The research on how conceptions of feedback relate to behaviours is largely limited to self-reported practices rather than to actual observed practices. This raises the possibility that such data are invalid because of memory failure or ego-protective responses. To minimise that threat, valid data collection uses multiple items for each potential latent cause, designed to present theoretically important stimuli that are analysed mathematically to determine fit to the theory (Brown, 2023). Reliance on observation alone cannot expose what the most knowledgeable informant knows about what lies in the beliefs, thoughts, ideas, emotions, and attitudes of that individual; hence, self-report (Brown, 2023).

Thus, survey research still produces limited information about how beliefs about feedback relate to feedback practices. For example, a survey of 54 Tanzanian mathematics teachers (Kyaruzi et al., 2018), found that endorsement of approaches to feedback that focus on monitoring (e.g., asking students to indicate what went well and what went badly with their assignments) and scaffolding (e.g., adjusting instruction whenever I notice that students do not understand a topic) had strong prediction on high-quality feedback delivery practices (e.g., being supportive when giving students feedback and encouraging students to ask for feedback whenever they are uncertain). In contrast, a small survey of 61 Ethiopian teachers found that the relationship of teacher beliefs about feedback had a statistically not significant correlation with their practices (Dessie and Sewagegn, 2019).

In contrast, based on responses to the Teachers Conceptions of Feedback inventory (TCoF; Harris and Brown, 2008), a nationally representative survey of 518 New Zealand teachers found that endorsement of feedback about learning processes and involving students in feedback predicted greater use of non-teacher feedback methods (Brown et al., 2012). The same study reported that the use of praise in feedback predicted feedback actions that protected students from negative evaluative consequences. A survey of 390 Pakistani teachers, using the TCoF, found that endorsement of feedback as encouragement led to greater use of protective evaluation practices such as giving positive messages to students and not making critical comments (Aslam and Khan, 2021).

According to the joint European Values Study and World Values Survey 2005–2022 (EVS/WVS, 2022), Sweden is a strongly secular-rational and individualistic country with a strong emphasis on equality and the individual’s freedom and wellbeing. A similar description is found in the six-dimensional Index of National Culture (INC) by Hofstede et al. (2010). This is reflected in teachers’ relatively high degree of freedom to interpret and concretize the objectives of the national curriculum and to decide on appropriate teaching methods to help students achieve these goals (Helgøy and Homme, 2007). The Swedish egalitarian ideal is also reflected in teacher–student relations. Teachers do not receive, nor demand, respect solely based on their position/role in society. In the classroom, the average student is the norm and discrimination between students in terms of, for example, special classes or educational tracks for gifted or underperforming students, is rare. Rating of students in public, whether explicit or implicit, based on their school achievement is not in line with the Swedish culture.

The Swedish curriculum is goal-oriented, with national standards for student learning in years 3, 6, and 9 (nominally aged 9, 12, and 15). Grades are given in school years 6–9, but only the year 9 grades are high stakes because they matter for admission to upper secondary school. The grades are criterion-referenced, meaning if the standards are achieved, any number of students can receive that grade. Legally, the municipality is responsible for providing adequate resources for education and to conduct systematic evaluations [Utbildningsdepartementet (Ministry of Education), 2010]. In order to support schools and teachers to fulfil their obligations, the Swedish National Agency for Education (SNAEd) provides national screening materials, assessment support material, and standardized national tests. These assessment materials and tests serve various purposes: to inform decisions about support and adaptations of teaching, grading, and, at an aggregated level, to provide point estimates of student achievement at school or system level to support between-school equivalence in grading and for trend analysis [Skolverket (National Agency for Education), 2020].

The national standardized tests (NSTs) in years 3, 6, and 9 are mandatory. When grading students, the teachers are required to use all available information about students’ knowledge and skills, with particular consideration of the results from the NSTs [Utbildningsdepartementet (Ministry of Education), 2010]. Thus, the SNAEd advises teachers to design and use different types of assessment situations for formative and summative purposes [Skolverket (National Agency for Education), 2022], and that teachers on all school levels choose, design and implement their own classroom assessment.

