Abstract
Objectives. We evaluated bias in estimated obesity prevalence owing to error in parental reporting. We also evaluated bias mitigation through application of Centers for Disease Control and Prevention’s biologically implausible value (BIV) cutoffs.
Methods. We simulated obesity prevalence of children aged 2 to 5 years in 2 panel surveys after counterfactually substituting parameters estimated from 1999–2008 National Health and Nutrition Examination Survey data for prevalence of extreme height and weight and for proportions obese in extreme height or weight categories.
Results. Heights reported below the first and fifth height-for-age percentiles explained between one half and two thirds, respectively, of total bias in obesity prevalence. Bias was reduced by one tenth when excluding cases with height-for-age and weight-for-age BIVs and by one fifth when excluding cases with body mass–index-for-age BIVs. Applying BIVs, however, resulted in incorrect exclusion of nonnegligible proportions of obese children.
Conclusions. Correcting the reporting of children’s heights in the first percentile alone may reduce overestimation of early childhood obesity prevalence in surveys with parental reporting by one half to two thirds. Excluding BIVs has limited effectiveness in mitigating this bias.
The US Institute of Medicine now highlights early childhood as a critical period for obesity-related public health prevention.1 Few studies of kindergarten age or preschool age obesity, however, offer national generalizability.2–11 Among those that do, more costly measurement protocols may be sacrificed for longer follow-ups and greater breadth of determinants.3,8–11 These trade-offs are frequently necessary in data sources used in multilevel and dynamic policy models, including those systems scientists have developed.12 In the only 2 nationally representative US panel surveys with continuous measurement of individual, family, household, and environmental data from birth through early adulthood—the Panel Study of Income Dynamics (PSID) and the National Longitudinal Survey of Youth (NLSY79)—height and weight in early childhood are assessed predominantly via parent report rather than direct measurement. Their accurate assessment is crucial because the body mass index (BMI), from which obesity prevalence is typically derived in population studies,13 is calculated from weight in kilograms divided by height in meters squared. Unfortunately, the accuracy of obesity prevalence estimated from parent-reported data on height and weight is known to be low, especially among young children.14–20 In the 1999–2004 National Health Interview Survey and the 2003–2004 National Survey of Child Health, both of which rely exclusively on the parental reporting of height and weight, obesity prevalence has been found to be overestimated by a factor of 5 for children aged 2 to 3 years and by a factor of 3 for children aged 3 to 7 years.16 Such extremely high biases led the National Survey of Child Health to cease releasing parent-reported height and weight and calculated BMI for children younger than 10 years.21
A useful first step toward understanding and correcting these very large biases is separating the contributions to bias of the misreporting of weight from the misreporting of height. This is our primary goal. It has been remarked that parents’ perceptions of their young children’s heights do not keep up with their rapid growth in this period of childhood and that the potential for height misreporting to generate bias in obesity prevalence is increased by the squaring of height in the BMI denominator.20 Errors in the reporting of very high values of weight, on the other hand, directly affect whether BMI crosses the obesity threshold. In a previous study,15 we used a graphical method of diagnosis and found implausibly high prevalence of very low height for age for children aged 2 to 5 years in the PSID and NLSY79.
We have extended that study by developing and applying a simulation method to quantify the resulting bias in obesity prevalence and then comparing it with the bias resulting from parental misreporting of weight. We do so alternately in conjunction with, and ignoring, the US Centers for Disease Control and Prevention (CDC) height-for-age, weight-for-age, and BMI-for-age biologically implausible values (BIVs).22,23
The National Longitudinal Survey of Youth 1979 (NLSY79) is an ongoing survey of a probability sample of men and women who were aged 14 to 21 years in 1979. All children born to NLYS79 women and who are still in their care (NLSY79-Child) are given biennial assessments that include height and weight.24 Although the NLSY79-Child survey protocol directed the interviewer to measure the child’s height and weight, frequently the mother instead reported either the child’s height or weight (14.0%) or both (31.0%; Table 1). Our sample comprises children from the 1996 through 2008 survey waves observed at aged 2 to 5 years (n = 2608 children contributing 4082 person-years), excluding observations in which height or weight was not assessed (1.8%).
