An R package to compute and decompose the Mutual Information Index (M) introduced to the social sciences by Theil and Finizza (1971). The M index is a multigroup segregation measure that is highly decomposable, satisfiying both the Strong Unit Decomposability (SUD) and the Strong Group Decomposability (SGD) properties (Frankel and Volij, 2011; Mora and Ruiz-Castillo, 2011).
The package allows for:
data.table and parallel libraries (which Windows does not permit to run with more than one CPU core).Rafael Fuentealba-Chaura
School of Computer Science
Universidad Católica de Temuco
Rudecindo Ortega 02351
Temuco, Chile
rafael.fuentealba97@gmail.com
Ricardo Mora
Department of Economics
Universidad Carlos III de Madrid
Getafe, Spain
ricmora@eco.uc3m.es
Julio Rojas-Mora
Department of Computer Science
Universidad Católica de Temuco
Rudecindo Ortega 02351
Temuco, Chile
and
Centro de Políticas Públicas
Universidad Católica de Temuco
Temuco, Chile
julio.rojas@uct.cl
You can install the stable version of mutualinf from CRAN with:
and the development version from GitHub with:
The package provides two functions:
mutual function. For more details see help(prepare_data).help(mutual).The library computes the M Index. Suppose you have 2016-2018 primary school enrollment Chile data. Each observation is a combination of, among other variables, school (school), school district (district), ethnicity (ethnicity), and socio-economic level (csep) in a tabular format object (data.frame, data.table, tibble). Variable nobs represents students frequencies in each of these combinations. In the first step, we load the package and use the prepare_data function (i) to declare the variable that includes the frequencies and (ii) to format the data for the mutual function:
library(mutualinf)
DT_Seg_Chile_1 <- prepare_data(data = DF_Seg_Chile,
vars = "all_vars",
fw = "nobs")
class(DT_Seg_Chile_1)
#> [1] "data.table" "data.frame" "mutual.data"If vars =" all_vars ", prepare_data uses all columns in the table. You may, nonetheless, use option vars with tables that have a large number of columns that are not needed in the analysis. For example:
prepares the data to conduct, as we see below, an analysis of socioeconomic segregation by school. If you want to additionally study segregation by ethnicity in the schools, the data preparation should collect all the relevant variables:
DT_Seg_Chile_1 <- prepare_data(data = DF_Seg_Chile,
vars = c("school", "csep", "ethnicity"),
fw = "nobs")If the data is originally fully disaggregated (i.e., one record represents one student), prepare_data computes the cell frequencies of the specified variables:
The mutual function can compute the index M in its simplest form, i.e., on a group dimension for a unit of analysis. For example, to compute socioeconomic segregation by schools:
and to compute ethnic segregation by schools:
The mutual function also allows the use of multiple group dimensions on which segregation is computed. For example:
computes socioeconomic and ethnic segregation in schools, effectively defining the groups as the combinations of socioeconomic and ethnic categories. As we can see, the segregation obtained considering, simultaneously, socioeconomic level and ethnicity (0.2610338) is larger than those obtained separately (0.1995499 and 0.06213906, respectively).
More generally, segregation analysis can be computed using multiple unit and/or group dimensions. For example:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "district"))
#> M
#> 1: 0.2610338computes socioeconomic and ethnic segregation in combinations of schools and districts. Note that the result is identical to that obtained in the previous case, 0.2610338. The reason is that each school only belongs to one district so that the combinations of schools and districts coincide with the set of schools. We can say that the districts are a partition of the schools and districts do not add a new source for socioeconomic and ethnic segregation.
Yet the variables that define the units may not have a hierarchical relationship between them. For example, if instead of district (district) we use type of school (sch_type, either private, charter, or public):
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "sch_type"))
#> M
#> 1: 0.2610865computes segregation in units defined by combinations of schools and types of schools. There is no hierarchical structure in the units as some schools change their type in the sample period. Consequently, the level of segregation is higher (0.2610865 vs. 0.2610338).
Option by computes the index for subsamples. The data used as an illustration include primary schools in the Chilean regions of Biobio, La Araucania, and Los Rios. Option by allows obtaining the level of segregation for each of the three regions in a single command:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "sch_type"),
by = "region")
#> region M
#> 1: Biobio 0.2312423
#> 2: La Araucania 0.2367493
#> 3: Los Rios 0.2125013In this case, the function displays the index for each region. We see that socioeconomic and ethnic segregation is greater in La Araucania (0.2367493) than in Biobio (0.2312423) and Los Rios (0.2125013).
Option within additively decomposes the total segregation index into a “between” and a “within” term:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "sch_type"),
by = "region",
within = "csep")
#> region M M_B_csep M_W_csep
#> 1: Biobio 0.2312423 0.2030819 0.02816039
#> 2: La Araucania 0.2367493 0.1906641 0.04608521
#> 3: Los Rios 0.2125013 0.1774420 0.03505928We get three terms for each region. The first, M, contains the total segregation and matches the values without option within. The second, M_B_csep, referred to as the “between” term, measures socioeconomic segregation in the combinations of schools and types of schools. The third, M_W_csep, referred to as the “within” term, is the weighted average of ethnic segregation (in the combinations of schools and types of schools) computed for each socioeconomic level (with weights equal to the demographic importance of each socioeconomic level). This “within” term can be interpreted as the part of total segregation, M, derived exclusively from ethnic differences. From this point on, we will refer to this term as “the contribution of” ethnicity.
