The video presents a step‑by‑step methodology for choosing and applying principal component methods (PCA, MCA, FAMD, MFA) using the R package FactoMineR. It begins by prompting analysts to answer a series of diagnostic questions: whether variables form groups, the data type (contingency tables, mixed quantitative/qualitative tables, or pure quantitative tables), which elements are active versus supplementary, and the nature of active variables (continuous, categorical, or mixed). It then discusses practical considerations such as standardizing variables, handling missing values with the missMDA package, and the role of imputation. After selecting the appropriate method, the speaker explains how to interpret the results using the dimdesc function to describe dimensions and how to perform clustering on the reduced dimensions for robust group identification. The overall message is that a systematic pre‑analysis checklist, combined with FactoMineR’s functions, leads to more reliable exploratory multivariate analyses.
Chapters
The purpose of this video is to propose a methodology in exploratory multivariate data analysis and how to perform a principal component method with FactoMineR. We can suggest some questions that are important to answer before performing a principal component method. So first question: are there groups of variables? So I mean, are there different sources of information? In this case, we want to take into account these different sources of information and consider different groups of variables. So in this case, we will perform a multiple factor analysis with FactoMineR. Second question, which kind of information do we have? Do we have a contingency table? So in this case, we have rows and columns, and in a cell we have the number of occurrences that a row is present in a column. So for example, we have in rows the species and in columns we have the sites, and in the cell we have the number of times a species is present in a site. So we can also have textual data, so in rows we have some texts and in columns we have some words. So in this case, we want to perform a correspondence analysis with the CA function of FactoMineR, or if we have several contingency tables and we want to compare the information from these different contingency tables, we can perform a multiple factor analysis. Another kind of information is tables with individuals in rows and variables in columns. So in this case, there are different principal component methods such as PCA, multiple correspondence analysis, factor analysis on mixed data, or multiple factor analysis if we have groups of variables. So what are the active elements? This is an important question because The active elements will participate to the construction of the dimensions of the principal component method. So these are the only elements that are used to perform the analysis, to construct the dimensions, and the other elements that are supplementary are just used to better interpret the dimensions, to better understand the information brought by the active elements. So we can have supplementary rows or supplementary columns if we have contingency tables, or supplementary individuals, or supplementary quantitative variables or qualitative variables, or supplementary groups of variables. So these are some arguments that can be used in the principal component methods with FactoMineR. Another question is, what is the type of the active variables? So I really mean the active variables only because whatever the method with FactoMineR, we can perform a principal component analysis, for example, with continuous and categorical variables, but for the principal component analysis, the active variables must be continuous. So we need to have continuous variables to perform the PCA, and then we can have supplementary variables that are continuous and/or that are categorical. So we can also have categorical variables to better interpret the dimensions of the PCA. So if all the active variables are categorical variables we can perform a multiple correspondence analysis. And if in the active variables there are both continuous and categorical variables, we will perform a factorial analysis on mixed data. And of course, if we have groups of variables, and with groups of variables that are continuous and groups of variables that are categorical, we can perform a multiple factor analysis. If we want to perform a principal component analysis or a multiple factor analysis with continuous groups of variables, we need to answer to the question, should the variables be standardized or not? This question is important because if we standardize, it means that all the variables will contribute at the same for the construction of the dimension. And if we do not standardize, then the variables that have a higher variance will contribute more than the others. So if the units of the variables are not the same, it is necessary to standardize. But if the units of the variables are the same, we can standardize or not. It depends on the problems. So if you want to standardize the variables, you will perform a standardized PCA and use the argument scale_unit= TRUE. Another question is, are there missing values? So if there are missing values, you can use the package missMDA to first impute the dataset. So the functions in this package will impute the dataset in order to perform a principal component method on a completed dataset. So you first impute the dataset and then you can perform the usual methods, PCA, MCA, or MFA, as usual. So the imputation is done in a way that the imputed values do not contribute to the construction of the dimensions. So once you have answered to all these questions, you can perform the principal component methods, so PCA, MCA, factorial analysis on mixed data or MFA. Then you want to interpret the results. So to better interpret the results, you can use some tools. The first tool is that you can describe the dimensions using the original variables, so the variables of the dataset, and you can use the function which is named dimdesc. So this function will describe the dimensions with the continuous and the categorical variables of the dataset, both the active and the supplementary variables. It is a tool to better interpret the results. And you can also perform a clustering on the individuals and then of course describe the clusters. So performing a clustering on the results of a principal component method is interesting because the principal component methods will concentrate the information on the first dimensions, and then the last dimension can be considered as noise. So principal component methods can be seen as denoising methods, and performing a clustering using the first dimension of principal component methods will contribute to have a clustering that is more stable, more robust than performing a clustering on all the data. And it is also interesting to perform clustering on the results of a multiple correspondence analysis because you can perform the clustering on continuous variables. The dimensions of the multiple correspondence analysis are continuous, so you can perform the clustering as usual, and at the beginning you have variables that are categorical. So the Multiple Correspondence Analysis is a way to transform categorical variables in continuous variables. This is a methodology that can be used before performing principal component methods and using the FactoMineR package. So you can see the different videos on PCA, MCA, or MFA.
