Reliability of Self-reported Weight and Height among State Bank Employees

Reliability of self-reported weight and height Chor, D. et al. Abstract Introduction Self-reported weight and height were compared with direct measurements in order to evaluate the agreement between the two sources.


INTRODUCTION
Obesity has been increasingly related to chronic health problems like cardiovascular disease, diabetes and cancer 13 .There is evidence that average body weight has been increasing during the last decades in several countries 12 .In a study carried out in 1989 8 , prevalence of overweight in Brazil was 27% for men and 38% for women over 18 years of age.In fact, changes in nutritional status in Brazil are impressive: between 1974/1975 and 1989, prevalence of overweight and obesity in men increased 56.3% and 92.0%, respectively; women showed increments of 39.7% and 69.6% in the same period 19 .As a consequence of these considerable changes, weight, height and body mass index (BMI) have been increasing by stressed in epidemiological research.
Direct methods such as body density and total potassium can be used to classify individuals as obese (excess of body fat) 6 .Nevertheless, most clinical and epidemiological investigations have used BMI, which can be calculated very easily.BMI can be considered a proxy measurement for the amount of fat, as body constitution can vary between individuals with the same weight and height 13 .Its limitations are related to the fact that BMI does not measure the body fat distribution, which has been related to the risk of coronary disease, and does not distinguish lean body mass, fat body mass and edema 10,13 .Taking these limits into account, individuals with BMI over a specific cut-off level are classified as obese, as there is a high correlation between this index and direct measurement of body fat.Another advantage of the BMI is that it is not related to height.According to Anjos 2 , BMI seems to be a valid index for measuring nutritional status in populational groups and has been widely used for this purpose.
Although direct measurements of weight and height can be easily obtained, they are not always collected.In this case, self-information is used.Comparing self-information with direct measurement is essential and several studies have shown its validity in Brazil and in other countries 5,15,16,20 .
In the present paper self-reported weight, height and estimated BMI are compared with the direct measurement of a group of employees in a State Bank in Brazil.Measures of reliability were estimated and their advantages and disadvantages are discussed.

Study Design
A cross-sectional study was carried out with employees of the bank, working in the state of Rio de Janeiro.As the main purpose was to evaluate the health status of these employees, questions were basically related to morbidity, risk factors for cardiovascular disease and use of health services.Probabilistic samples were obtained for three different strata.Data were collected at work between August and December 1994, using a selfadministered questionnaire.Weight was reported in kilograms and height in centimeters.Direct measurements of weight and height were carried out on a subsample of employees who answered the questionnaire in four different units of the bank.

Direct Measurements
Weight was measured after the questionnaire has been filled out.Portable electronic scales (Filizola) were used on flat, hard surfaces.Individuals took off their shoes and removed all objects from their pockets.Measurements were recorded in kilograms and grams (intervals=100 g).Height was measured using a metallic tape (Stanley) fixed on the wall according to instructions.Height was recorded in meters and centimeters (intervals=1 cm).

Training of Examiners
Measurements were taken by two researchers trained according to the methods proposed by Habicht 9 .Training sessions included repeated measurements made by each research assistant at intervals of 15 min.These measurements were compared to those performed by the instructor, used as the gold standard.Intra-rater comparisons evaluated reliability, and differences between research assistants and instructor were used to evaluate validity.

