-23rd), the execution of business rules
to determine each case’s priority classification will occur in two steps. First, we will execute the initial
model-based prioritization, and then we will evaluate those initial priorities against four specific
exceptions to determine the final prioritization. On the 15th-16th, we will assign 20% of open cases a
high priority, and 20% of cases a low priority. At the 22nd-23rd prioritization, we will evaluate remaining
cases against the thresholds set on the 15th-16th to update priorities.
As previously described, each case will be assigned a response propensity score and a balancing
propensity score. The response propensity score estimates the likelihood that a case will be completed
(high versus low likelihood) and the balancing propensity score estimates how unlikely a case is to be in
the respondent population (high versus low value). Because these scores are bounded by zero and one
15

and are the result of logistic regressions, the distributions of estimated propensities are often skewed.
As a result, we transformed the propensity estimates to obtain distributions closer to standard normal
distributions, and then we standardized them so the 50th percentiles of each propensity lie at zero.
Creating a scatterplot of the standardized response and balancing propensities for cases still in the field
results in a plot with four quadrants, as follows: TR (top right: higher value, higher likelihood); TL (top
left: higher value, lower likelihood); BR (bottom right: lower value, higher likelihood); and BL (bottom
left: lower value, lower likelihood). Again, the axes are the “average” value and likelihood. Figure 5
shows these scatterplots for February and March of 2016. Approximately, though not exactly, 25% of
the sample resides in each quadrant.

Figure 5. Scatterplots for February and March Standardized Response and Balancing Propensities
Given the fact that we want to assign high priority to cases that have higher than average value and
higher than average likelihood of response, we want to focus our efforts on the TR (Top Right) Quadrant.
Within that quadrant, cases farthest away from the origin (farthest top and farthest right) are going to
be the cases with the highest value and the highest likelihoods, so we want to select cases farthest away
from the origin first for prioritization. Similarly, we want to assign low priority to cases that have lower
than average value and lower than average likelihood of response, so we want to focus on the BL
(Bottom Left) Quadrant, first selecting cases farthest away from the origin in the negative direction.
To operationalize the selection, we can use a simple Cartesian distance metric to estimate the distance
of each case from the origin and rank cases within their quadrant:

 = 
+ 

,

where
 is the distance we can rank on (where larger distances are ranked higher),  is the
standardized estimate of the response propensity and 
is the standardized estimate of the balancing
propensity. We can then rank cases based on their distances in the TR and BL quadrants and prioritize
so that 20% of cases are assigned a high priority and 20% of cases are assigned a low priority. Figure 6
shows the same scatterplots above with the high and low priority cases identified.

16

Figure 6. Figure 4 Scatterplots with Assigned Prioritization
Based on the Day 16 prioritization, the minimum distance for which a case was assigned a high or low
priority will be retained for use in the Day 23 reprioritization that will be described later.
After prioritization has been assigned with respect to the models, as shown above, five business rules
will be evaluated which could override this priority assignment for individual cases. Those rules are:
(1) If a CHI and or a NOI entry has not been completed at the time of prioritization, these cases will
receive an automatic high priority. CHI and NOI data are crucial to our ability to prioritize, since
they comprise the variables in our propensity models. Further, NOI and CHI data provide insight
into how a case is progressing in the field. Lacking these data means no progress has been
made, and the case should be prioritized to rectify that situation. FRs will be evaluated partially
based on their NOI and CHI completion rates by the 15th of the month, reducing the number of
cases that should be affected by this rule.
(2) If there are multiple sample units in a single Group Quarter (GQ) or Multi-Unit Structure (MU),
and an access barrier is encountered, all cases in that GQ or MU will receive the highest priority
of all cases in the GQ or MU. In other words, if there are 3 cases within a GQ, and two are
medium priority and one is high, we will assign all three cases a high priority. Due to the
additional effort commonly required by FRs to gain entry to GQs and MUs, we want to be as
efficient about this additional effort as possible. We expect this rule to be executed rarely given
the low level of clustering in the NHIS. In addition, not all sample months have GQ sample
cases, further reducing the frequency of this rule being executed.
(3) Extra units, additional units, and spawns will all inherit the priority of their parent case until we
have a scheduled reprioritization. These cases are usually geographically co-located with the
parent unit, and similarly to the units in GQs and MUs, we want the FRs to be efficient about
their effort. Therefore, we will assign the same priorities so the cases are worked with the same
level of effort. Their priorities may diverge when we reprioritize cases at a later date, given
information in the NOI or CHI.
(4) Partial interviews at the time of prioritization will not be assigned a low priority (they might be
medium or high). These cases have already been worked, and the interviewer has collected
partial data from the household. In order to remain efficient and obtain completed interviews
17

where possible, these cases will not be assigned a low priority, and if we pull cases in Week 4
from interviewer laptops, these cases will not be pulled.
(5) If an appointment has been scheduled with a household, we will not allow the case to be a low
priority in Week 4. We do not want to pull a case from the interviewer’s laptop if an
appointment has been scheduled.
Once these rules have been evaluated for all cases, priorities will be pushed down to interviewer
laptops. Interviewers in the experiment will receive the “true” priorities, while interviewers in the
control will receive only “medium” priorities.
Reprioritization will happen on the 22nd-23rd. Rather than prioritize from scratch by selecting a new
minimum distance for prioritization as described earlier, we will use the thresholds from the
prioritization that occurred on the 15th-16th. This ties closely with the idea of balance. If interviewers
work cases according to priorities, spending more time in Week 3 on high priority cases and no to very
little effort on low priority cases, more high priority cases will be completed. This will make the low
priority cases which were over-represented less over-represented, and the under-represented cases less
under-represented. By improving sample balance, more cases may be considered medium priority
during Week 4, and those cases can be worked normally for the remainder of data collection. This will
reduce the number of cases eligible to be pulled from interviewer laptops, while also reducing the
number of cases that require additional effort late in data collection. From the perspective of contact
burden, having a flexible prioritization that allows cases to move from high priority to medium priority
will help avoid cases being overworked. Similarly, allowing cases to move from low priority to medium
priority, or medium priority to high priority ensures cases that will help meet dynamic data collection
goals are targeted for additional effort.
To understand how cases might be reassigned to different priorities during Week 4, and the effect this
would have on resolution rates during Week 3, we simulated how cases could move across different
priority classifications as interviewers complete more high priority cases successfully. To do this, we
made some simple but reasonable assumptions:
•

