2.2. Questionnaire Content and Pre-Testing
The questionnaire was developed in English and translated into the respective national languages, beside English for the UK; Dutch for the Netherlands, Finnish for Finland, Spanish for Spain and Polish for Poland by a contracted professional translation office. The translated versions of the questionnaire were proofread by native speakers of the respective languages who were affiliated with the research consortium. The questionnaires were pretested by the market research agency and the involved researchers in a sample of about 30 consumer panel members for clarity of content, language and wording, overall understanding and length of the survey. Based on this pre-test and feedback the questionnaire was refined and finalised.
The survey began with a short description of the EC-funded project—PROMISS (“PRevention Of Malnutrition In Senior Subjects in the EU”) and the informed consent; this was followed by a screening for sample selection based on gender, age, region and current living condition. The core questions consisted of eight sections: appetite, dietary habits, physical activity, food-related and physical activity attitudes, knowledge and perception of protein and food in the diet, socio-demographics and personal information. Order bias was avoided by rotating items within a question.
Appetite was measured using the simplified nutritional appetite questionnaire (SNAQ) by Wilson et al. [
39]. The SNAQ had been validated by Young et al. [
40] and Hanisah, Shahar and Lee [
41]. It consisted of four questions that participants had to indicate the answer that could best apply to their current situation e.g., “My appetite is…”, with responses on a five-point-scale ranging from “very poor” (=1) to “very good” (=5). SNAQ score was calculated based on the numeric scale (i.e., from 1 to 5) assigned to the choices of each question. Poor appetite was defined as having a total SNAQ score below or equal to 14, as an individual with SNAQ ≤14 has a significant risk of weight loss of more than 5% within 6 months with a sensitivity of 81.5% and a specificity of 76.4% [
42].
Protein intake was estimated using the Protein Screener 55+ (Pro55+) (Vrije Universiteit Amsterdam, Amsterdam, the Netherlands), which is a 14-item questionnaire based on the HELIUS (HEalthy LIfe in an Urban Setting) food frequency questionnaire [
43,
44]. The questions include the consumption frequency and portion size in relation to nine food items, e.g., “In the last 4 weeks, how many slices/pieces of bread did you eat on an average day?” where participants could indicate the amount from none/less than 1 to more than 12. A lower level of protein intake was defined as having a probability higher than 0.3 that the protein intake was lower than 1.0 g per kilogram of adjusted body weight per day (g/kg adjusted BW/day) based on recalibrated models [
44].
The consumption pattern of 10 food items (cereals like cornflakes or muesli, dairy or plant-based milk or yogurt, soup, warm meal, cold meal, dessert, biscuits or cookies, fruits, nuts or seeds) was recorded by means of check-all-that-apply (CATA) questions for seven eating occasions (“breakfast”, “mid-morning snack”, “lunch”, “mid-afternoon snack”, “dinner”, “evening snack”, “nocturnal eating”) or “I do not consume this food”.
Appendix A Table A1 presents the frequency table for the consumption pattern of the 10 food items.
Physical activity (PA) was measured using the short version of the International Physical Activity Questionnaire (IPAQ) [
45], in which the questions were adapted for older adults and validated by Hurtig-Wennlöf, Hagströmer and Olsson [
46]. For example, “Think about the time you spent sitting during the last 7 days, including time spent at home, while doing housework and during leisure time. This may include time spent sitting at a desk, visiting friends, reading, or lying down or sitting to watch television. During the last 7 days, how much time did you spend sitting during a day?” and participants were asked to indicate the hours and minutes. The IPAQ scores were calculated following the guidelines given by the IPAQ Research Committee [
47]. Apart from the IPAQ, participants also had to indicate the moment of the day for the physical activity with CATA questions (“before breakfast”, “between breakfast or and lunch”, “between lunch and dinner” and “after dinner”), whether they were willing to change their daily physical activity pattern and duration of sleeping. Short sleepers were defined as individuals who sleep less than or equal to 6 h per night and long sleepers were people who reported to sleep more than 8 hours per night [
48].
The definition of physical activity according to the National Institutes of Health was shown to the participants prior to the question of self-efficacy to engage in PA, i.e., “physical activity is any body movement that works your muscles and requires more energy than resting (lying down or sitting). Walking, running, swimming, yoga, all types of sports, and housework like cleaning and gardening are a few examples of physical activity”. Self-efficacy is defined as one’s confidence in their ability to engage in physical activities under various situations, and it appeared to be a stronger predictor to actual behaviour than attitudes towards physical activity [
49]. Self-efficacy was measured using 15 items modified based on Bandura [
50] with the validation and item combination suggested by Everett, Salamonson and Davidson [
51]. For example, participants could rate how confident they were on a five-point-scale from “not at all confident” (=1) to “extremely confident” (=5), or “not applicable” (=missing value) to a statement such as “I am confident in my ability to engage in physical activities when I am feeling tired”.