Data Analysis
Errors in reporting weight and height were calculated by the difference between self-reported and direct measurements.Positive values represent an overestimation and negative values represent underestimation of the true value.
Self-reported and direct measurements were also compared using the methodology proposed by Bland and Altman 4 and Altman 1 , which consists of a graphical display of the differences (self-reported -measured) against their mean (self-reported + measured / 2), the limits of agreement and respective confidence intervals.The limits of agreement are estimated by the mean difference ± (1.96 * standard deviation of the differences), assuming the differences are normally distributed.Confidence intervals for these limits are calculated by the limits of agreement ± (t * standard error of the limit), where t-value is provided by the t-Student distribution with n -1 degrees of freedom and significance level of 95%.The standard error of the limits of agreement is approximated by (3s 2 /n) 1/2 , where n represents the sample size and s is the standard deviation of the differences.
Agreement between information and direct measurement was also investigated using the intraclass correlation coefficient (ICC), which is appropriate for continuous variables, like weight and BMI.It has a simple and direct interpretation and can vary from -1 to + 1 17 .The coefficient was estimated using information on mean squares obtained from the analysis of variance (ANOVA) table, where each individual was regarded as a level of the factor "INDIVIDUAL".Two-way ANOVA was used as the information on weight and BMI was obtained using the same sources (questionnaire and direct measurement) for every participant.Individuals and researchers were regarded as a random factor (the opposite of fixed) as they were selected to represent the target population.ICC was calculated using the following expression 18 .

IC =
where BMS = between targets mean square EMS = mean square expectations JMS = between judges mean square n = number of targets k = measurement sources BMS, EMS and JMS were obtained using SPSS for Windows 6.12 14 for: statistics/scale/reliability analysis/statistics/F test.

RESULTS
Weight and height were measured in a subsample of 322 employees (51% of men) in four units of the bank, of the 1,183 who filled out the self-reported questionnaire (Table 1).Mean age was 38.5 years with a median of 39.3.Educational level was high: 89.0% of the population had been to university, although not all had finished it.
Self-reported weight was underestimated as compared to direct measurement for men and women from all four units.Maximum mean differences were -1.016 kg and -1.275 kg for women and men, respectively (Table 2).There was no uniform pattern for height, and differences were considered insignificant.

BMS-EMS BMS+(k-1)EMS+k(JMS-EMS)/n
Figures 1-4 present the differences between selfreported and directly measured weight plotted against their means (self-reported + measured / 2).All units showed a trend for underestimation as there were     more points under the horizontal line representing the zero value (which means no difference between self-reported and measured values).Moreover, it was possible to identify distinct patterns of agreement for men and women, which were also dependent on weight values.At Central Branch (Figure 1) most women underestimated their weight, regardless of its value.For men, there was no trend below 80 kg.Above this value, men tended to underestimate their weight.In the Ilha do Governador Branch the same trend was observed for men.
Although the mean difference between selfreported and measured weight was never greater than 1.275 kg (see Table 2), the limits of agreement (mean difference ± 1.96 * sd) ranged from -5.8 kg to +3.8 kg.Thus 95% of the differences must be within these limits, meaning that self-reporting can underestimate weight by almost 6.0 kg or overestimate it by almost 4.0 kg (Table 3).If one takes into account the confidence intervals of the limits of agreement, the difference may be even greater.
When the BMI was studied instead of weight, the mean differences, limits of agreement and confidence intervals were found to be smaller (Tables 4 and 5).
The BMI estimated on the basis of self-reported information was also understated, and the mean differences in the four units ranged from -0.621 kg/ m 2 to -0.194 kg/m 2 (Table 4).The limits of agreement ranged from -2.29 kg/m 2 to 1.74 kg/m 2 (Table 5).
Figures 5 to 8 present closer agreement for the BMI than for weight.Actually, 75% of the disagreements were under 0.237 kg/m 2 at the Center Branch, 0.292 kg/m 2 at the Internal Services unit, 0.302 at the General Board, and 0.173 kg/m 2 at the Ilha do Governador Branch.Differently from the trend observed for self-reported weight in some units, no systematic error concerning the information on BMI was found.
Agreement between measured and self-reported weight was almost perfect (Table 6).The intraclass correlation coefficients (ICC) were over 0.98 at all units.These coefficients were also very high for BMI (over 0.95).Although mean differences were smaller for BMI than for weight, ICC were smaller for BMI, employing lower levels of reliability.This difference is probably to be explained by the fact that variability was smaller for BMI than for weight, which influenced ICC values.

