Rounding policies and real data
To construct a measure of inflation that is free from rounding
error, this section uses the CPI’s Research Database index data
files. The data employed here include all of the major indexes
from January 1986 to July 2005 at the full level of precision used
internally at the Bureau. For this article, the CPI all-items index
and its top-level components are considered. In addition, the
information technology and personal-computer indexes are
included because they have seen rapid declines in price and
are probably the worst-case scenario for rounding error in the
post-1986 period.
A monthly benchmark inflation series is calculated from the
unrounded data and then is rounded to the one-tenth-of-apercent
level to match the published inflation series, as in the
ideal method presented in the previous section. To copy current
and possible BLS procedures, the data from the Research Database
also are rounded to one, two, and three decimal places
initially, and inflation rates are calculated. The resulting inflation
series is then rounded to the tenths place in percentage terms.
The only difference between the benchmark and rounded series
is the precision in the first stage of rounding.
Table 1 reports the percentage of the sample for which the
inflation rates in the rounded data differ from the benchmark
series at a 0.1-percent level of precision. Results are presented
for both the non-seasonally-adjusted series and the seasonally
adjusted series. Because the two series are similar and the
rounding errors should be independent between them, the differences
in the percentages shown give an indication of the
variability of the estimated percentage.
The table shows that following the current practice of rounding
the CPI index to the tenths place results in a derived monthly
inflation that is materially different from the benchmark inflation
rate roughly 25 percent of the time. This finding is basically consistent
across the various series, with a few exceptions. The
relatively low percentage of differences in the medical index and
the index for other goods and services is due to the fact that
those sectors saw high inflation over the 1986–2005 period and
consequently have large index values for most of the period. In
contrast, information technology and personal computers decreased
in price dramatically over the same period and so have
very small index values, making the first-stage rounding error
large enough to change the monthly inflation rate as often as 75
percent of the time.
A look at the columns corresponding to retaining two and
three decimals in the CPI indicates that the frequency of
discrepancies between the inflation series can be reduced to
nearly zero for most series (though not the problematic personalcomputer
series) by reporting the index rounded to three, rather
than two, decimal places.
If the inflation series created from CPI data rounded to the
tenths place differs from the benchmark series roughly 25 percent
of the time, by how much is it off? Fortunately, the rounded data
are precise enough that the difference is always limited to ±0.1 percent from 1986 to the present. In recent times, however,
monthly inflation rates have been around 0.2 percent, which
makes the rounding error as a percentage of the actual monthly
change quite large indeed. Table 2 summarizes the distribution of
the magnitude of the rounding errors relative to the unrounded
inflation rate for the all-items index.
The first column of the table indicates that, of the 234 total
observations of the rounded CPI all-items inflation index, 19
(8.1 percent) are in error by between 25 percent and 50 percent
of the magnitude of the unrounded monthly change. Summing
down the columns reveals that 62 observations (26.5 percent)
differ by more than 5 percent of the benchmark inflation rate.
Slightly more than 21 percent of the time, the reported CPI
inflation rate differs from the benchmark inflation rate by 25
percent or more. More than 6 percent of the time, the inflation
rate derived from the CPI rounded to one decimal place is off
by 100 percent or more.
Reading across the table makes it clear that raising the
initial level of rounding to the hundredths place eliminates all
of the very large relative errors. Reporting the index rounded
to the thousandths place would reduce the frequency of discrepancies
to under 1 percent, and the magnitude of the error
would be greatly diminished.
An alternative measure of the importance of rounding
error for CPI inflation is a comparison of rounding error
variance with the intrinsic sampling error variance. Sampling
error arises because the Bureau is unable to collect
all prices on all goods in the market and instead takes a
sample of these prices. To assess the reliability of the sample
of prices collected, the Bureau reports an estimate of
error variance due to its sampling procedure. Currently, the
monthly sampling error variance of all-items CPI inflation
is about 0.0036 percent.
The estimates of the sampling error variance were created
from unrounded figures, so adding rounding error to the CPI
increases the variance of the reported inflation series relative to an unrounded series. The following tabulation compares the contributions to total error variance made by sampling
error and by rounding error:
Number of decimal digits reported
Type of error variance One Two Three
Total .................................. 0.0062 0.0038 0.0036
Sampling ................................ .0036 .0036 .0036
Rounding ............................... .0026 .0002 .0000
One can see that rounding error variance is approximately 72
percent as large as sampling error variance. Reducing the
rounding error variance would reduce the total error variance
by 42 percent.