Pivot table

A pivot table contains label rows and columns, and an array of values, usually numeric data. It can contain additional information, such as table header or footer.

In the table above, columns c1 and c2 contain labels, also rows r2 and r3. The header is made up of row r1 and the intersection between rows and columns of labels (cells of c1 and c2 with r2 and r3). Let us suppose that the content of the cell (r1, c1) is especially relevant because it serves to identify the content of this pivot table with respect to others.

The matrix of values is made up of rows and columns after those containing labels (rows from r4 on, and columns from c3 on). Each value of this matrix is characterized by the combination of labels of the corresponding row and column. It is common to find null values in that matrix because the data is not produced or recorded for the combination of labels that define it.

Flat table

A flat database or flat-file database is a database that only contains a single table. A flat table is a generally denormalized table that is not related to other tables.

The table above is a flat table whose data has been obtained from the pivot table in the previous section. It only contains raw data and the labels that characterize it. An additional label has been added with the value that identifies the pivot table, the pivot table page. NA values have not been included. Aggregated data can be obtained from raw data.

Transforming a pivot table

The transformation to obtain the flat table from the pivot table has been carried out in R, using flattabler package. The code is shown below. For clarity, the pipe operator syntax provided by tidyr package has been used.

library(flattabler)
library(tidyr)

ft <- pt %>%
  set_page(1, 1) %>%
  define_labels(n_col = 2, n_row = 2) %>%
  remove_top(1) %>%
  fill_labels() %>%
  remove_agg() %>%
  fill_values() %>%
  remove_k() %>%
  replace_dec() %>%
  unpivot()

Starting from the pivot table pt:

1. We define that the value that identifies it (the pivot table page) is in cell (r1, c1), by means of set_page(1, 1).

2. The pivot table only contains labels and values, therefore rows or columns with other information have to be removed. The cells between the rows and columns of labels can be ignored. We delete the first row using remove_top(1) because it does not contain labels.

3. When we only have the cells of the pivot table, we define the number of rows and columns containing labels using define_labels(n_col = 2, n_row = 2). There are two columns and two rows of labels.

4. The example is a Spanish report that uses thousands and decimal separators in the value matrix. We need to adapt them to the R syntax for numbers. This operation is carried out using remove_k() to remove the thousands separator and replace_dec() to replace the decimal separator.

5. The array of values has gaps that, instead of having a numeric value, have an empty string. In R it is more appropriate to have NA if the data is not available. This operation is performed through fill_values().

6. Since there are more than one row or column with labels, the values of the labels of the first row and column have not been repeated. They are filled using fill_labels().

7. The pivot table contains aggregated data that can be obtained from the raw data. It is removed by remove_agg().

8. Finally, it is transformed into a flat table by unpivot(): each row corresponds to a value with its combination of labels. NA values have not been considered.

Transforming a list of pivot tables

Once we have defined the necessary transformations for a pivot table, we can apply them to any other with the same structure. Candidate tables can have different number of rows or columns, depending on the number of labels, but they must have the same number of rows and columns of labels, and the same number of header or footer rows, so that the transformations are the same for each table.

To easily perform this operation, we define a function f from the transformations, as shown below.

f <- function(pt) {
  pt %>%
    set_page(1, 1) %>%
    define_labels(n_col = 2, n_row = 2) %>%
    remove_top(1) %>%
    fill_labels() %>%
    remove_agg() %>%
    fill_values() %>%
    remove_k() %>%
    replace_dec() %>%
    unpivot()
}

ft <- flatten_table_list(list_pt_ie, f)

Given a list of pivot tables, list_pt_ie, flatten_table_list() applies the transformation defined by f to each of them, and merges the results into a flat table. In this case, the full result is not shown in this document because it takes up too much space, but a sample is shown below.

Flat table transformation

Once we have a flat table, implemented using tibble, we can use tidyverse package components to transform it, as shown below. In this case all results are displayed.

t <- ft %>%
  tidyr::pivot_wider(names_from = page, values_from = value) %>%
  dplyr::rename(A = col1, B = col2, E = row1, D = row2) %>% 
  dplyr::arrange(A, B, E, D)

Pivot table import

The objective of import operations is to obtain external data that contains one or more pivot tables to transform them. Two formats have been considered: text file, and data frame.