We assumed the response rate of the low priority cases would not increase during Week 3 (as
the cases are not being worked);

•

We assumed the response rate of the medium priority cases would increase by 15% during
Week 3; and

•

We assumed the response rate of the high priority cases could increase by 20%, 25%, or 30%,
depending on how successful interviewers’ efforts were at converting high priority cases.

Table 6 below includes cases that were open in Week 4, so cases that would not have been completed
during the prioritization of Week 3. It is clear that there is some movement between the different
strata. In February, across the two prioritization times (Week 3 and Week 4), both the high and low
priority strata had net losses of unresolved cases, while the medium priority experienced a net gain.
Conversely, in March, while cases still shifted from the low priority into medium, some cases in the
medium strata actually moved to the high strata. These differences may be due to varying compositions
of nonresponders at weeks 3 and 4 for the two months.

18

Table 6. Week 4 Reprioritization - Net Case Movement by Strata for Open Cases in Week 4
Case Totals by Priority Category and Completion Pattern
Completion
Pattern
High Increases
20%
High Increases
25%
High Increases
30%

Priority
Level
High
Medium

February, 2016
Week 3
Week 4
Priority
Priority
462
314
1483
1819

March, 2016
Week 3
Week 4
Priority
Priority
662
756
2150
2118

Low
High
Medium

588
429
1483

400
321
1789

848
621
2150

786
739
2105

Low
High
Medium

588
396
1483

390
298
1776

848
572
2150

775
714
2097

Low

588

393

848

759

Of particular interest is the fact that the number of cases in the low priority strata in Week 4 exceed, if
only slightly, the number of cases in the high priority strata in Week 4. This is important for cost control,
as we need to reduce contact attempts on some cases in order to reallocate resources to other cases.
Cases that were assigned a high priority status in Week 3 and were considered completed for the sake of
our simulations are not included in these tables, as we are showing the movement of open cases across
strata.
Finally, we can estimate the effect of each of these completion patterns. Table 7 shows that the biggest
increase in response rate that we achieve is when we prioritize cases and the high priority cases are
completed at a rate of 20%. After that, the response rate does not appreciably increase, likely because
the number of high cases remaining is diminishing. This also supports our theory that the low priority
cases are less likely to respond – assuming they will not respond does not diminish the response rate
significantly.
Table 7. Week 3 Effect on Resolution Rates by Completion Pattern
Resolution Rate (end of Week 3)
Response Pattern
Actual Response
Pattern
High Increases 20%
High Increases 25%
High Increases 30%

February, 2016
53.87%
60.33%
60.75%
61.28%

March, 2016
51.99%
58.66%
59.20%
59.74%

19

Evaluation of the Experiment
Both process and outcome measures will be examined as part of the evaluation of the experiment.
Distribution of Interviewer Effort
The success of the experiment is predicated on interviewers adjusting their effort according to the
prioritizations assigned to their cases. Once the case prioritizations are assigned and pushed out to FR
laptops on the 16th of each month, the distribution of contact attempts by case prioritization will be
monitored. Focusing on the experimental group, for example, are interviewers shifting their contact
attempts away from low priority cases toward high priority cases? When interviewers make their
contact attempts will also be monitored. Are interviewers in the experimental group moving attempts
on high priority cases to more lucrative time slots (e.g., evening hours and weekends)?
Sample Representativeness
As noted previously, the primary objective of this experiment is to improve sample balance or
representativeness (minimize nonresponse bias) within current cost constraints while maintaining
current response rates.
During each month of the quarter, the R-indicator (a measure of variation in response propensities that
ranges between 0 and 1) will be monitored for both the control and experimental groups using the
aforementioned balancing model [19, 20]. The R-indicator provides an assessment of sample
representativeness conditional on the covariates included in the balancing model. (Again, the balancing
model will include covariates that are a mixture of interviewer observational measures along with
decennial Census and ACS measures captured at the Census block group or tract level. The covariates
selected for the model are those that are highly correlated with key health outcomes.) A decreasing Rindicator represents an increase in the variation in response propensities, suggesting less sample
balance or representativeness. An increasing R-indicator indicates less variation in response propensities
and better sample balance. Comparing R-indicators for the experimental and control groups during and
at the end of data collection will provide an assessment of whether the experimental case prioritizations
led to improved representativeness (indicated by a higher value on the R-indicator for the experimental
compared to the control group).
Post-data collection, demographic characteristics of the two groups, including age, sex, race/ethnicity,
and education level, will be compared. Next, point and variance estimates of 16 select health variables
included in NHIS’s Early Release (ER) indicator reports (e.g., health insurance coverage, failure to obtain
needed medical care, cigarette smoking and alcohol consumption, and general health status) will be
compared between the two groups. If increased representativeness of the sample has been achieved,
differences in the estimates across the two groups should be observed. Using past analyses of
nonresponse bias (see Section 2), statements can be made as to whether the experimental group
estimate is moving in a direction that minimizes nonresponse bias.
Reductions in variance would also suggest increased representativeness. A more representative sample
should result in less extreme nonresponse and poststratification adjustments, reducing the variability of
sample weights. Design effects and standard errors for all estimates examined will be compared
between the control and experimental groups.

