The Principal Components of Anthropometry

The Principal Components of Anthropometry#

In a previous post I explored the correlations between measurements in the ANSUR-II dataset, which includes 93 measurements from a sample of U.S. military personnel. I found that measurements of the head were weakly correlated with measurements from other parts of the body – and in particular the protrusion of the ears is almost entirely uncorrelated with anything else.

A friend of mine, and co-developer of the Modeling and Simulation class I taught at Olin, asked whether I had tried running principal component analysis (PCA). I had not, but now I have. Let’s look at the results.

Click here to run this notebook on Colab.

Hide code cell content

 
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

The data#

Here’s the ANSUR data, originally downloaded from The OPEN Design Lab.

Hide code cell content

 
ansur_female = pd.read_csv("ANSURIIFEMALEPublic.csv", encoding="ISO-8859-1")
ansur_female.shape

(1986, 108)

Hide code cell content

 
ansur_male = pd.read_csv("ANSURIIMALEPublic.csv", encoding="ISO-8859-1")
ansur_male.shape

(4082, 108)

Since heights are in mm, I converted to cm.

Hide code cell content

 
height_male = ansur_male["stature"] / 10
height_female = ansur_female["stature"] / 10

Hide code cell content

 
def add_earratio(df):
    df['earratio'] = df['earprotrusion'] / df['earbreadth']

Hide code cell content

 
measurements = [column for column in ansur_female.columns if column.lower() == column]
measurements

['abdominalextensiondepthsitting',
 'acromialheight',
 'acromionradialelength',
 'anklecircumference',
 'axillaheight',
 'balloffootcircumference',
 'balloffootlength',
 'biacromialbreadth',
 'bicepscircumferenceflexed',
 'bicristalbreadth',
 'bideltoidbreadth',
 'bimalleolarbreadth',
 'bitragionchinarc',
 'bitragionsubmandibulararc',
 'bizygomaticbreadth',
 'buttockcircumference',
 'buttockdepth',
 'buttockheight',
 'buttockkneelength',
 'buttockpopliteallength',
 'calfcircumference',
 'cervicaleheight',
 'chestbreadth',
 'chestcircumference',
 'chestdepth',
 'chestheight',
 'crotchheight',
 'crotchlengthomphalion',
 'crotchlengthposterioromphalion',
 'earbreadth',
 'earlength',
 'earprotrusion',
 'elbowrestheight',
 'eyeheightsitting',
 'footbreadthhorizontal',
 'footlength',
 'forearmcenterofgriplength',
 'forearmcircumferenceflexed',
 'forearmforearmbreadth',
 'forearmhandlength',
 'functionalleglength',
 'handbreadth',
 'handcircumference',
 'handlength',
 'headbreadth',
 'headcircumference',
 'headlength',
 'heelanklecircumference',
 'heelbreadth',
 'hipbreadth',
 'hipbreadthsitting',
 'iliocristaleheight',
 'interpupillarybreadth',
 'interscyei',
 'interscyeii',
 'kneeheightmidpatella',
 'kneeheightsitting',
 'lateralfemoralepicondyleheight',
 'lateralmalleolusheight',
 'lowerthighcircumference',
 'mentonsellionlength',
 'neckcircumference',
 'neckcircumferencebase',
 'overheadfingertipreachsitting',
 'palmlength',
 'poplitealheight',
 'radialestylionlength',
 'shouldercircumference',
 'shoulderelbowlength',
 'shoulderlength',
 'sittingheight',
 'sleevelengthspinewrist',
 'sleeveoutseam',
 'span',
 'stature',
 'suprasternaleheight',
 'tenthribheight',
 'thighcircumference',
 'thighclearance',
 'thumbtipreach',
 'tibialheight',
 'tragiontopofhead',
 'trochanterionheight',
 'verticaltrunkcircumferenceusa',
 'waistbacklength',
 'waistbreadth',
 'waistcircumference',
 'waistdepth',
 'waistfrontlengthsitting',
 'waistheightomphalion',
 'weightkg',
 'wristcircumference',
 'wristheight']

Hide code cell content

 
len(measurements)

