A Method to Create Body Sizing Systems

Abstract

Literature reviews show that consumers are not satisfied with the fit of garments sold in stores, primarily due to outdated sizing systems used by companies and the limited number of sizes being produced. The purpose of this article is to present a sizing system creation method, which can be applied on body measurement data to improve the overall fit of garments. Data from the SizeUSA study was used. The sizing system creation process included natural log-transformation, principle component analysis, multivariate linear regression, size range determination, and measurements calculation. The created sizing system was then compared with ASTM D5585-11e1. Analysis of the results showed that the method used to create the sizing system was reliable and repeatable. This was an important research effort for apparel manufactures, as it demonstrated a method to create a sizing system that is flexible and can be altered to fit target consumer groups!

Keywords

sizing anthropometry American Society of Testing and Materials (ASTM)principle component analysis natural log-transformation multivariate

Linear Regression

Sizing systems originated from the development of ready-to-wear (RTW) and mass production. There are two types of sizing systems, namely, the body sizing system and the garment sizing system. A body sizing system is developed from close-to-body measurements, while a garment sizing system begins with a body sizing system and is tested on prototypes for fitting. Scholars began studies of standard body sizing systems in the first half of the 20th century, and they bloomed in the late 20th century. A good sizing system benefits both consumers and manufacturers (Chen-Yu, Williams, & Kincade, 1999). However, researchers have shown that consumers were not satisfied with the fit of apparel (Bickle, Burnsed, & Edwards, 2015; Giovis, 2007; Murray, 2016). One reason for this issue is that the grade rules companies have used are outdated (Ashdown & DeLong, 1995; Ashdown & Loker, 2010; Shin & Istook, 2007; Workman & Lentz, 2000). Another reason is that the number of sizes provided by RTW companies is limited, especially for the plus-size market (The NPD Group, 2017). It is important to develop an updated and more flexible sizing system.

The clothing industry has been turning toward mass customization. A goal of mass customization is to give a better fit at a low cost and fast speed (Loker, 2007). A strategy for this is mass customized sizing. There are three ways of making customized fit patterns: (a) three-dimensional to two-dimensional flattening (Hinds, McCartney, & Woods, 1991; Kwong, 2004; Okabe, Imaoka, Tomiha, & Niwaya, 1992; Yang & Zhang, 2007); (b) automatic two-dimensional drafting (Kang & Kim, 2000); and (c) altering from standard sizes (Istook, 2002; Song & Ashdown, 2012). For the final option, it is critical to have a good method of creating standard sizes based on anthropometric data.

The development of Kinect technology (a line of motion-sensing input devices developed by Microsoft that can measure distance and therefore form 3D models) has raised the potential of breaking the space limitation of three-dimensional scanning. In the future, body measurement data could be more accessible than today, which will bring opportunities as well as challenges. How to analyze the data and make full use of it is a question that the apparel industry has to face in the near future.

Consumer dissatisfaction with fit has been well documented, and measurements captured by 3D body scanning techniques have become more accessible (Bickle et al., 2015; Giovis, 2007; Murray, 2016). This raises questions about how a sizing system might be better developed to meet consumers’ needs. Therefore, this research was conducted to explore and develop a method for the creation of a sizing system that implements more current anthropometric data.

Literature Review

The purpose of creating good sizing systems is to help apparel companies produce RTW clothes that can fit potential consumers as well as possible and allow them to make profits at the same time (Bickle et al., 2015; Giovis, 2007; Murray, 2016; Petrova, 2007; The NPD Group, Inc., 2017). The earliest sizing systems were created based on the tailor’s own experience and first appeared in pattern books by the end of the 18th century (Aldrich, 2000). These early sizing systems used the proportional scaling method.¹ The demand for army clothing brought by wars during the 19th and 20th centuries accelerated the production of RTW clothing and forced the improvement in size design (Aldrich, 2007). The proportional scaling method did not accurately meet the needs of the market. During the second half of the 20th century, sizes began to be generated from body measurements using statistical methods (Aldrich, 2007). Since then, anthropometric surveys have been conducted, and sizing standards have been published in different countries. The ASTM (formerly the American Society of Testing and Materials) Committee D13 was formed in 1914 and is the industry agency that publishes national voluntary sizing standards in the United States (www.astm.org/COMMITTEE/D13.htm).

Because ease is closely related to the clothing style and the consumer’s personal preference, recently published sizing systems are body sizing systems rather than garment sizing systems. In patternmaking, ease is the extra amount that is added to or reduced from (often known as negative ease) the exact measurements of a body to allow for design variation, movement, and comfort fit (Lo, 2011). Creating a body sizing system often follows the process of collecting and preparing anthropometric data, picking control variables, deciding a size range for each control dimension, developing a subgroup population based on intervals and size numbers, calculating secondary dimensions, and, finally, labeling the sizing system (Petrova, 2007). The following discussion addresses methods used in data preparation, selection of control variables, determination of proportions, setting intervals, and calculation of secondary dimensions.