20

All comparisons in this section would be performed overall and by age, sex, and race/ethnicity
subgroups.
Response Rates and Survey Costs
Since the goal is to improve representativeness within current cost constraints while maintaining current
response rates, overall and module-specific response rates, completed interview rates, sufficient partial
interview rates, and refusal rates will be compared by control and experimental groups. In addition, the
total number of contact attempts (in-person visits and phone calls) for the two groups will serve as a
proxy for survey costs.
Other Data Quality Indicators
Additional measures of data quality will also be examined by adaptive design group, including, but not
limited to, item nonresponse rates (“don’t know” and “refused” responses), item response times, and
survey breakoffs.

A12. Estimates of Annualized Burden Hours and Cost
Requiring no additional data collection nor changes to data collection procedures experienced by NHIS
respondents, the planned adaptive design experiment would not alter the previously-approved
estimates of annualized burden hours and survey administration costs. Costs associated with
implementing the experiment are limited to interviewer training expenses, and are covered by funds
designated to Methodological Projects listed in Line 5 of the previously-approved burden table.

21

References
[1] Curtin, R.; Presser, S.; Singer, E. (2000). “The Effects of Response Rate Changes on the Index of
Consumer Sentiment.” Public Opinion Quarterly, 64: 413-428.
[2] Olson, K. (2006). “Survey Participation, Nonresponse Bias, Measurement Error Bias, and Total Bias.”
Public Opinion Quarterly, 70: 737-758.
[3] Groves, R.; Heeringa, S. “Responsive design for household surveys: tools for actively controlling
survey errors and costs.” Journal of the Royal Statistical Society, Series A. 169 (2006): pp 439-457.
[4] Axinn, W.; Link, C.; Groves, R. “Responsive Design Demographic Data Collection and Models of
Demographic Behavior.” Demography. 48 (2011): pp 1127 – 1149.
[5] Calinescu, M.; Bhulai, S.; Schouten, B. “Optimal Resource Allocation in Survey Designs.” European
Journal of Operations Research. 226:1 (April, 2013): pp 115-121.
[6] Mercer A.; Caporaso, A.; Cantor, D.; Townsend, R. “How Much Gets You How Much? Monetary
Incentives and Response Rates in Household Surveys.” Public Opinion Quarterly. 79: 1 (2015): pp 105129.
[7] Coffey, S.; Reist, B.; Zotti, A. “Static Adaptive Design in the NSCG: Results of a Targeted Incentive
Timing Study.” Presentation at the Joint Statistical Meetings, 2015. Seattle, WA.
[8] Konicki, S. “Adaptive Design Research for the 2020 Census.” Presentation at the Joint Statistical
Meetings, 2015. Seattle, WA.
[9] Wagner, J.; West, B.; Kirgis, N.; Lepkowski, J.; Axinn, W.; Ndiaye, S. “Use of Paradata in a Responsive
Design Framework to Manage a Field Data Collection.” Journal of Official Statistics. 28:4 (2012): pp. 477499.
[10] Erdman, C. and Dahlhamer, J. (2013) “Evaluating Interviewer Observations in the National Health
Interview Survey: Associations with response propensity.” Proceedings of the Joint Statistical Meetings,
Section on Survey Research Methods. Montreal, CN.
[11] Walsh, R.; Dahlhamer, J.; Bates, N. (2014) “Assessing Interviewer Observations in the NHIS.”
Proceedings of the Joint Statistical Meetings, Section on Survey Research Methods. Boston, MA.
[12] Little, R. J.; Vartivarian, S. L. “Does Weighting for Nonresponse Increase the Variance of Survey
Means?” Survey Methodology. 31 (2005): pp. 161-168.
[13] Royall, R. M.; Herson, J. “Robust estimation in finite populations I.” Journal of the American
Statistical Association 68. (1973): pp. 880-889.
[14] Royall, R. M.; Herson, J. “Robust estimation in finite populations II: Stratification on a size
variable.” Journal of the American Statistical Association 68. (1973): pp. 890-893.
[15] Royall, R. M. “Robustness and optimal design under prediction models for finite populations”
Survey methodology 18 (1992): pp. 179-185.
[16] Sarndal, C-E; Lundquist, P “Accuracy in estimation with nonresponse: A function of degree of
imbalance and degree of explanation.” Journal of Survey Statistics and Methodology. 2 (2014): 361-387.
[17] Schouten, B.; Cobben, F.; Lundquist, P.; Wagner, J “Does more balanced survey response imply less
non-response bias?” Journal of the Royal Statistical Society: Series A (Statistics in Society) (2015) doi:
10.1111/rssa.12152.
[18]Valliant, R.; Dorfman, A. H.; Royall, R.M. Finite Population Sampling and Inference: A Prediction
Approach (2000) John Wiley & Sons, New York
[19] Schouten, B. Cobben, F. Bethlehem, J. “Indicators for representativeness of survey response.”
Survey Methodology. 35.1 (June 2009): pp 101 – 113.
[20] Schouten, B. Shlomo, N. Skinner, C. “Indicators for monitoring and improving representativeness of
response.” Journal of Official Statistics. 27.2 (2011): pp 231 – 253.
[21] Snedecor, G. W.; Cochran, W.G. (1989). Statistical Methods. 8th ed. Ames, IA: Iowa State University
Press; 8ed.
22

[22] Dahlhamer, J; Gindi, R.; Bates, N.; Walsh, R. (2014) “Judgments under Uncertainty: Assessing the
Quality of Interviewer-Generated Paradata.” Paper prepared for the 26th International Workshop on
Household Survey Nonresponse, Leuven, Belgium.

