We followed a process model approach to instrument design and validation (Chatterji, 2003) to create and validate a measure of student trust in their STEM instructor. This approach involved several progressive steps from building a codebook of themes defining the domain of college student trust in STEM instructors and then using the codebook to write items and iteratively design an instrument. The advantage of this approach is that it leads to a better instrument factor structure that necessitates fewer iterations (Chatterji, 2003). Following the process model approach, we developed a comprehensive codebook of college students’ trust in instructors (Zhang et al., 2025). This codebook was developed by reviewing the social and educational literature on the construct of trust and interviewing current STEM students to understand and identify instructor behaviors or characteristics that cultivate trust. Based on qualitative analysis of student interviews, we identified 28 trustworthy instructor characteristics organized into three domains: cognitive, affective, and behavioral trust (Felten et al., 2023; Zhang et al., 2025). The codes from the codebook were divided between two members of the research team, who drafted survey items which were then reviewed by college student research assistants for tone and clarity. Final inclusion of each item was determined based on a vote between the two item creators and a third member of the research team. The result was a 38-item survey, which was content validated with a group of reviewers (including STEM undergraduate students). The initial items represented the 28 characteristics emerging from student interviews as well as the identified in broader literature (Clark and Lemay, 2010; Felten et al., 2023).

After the 38-items of the instrument were created, the items were content validated with a group of external reviewers (N = 6). These reviewers included undergraduate students (n = 3) and faculty (n = 3). The undergraduate students were close in age and experience to the target population of the survey. The faculty members gave feedback on the clarity, wording, and meaningfulness of the items. Feedback was collected in written (i.e., email) form. Following their feedback, we rectified suggested edits to the literature and updated the drafted instrument. The empirical validation of the instrument encompassed an exploratory and, separately, a confirmatory validation study that involved two new datasets (N = 453 and N = 1,438 respectively). The purpose of the exploratory validation study was to understand the underlying structure or properties of the instrument, while the confirmatory validation study intended to validate the specified model based on these results. Students were contacted via email from class lists provided by their instructor around mid-semester and were sent email reminders no more than three times. Students were incentivized with a small extra credit score or lottery to win a digital gift card to participate in the study. Data for the exploratory and confirmatory studies were acquired around the mid-point of two semesters. Analysis was conducted using Jamovi Project (2022), which utilized the lavvan (Rosseel, 2012) and semPlot (Epskamp, 2015) packages in R V4.1 (R Core Team, 2021). The survey participation request and online informed consent process were conducted with approval from the institutional review board at Yale University (#2000028325).

The initial 38-item instrument used for the empirical validation had a Likert scale (1–5, Strongly Disagree to Strongly Agree). An example item includes: “My instructor can connect with students.” This initial instrument also included an item probing overall trust that the student has in the instructor on a 0–10 scale from 0 (not at all) to 10 (to a great extent) and the survey collected student information such as first-generation status, and self-reported GPA,

For the exploratory validation study, participants (n = 453) were undergraduate students enrolled in five courses at a large public university within the physiology and neurobiology department. The course enrolled major and non-major students, offering variation in student type. The size of these courses ranged from 12 to 304 students. The goal was to have a large sample of students test the items rather than analyze them at the classroom level. For the confirmatory validation study, participants completing the survey were undergraduate students (n = 1,438) from seven instructors teaching one of eight courses at one of three public universities. These courses were characterized by large enrollments, and covered STEM topics such as physiology, biology, and neurobiology. The size of these courses ranged from 60 to 597 students. The overall participation rate (74.5%) was high due to an extra credit incentive offered to students for completing the survey. Information on participants’ characteristics for the exploratory and confirmatory validation studies are located in the Supplemental Materials.

An Exploratory Factor Analysis (EFA) using Principal Axis Factoring (PAF) with Direct Oblim rotation was calculated using SPSS 29 software (IBM Corp, 2023) to investigate the latent structure underlying student perceptions of the trust scales. PAF was chosen because it is a robust method for conducting an EFA, and an oblique rotation was used to account for possible correlations between factors. Based on the outcomes of the exploratory factor analysis, a confirmatory factor analysis (CFA) for a three-factor solution was calculated using Maximum Likelihood estimation. The R statistic package (R Core Team, 2021) with the windows-based interface Jamovi Project (2022) was used for this analysis.

An Exploratory Factor Analysis (EFA) using Principal Axis Factoring (PAF) with Direct Oblim rotation was calculated using SPSS 29 software to investigate the latent structure underlying student perceptions of the trust scales (Table 1). PAF was chosen as it is a robust method for conducting an EFA and an oblique rotation was used to account for possible correlations between factors. To test the appropriateness of the data for factor analysis, a Kaiser-Meyer-Olkin (KMO) Measure of Sampling Adequacy was calculated. The KMO had a score of 0.931 which satisfies the necessary assumption. Bartlett’s Test of Sphericity was calculated, and the results were significant, χ²(703) = 6524.77, p < 0.001. The KMO and the Bartlett’s Test of Sphericity suggest that an EFA approach is appropriate with this data sample.

