Introduction
The prevalence of eating disorders is getting higher as of now 6% to 8% of adolescence is at
risk for developing an eating disorder (Batista et al., 2018). Women are especially vulnerable, with
men accounting for only 10% of anorexia nervosa or bulimia nervosa disorders (Smink, van
Hoeken & Hoek., 2012). Prevention of eating disorders is the most effective way to decrease their
global burden. Common ways of assessing the risk of developing eating disorders are self
questionnaires. This study aims to investigate the dimensionality and validity of the length of the
Girls At Risk of Eating Disorders (GARED). In addition, we look at if the obtained score indeed
correctly predicts the risk of developing an eating disorder. This will be done by assessing the
communalities, conducting a factor analysis and a logistic regression analysis. The GARED is an
online questionnaire consisting of fifteen items, our aim is to shorten it by five items so that it
becomes more accessible and easier to complete while remaining true to the objective of the
questionnaire. The first goal is to choose 5 items, which predict the development of an eating
disorder, the least. We will delete items which least represent the underlying construct, this is done
by choosing items with the lowest communality. Our second aim is to investigate the dimensionality
of the remaining ten items, this is done by the means of a factor analysis. Lastly, we investigate if a
higher total GARED score is predictive of actually being at risk of developing an eating disorder. A
risk score will be individually ascribed by an independent psychologist. This score will be the true
measure to which the total GARED score will be compared. To determine how predictive the total
GARED score is, a logistic regression will be used.
By combining these three goals we aim to confidently recommend the shortened GARED
questionnaire for providing accurate predictive risk of developing an eating disorder on the internet.
Sample and variables
The questionnaire was completed by 908 adolescent girls, of which the age varies between
13 and 17 years of age (M = 14.98, SD = 1.45). All 15 items were answered on a five point likert
scale ranging from (1) ’no, definitely not’ to (5) ‘yes, definitely’ (see table 1). The minimum
obtained total score on the GARED questionnaire is 10 and it can range until the maximum score of
50. The girls were individually assessed to be at risk or not at risk for an eating disorder by a
psychologist. This external measure shows that 88 (9.7%) girls are at risk and 820 (90.3%) we are
not at risk of developing an eating disorder. The measurement level of this risk variable is nominal/
categorical as there is only an option of being at risk or not being at risk. In Table 1, all items, their
means and standard deviations, are depicted.
, Reducing the GARED test from 15 to 10 items
The first goal of this study is to shorten the GARED from 15 to 10 items. Five items that are
least predictive of an eating disorder will be deleted. This is done by looking at the communities
which are derived by using factor analysis on the 15 items. The communality describes the variance
in one item explained by the underlying factor in this case ED. The more variance it explains the
better the quality of the item. Low communality means that less variance is explained, which in turn
means that the item is of lower quality. Communalities of the 15 items of the GARED test can be
found in table 2. The five items with the lowest communality were deleted from the data set. The
five lowest items are: item 5 ‘sport’ (.082), item 6 ‘fat’ (.059), item 8 ‘diet’ (.094), item 10
‘bad’ (.131) and item 12 ‘decide’ (.099). The communalities of the five deleted items lie between
0.059 and 0.131, the communalities of the remaining 10 items range between .280 and .522. A
possible explanation for low communalities might be that is is not specific for only girls at risk of
developing an eating disorder. So, the items with low communalities are often scored high on even
when a person is not at risk. For example, item 6 (‘I think it is important not to be fat’), might be a
general thought in adolescent girls. Thus many girls would score high on this item, despite not
being at risk of developing an eating disorder.
Factor analysis of the 10 item GARED test
The second goal of this study is to assess the dimensionality of the GARED test. This is
done by conducting a factor analysis on the 10 remaining items. When assessing dimensionality one
looks and whether items of the test load or one or multiple dimensions or factors. For this, we look
at the correlation matrix (see table 3), scree plot (figure 1), factor loadings and communalities (table
4) and the total explained variance of the items.
The correlation matrix shows how much items correlate. All correlations are positive and
significant indicating a weak to average relation between the items themselves. In addition there are
no visible patterns between the items. Based on this correlation matrix it is acceptable to assume
that the items depend on one factor. In other words that the questionnaire seems unidimensional.
When looking at the scree plot there are two rules of visual inspection. Firstly, the number
of eigenvalues > 1 and secondly, the number of eigenvalues before the “joint” / “breaking point”.
Looking at our scree plot, there is a ‘joint’ at the second eigenvalue. Before the joint we see 1
eigenvalue, this indicates that there is one factor present. Looking closer, the first eigenvalue is also
the only value above one. Again, this indicates a singe factor present. Based of the two rules of
thumb it is likely that there is one factor present and so, the test seems unidimensional.