The research on Swedish teachers’ conceptions of feedback is limited in both number and scope. However, by interviewing and surveying approximately 70 teachers and principals with different qualifications and experiences at seven schools in four of the largest cities in Sweden and Norway, Helgøy and Homme (2007) found that Swedish teachers to a higher extent than Norwegian teachers perceived NSTs as valuable tools in grading and in the improvement of their teaching. Moreover, unlike the Norwegian teachers, the Swedish teachers did not perceive NSTs as limiting their autonomy in the interpretation of national goals and how to organize the teaching to help students reach those goals.

A total of 461 teachers working between years 1 and 9 in a northern city in Sweden responded to the survey. This is a 62% response rate from the municipality. Prior to analysis, data preparation involved identifying and removing from consideration participants who had more than 10% of responses missing per instrument. This sample was chosen because the research group and the municipality had previously decided to initiate a larger research project on improving assessment practices in these school years. This meant investigation of teacher conceptions was included as part of a multi-method, multi-study project. After deletion of 11 cases for high number of missing values, 450 teachers were retained (82% women, 17% men, 1% missing). Most participants had a teaching degree (78%), with 5% not having such a degree, and 17% not answering. Length of teaching experience was grouped by year ranges: 5% <2 years, 16% 2–5 years, 12% 6–10 years, and 66% >10 years. Distribution across the grade levels taught was almost equal (Years 1–3, n = 149; Years 4–6, n = 156, Years 7–9, n = 141).

While the TCoF focused on teacher conceptions of the nature and purpose of feedback, a close reading of the inventory suggested that embedded within the TCoF, there were eight statements that described specific behaviours teachers might enact. Six of these practices were from the Hattie and Timperley (2007) process (Process1, 2, 5), self-regulation (SRL1, 3), and praise (Praise6) factors. The two other possibilities were from Timeliness (Time1, 4). To test the possibility that these items formed a Practices factor they were disaggregated from their original scale and aggregated into a new separate scale of Feedback Practices, which theoretically would be predicted by the remaining TCoF conceptions of feedback factors. Hence, a major contribution of this study is to examine whether this adaptation had validity. An advantage to this approach would be to minimise the number of items needed to elicit both beliefs about feedback and practices.

After deleting participants with more than 10% missing responses, we imputed missing values with the expectation maximisation algorithm (Dempster et al., 1977). The imputation had a statistically not significant result (χ² = 1508.122, df = 1,453, p = 0.153) for Little’s (1988) Missing Completely at Random (MCAR) test showing the distribution of missing responses was random. Hence, all analyses were conducted without missing values. All but one variable met accepted standards for skew (<2.00) and kurtosis (<7.00), meaning variables were normal (Kim, 2013). Because item Irr4 had kurtosis =11.40, it was transformed using a Box-Cox transformation (Courtney and Chang, 2018) in the normalr ShinyApp.¹ This produced kurtosis = −0.40 and that version of the item was used in all analyses.

Coherent with the MIMIC framework, a two-step process (Anderson and Gerbing, 1988) of testing measurement models for each construct (i.e., conceptions of feedback and feedback practices) through confirmatory factor analysis (CFA) was implemented before testing a structural model that linked the beliefs to practices. Once a well-fitting model for each construct was found, a structural equation model (SEM) was tested in which beliefs about feedback were positioned as predictors of feedback practices on the assumption that beliefs are a predictor of behaviours (Ajzen, 1991).

We tested a correlated model of eight conceptions of feedback factors but this had poor fit. Consequently, we inspected modification indices to identify items that violated simple structure (Revelle and Rocklin, 1979) or independence of residual assumptions (Barker and Shaw, 2015), while aiming to retain the eight conceptions factors. Items with weak loadings on their intended factors (i.e., <0.30) were candidates for deletion. Items with strong modification indices (i.e., MI > 20) to other factors or whose residuals are strongly attracted to those of other items were also candidates for deletion (Bandalos and Finney, 2010). Although, most researchers expect three items per factor, it is possible in multi-factorial inventories to estimate factors that have only two items (Bollen, 1989).

