Exemplar study from InterConnect consortium

The target variable is derived from harmonisation of the raw data in different studies and should be specified in terms of units.

This variable should be appropriate to answer the research question as well as being dependent on the data and methods used in the different datasets.

In InterConnect, they aimed to harmonise a total of eight variables (total fish, fatty/oily fish, lean fish, shellfish, saltwater fish, freshwater fish, fried fish, smoked/salted fish), all in g/d, across 12 studies.

It is important to know if the existing data have the ability to capture the same latent constructs.

Understanding the specific methods and instruments used in each study as well as the format of the data, the overall study design and any assumptions made during processing within-study data are essential.

In InterConnect, ten studies assessed fish intake using FFQs with two using diet history (a retrospective structured interview method consisting of questions about habitual intake of foods from the core food group).

While all studies could create a “total fish” variable, not all could contribute to the seven other variables.

A number of different methods can be applied to derive a common format for the target variable within each study, for example, using a conversion factor or collapsing to the least common denominator.

Applying a conversion factor can be straightforward when the relationship between two units is known, as is the case for converting kilocalories per day to kilojoules per day.

Collapsing to the least common denominator can include recoding or transforming existing data and would involve applying an agreed set of rules or algorithms depending on within-study data availability.

External data can also be used to support deriving a common format. For example, data on average portion sizes could be used in combination with frequency and food type to derive food quantities.

However, this should be applied with caution as the degree to which these values can be generalised depends on the specific study population.

When considering the harmonisation of dietary patterns (DPs), the food groups within each study and the items within these groups should be as similar as possible between the studies.

If using PCA to determine a DP, the coefficients from study will need to be applied to the other to ensure the same DP is being compared.

All of these suggested approaches have limitations which might make it difficult to compare absolute levels of dietary intake across studies. However by ranking individuals in quartiles according to intake or adherence to a DP, a comparison of associations between diet and health outcomes between studies can be made.

For the InterConnect consortium, methods to transform variables from each study to the common target variable were created and agreed with each study.

The two tables – of an example of pre-existing data used to derive target variables (FFQ) and an example of pre-existing data used to derive target variables (diet history) – outline the harmonisation approach taken. There were some specific challenges related to this study. For instance, for some types of fish it was unclear if they should be classified as lean or fatty.

Furthermore, the fat content of certain fish and portion sizes can vary depending on location; therefore local knowledge was required to make these decisions.