Preparing Data

Anthropometric surveys are designed and conducted to collect body data. The number of large-scale anthropometric surveys ever conducted worldwide is countable because of their high cost ([TC]², 2004). Because anthropometric surveys have many steps and the efforts of many people are required over a long period, there are usually missing values and errors within the data. Listwise deletion is one way of dealing with missing values. In this method, only cases with available data on each variable are analyzed. Listwise deletion is simple, but because anthropometric data always contain many variables, the proportion of deleted cases can be large, which can reduce statistical significance (Humphries, n.d.). Hsu and Wang (2005) applied the listwise deletion method and reduced the sample size from 610 to 590. Another way of dealing with missing values is to replace the missing value with the mean of the variable. This is called the series mean method. Esfandarani and Shahrabi (2012) used the series mean method on data preparation for a suit sizing system.

Control Variables

The control variables were the body measurements used to classify size groups (Petrova, 2007). These measurements best describe the body size for each individual. Control variables should be easy to measure and have little or no intercorrelation. They may change when a sizing system is designed for different styles. Principal component analysis (PCA) is a statistical analysis method usually used to find the control variables in a study. A method of factor analysis, its purpose is to replace all data with a smaller number of uncorrelated variables. Guan et al. (2012) used PCA to represent 12 body measurements with three principal components (PCs) in a truck driver anthropometric study. Sometimes, PCs with small variance (also denoted as the eigenvalue) may be as important as those with large variance (Jolliffe, 1982). In Esfandarani and Shahrabi’s (2012) study of the suit sizing system, different numbers of PCs were tested, and it was found that two PCs were the best choice in their case. PCs extracted from the PCA can sometimes be represented by body measurements, which in this case are called the principal measurements. When selecting the principal measurements for the PCs, practical situations must be considered. For example, even though hip circumference was more closely related to the girth factor, Hsu and Wang (2005) used waist circumference as the principal measurement to represent the girth factor (the PC), because waist is easier to measure. Newcomb (2006) also used waist circumference to represent her first PC, with the consideration of predicting both the top and bottom parts of the body. Varimax rotation can be used to provide independence among PCs (Song & Ashdown, 2011). In statistics, a varimax rotation is used so variables can associate with as few factors as possible, and the results of PCA can be more easily understood (Tanagra, 2009).

Proportion

A sizing system is often designed to cover a certain range of people, called the size range. The proportion of the population covered by the sizing system is called the accommodation rate of the sizing system (Petrova, 2007). Not all sizes within a sizing system will be produced. Only garments in sizes that can represent a majority of target consumers will be manufactured. The actual production accommodation rate is between 65% and 85% (Petrova, 2007). Assuming the complete sizing system can cover the population very well, because of the production accommodation rate, at least 15–35% of the population may not be able to find the right size. This opens up the market place for customization. Percentiles of the control variables or the PCs can be used to set the size range. Song and Ashdown (2011) used the 90th percentile of body mass index to categorize different body shapes. However, percentile values of different variables are not additive. A person with a 95-percentile height does not necessarily have a 95-percentile waist circumference. Guan et al. (2012) used 95th percentiles for three orthogonal PCs.

Intervals

The control variables are divided into small ranges for each size. The increments are called the size intervals (Petrova, 2007). The interval can be either a constant or a variable. The size number is determined by the size range and the intervals (Petrova, 2007). For a garment sizing system, the value of the interval depends on the absolute value of the control variables, the fabric properties, and the tolerance level of consumers for the control variables (Ashdown & DeLong, 1995; Petrova, 2007). Mpampa, Azariadis, and Sapidis (2010) used the intervals of the drop value between chest girth and waist girth to classify people into seven body types. Gupta and Gangadhar (2004) used the standard deviation as the interval for height. While intervals are set according to convenience, common practices, and fit consideration in these examples, they can also be calculated using a statistical procedure such as K-means cluster analysis (Song & Ashdown, 2011). IBM Corporation (2012) defines K-means cluster analysis as a clustering tool designed to assign subjects to a predefined number of clusters where similarities are determined by a set of specified variables. However, the K-means cluster center is very sensitive to the initial cluster centers (Bradley & Fayyad, 1998). The number of sizes has to be predefined. Sometimes a researcher has to run the K-means analysis multiple times with different initial centers and size numbers before a satisfactory result is achieved. Song and Ashdown (2011) did K-means cluster analysis on three PCs and two z-scores (a measure of how many standard deviations below or above the population mean) with the number of clusters set to two, three, and four. The three clusters result was picked as the best one. Discriminant analysis (DA) is sometimes implemented to find discriminants among variables that provide maximum separation between clusters (Taylor, 1998).

