• Dimension reduction refers to transforming data with many related variables into data with fewer variables for the purpose of visualization and analysis.

• Common methods we will discuss today:
  • Principal Components Analysis (PCA)
  • Factor Analysis (FA)
  • Multidimensional Scaling (MDS)

• Participants will be familiar with common techniques for dimension reduction
• Participants will understand the similarities and differences among the presented techniques
• I will use case-studies and examples to introduce methods while keeping mathematical details to a minimum.

• In a principal components analysis we create a new set of uncorrelated variables which are linear combinations of the original variables. These new variables - the principal components - are chosen sequentially to maximize the variance explained.

• In a factor analysis we look for a small number of latent factors which explain the correlation structure among the original variables.

• Multidimensional scaling is a technique for converting a matrix of pairwise distances into a low-dimensional map that preserves the distances as well as possible.

• Our first toy example uses data on 145 adults from three groups: controls, 'chemical' diabetics, and overt diabetics.
• Here is a correlation matrix for the three variable, glucose, insulin, and steady state plasma glucose (sspg):

##         glucose insulin  sspg
## glucose    1.00    0.96 -0.40
## insulin    0.96    1.00 -0.35
## sspg      -0.40   -0.35  1.00

• Here is a scatter plot matrix.

• The high correlation between insulin and glucose indicates that most of the variation in the data occurs along two directions.

• The goal of PCA is to find the directions of maximum variation.

• Consider a data matrix \(D\) consisting of continuous variables.
  Mathematically, PCA solves the following problem:

\[ w_1 := \arg\max_{||w||=1} \textrm{var}(Dw) = \arg\max_{||w||=1} w'X'Xw \]

\[ w_k := \arg\max_{||w||=1} \textrm{var}(D_kw) \]

where \(D_k := D - \sum_{j<k} Dww'\) is the residual relative to the first \(k-1\) components.

• In practice, this is accomplished using:
  • The eigen-decomposition of the covariance matrix cov(\(X\)) or correlation matrix cor(\(X\))
  • The singular value decomposition of \(X\)

• This is the eigen-decomposition of a (symmetric) matrix: \[ \Sigma = \Gamma \Lambda \Gamma' \]
• The eigenvectors \(\Gamma\) (loadings) can be viewed as weights for creating new orthogonal variables. from the original variables: \(D_{new} = D\Gamma\).
• The eigenvalues for each component are proportional to the explained variance.

• The eigenvalues for each component are proportional to the explained variance.
• The diabetes correlation matrix has eigenvalues: 2.196, 0.769, 0.035

• Each of the new components is a linear combination of the original variables
• These are the weights for the first two variables:

• After dimension reduction, we can create a scatter plot of the new variables.

• You will often see the component loadings and new variables portrayed as a 'biplot'.

• In the diabetes example, we use the eigen-decomposition of the correlation matrix because the have different levels of natural variation.
• Working with correlations is equivalent to first transforming all variables into z-scores.
• This means that each of the original variables is weighted equally.
• When the original variables are on the same scale, it can make sense to use the eigen-decomposition of the covaraince matrix so that variables with higher variance are weighted more heavily.

• In this example, the variables are eight genes (measured by qPCR) known to be epithelial markers.
• The samples are the results of an assay for capturing circulating tumor cells (CTCs) in men with metastatic prostate cancer.
• The goal of this analysis is to use data from positive and negative controls to create a predictive model for identifying samples with CTCs present.

• Here are the data as a scatter plot matrix:

• In this example, I chose not to scale the variables since the values were already on a normalized scale.
• Here are the results, after flipping the sign for both components:

• In this case, we can use the first component as an 'epithelial score' with weights:

• In the actual application, this score is used to classify which samples contain CTCs.

• Factor analysis is similar to PCA, but takes a different point of view.
• Rather than looking for directions of maximum variation, in a factor analysis we look for a small number of latent factors to explain the relationships among the observed variables.

• For example, we might think grades in 5 high school subjects are primarily explained by students' reading comprehension and quantitative preparation.
  

