Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Open Access

Peer-reviewed

Research Article

Prediction of Body Mass Index Using Concurrently Self-Reported or Previously Measured Height and Weight

  • Zhaohui Cui ,

    cuiz@email.unc.edu

    Affiliation Department of Nutrition, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, United States of America

    ⨯
  • June Stevens,

    Affiliations Department of Nutrition, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, United States of America, Department of Epidemiology, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, United States of America

    ⨯
  • Kimberly P. Truesdale,

    Affiliation Department of Nutrition, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, United States of America

    ⨯
  • Donglin Zeng,

    Affiliation Department of Biostatistics, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, United States of America

    ⨯
  • Simone French,

    Affiliation Division of Epidemiology and Community Health, School of Public Health, University of Minnesota, Minneapolis, Minnesota, United States of America

    ⨯
  • Penny Gordon-Larsen

    Affiliation Department of Nutrition, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, United States of America

    ⨯

Abstract

Objective

To compare alternative models for the imputation of BMIM (measured weight in kilograms/measured height in meters squared) in a longitudinal study.

Methods

We used data from 11,008 adults examined at wave III (2001–2002) and wave IV (2007–2008) in the National Longitudinal Study of Adolescent to Adult Health. Participants were asked their height and weight before being measured. Equations to predict wave IV BMIM were developed in an 80% random subsample and evaluated in the remaining participants. The validity of models that included BMI constructed from previously measured height and weight (BMIPM) was compared to the validity of models that used BMI calculated from concurrently self-reported height and weight (BMISR). The usefulness of including demographics and perceived weight category in those models was also examined.

Results

The model that used BMISR, compared to BMIPM, as the only variable produced a larger R2 (0.913 vs. 0.693), a smaller root mean square error (2.07 vs. 3.90 kg/m2) and a lower bias between normal-weight participants and those with obesity (0.98 vs. 4.24 kg/m2). The performance of the model containing BMISR alone was not substantially improved by the addition of demographics, perceived weight category or BMIPM.

Conclusions

Our work is the first to show that concurrent self-reports of height and weight may be more useful than previously measured height and weight for imputation of missing BMIM when the time interval between measures is relatively long. Other time frames and alternatives to in-person collection of self-reported data need to be examined.

Citation: Cui Z, Stevens J, Truesdale KP, Zeng D, French S, Gordon-Larsen P (2016) Prediction of Body Mass Index Using Concurrently Self-Reported or Previously Measured Height and Weight. PLoS ONE 11(11): e0167288. https://doi.org/10.1371/journal.pone.0167288

Editor: Rebecca A. Krukowski, University of Tennessee Health Science Center, UNITED STATES

Received: September 11, 2016; Accepted: November 11, 2016; Published: November 29, 2016

Copyright: © 2016 Cui et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: The data were from a third party. The authors are not allowed to release the data according to the Data Use Agreement. To obtain the data, please contact addhealth@unc.edu.

Funding: This work was supported by Award Number 1U01HL103561 from the National Heart, Lung, And Blood Institute, the Eunice Kennedy Shriver National Institute of Child Health and Development and the Office of Behavioral and Social Sciences Research. The funder had no role in the design, analysis, or writing of this article. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Heart, Lung, and Blood Institute or the National Institutes of Health.

Competing interests: The authors have declared that no competing interests exist.

Introduction

Longitudinal studies of body weight in humans almost always have missing values due to loss to follow-up or incomplete examinations. These missing values can bias estimates and reduce statistical power. A recent review of pharmaceutical randomized trials with weight as the primary outcome indicated that in most studies missing data for body mass index (BMI) was imputed by carrying forward BMI calculated from previously measured height and weight (referred as BMIPM) or by predicting the missing values using a model that included BMIPM [1]. However, BMI can change importantly over time, and in some cases those changes are the primary interest of the study. It is well established that BMI calculated from self-reported height and weight (BMISR) is highly correlated with concurrently assessed BMI calculated from measured height and weight (BMIM) with r values of approximately 0.90 [2]. Nevertheless, BMISR tends to underestimate BMIM, and this bias is exaggerated in individuals who are female, white, older and heavier [3–6]. Thus both BMIPM and BMISR have potential to the imputation of BMIM, but both have limitations.

