Time trend estimation of food consumption in repeated studies with different versions of food questionnaire among Brazilian schoolchildren aged 7 to 11 years

ao longo dos anos


introduction
One of the most frequent complaints of researchers in nutritional epidemiology is the impossibility of comparing the findings from the studies based on different methodologies.The latter include different dietary assessment methods (e.g.food frequency questionnaire, 24h dietary recall/ record, food diary), different definitions of the period analysed (e.g. last year, last month, previous day, typical day), varying number of days, weekdays and/or weekend days, the source of diet report (e.g.parents and children, only children), varying number and definition of food items or food groups, the differences in the target population and sampling, in questionnaire presentation (e.g.pen-and-pencil, computer/tablet/smartphone screen), in instructions to the respondents, and in the choice of statistical methods used to analyse the data [1][2][3][4] .
A special case of methodological differences emerges from evolving versions of the same questionnaire over time to accommodate the changes in the food choices and eating behaviour.With nutritional surveillance and monitoring becoming more prominent as an essential tool in combating obesity worldwide [5][6][7] , addressing the above issues in a quantitative analysis becomes even more important.In particular, time trend analysis over a period sufficiently long (e.g. over a decade) to capture significant changes in dietary intakes is very likely to encounter aforementioned methodological differences as survey means and presentation mode change rapidly with scientific and technological advances.
Our research team has developed and validated the Typical Day Food and Physical Activity questionnaire (acronym DAFA in Portuguese) 8 , the Previous Day Food Questionnaire (PDFQ) (Versions 2 and 3) 9,10 , and the Food Consumption and Physical Activity for Schoolchildren (acronym WebCAAFE in Portuguese) [11][12][13][14] which evolved from a pen-and-pencil food frequency questionnaire (FFQ) to a web-based food and physical activity questionnaire.
Funding agencies encourage collaboration between the researchers working in the same area to use financial and human resources more rationally 15 .In this context, quantitative data synthesis strategies such as meta-analysis have been highly appreciated, although a more flexible framework has been put forward, capable of evaluating time trend even for relatively small and heterogeneous groups [15][16][17][18][19] .
The objective of this work is to present regression framework for estimating time trend in food consumption and eating behaviour with different versions of a food questionnaire and apply it to analyse food/beverage consumption trends over the 2002-2015 period among 7-11-year-old schoolchildren in Florianopolis, southern Brazil.

methods
Two longitudinal studies provided the data for this analysis: one from the 2002 and 2007 surveys, and the other from the annual surveys over the 2002-2015 period (Table 1).The studies were conducted according to the guidelines set out in the Code of Ethics of the World Medical Association (Declaration of Helsinki) and all procedures involving human subjects were approved by the Human Studies Committee of the Federal University of Santa Catarina (protocols 037/02, 028/06, 2,250/11).Written informed consent was obtained from the parents and oral assent was obtained from the children.
In 2002, 16 elementary schools stratified by type (public, private) and city region (central, coastal) were randomly sampled from 122 eligible schools with children enrolled in the 2nd to 5th grades.In the second stage of the sampling, the school classes were randomly sampled within selected schools and all children in those classes were invited to answer the DAFA FFQ.The final sample comprised 2,936 schoolchildren (1,988 from public and 948 from private schools) 20 .
All 16 schools were invited to participate in the second survey in 2007.Fourteen schools accepted the invitation and the two that refused were replaced by other similar schools; one more public school was included in the sample.The target population was all schoolchildren between the second and the fifth grade in municipal elementary schools.The final sample obtained consisted of 1,232 children from 17 schools (782 from public and 450 from private schools) 21 .
The 2013-2015 surveys were restricted to public schools and included about 95% of these.Few schools without a computer room were excluded from data collection.Primary sampling units were 2nd to 5th-grade classrooms, all of which participated in the surveys.Mentally handicapped and visually impaired children were excluded from the survey.Among the children with informed consent and complete information on age, sex, weight, height, and food consumption, 9.1% were excluded because of implausible dietary data such as reporting less than four food items per day or out of the mean ± 3 standard deviations (SD) interval 14 .
In 2002 the children responded the DAFA FFQ.It was a pen-and-paper pictorial questionnaire which asked the children to check all of the 16 food drawings (Table 2) consumed on a typical day.The DAFA FFQ was validated against the 24h recall method in a sample of schoolchildren from the city of Florianopolis and showed moderate agreement 8 .
In 2007, the children answered the third version of the Previous Day Food Questionnaire (PDFQ-3), a new version of the questionnaire applied in 2002.New elements were the following: a) children were asked to mark the food/ food group items consumed on the previous day instead of on a typical day, b) evening snack was added as the sixth eating event and included the period between dinner and going to sleep, and c) four more drawings were added (green leaves, coffee with milk, vegetable soup, salty crackers) and one was withdrawn (egg).Trained researchers instructed the children to mark their typical food/beverage consumption on the previous day by crossing or circling the drawings.PDFQ-3 contained 32 food/beverage items (Table 2) and was previously validated through direct observation of school meals, demonstrating reasonable average sensitivity (probability of correctly reporting a food intake) of 70.2% and excellent average specificity (probability of correctly not reporting a food intake) of 96.2% 10 .
Since 2013, the questionnaire expanded the number of food/beverage items to 32, substituted the pen-and-paper mode with computer screen and used the Internet to access the questionnaire [11][12][13] .An animated robot-like Avatar guided children while answering the questions.Before closing the block on food consumption, the pupils were presented with a tray of the foods/beverages they selected for each meal and asked to check and revise their answers if necessary.The validation study showed 43% of matches between WebCAAFE and observed dietary intake in school, as well as 29% of intrusions and 28% of omissions, putting this questionnaire's accuracy close to that of other similar instruments 22 .
The food selection for these questionnaires was motivated by the food patterns of children in this age group already consolidated in the literature but also included the foods presented in school menus and those recommended in the guidelines for Brazilian population 23,24 .