In the case of working with files, the situation of jointly treating all the files in a folder has also been considered.

The S3 pivot_table class has been defined in the package. Transform operations are defined for objects of this class. Import operations can be classified into two groups: those that return a pivot_table object, and those that return a list of pivot_table objects. Objects in a list can be transformed together.

df <- data.frame(unclass(pt_m4)[c(1:7)])
pt <- pivot_table(df, page = "M4")

file <- system.file("extdata", "csv/set_v_ie.csv", package = "flattabler")
pt_set_v_ie <- read_text_file(file)

folder <- system.file("extdata", "csvfolder", package = "flattabler")
lpt <- read_text_folder(folder)

Pivot table definition

Once we have a pivot_table object or list of objects, pivot tables have to be defined. Each object generated by import operations contains a text table, it is expected to contain a pivot table, but may also have more information, generally in the form of a table header or footer. Through this set of operations we transform the text table in the object into a pivot table and define its characteristics.

A pivot_table object should only contain label rows and columns, and an array of values, usually numeric data. Additional information can be used to identify the pivot table relative to other similar tables: can be used to define the pivot table page.

Page

We consider the page of the pivot table as the literal that identifies it with respect to other homogeneous tables; generally it is the value of an attribute (i.e., 2020, 2019,…) or the name of a variable (i.e., amount, profit,…). When multiple pivot tables are integrated into a flat table, the page is essential to distinguish the origin of the data. It is considered as an additional label.

The workflow is generally as follows:

1. Explore the table to determine its distribution and characteristics.
   • If we start from a list of pivot_table objects, we will explore each one of the tables. In order to transform them together, they should have homogeneous structure. We will use member reference (instead of list slicing) to access the objects in the list. Example: pt <- list_pt_ie[[1]].
   • Exploration can be carried out using utils::View(). Additionally, view_table_attr() can be used since it also shows the attributes of the object.
2. In case the text table contains multiple pivot tables, they can be obtained using divide(), which returns a list of pivot_table objects; therefore, we return to the first step.
3. Define the characteristics of the pivot table: Number of rows and columns with labels, and page value.
4. Remove the rows and columns that are not part of the pivot table: It should only contain the rows and columns of labels and an array of values.

Therefore, four groups of functions are distinguished: Functions to view the object, to split a text table into several, to define the pivot table characteristics, and to remove the rows and columns that are not part of it.

file <- system.file("extdata", "csv/set_v_ie.csv", package = "flattabler")
pt_set_v_ie <- read_text_file(file)

pt_set_v_ie %>% view_table_attr()

list_pt_ie <- pt_set_v_ie %>% divide()

pt <- list_pt_ie[[1]]

pt <- pt %>% set_page(1, 1)

pt <- pt %>% define_labels(n_col = 2, n_row = 2)

pt <- pt %>% remove_top(1)

Pivot table transformation

Once a pivot_table object only contains pivot table data, and its attributes have been defined, it could be transformed into a flat table. However, we can take advantage of the table structure to modify and complete it. Therefore, optionally, we can complete and transform the components of the pivot table: Labels and values.

pt <- pt %>% fill_labels()

pt <- pt %>% remove_agg()

pt2 <- pt_m4_compact %>%
  extract_labels(col = 1, labels = c("b1", "b2", "b3", "b4", "Total general"))

df <- get_col_values(list_pt_compact, start_row = 4)
labels <- sort(unique(df$label))

pt <- pt %>% fill_values()

pt <- pt %>% remove_k()

pt <- pt %>% replace_dec()

Flat table generation

In order to generate a flat table from a pivot_table object, it is an essential requirement to have properly defined its attributes and that it only contains the pivot table data. Optionally, if the table has a usual structure, we could have transformed the values and labels, if necessary.