To determine the number of factors in this data set a scree plot and a pattern matrix with factor loadings was calculated. Figure 1 presents the scree plot for this data set. As can be seen in the figure, the first and second factors explain a large proportion of the data and the eigenvalues for the other factors flatten out at 6 or 7 factors. The first factor explains 43.44% of the variance with factor 2 adding 9.38% of the variance. Together, factors 1 and 2 explain 52.73% of the variance. Using the Kaiser criterion of accepting factors with eigenvalues above 1, 6 factors can be defined. However, factors 3, 4, 5 and 6 add relatively small increases in explained variance. Factor 3 adds 6.22%, factor 4 adds 4.49%, factor 5 adds 3.27%, and factor 6 adds 2.67%. Together, all 6 factors explain 69.32% of the variance.

The pattern matrix converged in 11 iterations and defined 6 dimensions of trust. Table 1 presents the factor loadings and structure of each factor. Both the scree plot and the factor matrix suggest a 6-factor solution for the current data. As can be seen in Table 1, factor 1 is the primary factor with the largest number of items and explanation of variance. For the 6th factor, there were no items with loadings above 0.29 and as such this factor was considered not viable for describing the dimensionality of the trust. Factor 5 had only two items with item loadings of 0.55 and 0.54. Since the item loadings were low and the amount of variance explained was limited to 3.27%, this factor was also removed for the next stage of analysis. Finally, to prepare the analysis of dimensionality for the confirmatory factor analysis, items with factor loadings below 0.65 were removed from the analysis. This left factor 4 with only 1 item which explains only 4.48% of the variance. Accordingly, this factor was also removed from the analysis.

The initial exploratory factor analysis resulted in a reduced version of the trust instrument with 3 factors. Table 2 presents the prepared version of the trust instrument with reliability statistics. As seen in Table 2, factors 1 and 3 have high reliability. Factor 2 has barely adequate reliability statistics.

Based on the outcomes of the exploratory factor analysis, a confirmatory factor analysis for a three-factor solution was calculated using Maximum Likelihood estimation. The R statistic package with the windows-based interface Jamovi was used for this analysis. Table 2 presents the prepared items with the unstandardized factor loadings ready for confirmatory factor analysis. Table 3 presents the unstandardized factor loadings for the first model. All items loaded were significant. As seen in Table 3, some items loaded below the 0.7 level and were considered candidates for removal. Table 4 presents the fit indices for this model. The model yielded a significant chi-square statistic, χ²(186) = 2,303, p < 0.001, which may be the result of a large sample size. The fit indices showed a marginal to acceptable fit: CFI = 0.905, TLI = 0.893, SRMR = 0.048, and RMSEA = 0.087. Although the model had acceptable fit, the model could be improved.

To improve the model, all items with factor loadings below 0.7 were removed from the analysis. The revised model was then tested in a confirmatory factor analysis. Table 5 presents the unstandardized factor loadings for the revised model. All item loadings were statistically significant. Table 4 presents the fit indices for model 2. As can be seen in the table, model 2 is an improvement on the first model with excellent CFI fit (0.956), acceptable TLI (0.946), excellent SRMR (0.029) and acceptable RMSEA (0.087). Comparing the Akaike Information Criterion for the 2 models reveals a lower score for model 2, indicating that this is an improved model when compared to the first. Nevertheless, in model 2 there is a factor with only two items which can make this latent construct unreliable (Worthington and Whittaker, 2006). Accordingly, it was decided to remove this factor and it’s two associated items from the model.

The emergent structure was tested in a confirmatory factor analysis. Table 6 presents the unstandardized factor loadings for this model, and Table 4 presents the fit indices. As can be seen in Table 4, the AIC is better for this model than the previous two models. But the RMSEA is quite high at 0.094 which indicates a poor approximation of the population covariance matrix. In order to improve the model, modification indices and residual covariances were calculated. The modification indices suggested that the item “My instructor cares about students’ educational success” had a very high value (45.31) and residuals with multiple items suggesting cross loading and redundancy. The items “My instructor gets me” and “My instructor helps me achieve my professional goals” had high modification indices values (22.1 and 23.68) and overlapping content or correlated errors. Accordingly, it was decided to remove these items and run a confirmatory factor analysis.

Table 7 displays the unstandardized factor loadings, and Table 4 shows the fit indices for Model 4. As indicated in Table 4, Model 4 demonstrates improved values for the AIC, and it yields the strongest results for CFI (0.975), TLI (0.965), and SRMR (0.02). The RMSEA value of 0.081 is slightly above the typical threshold of 0.08 for a “reasonable” fit but is still considered within the acceptable range of 0.05 to 0.10 in structural equation modeling (Netemeyer et al., 2003). The covariance between Factor 1 and Factor 2 is 0.779. This suggests that the two factors, while reasonably distinct, are part of the same overall scale (Rönkkö and Cho, 2022). Comparisons with prior models show that Model 4 offers the best fit and can be considered a suitable solution for the CFA. The final model consists of 2 factors with 9 items, each demonstrating excellent reliability. Table 8 presents the TRUST instrument in a survey response format.