Participants were asked whether they know what dietary protein is. If “no” was indicated, the definition of dietary protein was given. If “yes” was indicated, they were directed to an objective knowledge test. Knowledge and perception of protein was evaluated by means of objective knowledge items, wherein correct and incorrect statements were presented (e.g., “You need protein in the diet for energy”) and participants had to answer one of the three choices “True”, “False” or “I do not know” (
Appendix A Table A2). The option of “I do not know” was included so as to reduce the level of bias induced by guessing.
Food fussiness was measured with six items adapted from Wardle et al. [
52] which were also previously tested in a sample with older adults [
53]. For example, “I enjoy tasting new foods” and the responses were on a five-point-Likert scale from “strongly disagree” (=1) to “strongly agree” (=5). Participants were also asked to indicate what they think about the amount of protein in their daily diet (i.e., on a five-point-scale ranging from “Too much” (=1) to “Too little” (=5) or “I do not know” (=missing value)), and whether they intend to change the amount (i.e., on a three-point-scale from “Yes, increase the amount” (=1) to “No, remain the same” (=3) or “I do not know” (=missing value)). In addition, participants were asked if they would increase the amount of protein in their diet if they were told by a health professional, food industry, family or friends (i.e., “Yes”, “No” and “I do not know”).
A series of 17 questions were used to assess the personal characteristics of participants, which included socio-demographics such as education level, being the main household (HH) grocery shopper, HH income, food expenses, lifestyle such as smoking and alcohol use, health characteristics such as risk of malnutrition, mobility, ability in preparing own foods, presence of various health problems, etc. The risk of malnutrition was measured using the Malnutrition Universal Screening Tool (MUST) [
54] with five questions and body mass index (BMI), which was evaluated and validated by Poulia et al. [
55] and Leistra et al. [
56] and found to have high sensitivity. For example, “Has your intake of food been poor for the last 5 days or likely to be poor for the next 5 days?” and participants were asked to indicate “Yes” or “No”.
2.3. Statistical Analysis
Statistical analyses were carried out with SPSS Statistics 25.0 (IBM SPSS, Armonk, NY, USA). Data processing and analysis included descriptive analysis (frequency distributions), bivariate ANOVA or chi-square tests (comparison of consumers’ characteristics across the appetite and protein intake strata) and multivariate analysis (calculation of the probability of lower protein intake and identification of its behavioural determinants).
2.3.1. Appetite and Protein Intake Strata
The appetite and protein intake strata were defined based on the SNAQ scores and probability of lower protein intake. The calculation of the probability of having a lower protein intake was performed based on multivariate logistic regression analysis, for which the protocol and algorithm have been reported in Wijnhoven et al. [
44]. There were 156 older adults who did not report their body weight (8.5% of the sample), while this information is required for the calculation of the probability of lower protein intake. Body weight was first adjusted for overweight and underweight individuals based on age and height (BMI), in which adjusted bodyweight was the nearest required for the individuals to have a BMI in the healthy range, i.e., 18.5–24.9 for adults aged ≤70 years and 22.0–27.0 for adults aged ≥71 years [
57]. Since adjusted body weight was used in the calculation of the probability of lower protein intake, the variation in body weight has relatively low leverage over the calculated probability. Therefore, instead of excluding the cases with missing body weight, the missing values were imputed with a fully conditional specification. Multiple imputation was performed using age, gender, country, education level, smoker status, height, physical activity level (IPAQ), self-reported presence of overweight and self-reported presence of underweight as predictor variables. The same analyses have been performed with missing data on the bodyweight excluded. The main conclusions remained unchanged, as the proportion of the strata and variables selected in the linear regression models did not differ substantially.