TABLE 1— Obesity Prevalence and Excess Estimated Obesity Prevalence Relative to the NHANES 1999–2008 in Children Aged 2–5: United States
| Study or Measurement | No. of Observations | Obesity Prevalence, % (95% CI) | Excess PSID or NLSY79 Prevalence,a % (95% CI) |
| Including BIVs | |||
| NHANES, 1999–2008 | 4145 | 11.0 (10.4, 11.6) | |
| NLSY79-Child, 1996–2008 | |||
| Total | 4082 | 19.9 (18.5, 21.4) | 9.0 (7.5, 10.6) |
| Measured weight and height | 2268 | 13.8 (12.3, 15.6) | 2.9 (1.5, 5.3) |
| Reported weight and measured height | 464 | 15.7 (12.4, 19.7) | 4.7 (2.1, 10.2) |
| Measured weight and reported height | 101 | 25.3 (17.1, 35.6) | 14.3 (7.2, 26.3) |
| Reported weight and height | 1249 | 32.3 (29.4, 35.3) | 21.3 (18.5, 24.5) |
| PSID-CDS-I, 1997 | |||
| Total | 1026 | 27.2 (23.6, 31.2) | 16.3 (12.8, 20.5) |
| Reported weight and measured height | 620 | 20.5 (16.6, 25.2) | 9.6 (6.0, 14.9) |
| Reported weight and height | 406 | 36.6 (30.9, 42.7) | 25.6 (20.2, 32.0) |
| Excluding height-for-age and weight-for-age BIVs | |||
| NHANES | 4111 | 10.4 (10.1, 10.8) | |
| NLSY79-Child | |||
| Total | 3802 | 18.2 (16.9, 19.7) | 7.3 (5.9, 9.0) |
| Measured weight and height | 2168 | 13.0 (11.5, 14.8) | 2.1 (0.9, 4.8) |
| Reported weight and measured height | 445 | 14.8 (11.5, 18.8) | 3.8 (1.4, 9.8) |
| Measured weight and reported height | 94 | 24.6 (16.3, 35.3) | 13.6 (6.6, 26.2) |
| Reported weight and height | 1095 | 29.4 (26.5, 32.5) | 18.4 (15.6, 21.7) |
| PSID-CDS-I | |||
| Total | 959 | 25.5 (21.9, 29.5) | 14.5 (11.1, 18.8) |
| Reported weight and measured height | 594 | 20.0 (16.0, 24.7) | 9.1 (5.5, 14.5) |
| Reported weight and height | 365 | 33.5 (27.8, 39.7) | 22.5 (17.1, 29.1) |
| Excluding BMI-for-age BIVs | |||
| NHANES | 4100 | 10.2 (9.8, 10.6) | |
| NLSY79-Child | |||
| Total | 3780 | 17.2 (15.8, 18.6) | 6.2 (4.8, 7.9) |
| Measured weight and height | 2166 | 12.8 (11.2, 14.5) | 1.8 (0.7, 4.7) |
| Reported weight and measured height | 437 | 13.2 (10.0, 17.1) | 2.2 (0.4, 10.5) |
| Measured weight and reported height | 91 | 22.6 (14.6, 33.3) | 11.6 (5.0, 24.7) |
| Reported weight and height | 1086 | 26.9 (24.1, 30.0) | 16.0 (13.2, 19.2) |
| PSID-CDS-I | |||
| Total | 928 | 23.7 (20.1, 27.7) | 12.7 (9.3, 17.1) |
| Reported weight and measured height | 579 | 19.2 (15.2, 24.0) | 8.2 (4.8, 13.9) |
| Reported weight and height | 349 | 30.3 (24.6, 36.6) | 19.3 (14.0, 26.1) |
Note. BIV = biologically implausible value; CI = confidence interval; NHANES = National Health and Nutrition Examination Surveys; NLSY79-Child = Children of the National Longitudinal Survey of Youth 1979; PSID-CDS-I = Panel Study of Income Dynamics, Child Development Supplement, wave I.
aWe calculated excess survey prevalence of obesity by subtracting the NHANES obesity prevalence from the obesity prevalence from the panel survey sample or subsample. Numbers may not add up exactly because of rounding. Percentages are weighted using sample weights for the respective surveys. CI estimates adjust for stratification and clustering in the NHANES sample design and for clustering in the NLSY79-Child and PSID-CDS-I sample designs.
