*Geek Box: Validity
We have discussed the concepts of reproducibility and validity in epidemiology in previous Deepdives, however, repetition always bears fruits in understanding. ‘Reproducibility’ is how well any instrument – in our case a food frequency questionnaire [FFQ] – is the consistency of the FFQ over time in the same individual. ‘Validity’ is the accuracy with which the FFQ compares to an independent measure of diet. Both are important concepts for nutritional epidemiology, but validity in particular given that a common criticism of nutritional epidemiology is that the dietary assessments are unreliable.
The ‘gold standard’ validation process is to have a randomly selected subgroup within the overall cohort complete a 7-day food record, where all food intake is weighed and measured. A food record is currently the most accurate way of measuring diet in a free-living population. As all food is weighed and measured, this is considered the closest to “true” intake as can be achieved. If more than a single 7-day food record is conducted, this also allows for any potential seasonal variation in dietary intake to be accounted for.
Also, day-to-day variation in food intake tends to be larger than dietary variation over the long-term [i.e., we tend to eat similar foods over time, but what we eat on Sunday may differ to Monday each week], which means that the 7-day food records may introduce variation when compared with the FFQ. To deal with this, the researchers can calculate the variation in an individuals’ food record [‘within-person’, and the variation between one individual and others [‘between-person’]. The ratio of the within-person and between-person variance can then be used to correct the day-to-day variability in the 7-day food records.
The answers from the FFQ can then be compared to the data from the 7-day food records, to obtain what is known as the ‘correlation coefficient’ or ‘coefficient of variation’ [CV]. The CV reflects the proximity with which the FFQ answers correlate to the 7-day food record data, and is expressed as a value between -1 to +1. For nutritional epidemiology, it is common for correlations to be in the 0.50 to 0.70 range. A CV of 0.70 or over would be quite high for nutritional epidemiology. For example, if the CV for saturated fat is 0.65, this indicates a relatively high correlation of the measure of saturated fat in the FFQ with saturated fat intake in the measured food record.
Now, two more points on FFQ, i.e., the ‘semi-quantitative’ design and the emphasis on frequency. In a semi-quantitative FFQ, portion sizes of foods are also specified in addition to the question on frequency of consumption, i.e., rather than give frequency options for ‘bread’ the questionnaire will state ‘2 slices of bread’ or ‘250mi milk’ or ‘half-cup rice’. These average quantities are provided within a question on frequency, with up to 9 frequency categories [from <1 serving/month to >6 servings/day]: the high categories on frequency is because research indicates that portion sizes vary less than frequency of intake. Having more frequency options therefore better captures variation of dietary intake, and more reflects average intake over time – which is the exposure that is most relevant for long-term diet and disease relationships.
To calculate levels of intake of a nutrient [or non-nutrient, as in the present study], the frequency of consumption of a food item is multiplied by the nutrient content of the specified portion size. Whenever you are reading a prospective cohort study, it is important to look to the validation method for the dietary assessment [most studies now use FFQ], and check the CV for the nutrient or food which is the exposure of interest in that study: a higher CV indicates more validity in the dietary assessment, and therefore in any relationship between that nutrient/food and the outcome of interest.