The use of modeling to predict missing data is currently recommended for data missing at random [7]. The use of BMISR in regression models that impute BMIM has been explored in cross-sectional data [3,8–11]. However, to our knowledge, modeling of missing BMIM values using BMISR has not been extended to longitudinal data. We know of no study that has compared equations that include BMIPM to equations that include BMISR for the prediction of BMIM. We hypothesized that in longitudinal studies in which the most recently measured weight is several years in the past, a regression model to predict BMIM that includes both BMIPM and BMISR would perform better (i.e. with greater accuracy) than a model that used only BMIPM. In addition, we anticipated that BMIPM would produce stronger predictions than BMISR. Since studies have shown that demographics and perception of weight status are associated with BMIM and misreport of BMIM [4,12], we also tested including demographic and perceived weight status variables in the regression models for prediction of BMIM.

Materials and Methods

Study population

The National Longitudinal Study of Adolescent to Adult Health (Add Health) [13] was selected to provide data for this analysis because both measured and self-reported height and weight were assessed in the same participants on two occasions. The Add Health study was designed to examine the relationship between social, behavioral and biological factors and health from adolescence to adulthood [13]. Samples of youth were drawn from 80 high schools and 52 middle schools selected to ensure representation of US schools in regard to region of country, urbanicity, school size, school type, and school ethnicity. The core sample is nationally representative of American students in grades 7–12 in the 1994–95 school year. Additional special samples provide enlarged subsamples for studies of ethnicity, disability and sibling and twin relations. We combined the data from the core sample and special samples at each wave in the present study. We chose to construct the sample in this way to include a more diverse participants. The first examination (called Wave I) included 20,745 students aged 11–21 years. Follow-up examinations were conducted in 1996–1997 (wave II), 2001–2002 (wave III) and 2007–2008 (wave IV), respectively. We focused the current work on adults and therefore used data from waves III (n = 15,197, age = 18–26 years) and IV (n = 15,701, age = 24–32 years). Survey procedures have been described elsewhere [14] and were approved by the institutional review board at the University of North Carolina at Chapel Hill. Written informed consent was obtained from all participants.

Measures

At each wave, participants were asked their height and weight in a home-based interview and were measured after the interview by a trained technician. Height was measured to the nearest 0.5 cm without shoes using a measure tape with a standard triangular headboard attached to a wall. Weight was measured to the nearest 0.1 kg without shoes and in clothing using a Health-o-meter 844KL High Capacity Digital Bathroom Scale (Jarden Corporation; Rye, NY). BMI is calculated as weight in kg divided by the square of height in meters. BMIPM was constructed from measured height and weight from wave III, while BMIM and BMISR were calculated from data collected at wave IV. In wave IV perceived weight status was obtained by asking the participant “how do you think of yourself in terms of weight?”: 1) very underweight; 2) slightly underweight; 3) about the right weight; 4) slightly overweight; 5) very overweight. Because only a few participants answered very underweight (0.7%), this category was combined with slightly underweight.

Participants reported their date of birth and age was calculated. Time intervals were calculated between examination dates. Race/ethnicity was self-reported by parents and adolescents at wave I and was classified as non-Hispanic white, non-Hispanic black, Hispanic, Asian American or Native American [15].

Analytic sample

Among the 15,701 participants at wave IV, we excluded those who reported pregnancy at wave III or IV (n = 967); had a part of their body amputated (n = 16); were missing race/ethnicity (n = 32), perceived weight status at wave IV (n = 7), or height or weight at either wave (n = 3,631); reported a weight less than 30 pounds (n = 11); or had either a BMIM, BMISR or BMIPM that was <12 kg/m2 or > 60 kg/m2 (n = 29). Our final analytic sample included 11,008 participants with weight and height data from both waves.

