Introductory example

Example without code

First some data to be rounded.

Table 1: Original data in tabular form with row and column totals
col1 col2 col3 col4 col5 Total
row1 6 0 1 3 4 14
row2 1 2 3 1 2 9
row3 0 1 1 0 2 4
Total 7 3 5 4 8 27

Some of the inner cells are rounded. Thereafter new totals are computed. The underlying algorithm tries to keep the values of these totals close to the original ones.

Table 2: All small inner cell values (1-4) are rounded using 5 as rounding base.
col1 col2 col3 col4 col5 Total
row1 6 0 0 5 5 16
row2 0 0 5 0 5 10
row3 0 0 0 0 0 0
Total 6 0 5 5 10 26

When the inner cells are not going to be published, the number of cells to be rounded can be limited.

Table 3: Assuming only row and column totals to be published, necessary small inner cell values (1-4) are rounded using 5 as rounding base.
col1 col2 col3 col4 col5 Total
row1 6 0 0 5 4 15
row2 1 0 5 0 2 8
row3 0 5 0 0 0 5
Total 7 5 5 5 6 28

The example dataset in Table 1

library(SmallCountRounding)
z <- SmallCountData("exPSD")
z
   rows cols freq
1  row1 col1    6
2  row2 col1    1
3  row3 col1    0
4  row1 col2    0
5  row2 col2    2
6  row3 col2    1
7  row1 col3    1
8  row2 col3    3
9  row3 col3    1
10 row1 col4    3
11 row2 col4    1
12 row3 col4    0
13 row1 col5    4
14 row2 col5    2
15 row3 col5    2

Rounding all small cells (Table 2)

To avoid any small values in the range 1-4 we can use 5 as rounding base.

a <- PLSrounding(z, freqVar = "freq", roundBase = 5)

The result is given in Table 2 and can bee seen in the output elements below.

a$inner  rows cols original rounded difference 1 row1 col1 6 6 0 2 row2 col1 1 0 -1 3 row3 col1 0 0 0 4 row1 col2 0 0 0 5 row2 col2 2 0 -2 6 row3 col2 1 0 -1 7 row1 col3 1 0 -1 8 row2 col3 3 5 2 9 row3 col3 1 0 -1 10 row1 col4 3 5 2 11 row2 col4 1 0 -1 12 row3 col4 0 0 0 13 row1 col5 4 5 1 14 row2 col5 2 5 3 15 row3 col5 2 0 -2 a$publish
    rows  cols original rounded difference
1  Total Total       27      26         -1
2  Total  col1        7       6         -1
3  Total  col2        3       0         -3
4  Total  col3        5       5          0
5  Total  col4        4       5          1
6  Total  col5        8      10          2
7   row1 Total       14      16          2
8   row1  col1        6       6          0
9   row1  col2        0       0          0
10  row1  col3        1       0         -1
11  row1  col4        3       5          2
12  row1  col5        4       5          1
13  row2 Total        9      10          1
14  row2  col1        1       0         -1
15  row2  col2        2       0         -2
16  row2  col3        3       5          2
17  row2  col4        1       0         -1
18  row2  col5        2       5          3
19  row3 Total        4       0         -4
20  row3  col1        0       0          0
21  row3  col2        1       0         -1
22  row3  col3        1       0         -1
23  row3  col4        0       0          0
24  row3  col5        2       0         -2

The output element publish contains the original and rounded versions of the all the 24 values in Table 2. The corresponding element inner contains only the 15 inner cells and is similar to the input data. The values in publish are additive. That is, marginal cells (Totals) can be computed straightforwardly from inner for both original and rounded counts.

