Limitations

View the most recent version.

The data in these analyses were obtained primarily by computer-assisted telephone interviews and were self- or proxy-reported.  Self-reported data may be affected by response biases such as social desirability; self-reported weight is generally an underestimate of measured weight.³  If an individual’s reporting bias changed over time, it could affect the results.  It is possible that the increased media attention obesity has received in recent years may have changed the magnitude of this bias; however, other analyses (US data) indicate no significant change in the extent of the bias associated with self-reports of weight and height during the 1988-to-1994 and 1999-to-2002 periods.¹⁴

As in all surveys, non-response may introduce bias into the survey results.  While the 1994/1995 longitudinal square weights adjust for non-response at the initial measurement, they do not adjust for subsequent non-response.  Differential non-response may have affected the results.  However, because regression using the person-period dataset does not require a respondent to answer at each cycle in order to include them in the analysis, this bias is somewhat attenuated.  Future analyses should take non-response patterns into account to investigate the possibility that selective attrition is affecting the results.

Because the data can be conceptualized as observations nested within individuals, a growth curve model would be an appropriate approach to analyzing the data.¹⁵   Initial analyses were conducted using a growth curve model in SAS; however, the estimates of SAS PROC MIXED have been reported to be biased when survey weights are used in the estimation¹⁶ and the bootstrapping procedure was not available to estimate variance.  Thus, an alternative approach using a person-period data set was adopted.  While it is less efficient than a growth curve model, it is unbiased and allowed for variance estimation using the bootstrap procedure.