Fit of both CFA and SEM models was established by inspection of multiple fit indices (Hu and Bentler, 1999). Because the chi-square measure of discrepancy between a model and its underlying data is sensitive to sample size and model complexity, we accepted not statistically significant values for the normed chi-square (i.e., χ²/df) as support for a model (Wheaton et al., 1977). Further evidence for non-rejection of a model arises when the comparative fit index (CFI) is >0.90 and the root mean square error of approximation (RMSEA) is <0.08. However, both the CFI and RMSEA indices are sensitive to models with more than three factors, with the CFI entering reject space and the RMSEA entering not reject space under those conditions (Fan and Sivo, 2007). Thus, greater reliance is put on the gamma hat >0.90 and the standardized root mean residual (SRMR) <0.08 because these are more robust against sample size, model complexity, and model misspecification than the CFI or RMSEA indices. Scale reliability was estimated using the Coefficient H maximal reliability index, which is based on an optimally weighted composite using the standardised factor loadings (Hancock and Mueller, 2001). Factor mean scores were calculated by averaging the raw score for each item contributing to the factor, an appropriate method when simple structure (i.e., items belong to only one factor) is present (DiStefano et al., 2009).

CFA and SEM were conducted in the Jamovi Project (2022) platform using the lavaan package (Rosseel, 2012). Because six-point ordinal scales function similarly to continuous variables (Finney and DiStefano, 2006), maximum likelihood estimation was used. Model syntax is provided in Supplementary material.

After analysis was completed, member checking (Tong et al., 2007) was conducted in two interview groups in the same city in which the survey had been administered. This was done to examine our interpretation of items that did not load as expected and the meaning of the unexpected paths. Ten teachers were recruited face-to-face on a volunteer, convenience basis to participate in two group interviews (n = 4 and 6, respectively) at separate occasions. In the interviews, the teachers were first given time to think individually about each issue and then shared their thinking in a joint discussion. The teachers had completed the survey, but their responses had been anonymous and so they were commenting on aggregate data results to which they had contributed. Groups were led by two of the authors. In these discussions, we explored how respondents had interpreted the items and how they understood the unexpected relations we had detected. The authors took field notes during the 1 h conversations and conversations were audio-recorded. The interviews were verbatim transcribed from which themes were identified. Aggregation of responses to the focus issues was carried out by authors 2 and 4 and manual coding of themes was carried out.

An inter-correlated factor model had promising fit, but still below expectations. Because of the high correlation values merging of SRL, Task, Process, and Improvement items into a single factor of Improvement helped fit. Further modifications removed items that violated simple structure or were strongly correlated with other items, resulting in improved fit to the data. Paths that were not statistically significant were also removed. These modifications created six correlated conceptions of feedback and one practices of feedback factor.

The teacher conceptions of feedback factors were:

The teacher Feedback Practices (Practices) factor consisted of three process items, two self-regulation of learning items, and one praise item. Together, these items create a set of formative practices that focus on giving students information and time to think about and improve their work, while taking responsibility for their own outcomes. Additionally, the feedback teachers provide includes commenting on the effort students put into their work, as well as noting how it can be improved.

Items, standardised loadings, and scale coefficient H statistics for all seven factors are given in Table 2.

The inter-correlation of conceptions of feedback factors is shown in Table 3. As expected, the Students Ignore factor had negative values to three other factors (i.e., Improvement, Required, and PASA) and non-significant values to Praise and Timely. In contrast, all other factors were moderately and positively inter-correlated with values ranging from 0.36 ≤ r ≤ 0.70. This suggests commitment to feedback for improvement is simultaneously weakly related to feedback being required, using praise, feedback from peers and self, and being timely, while not being something that students ignore. This suggests teacher beliefs are generally adaptive and in line with feedback theory.