Poor appetite was defined as having a total SNAQ score below or equal to 14; low level of protein intake was defined as having a probability of protein intake lower than 1.0 g/kg adjusted BW/day equal to or higher than 0.3. The strata were profiled and compared in terms of older adults’ socioeconomic and demographic background, health characteristics, presence of health problems, knowledge and attitude related to protein, food and diet and attitude towards physical activity. Effect sizes, Cramer’s phi (ϕc) or partial eta-squared (η
p2) were included in the analyses to support the interpretation of the
p-value for which very low values can be obtained as a result of large samples sizes as in this study. Effect sizes indicate the proportion of variance in the variable (e.g., strata) explained by another variable (e.g., socio-demographics) and as such, indicate the strength of a relationship between variables and the significance of differences between strata [
58]. Cramer’s phi was computed for chi-square tests and considered small when in the range 0.1–0.3, medium from 0.3 and large from 0.5 [
59]. Partial eta-squared was computed for Kruskal–Wallis one-way analyses of variance and considered small from 0.01, medium from 0.06 and large when equal to or greater than 0.13 [
60]. For chi-square association tests, the test was not considered reliable if more than 20% of the cells had expected counts of less than five, as a large amount of sparse cells does not allow a valid comparison [
61].
2.3.2. Behavioural Determinants of Protein Intake
Two multivariate linear regression models were used to identify the behavioural determinants that might explain the probability of lower protein intake (continuous scale), while stratifying by appetite. Model 1 considered older adults who have reported poor appetite (SNAQ score ≤ 14; n = 493), while Model 2 included older adults with good appetite (SNAQ score > 14; n = 1332).
The explanatory variables included a series of possible behavioural determinants in terms of dietary habit (e.g., food expenses, consumption frequency and moment of certain food groups or products, diet status, etc.) and physical activities (e.g., physical activity levels and pattern). Categorical variables were coded as dummy variables for comparison. Expenses on food per week (for consumption at home and out-of-home) had four levels (i.e., less than 60 EUR, between 60 and 119 EUR, 120 EUR or above, prefer not to say or do not know); the level “between 60 and 119 EUR” was set as the reference category while the others were coded as dummy variables. Physical activity level based on IPAQ had three levels (i.e., high, medium and low). For example, a high level of physical activity denotes at least five days of any combination of walking, activities of moderate or vigorous intensity per week; moderate level denotes at least three days of vigorous activity of at least 20 min per day; low level denotes no or very low (insufficient to meet moderate or high level of) physical activity reported. The scoring protocol and categorical levels are reported in the guideline published by the IPAQ Research Committee [
47]. High physical activity level was set as the reference category while the other levels were coded as dummy variables.
All major assumptions have been tested for the multivariate linear regression model. There was no issue of multicollinearity. Collinearity diagnostics included: no values of variance inflation factor (VIF) larger than 10; no average VIF-values substantially larger than 1 (the largest VIF value: 1.234 in Model 1 and 1.248 in Model 2); no tolerance value below 0.2 (the smallest tolerance value: 0.810 in Model 1 and 0.801 in Model 2); no correlation coefficient between two explanatory variables in the model larger than 0.50. The plot of standardized residuals against standardized predicted values for both models showed a slight tendency of funnelling out but no curve formation, signalling there could be heteroscedasticity in the data. Yet, the assumption of linearity has been fulfilled. The distribution of errors was close to normal. The results of the assumption tests have shown that multivariate linear regression is a suitable statistical analytical method for this study and its data. In order to account for possible issues of assumption violation, a bootstrap method was applied to provide more robust statistics.
There were eight potentially influential cases identified in Model 1 and 69 potentially influential cases in Model 2 based on several parameters: covariance ratio (CVR) (cases with CVR >1 + (3(k + 1)/n) or <1 − (3(k + 1)/n); Model 1: CVR >1.073 = 2 cases and <0.927 = 6 cases; Model 2: CVR >1.029 = 24 cases and <0.970 = 45 cases); standardized residuals (2.64% (Model 1) and 5.03% (Model 2) of cases have absolute values above 2, and 0.20% (Model 1) and 1.28% (Model 2) have absolute values above 2.5, any case with the value above about 3, could be an outlier: 0 cases (Model 1) and 1 case (Model 2)); all cases have Cook’s distance value lower than 1; the average leverage (cases >3 × ((k + 1)/n) = 0.073:0 cases (Model 1) and 0.029:0 cases (Model 2)); no absolute values of DF-Beta was greater than 1. Linear regression models were run with identical outcome and explanatory variables, both with and without the influential cases removed. The resulting models were similar in terms of retained variables; however, the adjusted R2 showed a reasonable improvement in the model fit after the removal of outliner (one case in Model 2) and influential cases (eight cases in Model 1 and 69 cases including the outliner in Model 2). The adjusted R2 increased from 17.7% to 20.6% in Model 1 and from 15.7% to 21.9% in Model 2. Therefore, all potentially influential cases were removed in the final models (Model 1: n = 485; Model 2: n = 1263).