93

Hide code cell source

 
ansur_male[measurements]
abdominalextensiondepthsitting acromialheight acromionradialelength anklecircumference axillaheight balloffootcircumference balloffootlength biacromialbreadth bicepscircumferenceflexed bicristalbreadth ... verticaltrunkcircumferenceusa waistbacklength waistbreadth waistcircumference waistdepth waistfrontlengthsitting waistheightomphalion weightkg wristcircumference wristheight
0 266 1467 337 222 1347 253 202 401 369 274 ... 1700 501 329 933 240 440 1054 815 175 853
1 233 1395 326 220 1293 245 193 394 338 257 ... 1627 432 316 870 225 371 1054 726 167 815
2 287 1430 341 230 1327 256 196 427 408 261 ... 1678 472 329 964 255 411 1041 929 180 831
3 234 1347 310 230 1239 262 199 401 359 262 ... 1625 461 315 857 205 399 968 794 176 793
4 250 1585 372 247 1478 267 224 435 356 263 ... 1679 467 303 868 214 379 1245 946 188 954
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
4077 235 1353 312 216 1263 228 193 384 318 241 ... 1557 442 279 816 210 347 1016 675 163 778
4078 247 1473 336 234 1346 253 196 428 374 284 ... 1715 475 351 978 235 385 1082 896 178 873
4079 264 1394 313 227 1280 245 193 407 367 271 ... 1682 483 333 991 258 353 1011 832 178 822
4080 203 1417 327 223 1314 250 196 419 365 271 ... 1589 430 293 783 192 350 1062 731 172 837
4081 327 1523 358 226 1408 269 225 442 379 275 ... 1764 511 354 1080 294 389 1119 985 182 894

4082 rows × 93 columns

Hide code cell content

 
df = ansur_male[measurements]

Principle Component Analysis#

As you might guess from the comments below, I used ChatGPT to generate a lot of the code. It took some revision to get it working, but the head start definitely saved me some time.

Hide code cell content

 
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA


# In this example, we will first standardize the data and then perform PCA.

# Step 3.1: Create a StandardScaler step
scaler = StandardScaler()

# Step 3.2: Create a PCA step
pca = PCA(n_components=10) 

# Step 4: Create the pipeline
steps = [
    ('scaler', scaler),  # Standardize the data
    ('pca', pca)        # Perform PCA
]

pipeline = Pipeline(steps)

# Fit the pipeline to your data
pipeline.fit(df)

# Now, you can access the transformed data and principal components if needed.
transformed_data = pipeline.transform(df)  # Transformed data after PCA
principal_components = pca.components_    # Principal components (eigenvectors)

# You can also access the explained variance ratio to see how much variance is explained by each component.
explained_variance_ratio = pca.explained_variance_ratio_
explained_variance_ratio

array([0.44234507, 0.1749688 , 0.04206542, 0.03248169, 0.02284759,
       0.02117402, 0.01887345, 0.01486669, 0.01423956, 0.01333269])

Here’s a visualization of explained variance versus number of components. I lifted this from The Python Data Science Handbook.

Hide code cell content

 
ys = np.cumsum(pca.explained_variance_ratio_)
xs = np.arange(len(ys)) + 1
ys

array([0.44234507, 0.61731386, 0.65937928, 0.69186097, 0.71470856,
       0.73588258, 0.75475603, 0.76962272, 0.78386228, 0.79719496])

Hide code cell source

 
plt.bar(xs, ys)
plt.xlabel('number of components')
plt.ylabel('cumulative explained variance')

plt.savefig('ansur_pca1.png', dpi=300)
_images/6dfeec57607b54db5731eee5a03e0f6b11b2fa6cbb7cf24ed96229a504d45e38.png

With one component, we can capture 44% of the variation in the measurements. With two components, we’re up to 62%. After that, the gains are smaller (as we expect), but with 10 measurements, we get up to 78%.

Loadings#

Looking at the loadings, we can see which measurements contribute the most to each of the components, so we can get a sense of which characteristics each component captures.

Hide code cell source

 
loadings = pca.components_  # Get the loadings (eigenvectors)