Secondary Dimensions

Patterns cannot be drawn only with the control variables. Secondary dimensions are necessary to describe the detail of a body shape, and they have strongly correlated with the control variables (Petrova, 2007). Because of the strong correlation, secondary dimensions can be estimated by control variables. The method often used to calculate secondary dimensions is called linear regression. Newcomb (2006) used linear regression to generate functions that can predict bust, high hip, hip, upper arm, and thigh max with waist circumference.

Objectives

Based on the literature review, consumer dissatisfaction with fit has been well documented (Bickle et al., 2015; Giovis, 2007; Murray, 2016). This raises questions about how a sizing system might be better developed to meet consumers’ needs. The primary objective of the research was to develop a method that can create a sizing system based on anthropometric data. In order to do that, the following needed to be defined: (a) key measurements, (b) control variables, (c) secondary dimensions, and (d) intervals between sizes. Once the system was created, it was then tested against ASTM D5585-11e1 (ASTM International, 2011) to check its performance.

ASTM D5585, titled “Standard Tables of Body Measurements for Adult Female Misses Figure Type, Size Range 00-20,” is a sizing standard developed by ASTM (formerly the American Society of Testing and Materials). The first version was published in 1995. The version used in this research (ASTM D5585-11e1) is a revision published in 2011. The sizing table was generated based on data published by the U.S. Department of Commerce, the CAESAR study, the SizeUSA study, industry studies, and documentation from Alvanon Inc. (ASTM International, 2011).

The research questions can be further defined as: (a) what are the key measurements, control variables, and secondary dimensions for a sizing system; (b) how are the intervals between each size created; and (c) how does the created sizing system work, comparing to ASTM D5585-11e1 (ASTM International, 2011).

Method

SizeUSA data collected by [TC]² from 2002 to 2003 was used in the study. It is the most recent large-scale American anthropometric data set. The data set has information of 6,310 female subjects. Two subjects were eliminated due to measurement errors. Therefore, 6,308 subjects were studied. In order to verify the accuracy of the sizing results, the data set of 6,308 subjects was randomly divided into two groups. Approximately 50% of the subjects were placed in each group. Group one was the training group, and group two was the testing group.

Choose Key Measurements

A combination of ASTM and ISO measurements were available in the SizeUSA data set. The selection of key measurements was made referring to measurements used in the past sizing standards. That is, if a measurement in SizeUSA was also defined in either ASTM 5219-15 (ASTM International, 2015) or ISO 8559 (ISO/TC 133, 1989) standard, then it was considered as a key measurement. In order to normalize measurements’ distributions, natural log transformations were done on height, girth, and width measurements.

Generate Control Variables

Control variables are used to guide the distribution of the sizes. A sizing system usually has one or two control variables. PCA in SPSS was conducted in this research to convert the large number of key body measurements to a smaller number of PCs, upon which the arranging of sizes would be based. SPSS is IBM’s commercialized integrated product that offers different kinds of statistical analysis tools (www.ibm.com/us-en/marketplace/statistical-analysis-and-reporting). Varimax rotation was applied. A target number of PCs had to be entered before PCA was conducted. To help determine the number of PCs, a pre-PCA was made before the final PCA. The decision was made with the consideration of the total number of sizes and the percentage of variance explained by the PCs. A coefficient matrix was calculated to check the correlation between different measurements. A scree plot was drawn to support the decision. Because PCA uses the listwise deletion method, descriptive analysis was made to check the numbers of missing values. To reduce the ratio of deletion, measurements with a lot of missing values were excluded from the final PCA.

Calculate Secondary Dimensions

Multivariate linear regressions (MLR) were conducted, with selected PCs as independents, to predict and calculate the secondary dimensions (the key measurements). The MLR assumed that the dependent variables could be calculated by a linear combination of all independent variables and a constant. Regressions were done on each of the secondary dimensions. The MLR process was used to estimate values of measurements for each size. It was not used to prove the existence of linear correlation between the dependent and the independent variables.

Determine Size Intervals

A common strategy used to decide size intervals is to use a larger interval between larger sizes and a smaller interval between smaller sizes. To understand it, if you consider a 1-inch increment on waist circumference, a more obvious change will be observed on a person with a 25-inch waist compared to a person with a 44-inch waist. The reason is because the base values of their waist circumference are different. K-means cluster analysis was tried for sorting sizes using the selected PCs together and separately. However, the resulting sizes were randomly distributed, and it was hard to control size centers. Therefore, it was concluded that the K-means cluster analysis was not a suitable method for this research. Mpampa et al. (2010) distributed sizes evenly along each control variable. This method is more controllable and fits this research. Because natural log transformations were made in the data preparation stage, intervals were divided evenly. This resulted in increasing intervals for the transformed back data. Mpampa et al. (2010) used standard deviation (SD) to determine the range of control variables by setting the range of height to mean ± 2SD. This idea was also applied in this research. Several rounds of tests were conducted to help determine the number of size groups for each control variable. The predicted value of waist was compared with ASTM D5585-11e1 (ASTM International, 2011).