Factor means are shown in Table 4, with between factor effect sizes (Cohen’s d; Cohen, 1992). Values d > 0.60 are considered large in educational research (Hattie, 2009). In general, teachers gave strongest endorsement to the conception of feedback for improvement, with large effects compared to all other feedback factors and practices. Praise and Prompt feedback, with scores less than moderately agree, had large differences to both Expected and Students Ignore feedback factors. Both Students Involved in PASA and Expected, above slightly agree, had large differences only to Students Ignore feedback, which was close to mostly disagree. Note that low score indicated that on the average teachers rejected the notion that students ignore feedback. The Formative feedback practices, with a score just above moderately agree, was much larger than Students Involved in PASA, Expected, and Students Ignore feedback factors, had medium to small differences with Prompt and Praise factors, respectively, and was much smaller than the Feedback Improvement conception.

Structural paths were retained in the model only if they were statistically significant. The only statistically significant predictors of Practice were the Improvement (β = 1.24) and Ignored (β = 0.61) (Figure 1). This model had acceptable to good fit indices (χ² = 465.334, df = 258, χ²/df = 1.80, p = 0.18; CFI = 0.90; gamma hat = 0.96; RMSEA = 0.042, 90%CI = 0.036–0.048; SRMR = 0.050) and so was not rejected. Note that the correlation between Improvement and Ignore was negative but their common positive loading on Practices is a non-transitive relationship, something possible given the relatively modest value (r = −0.65) (Kim and Mueller, 1978). This suggests that the implementation of the formative feedback practices depends both on seeing feedback as something students tend to ignore and thus may need to be overcome through these practices and something which will help them improve their learning. This conclusion was supported by the member checking process, in that all teachers expressed the view that feedback is an inherent part of good teaching and something they are expected to provide. Moreover, in case students ignore feedback, the teachers saw this as an indication that they need to either improve the characteristics of their feedback itself or the learning situation in which the feedback is provided.

Teachers generally agreed with the conception that the purpose of feedback is to enhance students’ self-esteem and, hence, should be full of encouraging and positive comments (i.e., Praise). Although this goal is commendable, a focus on giving praise may hamper learning. Studies have shown that focussing on praise may come at the expense of identifying students’ learning needs and suggestions on how to improve learning (Brown et al., 2012). Indeed, research has shown that praise commonly does not have a positive effect on students’ achievement (Hattie and Timperley, 2007). These results are corroborated by our model (Figure 1), where praise is not directly associated with the formative feedback practices that focus on the improvement of students work and self-regulation in learning. Hence, the goal of caring about the student and giving formative feedback may be in conflict. Since both goals are present in most curriculum statements, the solution is not to exclude either one, but instead to learn how to accomplish both. Thus, teacher education needs to address this tension and discuss how to circumvent it.

This could, for example, involve prospective teachers having to learn how to establish a classroom culture in which shortcomings (e.g., failure or not knowing) are seen as a natural part of learning and that attending to them is essential for learning. A prerequisite for the establishment of such a culture would be to find strategies to counteract the tendency of many students to link their school achievement with their self-esteem. Then, person-centered praise aiming at making students feel good about themselves could be replaced by positive comments aimed at linking positive outcomes to causes controllable by the student. Indeed, helping students to make adaptive attributions and to experience learning progress is associated with wellbeing (Winberg et al., 2014). Thus, teacher education must ensure that prospective teachers understand and enact caring for students by helping them to develop competence, rather than simply protect them from “bad” news.

We consider that the results presented here are likely to be typical of Swedish teachers rather than just the teachers participating in this study. While school administration is very localised, the policy and resource constraints exist equally for teachers elsewhere in the nation working in publicly funded primary and junior secondary schools. Nonetheless, a national survey would be needed to assure of generalisability claims made here. The relatively modest coefficient H values (i.e., all H < 0.80) suggest that the stability of these results is less than ideal. The stability of the factor measurement models needs to be tested in a further sample of Swedish teachers.

Of course, given the data are from a survey, we have made informed interpretations, corroborated by a small-scale member checking exercise, of what the factors mean and why the path values are what they are. Follow-up qualitative studies with teachers exploring their understanding of the results may provide further confidence in our explanations. Potentially, providing teachers with their own factor scores may provide further insights as to the meaning of teacher confidence in our findings.