Mapping the press in 1930s China: A Multiple Correspondance Analysis
This is the continuation of our Newspaper Directory of China series. In this new instalment, we rely on Multiple Correspondence Analysis (MCA) to summarize and go beyond the previous analyses based on single attributes or pairs of attributes (periodicity, language, size, circulation, etc). MCA is a method for exploratory data analysis. In contrast to simple statistical analyses, MCA is concerned with correlations between variables. Based on MCA, we can further cluster individuals (periodicals) based on the attributes they have in common. Introduced by the French sociologist Pierre Bourdieu in his famous study of French educated elites (The State Nobility)1, the method has inspired a wide range of applications, including not only sociological studies but also text analysis of word co-occurrences.
In this essay, we conduct two MCA on two datasets derived from the 1931 and 1935 editions of the Directory. The objective is to reveal complex patterns of correlation among variables such as periodicity, language, age, format, layout, circulation and publisher, which we can not clearly perceive through simple statistical analyses. On this basis, we seek to group periodicals that present similar profiles, i.e. similar correlation patterns.
The datasets and the replicable code used for performing MCA and creating the visualizations are available as RMarkdown documents.
For this experiment, we start from the complete dataset containing the 1060 periodicals listed in the two directories. In the first preprocessing step, we lump together and recode the variables with rare values (e.g. language) or with too many different values (e.g. city of publication, year of establishment, publisher). In addition, we transform the continuous (numerical) variables into categorical ones (circulation, number/size of page, number/size of columns) based on the thresholds defined in our previous analyses. Next we convert all variables into factor variables, as required for MCA.
Since we intend to perform one MCA for each directory, we need to split the original dataset into two samples, one for each directory. Finally, we attribute a unique id to each observation (periodical), and we read the first column as row names.
In the final datasets, each row corresponds to a unique individual (periodical) and each column to a single variable. Both datasets contain 11 variables. We have 360 individuals in the 1931 dataset and 703 in 1935.
The pre-processing workflow is described in detail in the R Markdown document.
Multiple Correspondence Analysis
In a first step, we consider all variables as active variables. Alternatively, it may be better to treat the place of publication (Province) and the period of founding (Founded) as supplementary variables (which means they do not participate in the analysis), since these two variables are poorly projected on the two first dimensions.
Altogether, the two first dimensions capture more than 18% of information (11.92% on the first dimension, 7.23% on the second in the 1931 dataset, 10.7% and 7% in 1935). This is quite satisfactory given the high number of variables and observations contained in the datasets. 12 dimensions are necessary to capture at least 50% of information, 22 dimensions to have 75%, and 48 for 100% in 1931. In 1935, 13 dimensions are necessary to capture at least 50% of information, 25 dimensions to retain 75%, and 48 for 100%. The graphs of variance and cumulative variance below focus on the 10 first dimensions.
Graph of Variables
The first set of graphs shows how well the variables are projected on the two first dimensions. Active variables are represented in red, supplementary variables in green.
The graph delineates four main groups of variables in 1931:
- Variables that are equally projected on the two dimensions : number of pages, publisher’s nationality (over 0.5), language (0.25-0.5).
- Variables that are clearly better projected on the first dimension than on the second (publisher’s profile, periodicity, variables related to the format and layout)
- Variables that are clearly better projected on the second dimension (circulation/audited)
- Variables that are poorly projected on the two dimensions (founded, province).
Similarly, we can delineate four groups of variables in 1935:
- Variables that are well projected on the two dimensions: number of pages, publisher’s profile and nationality (close to 0.75 on the first, close to 0.5 on the second)
- Variables that are better projected on the first dimension than they are on the second (language, periodicity, size of page)
- Variables that are better projected on the second dimension (number of columns)
- Variables that are poorly projected on the two dimensions (audited, founded, circulation, province).
From these preliminary observations, we shall beware of over-interpreting poorly projected variables (related to circulation) and supplementary variables (province, founding period) which do not participate in the analysis.
Graph of MCA
The second set of graphs shows how well the variable categories are projected on the two first axes. On the visualizations below, the size of points is proportionate to their contribution. Each variable is represented by a distinct color. Note: All variables on these plots are considered as active.
In the 1931 dataset, the first dimension clearly opposes Chinese-language daily broadsheets (on the left) and foreign-language (English) small-size (octavo) periodicals with less frequent periodicity (monthly, quarterly, annual) (on the right). The latter group refers to earlier, pioneering publications (1829-1903), whereas the former includes the most recently established periodicals (1928-1935). Furthermore, we can see a gradient from the most to the least frequent publications as we move along the x axis from left to right.
The second dimension further separates foreign commercial publishers (newspaper, newsgroup, private entreprises) above and Chinese publishers (organization, academic, publishing house) below. As we move downward, we see a gradient from less (10-50 pages) to more substantial periodicals (100-500 pages). The former group is associated with weeklies, the latter with monthlies.