Finalize Sizing System

Means of the PCs were used as center values of the middle size; other sizes were arranged evenly across the cover range. Once the PC values were determined for all size centers, center values of key measurements were calculated for each size using formulas from the multilinear regression. A sizing system was then created. Steps that were followed to create the sizing system in this article are listed in Figure 1.

Figure 1.
The sizing system creation steps.

Test Performance

In order to check the performance of the created sizing system, two rounds of comparisons were made between the newly created sizing system and the sizing system from ASTM D5585-11e1 standard (ASTM International, 2011), the most recent ASTM sizing standard designed for adult female misses’ figures when the research was conducted. The ASTM standard contains two body shapes, and there are 12 sizes for each body shape (24 sizes in total). Twelve sizes were selected from the created sizing system to be compared with the ASTM standard. The purpose of these comparisons was to see how many subjects’ measurements fell within each size range and how different the subjects’ measurements were from the center of the closest size. Subjects from the testing set were used in the comparison to avoid bias.

The first set of comparisons was designed to see how many people could find a size that fits them “perfectly,” meaning that all of the selected measurement criteria for the subject should fit within the defined size range. For example, in the first comparison, only height (stature) and waist girth measurements were used as the criteria measurements. If a person’s height and waist measurements both fell into the ranges for a size, she was assigned with the label for that size. The total number of subjects that had assigned sizes was counted for each sizing system. A larger number meant better fit. The same steps were done with three criteria measurements (height, waist, and bust), four criteria measurements (height, waist girth, bust girth, and hips girth), and five criteria measurements (height, waist girth, bust girth, hips girth, and high hip girth). The reason for choosing these measurements as criteria measurements was because they were important for patternmaking and general size selection.

The second set of comparisons was added to see how different a subject’s measurement was from the center of the closest size. The difference was measured by the Euclidean distance (straight line distance). Instead of trying to find the “perfect” size, subjects were assigned with the size that had the smallest Euclidean distance value. Comparisons were done with the same setup of the criteria measurements. Paired right tail t tests were conducted to see if the mean distance of the newly developed sizing system was significantly smaller than the mean distance of ASTM D5585-11e1 (ASTM International, 2011).

Results

Key Measurements

Based on the measurements used in past sizing standards, a total of 64 body measurements were chosen as the key dimensions for creating a sizing system. Natural log transformations were applied on these measurements, with the exclusion of body mass index and degree of shoulder slope (62 measurements). The data were then analyzed in SPSS. Histograms demonstrated a decrease in skewness after the natural log transformation. The training data set contained 3,112 subjects with complete measurement data after the listwise deletion.

Control Variables

Descriptive analysis was done on all 64 measurements. It was found that the ankle height and the wrist girth measurements had many missing values. Because these were not the primary measurements for patternmaking, they were excluded from the final principal component analysis (PCA). All of the 60 natural log-transformed variables, along with the nontransformed body mass index and degree of shoulder slope, were used in both the pre-PCA and the final PCA.

The PCs extraction was based on the eigenvalue. Factors with eigenvalues larger than one were exported as the PCs for size design. The varimax rotation method and the listwise deletion method were applied. A coefficient matrix and a scree plot were drawn (Figure 2). Eight pre-PCs were automatically extracted from the pre-PCA. The first pre-PC was primarily related to the weight and girth measurements, and it explained over 37% of the total variance. The second pre-PC was primarily related to the height measurements. It explained around 22% of the total variance. The third pre-PC, primarily related to the waist length measurements, explained 7.4% of the total variance. And the fourth pre-PC was primarily related to the shoulder slope measurements and explained 4% of the total variance. By considering the variance explained by the PCs and the total number of possible sizes for the final standard, it was decided to set the number of PCs to two for the final PCA. These two PCs represented most of the height and girth measurements.

Figure 2.
Scree plot of pre-principle component analysis (PCA).

The final PCA was done on the training set using the same settings in the pre-PCA, except for the number of target PCs. Factor scores for PC1 and PC2 were saved. PC1 represented the majority of the girth measurements and PC2 represented the majority of the height measurements. They were orthogonal (noncorrelated) with each other. Because the PCs were standardized, the mean value for them was zero and the standard deviation was one.

