Skip to content

Settings dictionary reference

This document enumerates all the settings and configuration options available when developing your data linkage model.

You can find an interative settings editor here.

Settings keys in the base setting dictionary

The type of data linking task. Required.

  • When dedupe_only, splink find duplicates. User expected to provide a single input dataset.

  • When link_and_dedupe, splink finds links within and between input datasets. User is expected to provide two or more input datasets.

  • When link_only, splink finds links between datasets, but does not attempt to deduplicate the datasets (it does not try and find links within each input dataset.) User is expected to provide two or more input datasets.

Examples: ['dedupe_only', 'link_only', 'link_and_dedupe']

probability_two_random_records_match

The probability that two records chosen at random (with no blocking) are a match. For example, if there are a million input records and each has on average one match, then this value should be 1/1,000,000.

If you estimate parameters using expectation maximisation (EM), this provides an initial value (prior) from which the EM algorithm will start iterating. EM will then estimate the true value of this parameter.

Default value: 0.0001

Examples: [1e-05, 0.006]

em_convergence

Convergence tolerance for the Expectation Maximisation algorithm

The algorithm will stop converging when the maximum of the change in model parameters between iterations is below this value

Default value: 0.0001

Examples: [0.0001, 1e-05, 1e-06]

max_iterations

The maximum number of Expectation Maximisation iterations to run (even if convergence has not been reached)

Default value: 25

Examples: [20, 150]

unique_id_column_name

Splink requires that the input dataset has a column that uniquely identifies each reecord. unique_id_column_name is the name of the column in the input dataset representing this unique id

For linking tasks, ids must be unique within each dataset being linked, and do not need to be globally unique across input datasets

Default value: unique_id

Examples: ['unique_id', 'id', 'pk']

source_dataset_column_name

The name of the column in the input dataset representing the source dataset

Where we are linking datasets, we can't guarantee that the unique id column is globally unique across datasets, so we combine it with a source_dataset column. Usually, this is created by Splink for the user

Default value: source_dataset

Examples: ['source_dataset', 'dataset_name']

retain_matching_columns

If set to true, each column used by the comparisons sql expressions will be retained in output datasets

This is helpful so that the user can inspect matches, but once the comparison vector (gamma) columns are computed, this information is not actually needed by the algorithm. The algorithm will run faster and use less resources if this is set to false.

Default value: True

Examples: [False, True]

retain_intermediate_calculation_columns

Retain intermediate calculation columns, such as the bayes factors associated with each column in comparisons

The algorithm will run faster and use less resources if this is set to false.

Default value: False

Examples: [False, True]

comparisons

A list specifying how records should be compared for probabalistic matching. Each element is a dictionary

blocking_rules_to_generate_predictions

A list of one or more blocking rules to apply. A cartesian join is applied if blocking_rules_to_generate_predictions is empty or not supplied.

Each rule is a SQL expression representing the blocking rule, which will be used to create a join. The left table is aliased with l and the right table is aliased with r. For example, if you want to block on a first_name column, the blocking rule would be

l.first_name = r.first_name.

To block on first name and the first letter of surname, it would be

l.first_name = r.first_name and substr(l.surname,1,1) = substr(r.surname,1,1).

Note that splink deduplicates the comparisons generated by the blocking rules.

If empty or not supplied, all comparisons between the input dataset(s) will be generated and blocking will not be used. For large input datasets, this will generally be computationally intractable because it will generate comparisons equal to the number of rows squared.

Default value: []

Examples: [['l.first_name = r.first_name AND l.surname = r.surname', 'l.dob = r.dob']]

additional_columns_to_retain

A list of columns not being used in the probabalistic matching comparisons that you want to include in your results.

By default, splink drops columns which are not used by any comparisons. This gives you the option to retain columns which are not used by the model. A common example is if the user has labelled data (training data) and wishes to retain the labels in the outputs

Default value: []

Examples: [['cluster', 'col_2'], ['other_information']]

bayes_factor_column_prefix

The prefix to use for the columns that will be created to store the bayes factors

Default value: bf_

Examples: ['bf_', '__bf__']

term_frequency_adjustment_column_prefix

The prefix to use for the columns that will be created to store the term frequency adjustments

Default value: tf_

Examples: ['tf_', '__tf__']

comparison_vector_value_column_prefix

The prefix to use for the columns that will be created to store the comparison vector values

Default value: gamma_

Examples: ['gamma_', '__gamma__']

sql_dialect

The SQL dialect in which sql_conditions are written. Must be a valid sqlglot dialect

Default value: None

Examples: ['spark', 'duckdb', 'presto', 'sqlite']

Settings keys nested within each member of comparisons

output_column_name

The name used to refer to this comparison in the output dataset. By default, Splink will set this to the name(s) of any input columns used in the comparison. This key is most useful to give a clearer description to comparisons that use multiple input columns. e.g. a location column that uses postcode and town may be named location

For a comparison column that uses a single input column, e.g. first_name, this will be set first_name. For comparison columns that use multiple columns, if left blank, this will be set to the concatenation of columns used.

Examples: ['first_name', 'surname']

comparison_description

An optional label to describe this comparison, to be used in charting outputs.

Examples: ['First name exact match', 'Surname with middle levenshtein level']

comparison_levels

Comparison levels specify how input values should be compared. Each level corresponds to an assessment of similarity, such as exact match, jaro winkler match, one side of the match being null, etc

Each comparison level represents a branch of a SQL case expression. They are specified in order of evaluation, each with a sql_condition that represents the branch of a case expression

Examples: [{'sql_condition': 'first_name_l IS NULL OR first_name_r IS NULL', 'label': 'null', 'null_level': True}, {'sql_condition': 'first_name_l = first_name_r', 'label': 'exact_match', 'tf_adjustment_column': 'first_name'}, {'sql_condition': 'ELSE', 'label': 'else'}]

Settings keys nested within each member of comparison_levels

sql_condition

A branch of a SQL case expression without WHEN and THEN e.g. 'jaro_winkler_sim(surname_l, surname_r) > 0.88'

Examples: ['forename_l = forename_r', 'jaro_winkler_sim(surname_l, surname_r) > 0.88']

label_for_charts

A label for this comparson level, which will appear on charts as a reminder of what the level represts

Examples: ['exact', 'postcode exact']

u_probability

the u probability for this comparison level - i.e. the proportion of records that match this level amongst truly non-matching records

Examples: [0.9]

m_probability

the m probability for this comparison level - i.e. the proportion of records that match this level amongst truly matching records

Examples: [0.1]

is_null_level

If true, m and u values will not be estimated and instead the match weight will be zero for this column. See treatment of nulls here on page 356, quote '. Under this MAR assumption, we can simply ignore missing data.': https://imai.fas.harvard.edu/research/files/linkage.pdf

Default value: False

tf_adjustment_column

Make term frequency adjustments for this comparison level using this input column

Default value: None

Examples: ['first_name', 'postcode']

tf_adjustment_weight

Make term frequency adjustments using this weight. A weight of 1.0 is a full adjustment. A weight of 0.0 is no adjustment. A weight of 0.5 is a half adjustment

Default value: 1.0

Examples: ['first_name', 'postcode']

tf_minimum_u_value

Where the term frequency adjustment implies a u value below this value, use this minimum value instead

This prevents excessive weight being assigned to very unusual terms, such as a collision on a typo

Default value: 0.0

Examples: [0.001, 1e-09]