We can generate a flat table from a pivot table (a pivot_table object) or from a list of pivot tables (a list of pivot_table objects).

ft_tmp <- pt %>% unpivot()

f <- function(pt) {
  pt %>%
    set_page(1, 1) %>%
    define_labels(n_col = 2, n_row = 2) %>%
    remove_top(1) %>%
    fill_labels() %>%
    remove_agg() %>%
    fill_values() %>%
    remove_k() %>%
    replace_dec() %>%
    unpivot()
}

ft <- flatten_table_list(list_pt_ie, f)

Complementary transformations

We start from a flat table ft generated in the previous section through the operation flatten_table_list().

• tidyr::pivot_wider(): It is needed when an observation is distributed among several rows. One column contains the name of the observation variables and another contains the values. In the example, we had a pivot table for each variable. The variable name is in page column. Example:

ft <- ft %>% tidyr::pivot_wider(names_from = page, values_from = value) 

• dplyr::rename(): It allows us to change the name of the columns of the flat table. Example:

ft <- ft %>% dplyr::rename(A = col1, B = col2, E = row1, D = row2) 

• dplyr::arrange(): It allows us to order the rows of the flat table according to the selected columns. Example:

ft <- ft %>% dplyr::arrange(A, B, E, D) 

Alternative transformations

We consider the same data from the illustrative example. This is not the case, but let us suppose a situation in which the transformations cannot be performed on the pivot table (because its structure does not coincide with the assumptions of the operations).

pt <- list_pt_ie[[1]]

ft_at <- pt %>%
  set_page(1, 1) %>%
  define_labels(n_col = 2, n_row = 2) %>%
  remove_top(1) %>%
  unpivot()

The only operations we perform are those for defining the pivot table and generating the flat table. The result obtained is shown below.

The functions that we need to apply to carry out the transformation are the following:

• dplyr::mutate_all(): Apply the indicated function to all the columns of the table. We replace empty strings with NA. Example:

ft_at <- ft_at %>% dplyr::mutate_all( ~ dplyr::na_if(., ""))

• tidyr::fill(): Fill in the missing values for the columns. The filling direction can be indicated. Example:

ft_at <- ft_at %>% tidyr::fill(col1, row1)

• dplyr::filter(): Subset the table, retaining all rows that satisfy the condition. We remove the aggregates (label of the row or column closest to the array of values is empty) and NA values in the array of values. Example:

ft_at <- ft_at %>% dplyr::filter(!is.na(col2) & !is.na(row2) & !is.na(value))

• dplyr::mutate_at(): Apply the indicated function to some of the columns of the table. We remove the thousands indicators and change the decimal indicator. Example:

ft_at <- ft_at %>% dplyr::mutate_at(c("value"),
                                    ~ stringr::str_replace_all(., pattern = "\\.", replacement = "")) %>%
  dplyr::mutate_at(c("value"),
                   ~ stringr::str_replace(., pattern = ",", replacement = "\\."))

As in the previous case, they can be combined to define a function and apply the transformations to all the objects in a list.

g <- function(pt) {
  pt %>%
    set_page(1, 1) %>%
    define_labels(n_col = 2, n_row = 2) %>%
    remove_top(1) %>%
    unpivot() %>%
    dplyr::mutate_all(~ dplyr::na_if(., "")) %>%
    tidyr::fill(col1, row1) %>%
    dplyr::filter(!is.na(col2) & !is.na(row2) & !is.na(value)) %>%
    dplyr::mutate_at(c("value"),
                     ~ stringr::str_replace_all(., pattern = "\\.", replacement = "")) %>%
    dplyr::mutate_at(c("value"),
                     ~ stringr::str_replace(., pattern = ",", replacement = "\\."))
}

ft2 <- flatten_table_list(list_pt_ie, g)

The result coincides with the one previously obtained. Below is a sample.

Finally, below you can see all the data once transformed.

t2 <- ft2 %>%
  tidyr::pivot_wider(names_from = page, values_from = value) %>%
  dplyr::rename(A = col1, B = col2, E = row1, D = row2) %>% 
  dplyr::arrange(A, B, E, D)

The result obtained is the same. From my point of view, transformations carried out taking advantage of the structure of the pivot table are simpler and more intuitive. However, if necessary, this set of operations can help resolve special situations not yet considered in flattabler package.