23

Appendix A
Minimum Detectable Differences
The power of this experiment to make generalizations about case prioritization rests on having sample
sizes that result in reasonable standard errors for our analytic outcomes of interest. We assume that
there are approximately 6,400 cases in the NHIS in a given month, and so approximately 3,200 would be
randomized (through interviewer selection) into each of two groups, experiment and control.
For the purposes of these calculations, we make a simple assumption that by the 15th of the month,
which is 50% of the way through data collection, interviewers will have resolved 50% of their cases. This
is a conservative assumption, resulting in a lower number of unresolved cases than we would expect to
see halfway through data collection.
This leaves us with 1,600 cases in each of the treatments. If we assign 20% of cases a high priority and
20% of cases a low in each of the treatments, we have the following case breakdown:
•
•
•

H: 320 cases
M: 960 cases
L: 320 cases

Minimum Detectable Difference Formula [21]:



 
 =  + 
 +  
 
We assume a design effect of 1.4 based on the 2015 NHIS. Given that the new design is less clustered
than the old design, the design effect we are using is conservative. In addition, we will use an alpha-crit
of 0.10 (Census Requirements), and a beta-crit of either 0.10 or 0.20. Finally, we will assume that there
is no difference between the two groups, and that the proportions we are comparing are p1=p2 = 0.50.
This will maximize the variance and make a conservative estimate of the required true difference
required to see a change.

Minimum Detectable Differences for Strata Response Rates
Here, we are looking only at cases that are unresolved in the second half of data collection. In addition,
note that these percent differences will hold for all items we may compare at the strata level including
attempts per case, variability of contact times, and others.
Low or High Priority Strata (n=320 cases):
)

1:  = 0.10:  = "1.96 + "−1.28)



0. 5 0.5
+

 1.4 = 10.95%
320 320


0. 5 0.5
2:  = 0.20:  = "1.96 + "−0.84) ) 
+
 1.4 = 9.97%
320 320

24

Medium Priority Strata (n=960 cases):



0. 5 0.5
.1:  = 0.10:  = "1.96 + "−1.28) ) 
+
 1.4 = 6.32%
960 960


0. 5 0.5
.2:  = 0.20:  = "1.96 + "−0. 84) ) 
+
 1.4 = 5.76%
960 960

We would need to see differences of around 8%-10% in the High and Low identified groups to be able to
make statistically significant comparisons between the treatment and control groups. This, however,
assumes a 50% response rate across all groups, which maximizes the variance in the formula and also
means that a larger difference is required to detect a statistical difference. If, say the response rates in
the high group were closer to 70%, the requirement for statistical significance in A1 above would drop
to: 10.05% and 8.48% depending on whether the design effect is incorporated.
Minimum Detectable Differences for R-Indicators (Full Sample or Variable Level):
Here, we are looking at all cases in each treatment. Note: These percent differences will hold for all
items we may compare at the full sample level.


0. 5
0.5

)
+
1:  = 0.10:  = "1.96 + "−1.28) 
 1.4 = 3.46%
3200 3200


0. 5
0.5
2:  = 0.20:  = "1.96 + "−0. 84) ) 
+
 1.4 = 3.15%
3200 3200

We are including all cases from the treatments in these comparisons, so we will be able to detect much
smaller differences at a statistically significant level.

25

Appendix B:
Identifying Variables for Use in a Balancing Propensity Model
Introduction
To identify variables for inclusion in the balancing propensity model, we performed a correlation
analysis involving all 15 of the available NOI variables and 35 Census 2010 or ACS 2009-13 variables
available on the 2015 CPD. We then correlated each of the 50 NOI and CPD variables with 84 health
variables from the 2013 NHIS: 19 person-level variables from the person file, 22 sample child measures,
and 43 sample adult measures. Included in the 84 health variables were all measures included in the
NHIS Early Release (ER) Program.
Methods
Selection of Health-Related Variables
Health outcomes from the person, sample child, and sample adult files with universes of “all persons,”
“all sample children,” or “all sample adults,” were included in the analysis, with some exceptions. Given
their importance, ER variables were included by default. As a starting point, this produced a total of 19
variables from the person file, 49 from the sample child file, and 86 from the sample adult file.
All 19 person file variables were retained for the analysis. To narrow the list of sample child and sample
adult variables, SAS PROC VARCLUS was used. (Note that the ER measures were not included in the
VARCLUS procedures as all were retained in the subsequent analysis.) Clusters of similar health
measures were identified using the VARCLUS procedure and the variable with the smallest 1 – R2 ratio
was selected from each cluster. This reduced the number of sample child measures for analysis to 22
(including ER measures), and the number of sample adult variables to 43 (including ER measures). In
total, 84 health outcomes were included in subsequent analyses exploring correlations with NOI and
CPD variables. Table A1 lists the 84 health variables.
Selection of CPD Variables
The 2015 CPD was first merged with the 2013 NHIS at the block group level, and then the tract if block
group was missing on the NHIS. Thirty-five (35) of the Census 2010 or ACS 2009-2013 measures available
on the CPD (or recoded from existing CPD variables) were selected for analysis. Within each of nine
Census Divisions, the 35 measures were recoded into deciles, quintiles, quartiles, or tertiles depending
on the observed distribution of the original measure. The list and description of the 35 measures can be
found in Table A2.
Selection of NOI Variables
All 15 of the available NOI variables were included in the analysis (see Table A3).
Statistical Procedures
Each selected CPD and NOI variable was correlated with the full set of 84 health measures. Given that
the variables in this analysis were nominal (most health outcomes were dichotomous) or ordinal,