Similarly, in 1935, the first dimension clearly opposes daily newspapers on the left with less frequent periodicals on the right (weekly, monthly, quarterly, annual). As we move along the axis from left to right, we see a gradient from the most frequent to the least frequent publications. Moreover, the first dimension clearly separates Chinese-language newspapers on the left from foreign-language periodicals on the right. It also dissociates pioneering English publications on the right (1829-1903) from the most recent ones, mostly Chinese, on the left (1928-1935). As we move from left to right, we see another gradient from large-size newspapers with few pages (broadsheets) to small-size (octavo) but lengthy periodicals. As in 1931, the second dimension further separates commercial publishers (newspaper, newsgroup, private enterprises) of foreign origin (mostly British and American) above, from Chinese publishers with diverse profile below (organization, official, publishing house), principally dealing with monthlies and less frequent periodicals. As we move downward along the second dimension, we also see a gradient from more complex layout (larger number of columns) associated with Japanese and minor language newspapers, to simpler layout (fewer columns) associated with less frequent, book-modeled periodicals (monthlies, quaterlies, annuals).
Graph of individuals
The set of graphs below illustrates the periodicals colored according to their periodicity and their language:
Based on these MCA, we further apply Hierarchical Clustering on Principal Component (HCPC) to identify clusters of periodicals based on shared variables or groups of variables. HCPC is a clustering approach that combines the three standard methods used in multivariate data analyses (Husson, Josse, and J. 2010):
- Principal component methods (PCA, CA, MCA, FAMD, MFA)
- Hierarchical clustering, used for identifying groups of similar observations in a dataset.
- Partitioning clustering, particularly the k-means method, , used for splitting a data set into several groups.
In 1931, the partition is determined by (by decreasing importance): the number of pages, the size of page, publishers’ profile and nationality, periodicity, layout (number of column), language, year of establishment, and circulation data (audited). The place of publication (province) is not significant in the partition. In 1935, it is mostly characterized by (by decreasing importance): the number of pages, publishers’ nationality and profile, the periodicity, the format (size of page), the language, the layout (number of columns), circulation data and the date of founding. The place of publication (province) is not significant in the partition.
On this basis, the algorithm detected 4 clusters in 1931 and 3 in 1935.
In 1931 :
- Chinese daily broadsheets with large but few pages, whose publisher remained unknown (e.g. L’impartial, Hongkong Times Evening News, Public News, South Daily Voice, Commercial Press)
- English-language periodicals published by British independent entrepreneurs or newsgroups (e.g. China Mail, Shanghai Times, South Central Evening Post, Current News, Harbin Commercial Post)
- Less frequent periodicals with no information on their format and circulation (e.g. Travelers’ Gazette, Oriental Traveler, Life Daily News)
- Chinese monthly magazines published by Chinese publishing houses or organizations, of small size (octavo) but substantial length (50 to 500 pages) (e.g. Movie Monthly, Juvenile Student, Modern Student, K.C. Medical Journal, Construction of China)
The three clusters in 1935 include:
- The bulk of Chinese daily broadsheets with few pages and circulation figures from publishers’ statement (Time Daily News, Haimen Daily News, PEople’s Livelihood Daily News, Wenchow Daily News…)
- Small-size but substantial Chinese periodicals (especially monthlies) published by non-professional publishers (organizations, official and academic publishers) (Weekly Critic, Commercial Engineer, United Pictorial, Health Weekly, Medical Questionnaire).
- English-language periodicals, especially from British publishers (General Public’s Daily News, Strength of the South Daily Press, South Central News, Shanghai Builder).
The two MCA support our previous findings and highlight correlations we could not see or strongly demonstrate based on separate analyses. In particular, the graphs show the polarization between large-size daily newspapers with multiple columns (broadsheets) and less frequent periodicals of smaller size but more substance and simpler layout (fewer columns). MCA also highlights the exponential growth of the Chinese native press during the Republic, in contrast to the pioneering British periodicals of the late imperial years. This first series of MCA suggests we should rearrange certain variables to obtain more significant results. In future experiments, it is suggested to discard the variables with too many missing values (format, layout, publisher) and focus on the best, most consistently documented periodicals.
Other options for improving these preliminary MCA include:
- Exploring additional dimensions to account for variables that are poorly represented on the two first axes (founding period, province)
- Recoding variables with too many different values (province) to create analytically more meaningful variables (for instance, socio-economic regions);
- Focus on Chinese periodicals and discard minor language periodicals that create huge distorsions;
- Focus on numerical variables and rely on Principal Component Analysis (PCA) to examine correlations between the number of pages (length), the size of pages (format), the number of columns (layout) and the circulation figures.
In the next essay we propose to experiment with the last option (PCA).
Bourdieu, Pierre. La noblesse d’Etat: grandes écoles et esprit de corps. Paris: Minuit, 1989.
Husson, François, and Julie Josse. “Multiple Correspondence Analysis.” In Visualization and Verbalization of Data. Chapman and Hall/CRC, 2014.
Husson, François, J. Josse, and Pagès J. 2010. “Principal Component Methods – Hierarchical Clustering – Partitional Clustering: Why Would We Need to Choose for Visualizing Data?” Unpublished Data. http://www.sthda.com/english/upload/hcpc_husson_josse.pdf.
Kassambara, Alboukadel. “HCPC – Hierarchical Clustering on Principal Components: Essentials”. 2017. http://www.sthda.com/english/articles/31-principal-component-methods-in-r-practical-guide/117-hcpc-hierarchical-clustering-on-principal-components-essentials/
Rossier, Thierry. “Prosopography, Networks, Life Course Sequences, and so on. Quantifying with or beyond Bourdieu?” Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 144, no. 1 (October 1, 2019): 6–39. https://doi.org/10.1177/0759106319880148.
- Bourdieu, Pierre. La noblesse d’Etat: grandes écoles et esprit de corps. Paris: Minuit, 1989.