In a second nationally representative panel survey, the PSID, a Child Development Supplement in 1997 (CDS-I), was administered to primary caregivers of up to 2 randomly selected children aged 0 to 12 years.25 Under the CDS-I protocols, the interviewer was to measure the child’s height and the mother (or other primary caregiver) was to report the child’s weight. Again, however, 40% of the mothers reported both height and weight. Our sample comprises children aged 2 to 5 years in CDS-I (n = 1026), excluding observations in which height or weight was not assessed (10.1%). A detailed description of the NLSY79 and PSID child height and weight assessment and the reasons for off-protocol mother-reporting (e.g., nonconsent and telephone interview) have been reported elsewhere.15
The National Health and Nutrition Examination Survey (NHANES) is a continuous, cross-sectional survey of the US civilian noninstitutionalized population.26 Children are sampled as part of a nationally representative multistage probability sample that oversamples African Americans, Hispanics, and those in low-income households. Response rates for the sample of children are between 81.0% and 88.0%.27 Our NHANES sample comprises children aged 2 to 5 years surveyed in 1999–2008 (n = 4145), excluding observations in which height or weight was not assessed (7.9%). The NHANES height and weight assessment protocol involves strict guidelines to ensure benchmark quality data. Trained health technologists measured height and weight during a separate physical assessment. Supervisors monitored quality control and calibration of the equipment. Data collection software range-checked each measurement and required confirmation before accepting values below the first or above the 99th percentiles of the NHANES data.28
We categorized children into percentiles of height for age, weight for age, and BMI for age on the basis of gender and age in months using a SAS version 9.3 (SAS Institute, Cary, NC) program from the CDC, which employs the 2000 CDC growth charts.23 We excluded a small number of observations in the NLSY79 (n = 16) with BMI outside the range of this program. We classified BMI for age at or above the 95th percentile as obese.29
We developed and implemented a simulation method to quantify bias resulting from parental reporting of extremely low height-for-age and extremely high weight-for-age children. Our use of a simulation methodology to diagnose bias in obesity prevalence owing to respondent reporting of height and weight is, to our knowledge, novel. Our use of the NHANES as a standard for evaluation of this bias, however, follows studies of reporting error in the NLSY79,30 NHIS,16,31 and National Survey of Child Health.16 Correction factors from the first of these studies have been applied in subsequent studies to data sources with reported height and weight only.32–37
We evaluated the separate contributions to bias in early childhood obesity prevalence of parental misreporting of height and of parental misreporting of weight. We did this by first specifying a model of obesity prevalence, 
, whose parameters describe children recorded at either the very lowest percentiles of height (h) or very highest percentiles of weight (w) versus children recorded at all other percentiles of height and weight. Second, we used NHANES data to simulate a correction for error at those very lowest height or very highest weight percentiles only. We thereby tested whether obesity prevalence was being falsely pushed up by too many children whose height is severely underestimated by their parents, by too many children whose weight was severely overestimated by their parents, or by both.
To evaluate the extent to which obesity was being falsely pushed up by too many children whose height was severely underestimated by their parents, we modeled obesity prevalence as a weighted sum of the prevalence of obesity among children below the first height percentile P(obese|h < 1) and the prevalence of obesity among children above the first height percentile P(obese|h ≥ 1). The distributional weights for the 2 summands are the proportions of children who were and were not, respectively, below the first height percentile. Hence,
Analogously, to quantify the biasing consequences of weight misreporting, obesity prevalence was expressed as a weighted sum of the prevalence of obesity among children who were and were not above the 99th weight percentile 
and P(w ≤ 99):
Because P(h ≥ 1) = 1 – P(h < 1) and P(w ≤ 99) = 1 – P(w > 99), in both (1a) and (2a), obesity prevalence is modeled by only 3 parameters.
We simulated correct parental reporting of very low height (and of the weight of any children who are very low height) by substituting for the NLSY79 or PSID values of P(obese|h < 1) and P(h < 1) the corresponding NHANES values. We simulated correct parental reporting of very high weight (and of the height of any children who are very high weight) by substituting for the NLSY79 or PSID values of P(obese|w > 99) and P(w > 99) with the corresponding NHANES values. We retained the NLSY79 and PSID values of P(obese|h ≥ 1) and P(obese|w ≤ 99), respectively. Because the NHANES had so few (only 29) cases with height below the first percentile, however, we substituted the NHANES percentage of children who were obese when their height was below the fifth percentile for P(h < 1) in equation (1a).
