Food consumption evaluation has been commonly used in epidemiological studies since it correlates with health and chronic disease determinants (1, 2). However, there are countless challenges related to this process, including intra- and interpersonal variabilities, interviewer and/or interviewee bias, and adherence to assessment instruments.

In recent years, face-to-face data collection has decreased in epidemiological studies worldwide (3), parallel with an increase in Internet access (4). Thus, the development of instruments in the virtual environment for data collection has become a promising trend (4), including assessing food consumption (5), mainly when large populations are involved.

Online data collection allows sending alerts to users, reducing the study load for participants, and maintaining a distance between researchers and subjects, limiting self-censorship in the interview. In addition, it facilitates processing and causes greater reliability of information due to the multimedia support and elimination of steps related to data entry or scanning of paper forms (6, 7).

However, there are still few epidemiological studies in Brazil that have used online questionnaires for food consumption assessment (8, 9). In this context, the present study, the Cohort of Universities of Minas Gerais (CUME project, Brazil), aimed to evaluate the relationship between the Brazilian food pattern and the nutritional transition on non-communicable diseases (NCDs) in higher education graduates of federal universities located in the state of Minas Gerais, Brazil (10).

To achieve the objective of the CUME project, an online and self-administered Food Frequency Questionnaire (oFFQ) has been used. This method demonstrates the practicality in obtaining and analyzing data, at a low cost, and the possibility of investigating food consumption over a long period (2, 11, 12).

On the other hand, the Food Frequency Questionnaire (FFQ) has limitations that include the use of a previously defined list of foods, dependence on the memory of previous eating habits, and difficulty in establishing the precise amount of food consumed (13). Besides, the FFQ must have good accuracy, requiring its validation for the target population; the use of data collection instruments previously validated in other studies is not recommended. In addition, the FFQ must have good reproducibility to increase the guarantee that the data provided by the interviewees are not due to chance.

Thus, we aimed to describe the construction of the oFFQ used at the CUME project and evaluate its validity and reproducibility.

Approximately 146 participants (66.4% women; 34.4 ± 8.6 years old) from CUME were included in this study (Table 1). Most of them declared to be white, with individual incomes of up to five minimum wages and full-time jobs. Regarding lifestyle, 8.2% reported smoking, 70.5% consumed alcoholic beverages, and 78.8% practiced physical activity at least once a week. In addition, overweight (BMI ≥ 25.0 kg/m²) was observed in 40.4% of the participants.

Regarding the validity of the oFFQ, most of the means of energy, nutrients, and food group consumption in the oFFQ1 were higher than those measured in the 24hRs. Overall, an agreement between the oFFQ1 and the 24hRs was moderate, with an average ICC of 0.44 and the exact + adjacent percentage agreement of 88.1%. Energy and all macronutrients also showed moderate agreement between the instruments, with the following ICCs: 0.41 for energy, 0.50 for carbohydrates, 0.51 for lipids, and 0.59 for proteins. Variations between micronutrients for ICC values were 0.28 (sodium) to 0.65 (vitamin B6). According to nutritional similarity, the ICCs ranged from 0.34 (fats and oils) to 0.62 (fruits) for food groups. Concerning the degree of industrial processing, the concordances were moderate for the ultra-processed foods (ICC = 0.60) and processed (ICC = 0.54) groups, and low for the group of in natura/minimally processed foods (ICC = 0.36) and the group of culinary ingredients (ICC = 0.36) (Table 2).

Moreover, the concordance analysis of Bland and Altman showed that data were homogeneous (Figure 2).

Regarding reproducibility, the means of consumption of energy, nutrients, and food groups were similar. Overall, an agreement between the oFFQ1 and the oFFQ2 was excellent, with an average ICC of 0.76 and the exact + adjacent percentage agreement of 92.5%. The energy and all evaluated items showed moderate-to-excellent agreement, with ICCs ranging from 0.60 (vegetables) to 0.91 (vitamin E and retinol) (Table 3).

In this study, we demonstrated that the oFFQ used at the CUME had moderate validity and excellent reproducibility for total energy consumption and most nutrients and food groups.

Regarding validity, compared with the 24hRs, the oFFQ presented moderate ICC values for most of the items evaluated, results that were similar to the previous national (2, 29) and international (30, 31) studies.

These ICC values may be influenced by the fact that most of the means of energy, nutrients, and food group consumption in the oFFQ1 were higher than those measured in the 24hRs. In this sense, 24hRs provide more detailed and less biased data than FFQ (32). In general, the FFQ overestimates food consumption (33–35).

Although the ICCs of some micronutrients and food groups were low for some of them (iron, vitamin B1, vitamin B12, meat and fish, cereals and legumes, fats and oils, other foods, fresh/minimally processed foods, and culinary ingredients), the values were very close to the limit considered as moderate (0.40); furthermore, the average of the exact + adjacent tertiles agreement was high (88.1%) with all items showing a percentage higher than 80%, a result also close to those observed in Brazilian studies on the subject (2, 36). Interestingly, the mean differences by the Bland and Altman method were minor, and the data were homogeneous. These findings were consistent with those evidenced in a cohort study conducted with a sample of Brazilian middle-aged adults (2).