Data analysis
Private school pupils were excluded from the analysis in the present study to reduce the sample heterogeneity, so that the analytical sample was representative of the 2nd to 5th-grade public school pupils in the municipality of Florianopolis, the Santa Catarina state capital, Brazil.Survey parameters included schools as primary sampling units and probability weights for selecting schools and children within schools in the twostage random sampling design.All pupils from selected classes were invited to respond to the food frequency questionnaires (FFQs).
The clustering of pupils within schools was addressed by including this source of co-variation within primary sampling units among the survey design parameters.Hubert-White robust maximum likelihood estimators of food consumption rate per day and corresponding 95% confidence intervals (CI) were calculated by statistical package Stata 13.0 25 .
Main outcomes included the mean frequency of food consumption (MFC) for selected single food/beverage items (fruits, beans, sodas, milk/ cheese) and food groups (dairy products, meat/ chicken/fish/seafood, fruits and vegetables, fast food, sweets, sugary drinks).Key exposure variables were questionnaire presentation mode (pen-and-pencil vs. computer screen) and the instruction given to the pupils regarding the period of food consumption (previous day vs. typical day).Control variables included child age (rounded to full year), sex, school shift, family income and the day of the week the survey was applied.By definition, main outcomes had no missing values as unmarked food items were considered not consumed.
Ordinary least squares regression was used to estimate the linear change in the MFC.The means were weighted by the inverse of their variance estimated for each survey year and were not initially adjusted for the control variables.Also, multivariate Poisson regression, adjusted for the control variables, was used to obtain survey year point estimates and post hoc contrast estimates of interest.The maximum number of foods/ beverages per day was used as an offset in Poisson regression, thus producing rate estimates, hereby denominated as food consumption rate (FCR) per day.The rate allows comparability of a varying number of meals (five in the 2002 questionnaire vs. six since 2007) and food group components (e.g.adding green leaves to the fruit & vegetable group).The offset was the product of the number of meals and the number of food components for food groups, or simply the number of meals per day for single food/beverage items, keeping in mind that respondents were asked only whether these items were consumed or not for each meal/snack.In both cases, the offset was the maximum number of items achievable per day.
Additional statistical analyses were performed to evaluate the time-related changes and were guided by two major issues: the effect of asking about typical versus previous day food/ beverage consumption and of pen-and-paper versus computer screen presentation of the questionnaire.The former question was addressed by comparing the 2002 and 2007 FCR estimates because both used the pen-and-paper version, thus holding this factor constant while varying the question form.The 2013-2015 versus 2007 contrast was used to evaluate the presentation mode effect as the question mode was held the same within the periods compared.All contrasts were Furthermore, interval regression was used to estimate likely MFC in 2002 had the respondents been asked to check out the food/beverage items consumed on a typical day instead of on a previous day.To define the upper bound for this regression, the reasoning was that the typical day question tended to expand the food report to the number of items ever consumed, at least for some children, because averaging a variety of these items over an undefined period was a daunting cognitive task.Without understanding the meaning of the word "typical", children are likely to ignore this part of the question and answer the remaining part of "What do you eat on a typical day?", which is probably interpreted as "any day" and is thus likely to stimulate intrusion (false positive) reports.Consequently, the upper bound was set to the reported 2002 food/beverage consumption, whereas the lower bound was set to the median 2007 consumption for each combination of the aforementioned independent variables.Fitted values from the interval regression were assigned precision equal to the inverse of their variance (so-called "analytical weights") in 2002 to account for their hypothetical nature as these values were not actually reported, while all other data were observed.Time trend and contrast analyses used these hypothetical 2002 estimates to align the FFQ period of food/beverage consumption to the previous day for all surveys.Statistical significance for the type I error was set at < 0.05.