We calculated bias owing to parental misreporting of very low height as the observed value of in the NLSY79 or PSID minus the simulated value of with P(obese|h < 1) and P(h < 1) substituted with NHANES values as we have described in equation (1a). We calculated bias owing to parental misreporting of very high weight as the observed value of in the NLSY79 or PSID minus the simulated value of in equation (2a). We calculated total bias simply as the observed value of in the NLSY79 or PSID minus the observed value of in the NHANES.
We extended our evaluation of bias to include parental misreporting at less extreme percentiles. In particular, we evaluated contributions to bias from overreporting of heights up to the fifth percentile:
and from overreporting of weights above the 95th percentile:
The NLSY79 and PSID values, respectively, of P(obese|h ≥ 5) for equation (1b) and P(obese|w ≤ 95) for equation (2b) are combined with NHANES values of the other terms in the equations. We again calculated biases owing to parental misreporting of the bottom 5 height percentiles and of the top 5 weight percentiles from the observed value of in the NLSY79 or PSID minus the calculated with these substituted NHANES values in equations (1b) and (2b).
BIVs are anthropometric measurements that are more than a fixed number of standard deviations below or above the values predicted by the CDC’s 2000 growth charts. The child’s modified z score for height for age must fall less than 5 deviations below the mean and 3 deviations above to be considered biologically plausible. For weight for age, the plausible range extends 5 deviations above and below the mean. For BMI for age, the plausible range extends 4 deviations below and 5 deviations above the mean. The CDC classifies all values outside these ranges as BIVs and notes: “Typically these outliers are the result of data entry errors or mismeasurement, rather than from true extreme growth.”22 We evaluated the CDC’s BIVs as a method to mitigate the biasing effect of parental misreporting on obesity prevalence. We alternately included and excluded from the estimation of obesity prevalence children who are classified as having height-for-age or weight-for-age BIVs and then did the same for BMI-for-age BIVs.
All prevalence estimates employed survey weights. We adjusted confidence intervals (CIs) for the stratified survey design of the NHANES and for the clustering of children in families in the NLSY79-Child and PSID-CDS and in primary sampling units in the NHANES by using the SVY procedure in Stata version 11.0 (StataCorp LP, College Station, TX).
We estimated obesity prevalence at aged 2 to 5 years from the NHANES at 11.0% (95% CI = 10.4%, 11.6%; Table 1). Compared with this widely accepted standard for the measurement of obesity prevalence,38 the NLSY79 and PSID estimates for those aged 2 to 5 years were upwardly biased by 9.0 and 16.3 percentage points, respectively (NLSY79 19.9%; 95% CI = 18.5%, 21.4%; PSID 27.2%; 95% CI = 23.6%, 31.2%). Note from the “Observations” column that for no child does our PSID sample have measured height and weight and that a smaller proportion of PSID than NLSY79 children have measured heights. The greater overall bias estimated for the PSID than for the NLSY79 is therefore expected. In both surveys, obesity prevalence exceeded NHANES obesity prevalence most for children with parent-reported height and weight—by 21.3 and 25.6 percentage points, respectively (NLSY79 32.3%; 95% CI = 29.4%, 35.3%; PSID 36.6%; 95% CI = 30.9%, 42.7%)—and least for children with measured height and weight—by 2.9 percentage points in the NLSY79 at 13.8% (95% CI = 12.3%, 15.6%).
The exclusion of height-for-age and weight-for-age BIVs achieved relatively little bias reduction, reducing the overall NLSY79 obesity prevalence from 19.9% to 18.2% and the PSID prevalence from 27.2% to 25.5%. Even among children with reported weight and height, excluding height-for-age and weight-for-age BIVs reduced obesity prevalence in the NLSY79 only from 32.3% to 29.4% and in the PSID only from 36.6% to 33.5%. The exclusion of BMI-for-age BIVs was more effective in reducing bias, with obesity prevalence falling to 26.9% and 30.3%, respectively, of NLSY79 and PSID children aged 2 to 5 years with parent-reported heights and weights. These are still, however, 2.5 to 3.0 times higher than the 11.0% NHANES obesity prevalence. Excluding BIVs also reduced the NHANES obesity prevalence, from 11.0% to 10.4% when excluding height-for-age and weight-for-age BIVs and to 10.2% when excluding BMI-for-age BIVs. Because of the rigorous NHANES measurement protocols, we interpreted these reductions in obesity prevalence as owing to false positive cases, that is, children whose true measurements of height and weight are incorrectly excluded by the CDC BIV cutoffs.