Statistical analysis

We constructed a training data sets by random selection of 80% of observations without replacement. The remaining participants formed the test dataset. We developed equations to predict BMIM using general linear models and different combinations of variables. The equations developed in the training dataset were applied to the participants in the test dataset to create a predicted BMIM. These predicted values were regressed against BMIM in the test dataset and R2 and root mean square error (RMSE) were calculated. To evaluate whether the accuracy of the predicted values differed by BMI category [16] we calculated mean signed difference (MSD) as the mean of the difference between predicted values and BMIM. To examine whether the results from the above analyses were by chance, we repeated the above analyses in four different sets of training and test datasets generated by randomly allocating the 11,008 participants and found all results were similar [11]. All analyses were conducted with SAS 9.4 software (SAS Institute, INC., Cary, NC).

Results

In the analytic sample a little over half of the participants were non-Hispanic whites (Table 1). Mean BMI increased from wave III to wave IV in groups with normal weight, overweight and obesity. On average, participants over-reported their measured height. Normal weight participants over-reported, while participants with overweight or obesity under-reported, their measured weight. This resulted in a relatively accurate estimation of BMIM in the normal weight group but underestimation of BMIM by 0.5 kg/m2 in the group with overweight and 1.6 kg/m2 in the group with obesity.

thumbnail
Download:
Table 1. Characteristics of the analytic sample.

https://doi.org/10.1371/journal.pone.0167288.t001

Using BMIM as the dependent variable, nine models with different combinations of independent variables were developed in the training data set. The resulting 9 equations (intercepts and coefficients) were applied to the test dataset to generate predicted estimates of BMIM. These estimates were then compared to BMIM in the test data using univariate regression. Table 2 shows the R2 and RMSE produced from this analysis. Model 1–1 with BMIPM as the only independent variable produced an R2 of 0.693 and an RMSE of 3.90 kg/m2. Model 2–1 with BMISR as the only independent variable produced a larger R2 (0.913) and a smaller RMSE (2.07 kg/m2) compared to the models 1–1. Performance was not substantially improved by adding BMIPM to the model with BMISR. Addition of demographics and perceived weight status also had little impact. Further, the addition of non-linear forms of BMI and age (i.e., inverse, square and cubic), and the interactions of gender or race/ethnicity with BMI did not increase the adjusted R2 by more than 0.01 or reduce bias between normal weight group and the group with obesity by as much 0.1 kg/m2. The intercepts and coefficients of the models developed in the whole analytic sample are shown in the S1 Table.

thumbnail
Download:
Table 2. R2 and RMSE from regression* of predicted BMIM against actual BMIM in the test dataset.

https://doi.org/10.1371/journal.pone.0167288.t002

The MSD in overweight adults was -2.52 kg/m2 for BMIPM, -0.54 kg/m2 for BMISR and close to zero for all nine models (0.03–0.17 kg/m2). Those values were intermediate between those seen in the normal weight group and the group with obesity. Fig 1 shows MSD’s calculated using the test dataset that contrast estimates of BMI obtained in normal weight participants and participants with obesity (categories assigned using BMIM). The MSD estimates in normal weight participants and participants with obesity were over triple as far from the null for BMIPM compared to BMISR. Model predicted values overestimated BMIM in normal weight participants and underestimated BMIM in participants with obesity with the magnitudes of bias consistently smaller for models that included BMISR (i.e., models 2–1 to 3–3) compared to models that used BMIPM (i.e., models 1–1, 1–2 and 1–3). As a result, the amount of differential bias (discrepancy between the MSD in normal weight participants and the MSD in participants with obesity for a given model) was lower in models that included BMISR (i.e., models 2–1 to 3–3). In those models, the range of the discrepancies was relatively narrow ranging from 0.83 to 0.98 kg/m2. Analyses conducted by gender and race-ethnicity subgroups also showed smaller discrepancies in the systematic error of estimates produced from models that included BMISR compared to those that included BMIPM.

thumbnail
Download:
Fig 1. MSD between predicted BMIM and actual BMIM by weight status in the test dataset (n = 2202).