Low ICCs of some micronutrients and food groups may have been influenced by the following factors: (a) the FFQ consists of a set list of food items. At the same time, the 24hRs allow quantifying all foods and beverages consumed in the period before the interview. This potentially causes consumption overestimation in the FFQ due to the tendency of the participants to indicate greater food intake and rare items in their daily diet (37). The higher averages of energy consumption can reinforce this statement. Most of the nutrients and food groups were measured by the oFFQ in relation to the 24hRs; (b) we used the Multiple Pass method (20) to apply the 24hRs and minimize errors in measuring the diet. This method allowed greater detail of the foods consumed and investigation of sugar or sweetener addition to coffee, tea, milk, and juice, as well as the use of spices, sauces, and olive oil in vegetables, which may have reflected in higher consumption averages of these items in the 24hRs; (c) a higher average consumption of fiber in the oFFQ compared with the 24hRs could be explained by the intake overestimation of source foods considered socially approved, such as fruits, vegetables, and legumes (38); and (d) 24hRs is a method often used to validate the FFQ; however, it presents memory bias and errors when estimating the portion size.

In addition to the validity and reproducibility of nutrients, the present study extrapolated the assessment to food groups since individuals do not consume only isolated nutrients but meals consisting of food and nutrients (39). Therefore, validity and reproducibility by food groups have been carried out, considering only nutritional similarity as a grouping criterion (40). Still, to our knowledge, this is the first study assessing the validity and reproducibility of food groups according to the degree of industrial processing (41).

The oFFQ used at CUME proved to be valid compared with the 24hRs, showing that such a questionnaire can measure what it is intended to measure (29, 42). This result is important not only for this study, but it also has an impact on the nutritional epidemiology and public health areas, since instruments for data collection in the virtual environment (Internet) are very useful and practical in countries with a large geographical extension, such as Brazil, dismissing face-to-face meetings between researchers and participants (6, 7).

Regarding reproducibility, when compared with the oFFQ2, the oFFQ1 presented excellent ICC values in most of the items evaluated, and these results are congruent with those observed in studies on the same subject (29, 42); furthermore, in the analysis based on tertiles, the average of the exact + adjacent agreement of the comparison between the oFFQ1 and the oFFQ2 was almost perfect (92.5%), being this result similar to those evidenced in other studies (2, 37).

Reproducibility is the ability of an instrument to produce similar estimates in two different moments with the same accuracy (29); therefore, the fact that the oFFQ showed to be reproducible is fundamental for the CUME project because evaluations of the diet and food intake of the participants will be carried out on different occasions over time due to its longitudinal design.

To ensure the quality of the information, some precautions were necessary for carrying out the study, such as assessment of seasonal food consumption, with data collection of the 24hRs carried out in two different seasons of the year; applying the second oFFQ in a timely manner to avoid fundamental changes in diet or remembering the answers given in the first questionnaire (43); prior training for all nutritionists with a standardized script using the Multiple-Pass method (20); sending a photo album with utensils and homemade measures to aid participants during the food survey (36). Additionally, it is an innovative study because the validity and reproducibility of food have been carried out according to industrial processing.

Nonetheless, this study also has limitations. We performed data collection of the 24hRs in two seasons of the year (winter and autumn), not considering dietary variability throughout all year. However, other validation studies also opted for the application of two 24hRs with shorter intervals (43, 44) to improve study adherence. The 24hRs interview, including two non-consecutive days, is recommended in the European Food Safety Authority guidance on the European Union menu methodology for nationwide individual food consumption studies (45) and endorsed by researchers of the European Food Consumption Survey Methods group (46). Besides, this study innovated by encompassing the collection of four 24hRs in two different seasons; 12 participants (8.2%) did not respond to the two 24hRs applied in the second moment of the validation study, which could be compromising the assessment of food consumption variability of these participants; there was a loss of participants between the stages of validity and reproducibility of the oFFQ, which could reduce the power of the statistical tests. On the other hand, this fact does not seem to have occurred since all the ICCs were significant and with higher values in the reproducibility stage.

Finally, it is worth highlighting some advantages of the oFFQ method, such as low cost, simple analysis, easy application, does not modify consumption over time and can classify individuals according to their usual eating patterns and to associate them to health conditions, which makes it feasible for their use in population studies (2, 47).

When evaluating the results presented, it is concluded that the FFQ developed by the CUME that could be completed online can be used with satisfactory validity and reproducibility to analyze the association between food consumption and NCDs in adults with a high level of education, the target population of the CUME project; however, correction factors must be applied to some nutrients and food groups in future data analysis of the project involving food consumption.

The datasets presented in this article are not readily available because, Data belongs to a still ongoing longitudinal project involving several academic institutions, and therefore cannot be shared just yet. Requests to access the datasets should be directed to Adriano Marçal Pimenta, adrianomp@ufmg.br.

The studies involving human participants were reviewed and approved by protocol n° 596.741-0/2013. The patients/participants provided their written informed consent to participate in this study.

AP, HH, and JB conceived the study, analyzed the data, and drafted and revised the final version of the manuscript. HA, JM-R, AM, and LS analyzed the data and drafted and revised the final version of the manuscript. AGD, AL, and FO drafted and revised the final version of the manuscript. All authors contributed to the article and approved the submitted version.

This research was funded by the FAPEMIG (State of Minas Gerais, Brazil), grant numbers (CDS-APQ00571/13, CDS-APQ-02407/16, CDS-APQ-00424/17). AP, HH, FO, and JB are research productivity fellows of the CNPq (Ministry of Science and Technology, Brazil). JM-R has received a postdoctoral fellowship from the CAPES (Ministry of Science and Technology, Brazil). AM and AGD has received postdoctoral fellowships from the CAPES. HA has received master's fellowship from the CAPES, and AL is a recipient of PhD fellowship from the CAPES.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.