We next used our simulation model to better understand the sources of upward bias identified in Table 1. We first graphed the 3 parameters of equation (1b), being prevalence of very low height and prevalence of obesity conditional on whether the child was of very low height. In Figure 1, we compared with the NHANES the prevalence of all PSID and NLSY79 children below the first percentile height for age before and after excluding BIVs. Whereas a highly plausible 0.8% of children in the NHANES were below the first height percentile, an implausibly high 11.8% and 7.7%, respectively, of all PSID and NLSY79 children were below the first height percentile. Excluding height-for-age and weight-for-age BIVs reduced the prevalence of such PSID and NLSY79 children by just under 2 percentage points; excluding BMI-for-age BIVs reduced this prevalence by between 3 and 4 percentage points, respectively.
FIGURE 1— Percentage of children aged 2–5 years with height below the CDC first percentile by whether biologically implausible values are excluded: PSID-CDS-I, 1997; NLSY-Child, 1996–2008; and NHANES 1999–2008; United States.
Note. BIV = biologically implausible value; BMI = body mass index; CDC = Centers for Disease Control and Prevention; Excl. = excluding; NHANES = National Health and Nutrition Examination Surveys; NLSY-Child = Children of the National Longitudinal Survey of Youth 1979 cohort; PSID-CDS-I = Panel Study of Income Dynamics-Child Development Supplement, wave I. Percentages are weighted using respective survey sample weights.
In Figure 2, the obesity prevalence of children with very low height and the obesity prevalence of all other children are shown for the PSID, NLSY79, and NHANES. Among PSID and NLSY79 children with height below the first percentile, obesity prevalence was entirely implausible at 87.9% and 82.2%, respectively. Excluding children with BIVs, moreover, does relatively little to reduce this, although again more reduction is achieved when excluding BMI-for-age BIVs than when excluding height-for-age and weight-for-age BIVs. We learn from the NHANES that obesity is, in truth, quite rare among children with very short stature, indeed so rare that children below the fifth percentile height for age had to be pooled to obtain reliable estimates. Only 4.0% of children aged 2 to 5 years below the fifth height percentile were estimated to be obese, compared with 11.1% of children with heights above the fifth percentile.
FIGURE 2— Prevalence of obesity among children aged 2–5 years by whether below or above the CDC first height percentile and by whether biologically implausible values are excluded (PSID-CDS-I, 1997 and NLSY-Child, 1996–2008), compared with NHANES, 1999–2008 (< and > CDC fifth height percentile): United States.
Note. BIV = biologically implausible value; BMI = body mass index; CDC = Centers for Disease Control and Prevention; Excl. = excluding; NHANES = National Health and Nutrition Examination Surveys; NLSY-Child = Children of the National Longitudinal Survey of Youth 1979 cohort; PSID-CDS-I = Panel Study of Income Dynamics-Child Development Supplement, wave I. Percentages are weighted using respective survey sample weights. In NHANES, because there are few children with height < CDC 1st percentile, only the conditional prevalence of obesity among children with height below and above the CDC 5th percentile are presented.
Contributions to upward bias in estimates of obesity prevalence owing to height and weight error in the NLSY79 and PSID are quantified in Table 2. We compared NLSY79 and PSID obesity prevalence before and after the substitution of selected NHANES values in equations (1a), (1b), (2a), and (2b). We did this for all children aged 2 to 5 years and for the subsamples with parent-reported heights and weights. We conducted these analyses first without reference to CDC BIVs and then alternately excluding children with CDC BIVs of height for age or weight for age and with BIVs of BMI for age.