It is also possible to obtain the decomposition of the index into a “between” ethnicity term and a “within” ethnicity term:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "sch_type"),
by = "region",
within = "ethnicity")
#> region M M_B_ethnicity M_W_ethnicity
#> 1: Biobio 0.2312423 0.02582674 0.2054156
#> 2: La Araucania 0.2367493 0.04840892 0.1883404
#> 3: Los Rios 0.2125013 0.03324738 0.1792539We get, again, three terms for each region. The first, M, captures total segregation as before. The second, M_B_ethnicity, is ethnic segregation in the schools and types of schools combinations. The third, M_W_ethnicity, is the socioeconomic contribution.
Option contribution.from displays the two contributions simultaneously:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "sch_type"),
by = "region",
contribution.from = "group_vars")
#> region M C_csep C_ethnicity interaction
#> 1: Biobio 0.2312423 0.2054156 0.02816039 -0.002333648
#> 2: La Araucania 0.2367493 0.1883404 0.04608521 0.002323710
#> 3: Los Rios 0.2125013 0.1792539 0.03505928 -0.001811897We get four terms for each region: M, C_csep, C_ethnicity, and interaction. M is total segregation, as we have already seen. C_csep is the socioeconomic contribution and matches the “within” ethnicity term,M_W_ethnicity. C_ethnicity is the ethnic contribution and matches the “within” socioeconomic term,M_W_csep. Finally, interaction is equal to M minus the sum of C_csep and C_ethnicity. It is the part of the total segregation in the combinations of schools and school types that cannot be exclusively attributed to the segregation effect of either ethnicity or csep. We can see that the socioeconomic contribution is largest in Biobio (0.2054156), while the ethnicity contribution is largest in La Araucania (0.04608521).
Option contribution.from may also display the contributions of a subset of variables. For example:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "sch_type"),
by = "region",
contribution.from = "csep")
#> region M C_csep
#> 1: Biobio 0.2312423 0.2054156
#> 2: La Araucania 0.2367493 0.1883404
#> 3: Los Rios 0.2125013 0.1792539returns M and C_csep, omitting C_ethnicity and interaction.
The display of contributions can also be performed for organizational units. For example:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "sch_type"),
by = "region",
contribution.from = "unit_vars")
#> region M C_school C_sch_type interaction
#> 1: Biobio 0.2312423 0.1293566 4.860549e-05 0.10183706
#> 2: La Araucania 0.2367493 0.1709480 8.563946e-06 0.06579272
#> 3: Los Rios 0.2125013 0.1351602 1.903942e-04 0.07715072The first of the four terms is total segregation, M, as before. The second term, C_school, contains the contribution of schools, while the third term, C_sch_type, captures the contribution of school types. The fourth term, interaction, is the part of socioeconomic and ethnic segregation that cannot be exclusively attributed to segregation by schools or by school type. Most schools types do not vary in the sample, so sch_type is almost a partition of schools. Hence, the type of school is a minor source of information compared to the school, and its contribution is minimal.
In the presence of a true partition, the analysis of contributions is simpler:
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity"),
unit = c("school", "district"),
by = "region",
contribution.from = "unit_vars")
#> region M C_school C_district interaction
#> 1: Biobio 0.2311937 0.1558457 0 0.07534802
#> 2: La Araucania 0.2367407 0.1635589 0 0.07318187
#> 3: Los Rios 0.2123109 0.1605696 0 0.05174127The contribution of districts, C_district, is zero since there is no segregation by districts within each school. Intuitively, all segregation by districts becomes segregation by schools.
The analysis of contributions is generalized to situations in which there are more than two sources of segregation by groups or units. For example, if we consider three sources of group segregation (csep,ethnicity and gender):
mutual(data = DT_Seg_Chile,
group = c("csep", "ethnicity", "gender"),
unit = c("school", "district"),
by = "region",
contribution.from = "group_vars")
#> region M C_csep C_ethnicity C_gender interaction
#> 1: Biobio 0.2731123 0.2143102 0.03438802 0.04191863 -0.01750455
#> 2: La Araucania 0.2718037 0.2017662 0.05742349 0.03506293 -0.02244892
#> 3: Los Rios 0.2836338 0.1941962 0.04642725 0.07132289 -0.02831253displays five terms: total segregation, the contributions of the three sources of segregation by groups, and the interaction term.
The only restriction of option contribution.from is that contributions of variables that define groups and variables that define units cannot be simultaneously computed since there is no single way to do this decomposition. However, option components allows retrieving all the elements of the linear combination of the “within” terms to compute the decomposition desired by an advanced user.
Rafael Fuentealba-Chaura and Julio Rojas-Mora acknowledge the financial support by the FONDECYT/ANID Project 11170583. Ricardo Mora acknowledge the financial support of MCIN/AEI/10.13039/501100011033 (Project no. PID2019-108576RB-I00). Cluster time was provided by the UCT VIP Project FEQUIP2019-INRN-03.
Frankel, D. and Volij, O. (2011). Measuring school segregation. Journal of Economic Theory, 146(1):1-38. https://doi.org/10.1016/j.jet.2010.10.008.
Guinea-Martin, D., Mora, R., & Ruiz-Castillo, J. (2018). The evolution of gender segregation over the life course. American Sociological Review, 83(5), 983-1019. https://doi.org/10.1177/0003122418794503.
Mora, R. and Guinea-Martin, D. (2021). Computing decomposable multigroup indexes of segregation. UC3M Working papers, Economics 31803, Universidad Carlos III de Madrid. Departamento de Economía.
Mora, R. and Ruiz-Castillo, J. (2011). Entropy-based segregation indices. Sociological Methodology, 41(1):159-194. https://doi.org/10.1111/j.1467-9531.2011.01237.x.
Theil, H. and Finizza, A. J. (1971). A note on the measurement of racial integration of schools by means of informational concepts. The Journal of Mathematical Sociology, 1(2):187-193. https://doi.org/10.1080/0022250X.1971.9989795.