26

Table B1. Health Variables Used in the Correlation Analysis
Person/Family-Level Variables
Person limited in the kind or amount of work he/she can do
Person receives special educational or early intervention services
Person needs help with routine needs
Person has difficulty walking without special equipment
Person has difficulty remembering because he/she experiences periods of confusion
Person needs help with personal care needs (ER measure)
Persons reported health status is excellent or very good (ER measure)
Person had a medically-attended injury or poisoning in the past 3 months
Person did not receive care due to cost in past 12 months (ER measure)
Person was hospitalized overnight in the past 12 months
Person received care at home in the last 2 weeks
Person received medical advice or test results care over the phone in the last 2 weeks
Person received care at a doctor’s office, clinic, ER, or other place in last 2 weeks
Person received care 10 or more times in past 12 months
Person has health insurance coverage (ER measure)
Family is deemed to be food insecure
Anyone in family has a flexible spending account
Family has problems paying medical bills
Family out-of-pocket expenditures for medical care
Sample Child Variables
Sample child took medications for difficulties with emotions, concentration, or behaviors in past 6
months
Seen or talked to a specialist about sample child’s health in past 12 months
Delayed getting care for sample child in past 12 months
Did not get care for sample child due to cost in past 12 months
Time since last saw or talked to a doctor
Sample child has many worries, or often seems worried
Sample child had hay fever in past 12 months
Told by doctor that sample child has any other developmental delay
Sample child has difficulties with emotions, concentration, behavior, etc.
Sample child had stomach or intestinal illness that started in past 2 weeks
Seen or talked to a nurse practitioner, physician assistant, or midwife about sample child’s health in
past 12 months
Sample child’s hearing without a hearing aid
Does sample child have any trouble seeing
Sample child’s health compared to 12 months ago
Sample child ever has chicken pox
Told that doctor’s office/clinic would not accept sample child’s health insurance coverage in past 12
months
Sample child has a regular source of care (ER measure)
Sample child received a flu shot in the past 12 months (ER measure)
Sample child had an asthma attack in the past 12 months (ER measure)
Sample child still has asthma (ER measure)
27

Time since last saw or talked to a dentist
Sample child had stuttering or stammering during the past 12 months
Sample Adult Variables
Sample adult ever told by doctor that he/she had coronary heart disease
Sample adult had an asthma attack in the past 12 months (ER measure)
Sample adult still has asthma (ER measure)
Sample adult ever been told by doctor he/she had an ulcer
Sample adult ever told by doctor that he/she had cancer
Sample adult ever told by doctor that he/she had cancer (ER measure)
Sample adult told by doctor he/she had hay fever in past 12 months
Sample adult told by doctor he/she had bronchitis in past 12 months
Sample adult had symptoms of joint pain in past 30 days
Sample adult had neck pain in past 3 months
Sample adult had head/chest cold in past 2 weeks
Sample adult had high cholesterol in the past 12 months
Sample adult’s hearing without a hearing aid
Does sample adult have any trouble seeing
Sample adult lost all upper and lower natural teeth
Number of days illness/injury kept sample adult in bed in past 12 months
Sample adult’s health compared to 1 year ago
Sample adult has movement difficulties
Sample adults has social limitations
Sample adult had 5 or more drinks in 1 day in the past 12 months (ER measure)
Sample adult is a current smoker (ER measure)
Sample adult met federal physical activity guidelines (ER measure)
Sample adult is obese (ER measure)
Sample adult worries over being able to pay medical bills
Sample adult ever been tested for HIV (ER measure)
Sample adult had serious psychological distress in past 30 days (ER measure)
Average number of hours of sleep sample adult gets
Sample adult has a regular source of care (ER measure)
Sample adult saw an eye doctor in the past 12 months
Sample adult saw a specialist in the past 12 months
Sample adult received home care in the past 12 months
Number of doctor visits in the past 12 months
Sample adult delayed care in the past 12 months
Sample adult did not receive care due to cost in the past 12 months
Sample adult was denied care in the past 12 months
Sample adult received a flu shot in the past 12 months (ER measure)
Sample child received a pneumococcal vaccination in the past 12 months (ER measure)
Sample adult ever had hepatitis
Sample adult ever traveled outside of the United States to countries other than Europe, Japan,
Australia, New Zealand or Canada, since 1995
28

Sample adult volunteers or works in a health-care facility
Sample adult used health information technology in the past 12 months
Sample adult tried to purchase health insurance directly in past 3 years
Sample adult’s health insurance coverage compared to 1 year ago
Table B2. List of Recoded 2015 Census Planning Database Measures Used in This Analysis
Prediction of low Census mail return rate
% of 2010 Census total population < than 5 years old
% of 2010 Census total population identifying as NH white
% of 2010 Census total population identifying as NH black or Hispanic
% of 2010 Census total population < than 18 years old
% of ACS pop. 5+ that speaks language other than English at home
% of ACS pop. 25+ that are not high school grads
% of ACS pop. 25+ that have a college degree or higher
% of 2010 Census occ. housing units where householder and spouse in same household
% of 2010 Census occ. housing units where householder lives alone
% of 2010 Census occ. housing units that were owner-occupied
% of ACS housing units in a multi-unit structure
Average aggregated household income of ACS occ. housing units
Average aggregated value for ACS occ. housing units
% of 2010 Census total population 65+
% of 2010 Census total population identifying as Hispanic
% of 2010 Census total population identifying as NH black
% of 2010 Census total population identifying as NH AIAN
% of 2010 Census total population identifying as NH Asian
% of ACS population classified as below the poverty line
% of ACS civilians 16+ that are unemployed
% of ACS population that is uninsured
% of 2010 Census occ. housing units where householder lives alone or with a non-relative
Average # of persons per 2010 Census occ. housing unit
% of ACS pop. 5+ that speaks Spanish/Spanish Creole at home
% of ACS population that was not born a U.S. citizen
% of ACS pop. that are not citizens of the U.S.
% of 2010 Census occ. housing units w/ female householder and no husband
% of 2010 Census family-occ. housing units with a related child under 6
% of ACS occ. housing units where the householder moved in 2010 or later
% of ACS occ. housing units that received public assistance
% of 2010 Census housing units with no reg. occupants on Census day
% of ACS occ. housing units with > 1.01 persons per room
% of ACS occ. housing units with no working telephone
% of 2010 Census occ. housing units with a child