Results
The analytical sample size was about twice as big over the 2013-2015 period compared to the average of the 2002 and 2007 survey (Table 3).Of note, there was a six-fold increase in the percentage of the implausible food reports excluded from the analysis, from about 0.5% in the first two surveys to approximately 3% in the subsequent surveys.
MFC per day and its variation were more homogenous in the 2013-2015 period compared to the 2002-2007 period when pen-and-pencil versions of the questionnaire were applied (Table 4).Higher mean values were predominant in the first two surveys (Table 5).Unadjusted linear trend for the MFC weighted by precision showed statistically significant reductions for the animal proteins which included beef/poultry, fish/seafood and eggs (6% p/year), milk/cheese (4.2% p/year), dairy products (10.3% p/year), fruits/vegetables (7.7% p/ year), sweets (9.6% p/year), sodas (6.5% p/year) and sugary drinks (12.7% p/year) (Table 3).The reductions of 4.1% p/year for pizza/hamburger and 2.1% p/year for fruits were marginally significant (p-values of 0.058 and 0.065, respectively).
There was a significantly higher FCR for all food/beverage items reported in the 2002 survey using the typical day question (flagged as 2002 a in Table 4) compared to the likely FCR had the question been asked of the previous day instead (flagged as 2002 b ).The 2002 and 2007 pen-andpaper questionnaires rendered significantly higher values than the 2013-2015 computer screen prompts for foods/beverages analysed except yoghurt (penultimate column in Table 4).This comparison used the 2002 interval regression estimate of the hypothetical question on the previous day food consumption.To avoid the uncertainty related to this potential outcome, another comparison was made excluding the 2002 data, i.e. contrasting the 2013-2015 computer screen versus 2007 pen-and-pencil reports, all of which referred to the previous day food consumption.Although the same FCR reduction trend for the later period was confirmed, its magnitude was  2 Fruits and green leaves/vegetable soup in 2002; in addition to fruits, green leaves, vegetables and vegetable soup presented as separate icons in 2007; fruits, vegetables and green leaves kept in 2013-2015. 3Hamburger/pizza and French fries in 2002; salty snacks added in 2007 and instant pasta, nuggets and canned food in 2013-2015. 4included chocolate bars and powder in 2002 and 2007; presented four separate icons in 2013-2015: chocolate drinks, biscuits without cream, biscuits with cream and cakes. 5Sodas and sugar-added fruit juice.Source: Authors.smaller and its statistical significance (P < 0.05) included fewer items, namely fruits, meats/fish/ seafood, milk/cheese and other dairy products, sweets, sugary drinks and sodas (last column in Table 4).Among the significant reductions, the largest one was observed for milk/cheese (62% with 95% CI 56-67%) and the smallest was found for the meats/fish/seafood (17% with 95% CI 11-23%).