(A) Abbreviations: MSD for mean signed difference; BMI for body mass index, BMIPM is derived from measured height and weight at wave III, BMIM is derived from measured height and weight at wave IV, BMISR is constructed from self-reported height and weight at wave IV. (B) MSD was calculated as the mean of predicted BMIM minus actual BMIM. The dashed lines in the Fig are at ±0.5 kg/m2. (C) Weight status was based on BMIM. n = 719 for normal weight group (18.5≤ BMIM <25 kg/m2) and n = 776 for the group with obesity (BMIM ≥30kg/m2). Results for underweight group (n = 45) and for overweight group (n = 662) were not shown.

https://doi.org/10.1371/journal.pone.0167288.g001

Discussion

Our analyses found that equations produced from regression models that included BMISR performed better than equations that used BMIPM. This is consistent with the stronger correlation of BMIM with BMISR compared to BMIPM in this study with an average of 6.5 years of interval between the two examinations. We found that BMISR (r = 0.95) was more strongly correlated with BMIM than BMIPM (r = 0.82). Over 100 studies have compared self-reported height and weight with measured values [4–6,17]. Generally these studies have found that adults over-report their height and under-report their weight, resulting in an underestimation of BMI. Despite this bias, BMISR is very strongly correlated (r >0.90) with concurrent BMIM in most studies [18,19]. BMIPM would be highly correlated with BMIM if weight changes were proportionately similar across individuals over time. However, as shown in the Coronary Artery Risk Development in Young Adults (CARDIA) study weight changes are heterogeneous over a 5-year interval [20].

Only a handful of studies have developed equations to predict BMIM using BMISR and reported model fit statistics [6,9,21,22]. All those studies used cross-sectional data and therefore did not examine models that used BMIPM as in our present work. Most of those studies stratified their analyses by sex and reported R2 ranging from 0.85–0.92 for men and 0.83–0.96 for women [9,21,22], which is consistent with our sex-specific findings. Stommel and Schoenborn [6] developed prediction equations in men and women combined using the 2001–2006 National Health and Nutrition Examination Survey (NHANES) and found an R2 of 0.922, which is consistent with our sex-combined results.

We emphasized bias between normal weight participants and participants with obesity in imputation of BMIM because this bias could result in an overestimation of disease risk associated with BMI when predicted values are used in study of BMI-disease relationship [21]. Using data from the NHANES II, Plankey et al. [11] found that predicted values from a model that used BMISR to predict BMIM tended to increase bias in normal weight adults and reduce bias in adults with overweight or obesity, resulting in an overestimate of BMIM in the normal weight and an underestimate in participants with obesity. In addition to this study, another study is consistent with our work. Using data from the 2001–2006 NHANES, Stommel and Schoenborn [6] also found the predicted values systematically overestimate BMI in leaner individuals and underestimate BMI in heavier individuals. The BMIM at which the predicted BMI is unbiased is 28.3 kg/m2.

Although demographics and perceived weight status are associated with misestimation of BMIM [3,4,6,21], we found that addition of those variables and their more complex forms only minimally improved prediction over models that included BMISR as the only variable. Age was not statistically significant in models that used BMISR, perhaps because of the narrow age range in our sample. However, despite a wider age of participants (30–64 years), Plankey et al. [11] found that addition of age and nonlinear forms of BMISR (i.e., quadratic and cubic) to a model containing BMISR alone only slightly improved prediction of BMIM (R2: 0.94 vs. 0.94 in women and 0.90 vs. 0.90 in men; RMSE: 1.37 vs. 1.36 in women and 1.22 vs. 1.22 in men). The authors concluded that the bias differed little between the simple and more complex models. Connor Gorber et al. [21] analyzed data from individuals aged ≥18 years who participated in the 2005 Canadian Community Health Survey. They developed a simple model and a full model in the training dataset. The simple model contained BMISR only and the full model additionally included variables that were significantly associated with the bias in BMI (i.e., age, education, dissatisfied with life and perceived weight status). The R2 from the full model is only 0.01 larger than the R2 from the simple model. Application of those equations to the test dataset found that the MSDs between BMI predicted by the models and BMIM are close to zero. The full model produced little advantage in the sensitivity and specificity over the simple model in determination of obesity. Similar results were also found in studies from Sweden [9] and Australia [22]. Again, those studies did not use longitudinal data as in our work presented here.