29

Table B3. Neighborhood Observation Instrument (NOI) Variables
Variable
GRAFFITI

ADDR_COND

ACCESS

YARDS

WALLS

BARS

LOCKS

CHILDREN

Question
Did you observe graffiti or painted-over graffiti on buildings,
sidewalks, walls, or signs in the block face of the sample unit
or building within which the sample unit resides?
How would you describe the condition of the sample unit or
the building within which the sample unit resides?
Based on your observation, does the sample unit or the
building within which the sample unit resides have: …a
security buzzer, key code, doorman, or any other barrier that
may prevent access (for example dogs, locked gate, etc.)?
Based on your observation, does the sample unit or the
building within which the sample unit resides have: …a welltended yard or garden?
Based on your observation, does the sample unit or the
building within which the sample unit resides have: …peeling
paint or damaged exterior walls?
Based on your observation, does the sample unit or the
building within which the sample unit resides have:
…window bars or grating on the doors or windows?
If this is a multiunit structure, answer based on the sample
unit, not the building within the sample unit resides. Based
on your observation, does the SAMPLE UNIT have: …3 or
more door locks?

Response Options
Yes
No
Very poor
Poor
Fair
Good
Excellent
Yes
No
Yes
No
Unable to observe
Yes
No
Yes
No
Yes
No
Unable to observe

Based on your observation, does the SAMPLE UNIT
Yes
have…indication that children under 6 (including babies) may live at
No
the unit (visible toys, car seat, strollers, outdoor swing/play set for
Unable to observe
example)?

If this is a multiunit structure, answer based on the sample
unit, not the building within the sample unit resides. Based
on your observation, does the SAMPLE UNIT have: …a wheel
WHEELCHAIR
chair ramp or other indicators that the residents of the
sample unit are handicapped, disabled, or may have a
chronic health condition (deaf, blind, use oxygen, etc.)?
If this is a multiunit structure, answer based on the sample
unit, not the building within the sample unit resides. Based
BICYCLE
on your observation, does the SAMPLE UNIT have: …an
adult-sized bicycle?
If this is a multiunit structure, answer based on the sample
unit, not the building within the sample unit resides. Based
SMOKER
on your observation, does the SAMPLE UNIT have: …any
indication that the residents of the sample unit are smokers
(cigarette/cigar butts, ashtrays, smell smoke, etc.)?

Yes
No
Unable to observe

Yes
No
Unable to observe
Yes
No
Unable to observe

30

HHINC

Relative to the general population and based on your
observations, would you judge this sample unit to have a
household income:
1. In the bottom third of the population
2. In the middle third of the population
3. In the top third of the population

Bottom third
Middle third
Top third

EMPLOYED

Based on your observation, would you say at least one adult
resident of the sample unit is employed? Yes or no.

Yes
No

HHLANG

Based on your observation, would you say that the residents
of the sample unit speak a language other than English?

Yes
No
The responses were
recoded to all
occupants over the
age of 65 versus
other.

OVER65

How old would you estimate the residents of the sample unit to
be? 1. All occupants under the age of 30; 2. All occupants over the
age of 65; or 3. Other age composition.

polychoric/tetrachoric correlations were performed in SAS. To evaluate the relative strength of
correlations, absolute values were taken. As a summary measure, the average absolute correlation
between a CPD/NOI variable and the health variables was produced. For example, the ACS measure of
uninsured from the CPD was correlated with each of 19 health variables from the person file. The
absolute value of each of 19 correlations was taken and then summed. The total was divided by 19 to
produce an average absolute correlation between the ACS uninsured measure and the 19 person-level
health variables.
With regard to the NOI variables, past analyses of select observations revealed that the outcome of the
attempt on which the observations were recorded was a strong predictor of measurement error [22].
Not surprisingly, if the observations were recorded after contact was made with a household,
agreement between the observation and survey data increased. To guard against inflating the
magnitude of the correlations between the NOI observations and health outcomes, the correlational
analysis involving the NOI variables was limited to cases where the observations were recorded on an
attempt coded as a noncontact. This most closely approximates the ideal data collection protocol in
which the observations would be recorded prior to ever making a contact attempt on the household.
All correlations were produced using SAS PROC FREQ with the POLYCHORIC option. Note that the
analyses used base weights but ignored the complex sampling design of the NHIS.

Results
Table A4 summarizes the results of all the correlations. Again, since we were more interested in
summarizing the magnitude as opposed to the direction of the correlations, the table presents absolute
correlations. As shown in Table A4, five NOI measures had the strongest average absolute correlation
with the 84 health variables: whether or not the interviewer observed evidence that one or more
residents are disabled, handicapped, or has a chronic health condition (WHEELCHAIR); interviewer
assessment of whether a household’s income falls in the bottom, middle, or top third of household
incomes in the larger area (HHINC); whether or not the interviewer observed evidence that all residents
of the household are over the age of 65 (OVER65); the interviewer’s assessment of the physical
condition of the address (ADDR_COND); and whether or not the interviewer assessed that one or more
31

adults at the residence are employed (EMPLOYED). CPD variables with the highest average absolute
correlations with the 84 health measures included the percent of persons in the block group/tract with a
college degree (PCT_COLLEGE_ACS); average aggregated household income of the block group/tract
(AVG_HHINC_ACS); the average aggregated value of occupied housing units in the block group/tract
(AVG_HOUSEVALUE_ACS); and the percent of adults 25 or older with less than a high school education
(PCT_LTHS_ACS).
Given that the average absolute correlation from all correlations was .0670 (bottom row of Table B4),
we selected variables with average absolute correlations of .0700 or greater as possible inputs to a
balancing propensity model. This reduced the original list of 50 NOI/CPD variables to 19.