The prevalence of eating disorders is getting higher as of now 6% to 8% of adolescence is at
risk for developing an eating disorder (Batista et al., 2018). Women are especially vulnerable, with
men accounting for only 10% of anorexia nervosa or bulimia nervosa disorders (Smink, van
Hoeken & Hoek., 2012). Prevention of eating disorders is the most effective way to decrease their
global burden. Common ways of assessing the risk of developing eating disorders are self
questionnaires. This study aims to investigate the dimensionality and validity of the length of the
Girls At Risk of Eating Disorders (GARED). In addition, we look at if the obtained score indeed
correctly predicts the risk of developing an eating disorder. This will be done by assessing the
communalities, conducting a factor analysis and a logistic regression analysis. The GARED is an
online questionnaire consisting of fifteen items, our aim is to shorten it by five items so that it
becomes more accessible and easier to complete while remaining true to the objective of the
questionnaire. The first goal is to choose 5 items, which predict the development of an eating
disorder, the least. We will delete items which least represent the underlying construct, this is done
by choosing items with the lowest communality. Our second aim is to investigate the dimensionality
of the remaining ten items, this is done by the means of a factor analysis. Lastly, we investigate if a
higher total GARED score is predictive of actually being at risk of developing an eating disorder. A
risk score will be individually ascribed by an independent psychologist. This score will be the true
measure to which the total GARED score will be compared. To determine how predictive the total
GARED score is, a logistic regression will be used.
By combining these three goals we aim to confidently recommend the shortened GARED
questionnaire for providing accurate predictive risk of developing an eating disorder on the internet.
Sample and variables
The questionnaire was completed by 908 adolescent girls, of which the age varies between
13 and 17 years of age (M = 14.98, SD = 1.45). All 15 items were answered on a five point likert
scale ranging from (1) ’no, definitely not’ to (5) ‘yes, definitely’ (see table 1). The minimum
obtained total score on the GARED questionnaire is 10 and it can range until the maximum score of
50. The girls were individually assessed to be at risk or not at risk for an eating disorder by a
psychologist. This external measure shows that 88 (9.7%) girls are at risk and 820 (90.3%) we are
not at risk of developing an eating disorder. The measurement level of this risk variable is nominal/
categorical as there is only an option of being at risk or not being at risk. In Table 1, all items, their
means and standard deviations, are depicted.
, Reducing the GARED test from 15 to 10 items
The first goal of this study is to shorten the GARED from 15 to 10 items. Five items that are
least predictive of an eating disorder will be deleted. This is done by looking at the communities
which are derived by using factor analysis on the 15 items. The communality describes the variance
in one item explained by the underlying factor in this case ED. The more variance it explains the
better the quality of the item. Low communality means that less variance is explained, which in turn
means that the item is of lower quality. Communalities of the 15 items of the GARED test can be
found in table 2. The five items with the lowest communality were deleted from the data set. The
five lowest items are: item 5 ‘sport’ (.082), item 6 ‘fat’ (.059), item 8 ‘diet’ (.094), item 10
‘bad’ (.131) and item 12 ‘decide’ (.099). The communalities of the five deleted items lie between
0.059 and 0.131, the communalities of the remaining 10 items range between .280 and .522. A
possible explanation for low communalities might be that is is not specific for only girls at risk of
developing an eating disorder. So, the items with low communalities are often scored high on even
when a person is not at risk. For example, item 6 (‘I think it is important not to be fat’), might be a
general thought in adolescent girls. Thus many girls would score high on this item, despite not
being at risk of developing an eating disorder.
Factor analysis of the 10 item GARED test
The second goal of this study is to assess the dimensionality of the GARED test. This is
done by conducting a factor analysis on the 10 remaining items. When assessing dimensionality one
looks and whether items of the test load or one or multiple dimensions or factors. For this, we look
at the correlation matrix (see table 3), scree plot (figure 1), factor loadings and communalities (table
4) and the total explained variance of the items.
The correlation matrix shows how much items correlate. All correlations are positive and
significant indicating a weak to average relation between the items themselves. In addition there are
no visible patterns between the items. Based on this correlation matrix it is acceptable to assume
that the items depend on one factor. In other words that the questionnaire seems unidimensional.
When looking at the scree plot there are two rules of visual inspection. Firstly, the number
of eigenvalues > 1 and secondly, the number of eigenvalues before the “joint” / “breaking point”.
Looking at our scree plot, there is a ‘joint’ at the second eigenvalue. Before the joint we see 1
eigenvalue, this indicates that there is one factor present. Looking closer, the first eigenvalue is also
the only value above one. Again, this indicates a singe factor present. Based of the two rules of
thumb it is likely that there is one factor present and so, the test seems unidimensional.
