Preliminary Evaluation of the Impact of 2019 NHIS Questionnaire Redesign and Weighting Adjustments Page | 11
U.S. Department of Health and Human Services ● Centers for Disease Control and Prevention ● National Center for Health Statistics
Health policy analysts and other public health professionals are generally interested in whether there are
changes over time in the official estimates for the key indicators. That is, they will want to know whether the
2019 estimates are higher, lower, or statistically similar to the 2018 estimates. However, as noted earlier, any
differences observed between estimates for 2018 (Column A) and 2019 (Column E) may be partly attributable to
the 2019 NHIS questionnaire redesign, the updated weighting approach, or both, or they may reflect an actual
change over time. If questionnaire design effects or weighting effects are observable, it may not be appropriate
to compare the 2018 and 2019 estimates and conclude that the observed difference (or lack of difference)
reflects an actual change (or lack of change) over time. Health policy analysts are also interested in longer term
trends in addition to year-to-year changes. While design or weighting changes may affect year-to-year
comparisons, these changes may have less impact on ascertaining longer term trends.
Table 3 highlights two comparisons of interest in assessing the potential effect of the questionnaire redesign and
updated weighting approach on the official estimates. First, for each indicator, the potential effect of the
questionnaire design was examined by comparing estimates across designs during October-December 2018, a
time when both the prior design and the field test of the redesign were in the field simultaneously. That is, the
2018 Quarter 4 Production estimate (Column B) was compared to the 2018 Quarter 4 Bridge estimate (Column
C), and the statistical significance was tested. The same weighting approach was used for both data sets.
Second, for each indicator, the potential impact of the updated weighting approach was examined by comparing
estimates derived with the two different weighting approaches for the same data set. That is, the 2019 Full-year
Early Release estimate based on the old weight (Column D) was compared to the 2019 Full-year Early Release
estimate based on the new weight (Column E), and the statistical significance was tested. In this analysis, both
estimates are derived from the same questionnaire design.
Third, for each indicator, the potential impact of both the questionnaire redesign and updated weighting
approach on longer-term trends was explored by examining changes in six-month interval estimates from 2015-
2019. These line graphs, along with the information presented in Table 3, help evaluate longer-term trends
through 2019. The figures show the observed estimates for each time period and the predicted estimate if no
methodological changes had been made based on the net effect of the design and weighting changes. Note that
the predicted estimates are subject to sampling error (not shown on the figures) and will differ from what would
have been observed had no changes been made in 2019. Conclusions as to the extent of the impact of the
design and weighting changes on trends or differences between 2018 and 2019 should take this into
consideration. Detailed information on the effect of the methodologic changes is provided so readers can reach
their own conclusions as to how to interpret trends. The comparison between the January-June 2019 and July-
December 2019 estimates describes any change occurring during the calendar year. This information can be
used in conjunction with changes during the previous period to determine whether an existing increasing or
decreasing trend may have continued through 2019 in light of possible methodological impacts.
All statistical tests used a two-sided alpha level of 0.10 rather than the conventional 0.05 level to determine
statistical significance. This decision leads to identification of more significant differences and more conservative
conclusions. That is, one is more likely to conclude that questionnaire design effects and weighting effects exist
for a greater number of variables. This approach is a reasonably cautious one, suitable for this look at
preliminary data and the relatively small sample sizes available for the 2018 Quarter 4 comparisons. Readers
who wish to use an alternative approach will find the actual p-values identified in the text and footnotes for
alpha levels of 0.05 and 0.01 in the tables.
Reports from the Early Release Program for key health indicators have generally presented both unadjusted
(crude) and age-adjusted estimates in tables, but only unadjusted estimates when examining trends over time.
Similarly, the estimates and comparisons presented here are based on unadjusted estimates only.