Rounding necessary small inner cell (Table 3)

Assuming only row and column totals to be published, the publishable cells can be defined by the formula ~rows+cols. Rounding can now be performed by:

b <- PLSrounding(z, "freq", 5, formula = ~rows + cols)

The result is given in Table 3 and can bee seen in the output elements below.

b$inner  rows cols original rounded difference 1 row1 col1 6 6 0 2 row2 col1 1 1 0 3 row3 col1 0 0 0 4 row1 col2 0 0 0 5 row2 col2 2 0 -2 6 row3 col2 1 5 4 7 row1 col3 1 0 -1 8 row2 col3 3 5 2 9 row3 col3 1 0 -1 10 row1 col4 3 5 2 11 row2 col4 1 0 -1 12 row3 col4 0 0 0 13 row1 col5 4 4 0 14 row2 col5 2 2 0 15 row3 col5 2 0 -2 b$publish
   rows  cols original rounded difference
1 Total Total       27      28          1
2  row1 Total       14      15          1
3  row2 Total        9       8         -1
4  row3 Total        4       5          1
5 Total  col1        7       7          0
6 Total  col2        3       5          2
7 Total  col3        5       5          0
8 Total  col4        4       5          1
9 Total  col5        8       6         -2

Set seed for unique output

The underlying algorithm is sequential. Within a loop, the next cell to be given the rounding base value is selected according to a criterion. Random draw is used when draw criterion. Therefore, the result may vary. Use set.seed to ensure unique output. In the examples above set.seed(12345) was called before calling PLSrounding.

The output object

The result of printing the output from PLSrounding is (a and b as above):

a

PLSrounding summary:

maxdiff      HDutility    meanAbsDiff rootMeanSquare
4          0.715         1.3333         1.6833

Frequencies of cell frequencies and absolute differences:

inn.0 inn.1-4 inn.5 inn.6+ inn.all pub.0 pub.1-4 pub.5 pub.6+ pub.all
original     3      11     .      1      15     3      14     1      6      24
rounded     10       .     4      1      15    12       .     6      6      24
absDiff      4      11     .      .      15     5      19     .      .      24
b

PLSrounding summary:

maxdiff      HDutility    meanAbsDiff rootMeanSquare
2          0.941              1         1.2019

Frequencies of cell frequencies and absolute differences:

inn.0 inn.1-4 inn.5 inn.6+ inn.all pub.0 pub.1-4 pub.5 pub.6+ pub.all
original     3      11     .      1      15     .       3     1      5       9
rounded      8       3     3      1      15     .       .     4      5       9
absDiff      7       8     .      .      15     2       7     .      .       9

First some utility measures are printet. For example maxdiff is the maximum difference between an original and rounded cells within publish. Thereafter a table of frequencies of cell frequencies and absolute differences are printed. Summary of inner and publish are shown in the left and right parts of the table, respectively. For example, row rounded and column inn.6+ is the number of rounded inner cell frequencies greater than or equal to 6. The last row (absDiff) is based on the differences without signs.

It is possible to compute manually the printed utility measures by:

f <- b$publish$original
g <- b$publish$rounded
print(c(
maxdiff        =  max(abs(g - f)),
HDutility      =  HDutility(f, g),
meanAbsDiff    =  mean(abs(g - f)),
rootMeanSquare =  sqrt(mean((g - f)^2))
))
       maxdiff      HDutility    meanAbsDiff rootMeanSquare
2.000000       0.940951       1.000000       1.201850 

These measures are also found in the output element metrics together with the same measures based on inner. See ?HDutility for more information about the utility measure based on the Hellinger distance.

Apart from printing, output is a usual list and summary works as usual.

summary(b)
          Length Class      Mode
inner      5     data.frame list
publish    5     data.frame list
metrics    9     -none-     numeric
freqTable 30     -none-     numeric

The output element freqTable is the table seen when the output object is printed (frequencies of cell frequencies and absolute differences).

Hierarchical data

Example without code

Below is a small data set to be used as input.

Table 4: Input data
geo eu year freq
Iceland nonEU 2018 2
Portugal EU 2018 3
Spain EU 2018 7
Iceland nonEU 2019 1
Portugal EU 2019 5
Spain EU 2019 6

The variables geo and eu is hierarchical related. This data set can be processed in several ways. In some cases, the entire table will be input and in other cases the eu column can be omitted. Then, the hierarchical information is sent as input in another way. One possibility is the table below, where the hierarchy is coded as in the r package sdcTable.

