Hierarchical Indexing, Data Cubes, and Multi-indexes in SDTL

Revised: 5 June 2020

This document shows how data with hierarchical indexes (aka data cubes aka Ncubes aka multi-indexes) may be described in SDTL. Properties have been included in the SDTL DataframeDescription type to describe hierarchical indexes. SDTL describes data with hierarchical indexes in a general way intended to facilitate sharing data among different statistical software applications. Functionality for assuring the integrity of these indexes must be in the software used to create and use them. Although operations on hierarchically indexed data often involve specialized syntax and processing procedures, these operations can be described by standard SDTL commands.

Data cubes are often used to speed the analysis of large databases. Data cubes are defined by “dimensions,” which represent attributes used for classification of the data, and “measures,” which are often aggregations like counts and means. Although data cubes may have any number of dimensions, we show below that they can be represented as two-dimensional data matrices.

The Python language allows dataframes to have multiple hierarchical indexes – multi-indexes. In Python both rows and columns can be multi-indexed, which differs from the Ncube model where only rows have hierarchical indexes. Some data transformations in Python, notably aggregation (SDTL Collapse), produce dataframes with multi-indexes.

Figure 1 shows a multi-index created by the Python commands:

fruitprice = prices.groupby([‘Fruit’, ‘Color’]).agg( { ‘Price’ : [‘mean’, ‘count’] } )

fruitprice[ (‘Total’,’sum’)] = fruitprice[‘Price’,’mean’] * fruitprice[‘Price’,’count’]

The dataframe produced by these commands has two row dimensions (Fruit, Color) and two column dimensions (Price, Total).

Figure 1. A multi-index in Pythonimage0

image1

Figure 2. Multi-index saved to CSV file

When this dataframe is saved as a CSV file, it is rendered as in Figure 2. Notice that the row dimensions (Fruit, Color) are now columns, and that “Price” has been added to the column headers for both “mean” and “count”.

image2

Figure 3. Multi-index saved to Stata dta file

When the Python dataframe is saved to a Stata .dta file, we get Figure 3. Since Stata does not allow hierarchical indexes, variables are given compound names that include both parts of the column index, such as “Price_mean” and “Price_count”.

This example shows that data with hierarchical indexes can be represented as rectangular datasets. Row indexes are variables. Column indexes can be replaced by compound variable names.

Hierarchical indexes do provide additional functionality to languages like Python. For example, this expression returns only the row where “Fruit”=”Apple” and “Color”=”Red”:

fruitprice.loc[‘Apple’, ‘Red’]

We can select the columns “mean” and “count” by referring to the “Price” index, like this

Fruit price[ ‘Price’]

Notice that selection of rows and columns operates differently. Rows are selected by specifying a value. Columns are selected by using one or more index names. In other words, as the translation to Stata shows, column indexes are variable names.

Solutions for SDTL:

  1. Syntax for referring to columns in a multi-index

The levels in a column index can be treated as a variable name

The recommended syntax for combining indexes into variable names is:

“index1.index2.variablename”.

Using the example above, the columns would be:

“Price.mean”, “Price.count”, “Total.sum”

  1. Defining hierarchical indexes

Use SDTL element: DataframeDescription

Properties:

  1. rowDimensions: an ordered list of variables used as row indexes
  2. columnDimensions: an ordered list of variable names created from
    column indexes

The example above would be represented as

{“$type”: “DataframeDescription”,

“dataframeName”: “fruitprice”,

“variableInventory”: [“Price.mean”, “Price.count”, “Total.sum”].

“rowDimensions”: [“Fruit”, “Color”],

“columnDimensions”:[“Price”, “Total”]

}

  1. Selecting by row indexes is the same as a selecting using the values
    of variables in an SDTL IfRows command
Python syntax SPSS SDTL
fru itprice.loc[‘Apple’, (‘varX’)] = 2 IF (Fruit=’Apple’) varX=2.

{ “$type”: “IfRows”,

“command”: “ifRows”,

“condition”: {

“$type”: “Func tionCallExpression”,

“function”: “eq”,

“isSdtlName”: true,

“arguments”: [

{type”: “FunctionArgument”,

“argumentValue”: {

“$type”: “Variab leSymbolExpression”,

“variableName”: “Fruit” }

},

{type”: “FunctionArgument”,

“argumentValue”: {

“$type”: “String ConstantExpression”,

“value”: “Apple” }

}

]

},

“thenCommands”: [

{ type”: “Compute”,

“command”: “compute”,

” originalSourceText”: “IF (Fruit=’Apple’) varX=2.”

},

“variable”: {type”: “Variab leSymbolExpression”,

“variableName”: “varX” },

“expression”: { type”: “Numeric ConstantExpression”,

“value”: “2”,

“numericType”: “int” }

}

]

}

  1. Selecting columns is the same as a VariableRange expression in
    SDTL
Python syntax SPSS SDTL
fruitprice[ ‘Price’] Price_mean to Price_count

{type”: “Varia bleRangeExpression”,

“first”: “Price.mean”,

“last”: “Price.count”

}