Multivariate methods

Stata 10 includes many new methods of multivariate analysis, and many existing methods have been greatly expanded.

• Stata now performs several discriminant analysis techniques, including linear, quadratic, logistic, and kth-nearest-neighbor discrimination. Postestimation tools make obtaining classification tables, error rates, and summary statistics a snap.
• Stata now performs modern as well as classical multidimensional scaling (MDS), including metric and nonmetric MDS. Available loss functions include stress, normalized stress, squared stress, normalized squared stress, and Sammon. Available transformations include identity, power, and monotonic.
• Stata can now perform multiple or joint correspondence analysis, allowing you to explore the relationship among categorical variables by projecting onto reduced spaces that may correspond to unobserved factors.
  

Here are all the details...

1. New estimation commands discrim and candisc provide several discriminant analysis techniques, including linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), logistic discriminant analysis, and kth-nearest-neighbor discriminant analysis. See [MV] discrim and [MV] candisc.
   
2. Existing estimation commands mds, mdslong, and mdsmat now provide modern as well as classical multidimensional scaling (MDS), including metric and nonmetric MDS. Available loss functions include stress, normalized stress, squared stress, normalized squared stress, and Sammon. Available transformations include identity, power, and monotonic. mdslong also now allows aweights and fweights, and mdsmat has a weight() option. See [MV] mds, [MV] mdslong, and [MV] mdsmat.
3. New estimation command mca provides multiple correspondence analysis (MCA) and joint correspondence analysis (JCA); see [MV] mca and [MV] mca postestimation. You can use existing command screeplot afterward to graph principal inertias; see [MV] screeplot.
4. Concerning existing estimation command ca (correspondence analysis),
   
   
   1. ca now allows crossed (stacked) variables. This provides a way to automatically combine two or more categorical variables into one crossed variable and perform correspondence analysis with it.
   2. ca’s existing option normalize() now allows normalize(standard) to provide normalization of the coordinates by singular vectors divided by the square root of the mass.
   3. ca’s new option length() allows you to customize the length of labels with crossed variables in output.
   4. New postestimation command estat loadings, used after ca and camat, displays correlations of profiles and axes.
   5. Existing postestimation command cabiplot has new option origin that displays the origin within the plot. cabiplot also now accepts originlopts(line_options) to customize the appearance of the origin on the graph.
   6. Existing postestimation commands cabiplot and caprojection now allow row and column marker labels to be specified using the mlabel() suboption of rowopts() and colopts().
   See [MV] ca and [MV] ca postestimation.
5. Existing commands cluster, matrix dissimilarity, and mds now allow the Gower measure for a mix of binary and continuous data; see [MV] measure_option.
6. Existing command biplot has new options. dim() specifies the dimensions to be displayed. negcol specifies that negative column (variable) arrows be plotted. negcolopts(col_options) provides graph options for the negative column arrows. norow and nocolumn suppress the row points or column arrows. See [MV] biplot.
7. New postestimation command estat rotate after canon performs orthogonal varimax rotation of the raw coefficients, standard coefficients, or canonical loadings. After estat rotate, new postestimation command estat rotatecompare displays the rotated and unrotated coefficients or loadings and the most recently rotated coefficients or loadings. See [MV] canon postestimation.
8. Existing commands pcamat and factormat now allow singular correlation or covariance structures. New option forcepsd modifies a matrix to be positive semidefinite and thus to be a proper covariance matrix. See [MV] pca and [MV] factor.
9. Existing commands rotate and rotatemat now refer to the “Kaiser normalization” rather than “Horst normalization”. A search of the literature indicates that Kaiser normalization is the preferred terminology. Previously option horst was a synonym for normalize. Now option horst is not documented. See [MV] rotate and [MV] rotatemat.
10. Existing command procrustes now saves the number of y variables in scalar e(ny); see [MV] procrustes.

Back to highlights