Table 5: Hierarchy, geo
levels codes
@ Total
@@ EU
@@@ Portugal
@@@ Spain
@@ nonEU
@@@ Iceland

Another possibility is TauArgus coding. More general coding is also possible. See ?AutoHierarchies for more information.

Below is output in the case were all possible combinations (including the inner cells) are to be published. Also in this example we use 5 as a rounding base. As can be seen below, this output can be generated in several ways. The inner cells are colored according to the rounding.

Table 6: Ouput data (publish)
geo year original rounded difference
Total Total 24 23 -1
Total 2018 12 12 0
Total 2019 12 11 -1
EU Total 21 23 2
EU 2018 10 12 2
EU 2019 11 11 0
nonEU Total 3 0 -3
nonEU 2018 2 0 -2
nonEU 2019 1 0 -1
Iceland Total 3 0 -3
Iceland 2018 2 0 -2
Iceland 2019 1 0 -1
Portugal Total 8 10 2
Portugal 2018 3 5 2
Portugal 2019 5 5 0
Spain Total 13 13 0
Spain 2018 7 7 0
Spain 2019 6 6 0

Data (Table 4) and hierarchies (Table 5)

e6 <- SmallCountData("e6")  # As Table 4
eDimList <- SmallCountData("eDimList")
eDimList
$geo levels codes 1 @ Total 2 @@ EU 3 @@@ Portugal 4 @@@ Spain 5 @@ nonEU 6 @@@ Iceland$year
levels codes
1      @ Total
2     @@  2018
3     @@  2019

As seen above, a hierarchy is specified for both variables. eDimList$geo is given in Table 5 and eDimList$year is a plain hierarchy with total code.

Four ways to produce Table 6

The four lines below produce the same results with element publish as in Table 6. Ordering of rows can be different.

PLSrounding(e6, "freq", 5)                                                      # a)
PLSrounding(e6, "freq", 5, formula = ~eu * year + geo * year)                   # b)
PLSrounding(e6[, -2], "freq", 5, hierarchies = eDimList)                        # c)
PLSrounding(e6[, -2], "freq", 5, hierarchies = eDimList, formula = ~geo * year) # d)
• In a) and c) the function uses hierarchies for the calculations. In a) the hierarchies are found automatically from the input data.
• In b) the cross-classifications are found from the formula. In addition, hierarchical relations in the input data are analysed so that geo and eu are combined into the same output column.
• In d), how to cross the hierarchies are defined by a formula.

Remarks and other parameters

A difference occur when all combinations are not contained in input data. Then c) above will limit output to combinations available in input. In the other cases zeroes will be added. The extra zeroes can be avoided by using removeEmpty=TRUE. Note also the parameter inputInOutput which can be used to specify whether to include codes from input. Below is an example with incomplete input data using both these parameters.

out <- PLSrounding(e6[-1, ], "freq", 5, removeEmpty = TRUE, inputInOutput = c(FALSE,TRUE))
out

PLSrounding summary:

maxdiff      HDutility    meanAbsDiff rootMeanSquare
1         0.8925            0.5         0.7071

Frequencies of cell frequencies and absolute differences:

inn.0 inn.1-4 inn.5 inn.6+ inn.all pub.0 pub.1-4 pub.5 pub.6+ pub.all
original     .       2     1      2       5     .       2     .      6       8
rounded      1       1     1      2       5     2       .     .      6       8
absDiff      4       1     .      .       5     4       4     .      .       8
out$inner  geo year original rounded difference 2 Portugal 2018 3 3 0 3 Spain 2018 7 7 0 4 Iceland 2019 1 0 -1 5 Portugal 2019 5 5 0 6 Spain 2019 6 6 0 out$publish
    geo  year original rounded difference
1 Total Total       22      21         -1
2 Total  2018       10      10          0
3 Total  2019       12      11         -1
4    EU Total       21      21          0
5    EU  2018       10      10          0
6    EU  2019       11      11          0
7 nonEU Total        1       0         -1
8 nonEU  2019        1       0         -1

In this case only a single inner cell needed to be rounded (Iceland, 2019). The original small value of (Portugal, 2018) could be retained.