Linear Regression Formulas

Within the training set, MLR were done using PC1 and PC2 as the independent variables (predictors) to predict the 62 natural log-transformed variables plus the nontransformed body mass index and shoulder slope. The stepwise model selection method was applied using F value as the stepping criteria with F = 0.05 as the entry level and F = 0.1 as the removal level. The generated coefficients were then used for calculating size centers. The MLR results showed that 48 models of 64 (corresponding to 64 key measurements) had adjusted R ² values larger than .5. Twelve models had adjusted R ² values smaller than .3.

Size Range and Intervals

The SD of the two selected PCs were used when deciding the size range. The range of the height-related PC (PC2) was set to mean ± 2SD. Because the SizeUSA data set tended to have a larger girth measurement compared to the selected sizing standard, the range of PC1 was set to (mean − 2.3SD, mean + 2SD). This added one extra size to the small side of PC1. PC2 was divided evenly into three groups. These groups were named petite, regular, and tall. PC1 was divided into 14 groups evenly, and these groups were labeled from size 1 to size 14. Because PC1 and PC2 were independent of each other, the total number of sizes was 42. The size boundaries on histograms for PC1 and PC2 in the training set are illustrated in Figure 3. The number of subjects that fell within each size is shown in Table 1. After the cluster centers (PC1 and PC2 values for each size) were decided, values of the key measurements for each size were calculated based on the formulas from the MLR analysis. The values were rounded to the nearest 1/8-inch. A body sizing system was then created. Each subject had a size label based on her scores for PC1 and PC2. Real means of the key measurements for all sizes were compared with the size centers from the generated sizing system. The difference between the real mean and the predicted mean for some key measurements was less than a half-inch.

Figure 3.
Size bounds on histograms of the training set for principle component (PC)-1 and PC-2.

Table 1.
The Number of Subjects Within Each Size Group Based on Principle Component Two (Height Related).

Groups Min 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Max Total

Group 1: Petite 1 5 7 35 56 70 80 81 93 88 51 51 44 24 16 22 724

Group 2: Regular 0 6 16 63 107 179 193 196 159 149 118 92 88 41 40 69 1,516

Group 3: Tall 0 2 3 24 47 90 96 97 91 79 66 38 40 25 16 29 743

Fit Performance

In order to test whether the newly created sizing system was better than the published standard, two sets of comparisons were made between the created sizing system and ASTM D5585-11e1 (ASTM International, 2011). The comparisons were performed on the testing set. The total number of subjects in the testing set that had no missing values was 3,147. The ASTM D5585-11e1 (ASTM International, 2011) standard has one height group, two body shapes, and 12 sizes for each body shape. Its waist girth measurement ranges from 23.785 inch to 40.5 inch. To make things comparable, only sizes from size 1 through 12 in the regular height group from the created sizing system were selected for comparison.

In the first set of comparisons described in the methodology, subjects were labeled if they could fit into the range of a size for all the criteria measurements. Four groups of comparisons were made with different numbers of criteria measurements. The total number of labeled subjects, or the number of subjects that had a size label, was used to quantify the performance of a sizing system. Results from the first set of comparisons are illustrated in Table 2. The number of assigned subjects over the total number of tested subjects was defined as the perfect fit rate. For the second set of comparisons, subjects were assigned with sizes that had the least Euclidean distance. The t tests were conducted between the mean distance of a created sizing system and the mean distance of ASTM D5585-11e1 (ASTM International, 2011). Results from the second set of comparisons are illustrated in Table 3. The number of sizes considered in ASTM D5585-11e1 (ASTM International, 2011) was 24 (because two shapes were considered), while it was 12 for the created sizing system.

Table 2.
Compare Created Sizing System With ASTM D5585-11e1 Based on Perfect Fit Rate.

List of criteria measurements New Sizing System ASTM D5585-11e1

Number of sizes Perfect fit rate (%) Number of sizes Perfect fit rate (%)

Height and waist 12 41.98 24 35.40

Height, waist, and bust 12 12.13 24 16.72

Height, waist, bust, and hips 12 2.96 24 3.87

Height, waist, bust, and high hip 12 1.23 24 1.20

Table 3.
Compare Created Sizing System With ASTM D5585-11e1 (ASTM International, 2011) Based on the Euclidean Distance.

List of criteria measurements New Sizing System ASTM D5585-11e1 p value

Number of sizes Means Euclidean distance (inches) Number of sizes Means Euclidean distance (inches)

Height and waist 12 2.66 24 3.03 <.00001*

Height, waist, and bust 12 3.21 24 3.50 <.00001*

Height, waist, bust, and hips 12 3.91 24 4.15 <.00001*

Height, waist, bust, and high hip 12 4.29 24 4.72 <.00001*

Note. t test was done on right tailed paired t test.

*p < .025 one-tailed significant level.