TABLE 2— Total Bias in Obesity Prevalence and Bias Remaining After Substituting NHANES Parameters: United States
| Bias After Substituting NHANES Parameters for Very Low Heightb,c | Bias After Substituting NHANES Parameters for Very High Weightb,d | ||||
| Study or Measurement | Total Biasa | (1) Substitute < 1st Percentile Height | (2) Substitute < 5th Percentile Height | (3) Substitute > 99th Percentile Weight | (4) Substitute > 95th Percentile Weight |
| Aged 2–5 y, including BIVs | |||||
| NLSY79-Child | |||||
| Total | 9.0 | 3.7 | 2.2 | 8.8 | 8.7 |
| Reported weight and height | 21.3 | 9.7 | 6.3 | 20.8 | 20.7 |
| PSID-CDS-I | |||||
| Total | 16.3 | 8.0 | 5.5 | 16.1 | 14.3 |
| Reported weight and height | 25.6 | 11.9 | 10.2 | 25.3 | 22.6 |
| Aged 2–5 y, excluding height-for-age and weight-for-age BIVs | |||||
| NLSY79-Child | |||||
| Total | 7.8 | 4.1 | 2.6 | 7.6 | 7.2 |
| Reported weight and height | 18.9 | 10.7 | 7.1 | 18.2 | 17.7 |
| PSID-CDS-I | |||||
| Total | 15.0 | 8.2 | 5.6 | 14.8 | 13.1 |
| Reported weight and height | 23.0 | 11.8 | 10.0 | 22.8 | 19.9 |
| Aged 2–5 y, excluding BMI-for-age BIVs | |||||
| NLSY79-Child | |||||
| Total | 6.9 | 3.9 | 2.6 | 6.8 | 6.9 |
| Reported weight and height | 16.7 | 9.6 | 6.5 | 16.5 | 17.1 |
| PSID-CDS-I | |||||
| Total | 13.5 | 8.1 | 5.6 | 13.5 | 12.1 |
| Reported weight and height | 20.1 | 11.3 | 9.4 | 20.5 | 17.7 |
Note. NHANES = National Health and Nutrition Examination Surveys; NLSY79-Child = Children of the National Longitudinal Survey of Youth 1979; PSID-CDS-I = Panel Study of Income Dynamics, Child Development Supplement, wave I. Percentages are weighted using sample weights for the respective surveys.
aWe calculated total bias by subtracting the NHANES obesity prevalence from the obesity prevalence observed in the panel survey. In interpreting these as bias, for the “reported weight and height” rows, bias may also be because of nonrandom selection into that measurement protocol condition.
bWe calculated bias remaining by subtracting the simulated obesity prevalence in the panel survey from the observed obesity prevalence in the panel survey in which we obtained the simulated obesity prevalence through substitution of NHANES values for the parameters for very low height or very high weight.
cSubstitutions are of NHANES values for the parameters for very low height. Because of small samples sizes of children in the NHANES with height below the first height percentile, we substituted the NHANES obesity prevalence of obesity among children of all heights below the fifth height percentile for all the PSID or NLSY79-Child low height obesity prevalence parameters in the simulations of both columns (1) and (2).
dSubstitutions are of NHANES values for the parameters for very high weight.
Turning first to the analyses that do not reference CDC BIVs, bias induced by height error below the first percentile contributed almost two thirds of the total upward bias in the NLSY79: substitution of NHANES first height percentile parameters reduced total upward bias in obesity prevalence from 9.0% (“Total Bias” column) to only 3.7% (“Substitute NHANES Low Height Parameters” column [1]). Bias induced by reporting error below the first height percentile contributed approximately half of the total upward bias in the PSID: substitution of NHANES first height percentile parameters reduced total upward bias from 16.3% to only 8.0%.
The amount of upward bias in obesity prevalence induced by height error below the first percentile was much larger in the parent-reported height and weight subsamples of the NLSY79 and PSID. In the NLSY79, excess obesity prevalence compared with the NHANES was reduced by 11.6 percentage points from 21.3% to 9.7% substitution of NHANES first height percentile parameters. The same substitution in the PSID parent-reported height and weight subsample reduced this measure of upward bias in obesity prevalence by 13.8 percentage points, from 25.6% to 11.9%. Substituting NHANES values up to the fifth percentiles of height for age induced substantial additional reductions in bias (column [2]): only 6.3%, or less than one third, of the total 21.3% upward bias remained in the NLSY79, and only 10.2%, or two fifths, of the total 25.6% upward bias remained in the PSID. Substituting NHANES values for parent-reported weight-for-age above the 99th and 95th percentiles reduced the total upward bias in obesity prevalence of children aged 2 to 5 years little in the NLSY79 and by 2 to 3 percentage points in the PSID (substitutions [3] and [4]).