# Create a DataFrame to store the loadings with feature names as columns
loadings_df = pd.DataFrame(loadings, columns=measurements)
loadings_df
abdominalextensiondepthsitting acromialheight acromionradialelength anklecircumference axillaheight balloffootcircumference balloffootlength biacromialbreadth bicepscircumferenceflexed bicristalbreadth ... verticaltrunkcircumferenceusa waistbacklength waistbreadth waistcircumference waistdepth waistfrontlengthsitting waistheightomphalion weightkg wristcircumference wristheight
0 -0.085451 -0.134108 -0.116021 -0.100365 -0.127055 -0.105897 -0.112045 -0.108305 -0.092030 -0.103360 ... -0.124746 -0.094822 -0.097267 -0.095846 -0.087184 -0.082944 -0.114016 -0.131211 -0.115607 -0.120198
1 -0.162882 0.099494 0.106095 -0.092119 0.123047 -0.037801 0.064844 -0.017398 -0.141551 -0.055853 ... -0.096331 -0.038278 -0.157748 -0.165992 -0.161084 -0.061621 0.150431 -0.126668 -0.062109 0.048192
2 -0.014415 -0.122852 0.012455 0.022595 -0.112583 0.073198 0.118903 0.059989 0.103379 -0.135286 ... -0.167965 -0.179761 -0.081045 -0.045152 -0.021020 -0.228031 -0.030049 -0.003644 0.042584 -0.216034
3 0.173835 0.015046 0.085668 -0.166709 0.009379 -0.247034 -0.099303 -0.044501 0.032602 0.021611 ... -0.000887 0.023748 0.131975 0.161440 0.181350 -0.043284 0.010540 0.055891 -0.153328 -0.019573
4 0.066929 0.015922 -0.052847 0.179214 0.008862 0.148688 0.102660 -0.291544 0.025702 0.022706 ... 0.014710 0.006024 0.053549 0.058204 0.067503 -0.048844 -0.011521 0.043992 0.064811 0.057644
5 0.068020 0.045207 0.009170 -0.079944 0.034725 -0.068642 -0.007852 -0.210508 -0.059574 0.028846 ... 0.014737 0.083359 0.029012 0.049868 0.070885 0.007065 -0.017563 0.003736 -0.105828 0.075317
6 0.093649 0.012374 0.031909 -0.108847 -0.004847 0.022886 0.039429 -0.041191 -0.045501 0.027304 ... 0.018137 0.283710 0.027891 0.070190 0.116803 0.201994 -0.123756 -0.022860 0.095987 0.009422
7 -0.052688 0.061738 -0.115124 -0.010012 0.048177 -0.005232 0.088680 -0.011161 0.095275 0.027243 ... -0.009393 0.042837 -0.063307 -0.051032 -0.029532 -0.145553 0.058143 0.006287 0.020673 0.200656
8 -0.002601 -0.041691 0.011918 0.128483 -0.038433 0.077109 0.143232 0.113983 -0.133804 0.253301 ... -0.066053 0.180681 0.109480 0.053116 0.019980 -0.059466 -0.048724 0.005821 -0.105042 -0.049366
9 0.088042 -0.015056 0.071590 -0.043072 -0.004563 -0.106276 0.033972 0.002206 -0.011959 -0.078814 ... 0.053774 0.227461 0.021029 0.072728 0.097588 0.268130 -0.129249 0.027296 -0.009578 -0.088460

10 rows × 93 columns

Hide code cell source

 
loadings_transposed = loadings_df.transpose()
loadings_transposed
0 1 2 3 4 5 6 7 8 9
abdominalextensiondepthsitting -0.085451 -0.162882 -0.014415 0.173835 0.066929 0.068020 0.093649 -0.052688 -0.002601 0.088042
acromialheight -0.134108 0.099494 -0.122852 0.015046 0.015922 0.045207 0.012374 0.061738 -0.041691 -0.015056
acromionradialelength -0.116021 0.106095 0.012455 0.085668 -0.052847 0.009170 0.031909 -0.115124 0.011918 0.071590
anklecircumference -0.100365 -0.092119 0.022595 -0.166709 0.179214 -0.079944 -0.108847 -0.010012 0.128483 -0.043072
axillaheight -0.127055 0.123047 -0.112583 0.009379 0.008862 0.034725 -0.004847 0.048177 -0.038433 -0.004563
... ... ... ... ... ... ... ... ... ... ...
waistfrontlengthsitting -0.082944 -0.061621 -0.228031 -0.043284 -0.048844 0.007065 0.201994 -0.145553 -0.059466 0.268130
waistheightomphalion -0.114016 0.150431 -0.030049 0.010540 -0.011521 -0.017563 -0.123756 0.058143 -0.048724 -0.129249
weightkg -0.131211 -0.126668 -0.003644 0.055891 0.043992 0.003736 -0.022860 0.006287 0.005821 0.027296
wristcircumference -0.115607 -0.062109 0.042584 -0.153328 0.064811 -0.105828 0.095987 0.020673 -0.105042 -0.009578
wristheight -0.120198 0.048192 -0.216034 -0.019573 0.057644 0.075317 0.009422 0.200656 -0.049366 -0.088460

93 rows × 10 columns

Hide code cell content

 
num_top_features = 5

# Iterate through each principal component
for i, col in loadings_transposed.items():
    print(f"Principal Component {i+1}:")
    
    # Sort the features by absolute loading values for the current principal component
    sorted_features = col.sort_values(key=abs, ascending=False)
    
    # Select the top N features with the highest absolute loadings
    top_features = sorted_features[:num_top_features].round(3)
    
    # Print the top correlated factors for the current principal component
    for name, val in top_features.items():
        print(-val, '\t', name)
    print()