Discussion

When comparing the ASTM D5585-11e1 standard (ASTM International, 2011) with the sizing system created in the study, it was found that the newly created sizing system performed better than the ASTM standard on predicting subjects from the testing set. The comparisons were done between the ASTM standard considering two shapes and the created sizing system considering one shape. Two sets of comparisons were conducted. The idea of the first set of comparisons was to see how many people could find sizes that fit them perfectly, while the purpose for the second set was to see how different a person’s measurements were from the closest size center.

For the first set of comparisons, when using height (stature) and waist circumference as the criteria measurements, we found that the created sizing system had 41.98% of subjects labeled, while only 35.40% of the subjects were labeled with ASTM sizes. This is because height measurement varied between different sizes in the created sizing system (although only one height group was used), while it remained constant in the ASTM standard (Xia, 2013). When using height, bust, and waist as the criteria measurements, 16.72% of the subjects were labeled using the ASTM standard, while 12.13% of the subjects were labeled using the developed sizing system. The reason why the selected ASTM standard had a higher rate was because it considered two body shapes (curvy and straight). This demonstrated that a sizing system with the inclusion of body shapes would fit the population better (Xia, 2013). The third comparison used height, bust, hips, and waist as the criterial measurements. The fourth comparison used height, bust, high hip, hips, and waist as the criterial measurements. When five criteria measurements were used, the created sizing system only covered 1.23% of the subjects. The ASTM standard only labeled 1.20% of the subjects. This demonstrated that the created sizing system did a better job on fitting multimeasurements (although only slightly; Xia, 2013). It also showed that with a limited number of sizes, it was very hard to fit subjects simultaneously for more than four measurements. Customization or pattern alterations were needed for better fit.

Based on results of the second set of comparisons, we found that the mean Euclidean distance between subjects’ measurements and their closest size centers in the developed sizing system was significantly smaller than the mean Euclidean distance calculated from ASTM D5585-11e1 standard (ASTM International, 2011), even when two shapes were considered in the ASTM standard. This demonstrated that the developed sizing system performed better than the ASTM D5585-11e1 standard (ASTM International, 2011) on the testing set.

It is important to clarify that comparisons between the created sizing system and ASTM D5585-11e1 (ASTM International, 2011) were for the purpose of verification. We are not suggesting that every female in the SizeUSA data is a missy size.

Conclusion

In conclusion, a method of generating a sizing system based on human body measurements was designed and then tested on the SizeUSA data set. The method involves a process of selecting key measurements based on published standards, normalizing measurements using natural log transformations, generating control variables using PCA, estimating secondary measurements using MLRs, and determining size intervals and ranges using SDs (Xia, 2013). The developed method was applied on SizeUSA data to test its feasibility. A sizing system was created. Subjects were divided into three height groups (namely petite, regular, and tall) and within each height group, 14 sizes were subdivided. The created sizing system was tested to serve the test subjects better than ASTM D5585-11e1 (ASTM International, 2011) based on p values from t tests. It was found that body shape information is critical for size design. Customization and pattern alterations would be needed to provide a better fit. The results of this study are very beneficial to the apparel industry. When the anthropometric data becomes more accessible, apparel companies can build their own sizing library by applying the developed method. With the created sizing systems, apparel companies can improve the fit of their clothing products and also build a mass customization business for fit improvement from it. Overall, the study resulted in a more complete understanding of sizes and how to create them.

Future Study

Several directions for future research are provided by this study. While the SizeUSA data set was used in this research, the method can also be used to analyze any other anthropometric data sets and help in creating sizing systems so as to improve the fit of apparel products. For example, the sizing creation method can be used to study the difference between populations of various countries and help set an international sizing system. All of the 64 measurements used in the PCA were weighted evenly, while they were not necessarily of the same importance in patternmaking. In a future study, fewer key measurements could be used to generate the PCs. A combination of different measurements could be applied and compared to see which combination works best. It is important to include shape information in size design. An additional aspect to study would be to determine the best way to apply the sizing system. Because sizes were divided based on measurements and body shapes, how size information can be explained to consumers clearly and easily is important. Also, because RTW cannot achieve the perfect fit, solving the problem of how a sizing system can be used in mass customization is beneficial.

Limitations

The SizeUSA data were collected 14 years ago. It may not represent the current state of Americans’ body information. Also, when measurements were extracted from 3D models, they may not have been measured where they should have been. For example, the computer may not be able to find the exact natural waist location when it is computing the waist girth. All these factors may influence results generated from the developed method. However, because the SizeUSA data set is the most recent large-scale anthropometric survey conducted in the United States and checking the waist location was not the main focus of the article, the SizeUSA data set was used.