From the lower 2 panels of Table 2, the magnitudes of bias remaining after substituting NHANES values are quite similar before and after excluding BIVs (see especially substitutions [1] and [2]). This means that our measure of total bias contributed by parental reporting error below the first height percentile, after excluding BIVs, is reduced by approximately the amount that the “Total Bias” column amounts are reduced: by about a tenth when excluding height-for-age and weight-for-age BIVs and by about a fifth when excluding BMI BIVs, as seen in Table 1. For example, in the NLSY79 sample with reported heights and weights, excess obesity prevalence relative to the NHANES is reduced from 21.3% to 9.7% by substituting NHANES values of proportions less than the first height percentile and of obesity given in the first height percentile before excluding BIVs, whereas it is reduced from 16.7% to 9.6% after excluding BMI-for-age BIVs.
We estimated that obesity prevalence at aged 2 to 5 years in the parent-reported subsamples of the NLSY79 and PSID was 2.5 to 3.0 times that of the measured NHANES. This magnitude of bias threatens the validity of policy recommendations drawn from research employing parent-reported data on young children3,8–11,39 and raises a critical data problem for analysts seeking to address the recommendations both for an increased focus on early childhood1,40 and for better modeling of the dynamic, multifactorial, life course processes thought to drive epidemic levels of obesity in the United States.12,41
A crucial step toward resolving such a major problem of measurement bias is quantitative evaluation of the sources of that bias. We developed and applied a simulation method to assess bias from using parent-reported data on heights and weights in 2 nationally representative US panel surveys, the PSID and NLSY79. A simulation model’s level of complexity and its degree of focus are often, as in this study, inversely related. Because obesity is defined by exceeding a high distributional threshold point, a BMI that exceeds the reference population’s 95th percentile, our model focused on extreme values of height and weight. This allowed us to specify a very simple simulation model requiring only 3 parameters to estimate the source of one half to two thirds of the total bias in obesity prevalence: height reporting error manifest in the first height percentile. The addition of parameters for prevalence of children in the second through fifth height percentiles and for obesity prevalence conditional on being in those percentiles increased further still the amount of overall bias contributed by low height misreporting. The hypothesis that upward bias in obesity prevalence resulted from overreporting of children’s weight was rejected after applying an analogous 3-parameter model that focused on the misreporting of extremes of the weight-for-age distribution. We found remarkable congruence of results across the 2 surveys. Our simulation model and its estimation, moreover, may easily be expanded to quantify the sources in the bivariate height and weight distribution that account for that portion of upward bias not explained by observations falling below the first height percentile.
We used this same simulation model to evaluate the CDC’s method for identifying individual cases of BIVs of gender-specific height for age, weight for age, and BMI for age. We found that the exclusion of height-for-age and weight-for-age BIVs eliminated around one tenth, and the exclusion of BMI-for-age BIVs one fifth of the upward bias in obesity prevalence at aged 2 to 5 years. Unfortunately, we also found nonnegligible reductions in obesity prevalence when applying the BIVs to the NHANES, reductions that we interpret as owing to false positives: children whose true measurements of height and weight are incorrectly excluded by the CDC BIV cutoffs. We are therefore reluctant to recommend the use of BIVs. If used, the BMI-for-age BIVs are preferable for their identifying implausible combinations of weight and height.
Our simulation approach provides a novel, to our knowledge, method to isolate the types of errors in height or weight reporting that can lead to dramatically inflated estimates of obesity. For primary data collectors seeking to balance measurement cost against bias, our findings underscore the importance of creative approaches to improve parents’ reports of young children’s height, such as providing respondents with a measuring tape and instructions for its use before a telephone interview.42 Identification of reported heights in the first percentile during interactive data collection procedures with respondents is another fieldwork procedure that could be considered. Public health researchers using data already collected, such as those of our study, may consider sensitivity analyses in which they truncate their samples to exclude children reported by parents to be in the first height percentile. For simulation modelers, our study’s bias estimates may also be incorporated in sensitivity analyses.
Acknowledgments
This work was supported by the US National Institute of Child Health and Human Development (grant R01-HD061967) and benefitted from the authors’ participation in the National Collaborative on Childhood Obesity Research Envision Network.
Human Participant Protection
No protocol approval was necessary because we used de-identified secondary data sources throughout.