Principal Component 1:
0.135 	 suprasternaleheight
0.134 	 cervicaleheight
0.134 	 buttockkneelength
0.134 	 acromialheight
0.133 	 kneeheightsitting

Principal Component 2:
0.166 	 waistcircumference
-0.163 	 poplitealheight
0.163 	 abdominalextensiondepthsitting
0.161 	 waistdepth
0.159 	 buttockdepth

Principal Component 3:
0.338 	 elbowrestheight
0.31 	 eyeheightsitting
0.307 	 sittingheight
0.228 	 waistfrontlengthsitting
-0.225 	 heelbreadth

Principal Component 4:
0.247 	 balloffootcircumference
0.232 	 bimalleolarbreadth
0.22 	 footbreadthhorizontal
0.218 	 handbreadth
0.212 	 sittingheight

Principal Component 5:
0.32 	 interscyeii
0.292 	 biacromialbreadth
0.275 	 shoulderlength
0.273 	 interscyei
0.184 	 shouldercircumference

Principal Component 6:
-0.34 	 headcircumference
-0.32 	 headbreadth
0.317 	 shoulderlength
-0.279 	 tragiontopofhead
-0.263 	 interpupillarybreadth

Principal Component 7:
0.373 	 crotchlengthposterioromphalion
-0.321 	 earbreadth
-0.3 	 earlength
-0.284 	 waistbacklength
0.254 	 crotchlengthomphalion

Principal Component 8:
0.472 	 earprotrusion
0.342 	 earlength
0.217 	 crotchlengthposterioromphalion
-0.201 	 wristheight
-0.197 	 elbowrestheight

Principal Component 9:
-0.299 	 tragiontopofhead
0.292 	 crotchlengthposterioromphalion
-0.253 	 bicristalbreadth
-0.226 	 shoulderlength
0.193 	 bitragionchinarc

Principal Component 10:
0.401 	 earbreadth
0.358 	 earprotrusion
-0.268 	 waistfrontlengthsitting
0.236 	 earlength
-0.227 	 waistbacklength

I won’t explain all of the measurements, but if there are any you are curious about, you can look them up in The Measurer’s Handbook which includes details on “sampling strategy and measuring techniques” as well as descriptions and diagrams of the landmarks and measurements between them.

  • Not surprisingly, the first component is loaded with measurements of height. If you want to predict someone’s measurements, and can only use one number, choose height.

  • The second component is loaded with measurements of girth. Again, no surprises so far.

  • The third component seems to capture torso length. And that makes sense, too. One you know how tall someone is, it helps to know how that height is split between torso and legs.

  • The fourth component seems to capture hand and foot size (with sitting height thrown in just to remind us that PCA is not obligated to find components that align perfectly with the axes we expect).

  • Component 5 is all about the shoulders.

  • Component 6 is all about the head.

After that, things are not so neat. But two things are worth noting:

  • Component 7 is mostly related to the dimensions of the pelvis, but…

  • Components 7, 8, and 10 are surprisingly loaded up with ear measurements.

As we saw in the previous article, there seems to be something special about ears. Once you have exhausted the information carried by the most obvious measurements, the dimensions of the ear seem to be strangely salient.

Using the pca module#

I found this module and decided to experiment with it.

Hide code cell content

 
model = pca(normalize=True, n_components=10)
model.fit_transform(df);

[pca] >Extracting column labels from dataframe.
[pca] >Extracting row labels from dataframe.
[pca] >Normalizing input data per feature (zero mean and unit variance)..
[pca] >The PCA reduction is performed on the [93] columns of the input dataframe.
[pca] >Fit using PCA.
[pca] >Compute loadings and PCs.
[pca] >Compute explained variance.
[pca] >Outlier detection using Hotelling T2 test with alpha=[0.05] and n_components=[10]
[pca] >Multiple test correction applied for Hotelling T2 test: [fdr_bh]
[pca] >Outlier detection using SPE/DmodX with n_std=[3]

Hide code cell source

 
model.plot();
_images/6a1b2e1e422c62752594c9fb4d51ddf71c9d2b11ba6b56d9c8b7003b4c9d1ddf.png

Unfortunately, the most interesting plots don’t seem to work well with the number of dimensions in this dataset. I’d be curious to know if they can be tuned to work better.

Hide code cell source

 
model.biplot3d();

[scatterd] >INFO> Create scatterplot

[pca] >Plot PC1 vs PC2 vs PC3 with loadings.
_images/ee7f9a1fb98faf7a67f8aa989ddf5ceb7c0266a41a1b3a180b59e8c318259b95.png

Copyright 2023 Allen Downey

The code in this notebook and utils.py is under the MIT license.

Contents