Table B4. Summary of Polychoric/Tetrachoric Correlations between 35 CPD Variables, 15 NOI
Variables (noncontacts only), and 84 NHIS Person/Family, Sample Child, and Sample Adult
Variables: NHIS, 2013 (base weights)
CPD/NOI Variable

wheelchair
hhinc
over65
addr_cond
employed
pct_college_acs
avg_hhinc_acs
avg_housevalue_acs
pct_lths_acs
hhlang
pct_femnohusbhhld
yards
walls
pct_poverty_acs
smoker
pct_uninsured_acs
pct_vacant
pct_unemployed_acs
pct_mrdcplhhld
pct_nhasian
pct_ownerocc
pct_hispblack
pct_nhwhite
pct_65plus
pct_famwchildu6
graffiti
children
bars
pct_forborn_acs

|Average
Corr.|

|.1000.1999|

|.2000.2999|

|.3000.3999|

|.4000
+|

|Strongest
Corr.|

0.1425
0.1202
0.1086
0.1029
0.1025
0.0983
0.0978
0.0908
0.0888
0.0857
0.0831
0.0824
0.0805
0.0799
0.0755
0.0736
0.0709
0.0705
0.0704
0.0681
0.0660
0.0656
0.0650
0.0634
0.0622
0.0607
0.0603
0.0590
0.0566

28
29
24
26
24
25
29
28
25
24
25
19
27
24
17
22
22
18
17
19
16
21
20
18
17
12
18
14
9

11
13
9
8
7
9
8
6
4
4
3
5
3
3
3
2
0
1
2
1
2
1
1
0
1
0
0
1
1

4
2
4
1
3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

5
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0.4919
0.3888
0.3838
0.3245
0.3545
0.2830
0.2988
0.2895
0.2652
0.2461
0.2908
0.2493
0.2372
0.2742
0.4189
0.2680
0.1911
0.2213
0.2266
0.2526
0.2264
0.2048
0.2005
0.1781
0.2140
0.1923
0.1775
0.2061
0.2124
32

pct_hispanic
0.0565
14
0
0
0
0.1911
pct_crowding_acs
0.0553
13
0
0
0
0.1946
pct_othlang_acs
0.0545
11
0
0
0
0.1598
pct_under5
0.0541
15
0
0
0
0.1808
pct_noncitizen_acs
0.0535
11
0
0
0
0.1570
pct_pubassist_acs
0.0532
9
0
0
0
0.1514
pct_spanish_acs
0.0498
10
0
0
0
0.1877
access
0.0494
10
0
0
0
0.1959
avg_personsinhhld
0.0487
3
0
0
0
0.1362
pct_nonfamhhld
0.0452
5
0
0
0
0.1375
bicycle
0.0439
6
0
0
0
0.1273
pct_nhblack
0.0431
6
0
0
0
0.1591
pct_children
0.0428
3
0
0
0
0.1177
pct_multiunit_acs
0.0425
6
0
0
0
0.1527
pct_singperhhld
0.0408
3
0
0
0
0.1302
locks
0.0402
5
0
0
0
0.1754
pct_hhldmovedin_acs
0.0377
4
0
0
0
0.1525
pct_hhldwchildren
0.0370
1
0
0
0
0.1033
pct_nophone_acs
0.0364
4
0
0
0
0.1124
pct_nhaian
0.0337
0
0
0
0
0.0838
TOTAL
0.0670
779
113
14
6
0.4919
Bold = NOI variable. Note that only the correlations with health variables were considered when the
NOI observations were recorded during an attempt resulting in a noncontact.

Response Rates for Groups Defined by the 19 NOI/CPD Variables
Table B5 presents response rates by the NOI observations and the CPD variables selected for the
balancing propensity model. For the NOI observations, response rates were available for 2013-2015. For
the CPD measures, we present response rates for 2013 and 2014. As noted earlier, a goal of the
experiment is to balance response across groups defined by the NOI observations and CPD variables
included in the balancing propensity model. By that, we are trying to achieve similar response rates, not
achieve similar numbers of cases. Take, for example, the smoking observation. For each household,
interviewers are asked to assess whether they observed any evidence that household residents may be
smokers. In 2013, 7.3% of in-scope households were observed to include smokers. The response rate for
these households was 83.4%, while the response rate for the households where no evidence of smokers
was observed was 75.6%. Hence, smoking households, as assessed by the interviewers, were overrepresented in the sample. Knowing that this observation is also correlated with several health
outcomes on the NHIS (see Appendix B), the potential for nonresponse bias increases. For example,
from the correlation analysis, we know that the smoking observation is positively associated with the
sample adult not receiving specific health care services due to cost. Since we are over-representing
smoker households in our sample, our estimate of the percentage of adults who did not receive specific
health care services due to cost in the past 12 months may be too high. Using this simple example, our
goal via adaptive design would be to achieve similar response rates for the smoking and non-smoking
(as defined by the interviewers) households to help protect against nonresponse bias.