Discussion
The present study provided an example of time trend analysis based on different FFQs, namely pen-and-paper versus computer screen presentation mode, varying number of items within food group over survey years and variations in probabilistic sampling procedures.Despite uneven spacing of the surveys over the 2002-2015 period and above methodological differences, both linear (per year) and contrast type changes were evaluated.To the authors' best knowledge, this is the first time a data synthesis of this kind was published in nutritional epidemiological studies of children and adolescents, although some general recommendations for trend analysis have been provided 26 .steps were taken to improve the time-related estimates.First, inclusion criteria (public schools only) and exclusion criteria (implausible frequency of total food consumption per day) were used to reduce the analytical sample heterogeneity and thus both known and unknown confounding effects.Second, control variables were used to further reduce the impact of known confounding variables.Third, offset by maximum food/beverage consumption per day allowed comparability of varying number of food components in a food group over survey years.Fourth, survey parameters accounted for unequal selection probability of pupils and higher similarity of their responses within the same school.Fifth, interval regression used the available FFQ data to estimate a likely value of the 2002 survey had it used the same reference period for the food/beverage consumption as the subsequent surveys (i.e. the previous day), thus allowing a more consistent time trend analysis.
The FCR estimation allowed several types of time effects to be analysed: a) overall time trend assuming constant linear change in both unadjusted and adjusted regression analysis, b) the 2007 versus 2002 survey year contrast to evaluate the effect of switching from the typical day to the previous day question of the food/beverage consumption and adding more food items and improving instruction to children before responding these pen-and-paper questionnaires, and c) 2013-2015 versus 2002-2007 contrast to evaluate the impact of changing the questionnaire presentation from pen-and-pencil to computer screen mode.
Non-overlapping CI for adjusted FCR comparing the 2007 versus 2002 survey year indicated a significant reduction in consumption of all foods/beverages analysed, except yoghurt (Table 4).Changing the instruction to answer about typical day food/beverage consumption in 2002 to a previous day consumption in 2007 may have reduced intrusion of some foods which could appear when "typical" was interpreted by child as "possible" over a longer period (e.g. a week) or confounded with desired foods/beverages or simply not understood and ignored.Also, the reduction may be attributed to improved training of the respondents in 2007, including a short reminder of the meaning of food drawings presented as a classroom poster, before answering the PDFQ-3.Another possible explanation is that expanding the number of drawings in the latter version facilitated pupils' understanding of the menu.
The overall unadjusted linear time trend analyses pointed to a downward direction of the MFC for most of the foods/beverages analysed.However, the linear trend assumption may be too restrictive for some foods/beverages, given the methodological changes made across surveys.For example, 2013-2015 versus 2002-2007 contrast captures the putative effect of changing both the presentation mode and expanding the number of food/beverage options in the questionnaire.Although adjusted FCR in principle accounts for these changes, the contrast estimates are more robust against residual confounding than annualised linear time trend.
The avatar may invoke a more playful than a task-oriented set of mind compared to a classroom teacher's instruction, which in turn may reduce attention and memory effort needed to retrieve information on food consumption on the previous day, thus increasing the omissions in the questionnaire report.Consistently lower MFC obtained by WebCAAFE compared to the penand-pencil versions of the questionnaire for most of the foods analysed pointed to this source of residual confounding.Also, a six-fold increase in implausible food reports in the 2013-2015 period survey compared to the 2002 and 2007 surveys pointed in the same direction as we hypothesize that the screen presentation mode with an avatar in the later period contributed to a less task-oriented environment, thus resulting in lower precision of food reports.We further hypothesize that food report omissions may be more likely than intrusions in such an environment.
Judged by statistical significance of the contrast between the 2013-2015 versus 2002-2007 period, strong evidence was found for reduction in both healthy (fruits, milk/cheese, meats/fish/ seafood) and unhealthy diet markers (sodas, sugary drinks, sweets), whereas such evidence was insufficiently strong for pizza/hamburger, fast food, yoghurt, beans, and fruit and vegetables consumption.The 2013-2015 FCR values, obtained with the same questionnaire presentation mode, showed stable point estimates.Also, expanding the number of food/beverage items from 21 in 2007 to 32 in 2013-2015 should have facilitated the recall of consuming these items the day before.More research is needed to clarify and quantify residual confounding factors influencing food report errors such as omissions and intrusions.