It is important that in our work, measurements of height and weight were collected over a relatively long time interval (6.5 years). Our results are likely not applicable to longitudinal studies with short time intervals between examinations. In addition, participants in our study could have anticipated that they would be measured after reporting their height and weight and this could have reduced bias in the reported data [23]. If investigators collect self-reported height and weight from participants separately from other study variables it could also impact accuracy [23]. Finally, bias would likely have been different if self-reported weight and height had been collected in telephone interviews [10] or mailed survey [24] rather than in-person interviews.

Despite all these issues, our study does provide new information that is relevant to researchers who study participants over relatively long time intervals. In those instances collection of self-reported height and weight could be considered and used with an appropriate approach (e.g., multiple imputation) to improve the imputation of missing BMIM. Our work also indicates that investigators should be cautious about relying on previously measured height and weight to indicate current BMIM. Since almost all obesity-related longitudinal studies collect demographic variables, those variables could be included in the imputation model, but it appears they improve prediction only very slightly. Perceived weight status was not helpful for imputation of BMI derived from anthropometric measures. Further study is needed to understand time intervals and circumstances in which self-report of height and weight is superior to the other.

Supporting Information

S1 Table. Final models developed to predict BMIM in the whole dataset (n = 11,008).

https://doi.org/10.1371/journal.pone.0167288.s001

(DOCX)

Acknowledgments

This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Information on how to obtain the Add Health data files is available on the Add Health website (http://www.cpc.unc.edu/addhealth). No direct support was received from grant P01-HD31921 for this analysis.

Author Contributions

  1. Conceptualization: ZC JS KPT DZ.
  2. Formal analysis: ZC.
  3. Funding acquisition: JS.
  4. Methodology: ZC JS KPT DZ.
  5. Resources: JS PG.
  6. Software: ZC.
  7. Supervision: JS.
  8. Validation: ZC.
  9. Visualization: ZC JS KPT.
  10. Writing – original draft: ZC.
  11. Writing – review & editing: ZC JS KPT DZ SF PG.