DISCUSSION
Agreement between self-reported and measured weight was very high among employees of the State Bank.The mean of the differences between the two sources ranged from -0.741 kg (sd=1.76kg) to -1.057 kg (sd=2.21kg).Agreement measured by the intraclass correlation coefficient (ICC) was also very high not only for weight, but also for body mass index (BMI).It is unlikely that the high reliability observed could be explained by the fact that employees knew that their information would be compared to direct measurements.This hypothesis is based on the lack of difference between mean BMI in the group that was investigated with direct measurements and the one without it (p>0.10)7.Those findings suggest that    self-reported weight and height can provide reliable information for epidemiological studies on overweight and obesity in groups with a high level of education and health care.
Despite the small magnitude, the means of the difference between self-reported and measured weight in our population were greater than that observed by Schmidt et al. 16 in a sample of the general population of Porto Alegre, in Southern Brazil, between 15 and 64 years old (mean=-0.06kg; sd=3.16kg).Another difference between these studies was that, in Porto Alegre, women tended to underestimate their weight and men to overestimate it, according to a pattern identified by others 15 .In this case, employees from the State Bank tended to underestimate their weight, independently of gender.
It is important to emphasize that, despite criticisms 3,4,11 , Pearson's correlation coefficient has been used as an index of reliability.Its main problem arised from the fact that Pearson's coefficient measures correlation or linear relatedness, and not agreement.Consequently, Pearson's coefficient can reach high values when there is disagreement between two measurements, since the bias is systematic.If one rater finds values that are twice as great as those of another rater (multiplicative bias), Pearson's coefficient can reach its maximum value of +1.The same will happen if one rater finds values that are always two units above the values found by the other (additive bias).It is not correct to say that there is perfect agreement between the two raters.
The graphical method proposed by Bland and Altman 4 provides the visualization of both multiplicative and additive bias by the existence of points plotted above or below the horizontal line passing through the zero on the y-axis.Moreover, the presence of a heterogeneous pattern of reliability through different levels of the measured variable can be more easily identified by a quick look at the graph.In Figure 1, the distribution of the points suggested that women were more likely to underestimate their weights, but men usually did so if they were over 80 kg.
One limitation of the method proposed by Bland and Altman 4 is the absence of a standardized index summarizing the agreement between the two sources of measurement.Nevertheless, one can deal with this limitation by using the intraclass correlation coefficient (ICC) which "combines a measure of correlation with a test in the difference of means" 20 .In regression language, the intraclass correlation coefficient "assesses not only similarity of slopes, but also similarity of intercepts" 11 .So, differently from Pearson's coefficient, the ICC is not affected by systematic bias.
Combining the graphical approach of Bland and Altman 4 with ICC allows the identification of heterogeneous patterns of agreement.The presence of heterogeneity indicates the need to estimate ICC for different level of the variable studied (for example, below and above 80 kg for men).The two methods may complement each other in pilot studies aiming to evaluate the agreement between selfreported and measured weight and height.Identifying groups with reliable information for those variables may justify the use of self-reported values, thus making field work cheaper and easier.

Figure 1 -
Figure 1 -Mean and difference between self-reported and measured weight (Center Branch).

Figure 2 -
Figure 2 -Mean and difference between self-reported and measured weight (Internal Services Unit).

Figure 3 -
Figure 3 -Mean and difference between self-reported and measured weight (General Board).

Figure 4 -
Figure 4 -Mean and difference between self-reported and measured weight (Ilha do Governador Branch).

Figure 6 -
Figure 6 -Mean and difference between self-reported and measured BMI (Internal Services Unit).

Figure 7 -
Figure 7 -Mean and difference between self-reported and measured BMI (General Board).

Figure 8 -
Figure 8 -Mean and difference between self-reported and measured BMI (Ilha do Governador Branch).

Table 1 -
Distribution of employees by unit and source of information.

Table 2 -
Mean differences between self-reported and measured weight and height by unit and under gender.

Table 5 -
Limits of agreement and confidence intervals for BMI by unit (kg/m 2 ).

Table 3 -
Limits of agreement and confidence intervals for weight by unit (kg).

Table 4 -
Mean differences between self-reported and measured BMI by unit and gender.

Table 6 -
Intraclass correlation coefficient for self-reported and measured weight and BMI by unit.