Groups	Min	1	2	3	4	5	6	7	8	9	10	11	12	13	14	Max	Total
Group 1: Petite	1	5	7	35	56	70	80	81	93	88	51	51	44	24	16	22	724
Group 2: Regular	0	6	16	63	107	179	193	196	159	149	118	92	88	41	40	69	1,516
Group 3: Tall	0	2	3	24	47	90	96	97	91	79	66	38	40	25	16	29	743

List of criteria measurements	New Sizing System	ASTM D5585-11e1
Height and waist	12	41.98	24	35.40
Height, waist, and bust	12	12.13	24	16.72
Height, waist, bust, and hips	12	2.96	24	3.87
Height, waist, bust, and high hip	12	1.23	24	1.20

List of criteria measurements	New Sizing System	ASTM D5585-11e1	p value
Height and waist	12	2.66	24	3.03	<.00001*
Height, waist, and bust	12	3.21	24	3.50	<.00001*
Height, waist, bust, and hips	12	3.91	24	4.15	<.00001*
Height, waist, bust, and high hip	12	4.29	24	4.72	<.00001*

Footnotes

Acknowledgments

Thank you to TC2 for the SizeUSA study and for making the female data available for analysis.

Declaration of Conflicting Interests

The author(s) declared no potential conflict of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Note

References

Aldrich

(2000). Tailors’cutting manuals and the growing provision of popular clothing 1770–1870: ‘Falling apart like a ready-made.’ Textile History, 31, 163–201.

Aldrich

(2007). History of sizing systems and ready-to-wear garments. In Ashdown

(Ed.), Sizing in clothing: Developing effective sizing systems for ready-to-wear clothing (pp. 1–56). Cambridge, England: Woodhead.

Ashdown

S. P.

DeLong

(1995). Perception testing of apparel ease variation. Applied Ergonomics, 26, 47–54.

Ashdown

S. P.

Loker

(2010). Mass-customized target market sizing: Extending the sizing paradigm for improved apparel fit. Fashion Practice, 2, 147–173.

ASTM International. (2011). ASTM D5585-11e1 Standard tables of body measurements for adult female misses figure type, size range 00-20. doi:10.1520/D5585-11E01

ASTM International. (2015). ASTM D5219-15 Standard terminology relating to body dimensions for apparel sizing. doi:10.1520/D5219-15

Bickle

M. C.

Burnsed

K. A.

Edwards

K. L.

(2015). Are U.S. plus-size women satisfied with retail clothing store environments? Journal of Consumer Satisfaction, Dissatisfaction & Complaining Behavior, 28, 45–60.

Bradley

P. S.

Fayyad

U. M.

(1998). Refining initial points for k-means clustering. In Shavlik

(Ed.), Proceedings of the 15th international conference on machine learning (pp. 91–99). San Francisco, CA: Morgan Kaufmann.

Chen-Yu

H. J.

Williams

Kincade

D. H.

(1999). Determinants of consumer satisfaction/dissatisfaction with the performance of apparel products. Family and Consumer Sciences Research Journal, 28, 167–192.

10.

Esfandarani

M. S.

Shahrabi

(2012). Developing a new suit sizing system using data optimization techniques. International Journal of Clothing Science and Technology, 24, 27–35.

11.

Giovis

(2007, 1 22). More fitting clothes urged. Tribune Newspapers: South Florida Sun-Sentinel. Retrieved from http://articles.chicagotribune.com/2007-01-22/business/0701220053_1_fit-technologies-apparel-industry-long-torso

12.

Guan

Hsiao

Bradtmiller

Kau

T.-Y.

Reed

M. R.

Jahns

S. K.

… Piamonte

D. P. T.

(2012). U.S. truck driver anthropometric study and multivariate anthropometric models for cab designs. Human Factors: The Journal of the Human Factors and Ergonomics Society, 54, 849–871.

13.

Gupta

Gangadhar

B. R.

(2004). A statistical model for developing body size charts for garments. International Journal of Clothing Science and Technology, 16, 458–469.

14.

Hinds

B. K.

McCartney

Woods

(1991). Pattern development for 3D surfaces. Computer-Aided Design, 23, 583–592.

15.

Hsu

C. H.

Wang

M. J. J.

(2005). Using decision tree-based data mining to establish a sizing system for the manufacture of garments. The International Journal of Advanced Manufacturing Technology, 26, 669–674.

16.

Humphries

(n.d.). Missing data & how to deal: An overview of missing data [PowerPoint slides]. Retrieved from https://liberalarts.utexas.edu/prc/_files/cs/Missing-Data.pdf

17.

IBM Corporation. (2012, 10 2). IBM knowledge center—k-means cluster analysis. Retrieved from www.ibm.com/support/knowledgecenter/SSLVMB_21.0.0/com.ibm.spss.statistics.cs/quickcluster_table.htm

18.

ISO/TC 133. (1989). ISO 8559:1989—Garment construction and anthropometric surveys—Body dimensions. Retrieved from the International organization for Standardization website www.iso.org/iso/catalogue_detail.htm?csnumber=15821

19.