References

  1. 1. Elobeid MA, Padilla MA, McVie T, Thomas O, Brock DW, Musser B, et al. Missing data in randomized clinical trials for weight loss: scope of the problem, state of the field, and performance of statistical methods. PLoS One 2009;4(8):e6624. pmid:19675667
  2. 2. Niedhammer I, Bugel I, Bonenfant S, Goldberg M, Leclerc A. Validity of self-reported weight and height in the French GAZEL cohort. Int J Obes Relat Metab Disord 2000;24(9):1111–1118. pmid:11033979
  3. 3. Jain RB. Regression models to predict corrected weight, height and obesity prevalence from self-reported data: data from BRFSS 1999–2007. Int J Obes (Lond) 2010;34(11):1655–1664.
  4. 4. Wen M, Kowaleski-Jones L. Sex and ethnic differences in validity of self-reported adult height, weight and body mass index. Ethn Dis;22(1):72–78. pmid:22774312
  5. 5. Connor Gorber S, Tremblay M, Moher D, Gorber B. A comparison of direct vs. self-report measures for assessing height, weight and body mass index: a systematic review. Obes Rev 2007;8(4):307–326. pmid:17578381
  6. 6. Stommel M, Schoenborn CA. Accuracy and usefulness of BMI measures based on self-reported weight and height: findings from the NHANES & NHIS 2001–2006. BMC Public Health 2009;9:421-2458-9-421. pmid:19922675
  7. 7. Schafer JL, Graham JW. Missing data: our view of the state of the art. Psychol Methods;7(2):147–177. pmid:12090408
  8. 8. Schenker N, Raghunathan TE, Bondarenko I. Improving on analyses of self-reported data in a large-scale health survey by using information from an examination-based survey. Stat Med 2010;29(5):533–545. pmid:20029804
  9. 9. Nyholm M, Gullberg B, Merlo J, Lundqvist-Persson C, Rastam L, Lindblad U. The validity of obesity based on self-reported weight and height: Implications for population studies. Obesity (Silver Spring) 2007;15(1):197–208.
  10. 10. Ezzati M, Martin H, Skjold S, Vander Hoorn S, Murray CJ. Trends in national and state-level obesity in the USA after correction for self-report bias: analysis of health surveys. J R Soc Med 2006;99(5):250–257. pmid:16672759
  11. 11. Plankey MW, Stevens J, Flegal KM, Rust PF. Prediction equations do not eliminate systematic error in self-reported body mass index. Obes Res 1997;5(4):308–314. pmid:9285836
  12. 12. Paeratakul S, White MA, Williamson DA, Ryan DH, Bray GA. Sex, race/ethnicity, socioeconomic status, and BMI in relation to self-perception of overweight. Obes Res 2002;10(5):345–350. pmid:12006633
  13. 13. Harris KM, Halpern CT, Whitsel E, Hussey J, Tabor J, Entzel P, et al. The National Longitudinal Study of Adolescent Health: Research Design. http://www.cpc.unc.edu/projects/addhealth/design.
    • 14. Popkin BM, Udry JR. Adolescent obesity increases significantly in second and third generation U.S. immigrants: the National Longitudinal Study of Adolescent Health. J Nutr 1998;128(4):701–706. pmid:9521631
    • 15. The NS, Suchindran C, North KE, Popkin BM, Gordon-Larsen P. Association of adolescent obesity with risk of severe obesity in adulthood. JAMA 2010; 304(18):2042–2047. pmid:21063014
    • 16. World Health Organization. Obesity: preventing and managing the global epidemic. Report on a WHO Consultation on Obesity. Technical Report 894. 2000:9.
      • 17. Pursey K, Burrows TL, Stanwell P, Collins CE. How accurate is web-based self-reported height, weight, and body mass index in young adults? J Med Internet Res 2014;16(1):e4. pmid:24398335
      • 18. Spencer EA, Appleby PN, Davey GK, Key TJ. Validity of self-reported height and weight in 4808 EPIC-Oxford participants. Public Health Nutr 2002;5(4):561–565. pmid:12186665
      • 19. Keith SW, Fontaine KR, Pajewski NM, Mehta T, Allison DB. Use of self-reported height and weight biases the body mass index-mortality association. Int J Obes (Lond) 2011;35(3):401–408.
      • 20. Burke GL, Bild DE, Hilner JE, Folsom AR, Wagenknecht LE, Sidney S. Differences in weight gain in relation to race, gender, age and education in young adults: the CARDIA Study. Coronary Artery Risk Development in Young Adults. Ethn Health 1996;1(4):327–335. pmid:9395577
      • 21. Connor Gorber S, Shields M, Tremblay MS, McDowell I. The feasibility of establishing correction factors to adjust self-reported estimates of obesity. Health Rep 2008;19(3):71–82. pmid:18847148
      • 22. Hayes AJ, Kortt MA, Clarke PM, Brandrup JD. Estimating equations to correct self-reported height and weight: implications for prevalence of overweight and obesity in Australia. Aust N Z J Public Health 2008;32(6):542–545. pmid:19076745
      • 23. Shields M, Connor Gorber S, Janssen I, Tremblay MS. Bias in self-reported estimates of obesity in Canadian health surveys: an update on correction equations for adults. Health Rep 2011;22(3):35–45. pmid:22106788
      • 24. Scribani M, Shelton J, Chapel D, Krupa N, Wyckoff L, Jenkins P. Comparison of bias resulting from two methods of self-reporting height and weight: a validation study. JRSM Open 2014;5(6).