33

There are clear differences in response rates for the majority of the 19 variables, with chi-square
analysis revealing only two non-significant differences in response rate distributions: whether or not the
interviewer observed a well-tended yard or garden in 2013, and whether or not the interviewer
observed peeling paint or damaged walls in 2015. Some of the observed differences in response rates
were sizeable. For example, the response rate for households where interviewers observed evidence
that a resident may be disabled, handicapped, or have a chronic health condition was 80.1% in 2015.
The response rate for households where this was not observed was 70.5%, representing a nearly 10
percentage point difference between the two groups. When looking at all three years of data, the gap in
response rates between these two groups of households has been growing.
Focusing on other NOI observations, households where the interviewer observed evidence of a language
other than English being spoken had higher response rates for all three years compared to households
where this was not observed. Also, households assessed to have one or more working adults had lower
response rates for all three years compared to households where this was not observed.
Among the CPD variables, it is clear that higher response rates are achieved among lower educated and
less affluent households. For both 2013 and 2014, response rates were higher for households in the top
decile, compared to households in the bottom decile, of the ACS measure of the percentage of adults
aged 25 and older that are not high school graduates. In addition, households in the bottom decile of
the ACS measure of average aggregated household income had higher response rates for both years
compared to households in the top decile.
In sum, we observe consistent response rate differences across the set of variables (available for both
responding and nonresponding households) found to be more highly-related to several NHIS health
outcomes. Again, a primary goal of the adaptive design experiment will be to reduce the response rate
differences across these variables in order to reduce the potential for or magnitude of nonresponse bias
in critical health estimates.

34

Table B5. Response Rates (standard error) by Select NOI, Census 2010 (block group/tract), and ACS 2009-13 (block group/tract) Measures:
NHIS, 2013-2015 (base weighted)
NOI Measures
Evidence that household speaks a language other than English
Yes
No
All household residents over the age of 65
Yes
No
Household income relative to general population
Bottom third
Middle third
Top third
At least one adult resident of the household is employed
Yes
No
Sample unit has a wheel chair ramp or other indicators that residents are
handicapped or disabled
Yes
No/Unable to observe the sample unit
Any indication that the residents are smokers
Yes
No/Unable to observe the sample unit
Sample unit has a well-tended yard or garden
Yes
No/Not applicable
Damaged walls or peeling paint
Yes
No

2013

2014

2015

79.6 (0.25)
75.7 (0.18)

78.5 (0.25)
74.1 (0.22)

75.4 (0.34)
69.8 (0.32)

81.1 (0.35)
75.5 (0.18)

81.0 (0.39)
73.8 (0.21)

77.9 (0.39)
69.6 (0.31)

79.9 (0.30)
75.0 (0.17)
73.7 (0.31)

77.3 (0.34)
73.8 (0.20)
73.0 (0.38)

73.1 (0.48)
69.5 (0.30)
69.5 (0.41)

75.2 (0.18)
80.0 (0.24)

73.5 (0.22)
78.9 (0.28)

69.4 (0.32)
74.4 (0.37)

82.1 (0.56)
76.0 (0.17)

83.6 (0.56)
74.3 (0.20)

80.1 (0.73)
70.5 (0.30)

83.4 (0.43)
75.6 (0.17)

81.2 (0.55)
74.2 (0.19)

77.9 (0.52)
70.3 (0.31)

76.1 (0.19)
76.4 (0.24)

75.1 (0.22)
73.8 (0.29)

71.3 (0.27)
69.0 (0.45)

78.8 (0.33)
75.9 (0.17)

76.1 (0.40)
74.4 (0.20)

71.0 (0.61)
70.4 (0.29)

35

Table B5. continued
2013
Physical condition of the sample unit
Very poor/poor
78.8 (0.54)
Fair
77.1 (0.30)
Good
75.8 (0.20)
Very good
75.3 (0.28)
Census 2010 (block group/tract), and ACS 2009-13 (block group/tract) Measures
% of ACS population 25+ that has a college degree (ACS 2009-13)
Bottom decile
79.0 (0.54)
Top decile
72.8 (0.27)
% of 2010 Census occupied housing units with a female householder and
no husband (Census 2010)
Bottom quartile
73.8 (0.30)
Top quartile
77.3 (0.26)
Average aggregated household income of ACS occupied housing units
(ACS 2009-13)
Bottom decile
79.6 (0.47)
Top decile
71.0 (0.21)
% of 2010 Census housing units with no registered occupants on Census
day (Census 2010)
Bottom quartile
74.1 (0.20)
Top quartile
78.4 (0.31)
% of 2010 Census occupied housing units where householder and spouse
in same household (Census 2010)
Bottom decile
75.7 (0.37)
Top decile
73.9 (0.34)
Average aggregated value for ACS occupied housing units (ACS 2009-13)
Bottom decile
78.5 (0.49)
Top decile
71.7 (0.25)

2014

2015

76.7 (0.57)
74.8 (0.26)
74.5 (0.25)
73.9 (0.32)

70.6 (0.63)
69.7 (0.49)
71.2 (0.30)
69.9 (0.34)

77.7 (0.46)
72.7 (0.24)

N/A
N/A

73.5 (0.27)
75.2 (0.24)

N/A
N/A

75.8 (0.48)
71.9 (0.31)

N/A
N/A

73.1 (0.25)
75.2 (0.37)

N/A
N/A

71.6 (0.32)
73.1 (0.39)

N/A
N/A

75.4 (0.35)
72.1 (0.28)

N/A
N/A

36

Table B5. continued
% of persons uninsured (ACS 2009-13)
Bottom quintile
Top quintile
% of ACS population 25+ that are not high school grads (ACS 2009-13)
Bottom decile
Top decile
% of ACS civilians 16+ that are unemployed (ACS 2009-13)
Bottom quintile
Top quintile
% of persons in poverty (ACS 2009-13)
Bottom quintile
Top quintile

2013

2014

2015

74.0 (0.25)
78.0 (0.31)

73.5 (0.26)
75.9 (0.41)

N/A
N/A

73.4 (0.28)
78.9 (0.41)

72.7 (0.29)
77.6 (0.53)

N/A
N/A

74.3 (0.33)
77.8 (0.33)

73.0 (0.30)
75.7 (0.46)

N/A
N/A

73.8 (0.26)
78.2 (0.34)

73.2 (0.26)
75.7 (0.34)

N/A
N/A

37