Istook

C. L.

(2002). Enabling mass customization: Computer-driven alteration methods. International Journal of Clothing Science and Technology, 14, 61–76.

20.

Jolliffe

I. T.

(1982). A note on the use of principal components in regression. Applied Statistics, 31, 300–303.

21.

Kang

T. J.

Kim

S. M.

(2000). Development of three-dimensional apparel CAD system: Part 1: Flat garment pattern drafting system. International Journal of Clothing Science and Technology, 12, 26–38.

22.

Kwong

M. Y.

(2004). Garment design for individual fit. In Fan

Hunter

(Eds.), Clothing appearance and fit: Science and technology (pp. 198–233). Cambridge, England: Woodhead.

23.

D. C.

(2011). Patternmaking. London, England: Laurence King.

24.

Loker

(2007). Mass customization and sizing. In Ashdown

(Ed.), Sizing in clothing (pp. 246–263). Cambridge, England: Woodhead.

25.

Mpampa

M. L.

Azariadis

P. N.

Sapidis

N. S.

(2010). A new methodology for the development of sizing systems for the mass customization of garments. International Journal of Clothing Science and Technology, 22, 49–68.

26.

Murray

(2016, 8 5). It’s not you—women’s clothing sizes don’t make sense. Today. Retrieved from www.today.com/style/jeans-don-t-fit-here-s-explanation-inconsistency-women-s-t100419

27.

Newcomb

E. A.

(2006). Body shape analysis of Hispanic women in the United States. Master’s thesis, North Carolina State University. Retrieved from www.lib.ncsu.edu/resolver/1840.16/1820

28.

Okabe

Imaoka

Tomiha

Niwaya

(1992). Three dimensional apparel CAD system. Computer Graphics, 26, 105–110.

29.

Petrova

(2007). Creating sizing systems. In Ashdown

(Ed.), Sizing in clothing: Developing effective sizing systems for ready-to-wear clothing (pp. 57–87). Cambridge, England: Woodhead.

30.

Shin

S. J. H.

Istook

C. L.

(2007). The importance of understanding the shape of diverse ethnic female consumers for developing jeans sizing systems. International Journal of Consumer Studies, 31, 135–143.

31.

Song

H. K.

Ashdown

S. P.

(2011). Categorization of lower body shapes for adult females based on multiple view analysis. Textile Research Journal, 81, 914–931.

32.

Song

H. K.

Ashdown

S. P.

(2012). Development of automated custom-made pants driven by body shape. Clothing and Textiles Research Journal, 30, 315–329.

33.

Tanagra. (2009, 12 24). Tanagra—data mining and data science tutorials: VARIMAX rotation in principal component analysis [Blog]. Retrieved from http://data-mining-tutorials.blogspot.com/2009/12/varimax-rotation-in-principal-component.html

34.

Taylor

S. E.

(1998). Multivariate statistical methods applied to sizing and fit testing. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 42, 732–735.

35.

[TC]² . (2004, 2). SizeUSA: The national sizing survey women body measurement and data analysis reports on the U.S. population. Cary, NC: [TC]².

36.

The NPD Group, Inc. (2017). NPD group reports size matters to American women. Retrieved from www.npd.com/wps/portal/npd/us/news/press-releases/pr_120910/

37.

Workman

J. E.

Lentz

E. S.

(2000). Measurement specifications for manufacturers’ prototype bodies. Clothing and Textiles Research Journal, 18, 251–259.

38.

Xia

(2013). Sizing systems created using SizeUSA data for three body shapes. Master’s thesis, North Carolina State University. Retrieved from https://repository.lib.ncsu.edu/

39.

Yang

Zhang

(2007). Prototype garment pattern flattening based on individual 3D virtual dummy. International Journal of Clothing Science and Technology, 19, 334–348.

List of criteria measurements	New Sizing System		ASTM D5585-11e1
List of criteria measurements	Number of sizes	Perfect fit rate (%)	Number of sizes	Perfect fit rate (%)
Height and waist	12	41.98	24	35.40
Height, waist, and bust	12	12.13	24	16.72
Height, waist, bust, and hips	12	2.96	24	3.87
Height, waist, bust, and high hip	12	1.23	24	1.20

List of criteria measurements	New Sizing System		ASTM D5585-11e1		p value
List of criteria measurements	Number of sizes	Means Euclidean distance (inches)	Number of sizes	Means Euclidean distance (inches)	p value
Height and waist	12	2.66	24	3.03	<.00001*
Height, waist, and bust	12	3.21	24	3.50	<.00001*
Height, waist, bust, and hips	12	3.91	24	4.15	<.00001*
Height, waist, bust, and high hip	12	4.29	24	4.72	<.00001*