• Consider a data matrix \(D\) with \(p\) variables (columns) that have been centered.
• In a factor analysis, we look for a small number \(k\) of latent factors \(F = [F_1, \dots, F_k]\) such that: \[ D \sim LF + \epsilon \]
• The loading matrix \(L\) has \(p\) rows and \(k\) columns. Each row of \(L\) explains how one of the original variables is related to the latent factors.
• The matrix of factor scores \(F\) has \(n\) rows and \(k\) columns meaning that each observation has a unique set of factor scores.
• We usually assume the factors \(F\) are uncorrelated and have mean zero, and that \(F\) and \(\epsilon\) are independent.

• In a factor analysis, we look for a small number \(k\) of latent factors \(F = [F_1, \dots, F_k]\) such that: \[ D \sim LF + \epsilon \]

• Another way to look at this is in terms of covariance matrices, \[ \textrm{cov}(D) = LL' + \Psi \] with \(\Psi\) describing the unique or unexplained variance for each original variable.

• For diabetics, blood sugar control has important health consequences.
• Continuous glucose monitors are instruments that measure someones glucose levels at regular intervals (~5 min).

• Many 'metrics' for glycemic control have been proposed for mapping this time series to a single summary statistic.

• The following example works with a collection of metrics computed on baseline data from a JDRF clinical trial (n=443) that have been normalized using Box-Cox transformations.

• A heat map of the correlation matrix shows there are 2-4 distinct groups of metrics.

• We will begin with a model using three factors.
• Here are the loadings (L):

  	Factor1	Factor2	Factor3
  DySF	-0.018	0.060	0.784
  Mean	0.920	-0.355	0.114
  SD	0.931	0.208	0.234
  SD.Slope	0.554	0.201	0.781
  Time.Hyper	0.942	-0.283	0.097
  Time.Hypo	-0.076	0.976	0.137
  AUC.Hyper	0.877	-0.057	0.102
  AUC.Hypo	0.126	0.871	0.126
  CONGA.4	0.876	0.252	0.329
  MODD	0.909	0.230	0.238
  ADRR	0.722	0.433	0.468
  HBGI	0.966	-0.197	0.149
  LBGI	-0.134	0.974	0.122
  M.Value	0.296	-0.198	-0.033
  MAGE	0.921	0.233	0.226
  MAG	0.518	0.256	0.772
  GRADE	0.959	-0.110	0.169

• Recall: Metric \(\sim ~ L_1F_1 + L_2F_2 + L_3F_3\)

• For each variable, the squared loadings tell us the percent of variance explained by each factor.

• The overall variance explained by each factor is simply the average across variables.

  	Factor1	Factor2	Factor3
  Variance Explained	52.3	20.4	14.4
  Cumulative Variance	52.3	72.7	87.1

• We can use the amount of variance explained by each factor to select how many to retain.
• There are goodness of fit test as well, but more appropriate when designing a single measurement scale.
  

• Adding a third factor increases the explained variance by 15%, while a fourth factor explains less than 2% of additional variance.

  	Factor 1	Factor 2	Factor 3	Factor 4
  2	56.7	79.2
  3	52.3	72.7	87.1
  4	52	72.5	86.4	88.9

• A bar chart of the explained variance (squared loadings) allows us to visualize the relations among the variables and latent factors.

• It is also useful to plot the raw loadings in factor space.

• The latent factor scores can be used in downstream analyses.

• Multidimensional scaling is a technique for converting a matrix of pairwise dissimilarities into a low-dimensional map that preserves the distances as well as possible.

• The first step in any MDS is choosing a similarity measure. Often this will be a metric or distance, i.e.:
  • Euclidean Distance for continuous variables
  • Manhattan distance or Jacaard dissimilarity for binary variables

• Similarity measures can often become dissimilarity measures by inverting \(x \to 1/x\) or subtracting from 1 \(x \to 1-x\).