A notable sparsity of time trend studies on food consumption in early school age makes it difficult to compare the present study findings with those of the other trend analysis, such as similar studies with adolescents.Nevertheless, the latter overlaps with the former age band, so an attempt can be made to juxtapose their findings.In Brazil, unhealthy eating habits showed a significant increase, especially among adolescents from low-income families 27 .The consumption of ultra-processed foods has increased at the expense of unprocessed foods such as rice, beans, and fruits, although sugary drinks' intake has decreased in the last decade 28,29 .A similar decreasing trend in the consumption of sugary drinks was found in the USA among children and adolescents 30,31 .European adolescents have increased their intake of fruits and vegetables over the decade of 2010 32 .In most developed countries, the consumption of dairy products decreases significantly in adolescence compared to childhood 33 .
Several important limitations of this work should be noted.First, it was impossible to clearly distinguish between period-related and the FFQ method-related changes as they largely coincide.For example, pen-and-paper FFQs with a lower number of food/beverage items were employed in the earlier surveys in 2002 and 2007, whereas the number of these items increased 50% (from 21 32) and their presentation was made on the computer screen in the later 2013-2015 period.Comparing the changes within and between presentation mode provided some clues as to the FCR secular trends adjusted for known confounding factors but it does not eliminate residual confounding.Also, food choices are often correlated and so are their reporting errors but no adjustment was made to accommodate this problem.Furthermore, all questionnaires applied in the present study have the limitations inherent to this dietary assessment method such as memory error 1 .Some of the limitations were dealt with statistically (FFQ presentation mode end the number of food items per survey) but those inherent to the questionnaire method remained (memory error).Although internal consistency of food reports was not available with a single FFQ application 34 , the latter covered four weekdays and a weekend day (Sunday).Taken together, these information point to a likely underestimate of time trend variance but not necessarily to its bias.
Among the strengths of this work, it is worth highlighting the use of the food questionnaires evaluated for both internal 8 and external validity 10,12,13 , as well as for usability 11 .Also, the twostage random probabilistic sampling provided a solid representation of the target population in the first two surveys whereas the subsequent surveys achieved almost full population coverage.A large number of participants resulted in high power of statistical tests.Furthermore, the regression approach employed here may have wide applicability for other data syntheses because it used a standard statistical software and procedures to achieve comparability of the impact measures of interest (incidence rates, odds ratios, contrasts).
Web-based surveys have reduced costs and time for the FFQ application and thus made it a viable tool for food consumption monitoring at the population level.Future research needs to refine these instruments regarding respondent cognitive capacity, reliability and validity, based on usability testing and calibration studies.
In conclusion, a regression framework can be used to adjust for methodological differences in the number and/or definition of food items included in a food group.Such differences often arise in longitudinal research with different versions of the same food questionnaire.Both overall time trend and specific contrast analyses may be applied to discern between method-related and time-related components.A significant decrease in consumption of both healthy (fruits, animal protein) and unhealthy foods/beverages (sweets, sodas, sugary drinks) was observed in Florianopolis over the 2002-2015 period.

Collaborations
All authors were involved in analyzing the studies, reviewing and interpreting the results and writing the manuscript.All authors takes responsibility for all aspects of the reliability and freedom from bias of the data presented and their discussed interpretation.

Table 2 .
Food and beverage items by survey year.

Table 3 .
Sociodemographic characteristics of 7-11-year-old schoolchildren from public schools by food survey year in Florianópolis.

Table 4 .
Mean frequency (standard error) of food consumption per day by survey year and average annual change assuming linear time trend over the 2002-2015 period.

Table 5 .
Food consumption rate per day: point estimates for each survey year, 2013-2015 vs. 2002 and 2007 post hoc contrast ratios from multiple Poisson regression.
1,22002 indicate Poisson and interval regression estimates, respectively, for that year; FCR = food consumption rate adjusted for sex, age, family income, school shift and day of the week the questionnaire was applied; CI = confidence interval; SE MFC = standard error of the mean food consumption; FCRR = food consumption rate ratio adjusted for age, sex, family income, school shift and day of week the questionnaire was applied.Source: Authors.

Table 5 .
Food consumption rate per day: point estimates for each survey year, 2013-2015 vs. 2002 and 2007 post hoc contrast ratios from multiple Poisson regression.