• Consider a matrix of dissimilarities \(D = \{d_{ij}\}_{i,j}\)
• Metric MDS finds new coordinates \(X = \{(x_{i1}, x_{i2})\}_i\) that minimize the "stress" or "strain", \[ \sum_{i,j} [d_{ij} - ||x_i - x_j||^2]^{1/2}. \]

• The steps to perform a classical MDS are:
  • Obtain a matrix of squared pairwise dissimilarities
  • Double center this matrix by subtracting the row/column mean from each row/column
  • Compute the eigen decomposition of the double-centered dissimilarities
• In metric MDS a (possibly non-euclidean) distance is used and the optimization problem is solved using an iterative majorization-maximization algorithm.
• Non-metric MDS can be used when the dissimilarities are obtained directly rather than computed from other variables or otherwise exhibit nonlinearity.

• The table below shows distances, in miles, between several major US cities.

  	Atl	Chi	Den	Hou	LA	Mia	NY	SF	Sea	DC
  Atlanta	0	587	1212	701	1936	604	748	2139	2182	543
  Chicago	587	0	920	940	1745	1188	713	1858	1737	597
  Denver	1212	920	0	879	831	1726	1631	949	1021	1494
  Houston	701	940	879	0	1374	968	1420	1645	1891	1220
  LosAngeles	1936	1745	831	1374	0	2339	2451	347	959	2300
  Miami	604	1188	1726	968	2339	0	1092	2594	2734	923
  NewYork	748	713	1631	1420	2451	1092	0	2571	2408	205
  SanFrancisco	2139	1858	949	1645	347	2594	2571	0	678	2442
  Seattle	2182	1737	1021	1891	959	2734	2408	678	0	2329
  Washington.DC	543	597	1494	1220	2300	923	205	2442	2329	0

• We can use MDS to obtain a 2-dimensional map preserving distances as well as possible.

• For interpretation, it can be helpful to change the sign on the axes.

• You can also aid in interpretation by assigning a name to each axis using subject matter knowledge.
• I recommend removing the scales as only the distances between points is meaningful.

• As a final example we will compare the defensive value of MLB shortstops from 2016.
• We will use a collection of advanced defensive metrics from www.fangraphs.com as our starting data.

##                   rGDP rGFP rPM DRS  DPR RngR ErrR  UZR  Def
## Brandon Crawford     0    3  16  19  1.4 16.2  3.7 21.3 28.0
## Francisco Lindor    -2   -2  21  17 -0.8 18.0  3.6 20.8 27.8
## Freddy Galvis        0    2   3   5  0.7  9.1  5.3 15.1 22.0
## Addison Russell     -1    0  20  19 -1.1 14.5  2.0 15.4 21.9
## Andrelton Simmons    2    0  16  18  0.6 12.9  1.9 15.4 20.8
## Jose Iglesias        2    2  -1   3  1.9  2.6  7.2 11.6 17.6

• All of these metrics are in units of 'runs', but have varying scales so we will work with z-scores.
• Here is a heat map of the transformed values:

• Our first step is computing the Euclidean distance between each pair of players using the z-scores:

• Given the distances, an MDS algorithm returns a set of coordinates which can be plotted.

• As before, it is generally helpful to use subject matter knowledge to create names or concepts for each coordinate.

• Below we look at the correlation of the first coordinate with each of the original variables.

• In this case, the first coordinate tracks overall defense value which is closely tied to scores based on range.
• The second coordinate tracks other aspects of value, primarily value from turning double plays.

• Plot aspects such as color, symbol, and marker size can be used with the new coordinate system to help tell a coherent story.

• The goal of dimension reduction is to reduce multivariate data to a smaller number of variables while maintaing and eluciditing structure or relations between variables.
• PCA and factor analysis summarize the relationships/similarities among variables using correlation or covariance.
• MDS make use of dissimilarities between the units of interest.
• In their classical forms, all methods make use of a matrix decomposition to form new variables from old.
• Often a small number of new variables suffice to account for most of the important information – this leads to dimension reduction.