For classroom trust building, this is a many-to-one relationship role where students in the class perceive whether their instructor is fostering a supportive educational environment. The components of classroom trust are attentiveness, awareness, care, communication, and connection. Higher scores indicate that the instructor has successfully established a good working relationship with the classroom. These five components serve as sensible and realistic guidelines for instructors to better understand trust building through their role in the classroom.

Beyond academic performance, a trusting classroom environment enhances students’ emotional well-being and sense of belonging. When instructors foster respect, care, and mutual support, classrooms become more than venues for content delivery. The classes become communities in which students feel seen, valued, and accepted, which can be especially important for students from underrepresented or marginalized backgrounds, who may otherwise feel alienated in academic settings. A trusting classroom environment reduces the quantity and impact of perceived threats (e.g., fear of failure, social judgement, etc.), enabling deeper student engagement in the learning process for all students.

When students view their classroom environment as predictable, respectful, and supportive, they demonstrate greater willingness to be vulnerable, ask questions, and persist through challenging tasks. Additionally, evidence from instructors that use active learning, group-based, and collaborative instructional techniques, indicates that trust supports motivation, satisfaction, collaboration, knowledge sharing, and idea generation (Huff and Kelley, 2003; Staples et al., 2008; Zhu et al., 2009). A trusting classroom environment also provides the foundation for individual students and instructors to build interpersonal trust.

Focus is at the individual level, a one-to-one relationship, and is centered on a student’s perception of an instructor’s level of acceptance, care, responsiveness, and understanding. This four-item factor measures the extent of personalized support that a student feels they receive from their instructor, including: listens very carefully to me; ‘gets me’, helps me overcome personal challenges, and treats me like a person. Higher scores indicate that the instructor has successfully established an interpersonal connection that goes beyond the classroom dynamic. Such an instructor is seen as a trustworthy figure in a student’s life, and the student feels that this is someone they can turn to when facing personal and academic challenges.

Students who perceive strong interpersonal connections with their instructor report a greater sense of self-efficacy and lower classroom anxiety (Jiang et al., 2024). These connections are also associated with performing well academically and engaging more deeply when compared with peers who find their instructors intimidating (Liu, 2024; Roorda et al., 2017; van Herpen et al., 2024). Interpersonal relationships between students and teachers’ plays a critical role in shaping academic experiences (Kim and Lundberg, 2016), and interpersonal trust has been identified as central to educational improvement and student well-being (Bryk and Schneider, 2003). For college students trust in their instructor, the missing link between perceptions of an interpersonal relationship and the trust level within that relationship has been a validated measure. The findings of the present study have addressed this with the TRUST instrument (Table 8).

Increased awareness of the Trust and Responsiveness in Undergraduate Student Teaching (TRUST) instrument’s contents enable instructors to cultivate genuine connections and interactions within the classroom and with students. Trust can also serve as a barometer for evaluating the effectiveness of new pedagogical strategies. For instance, if students perceive that the instructor can connect with them, they may be more inclined to engage in that instructor’s recommended study practices, thereby enhancing their learning experience.

Implementing classroom trust and interpersonal trust behaviors effectively often requires instructors to place greater emphasis on, or a reframe of, their in-class interactions with students (Bossér and Lindahl, 2019). These actions shape undergraduate students’ perceptions of their instructor’s classroom personality, behaviors, and interactions. A positive perception of these traits suggests that students find their instructor approachable and trustworthy. Importantly, these perceptions do not necessarily indicate a close personal connection, but rather a favorable impression of the instructor’s overall classroom interactions.

In practice, instructors may already engage in some of these trust-building actions. Enhancing student trust through any of these actions is an achievable goal for most instructors, and it is complementary to sound educational practices (Handelsman et al., 2004). The ability to listen attentively, understand students’ multifaceted lives, and provide personalized support fosters an engaging learning environment. To enable measurement of these abilities, the TRUST instrument now provides educators with a valuable tool for assessing and enhancing trust in higher education. Understanding college students’ trust in their instructors can lead to improved instructional strategies, enhanced student learning outcomes, and a more engaging classroom environment.

This study included students from various disciplines and institutions who were taking a life science course. Such participants introduce variations in instructors, course content, and assessment structures. Additional investigations should explore these factors in greater detail. Data collection for this study was during the covid-19 outbreak. This could be a limitation of the research because students were sometimes in novel academic learning contexts (e.g., remote learning). This could also be advantageous because student trust in their instructor was heightened during the pandemic and perhaps it exposed trust at the core. In addition, information regarding overall trust or demographics was not used in the present study because our focus was to validate an instrument. Though beyond the scope of this study, the datasets also included an item probing overall trust that the student has in the instructor on a 0–10 scale from 0 (not at all) to 10 (to a great extent). The survey collected student information such as first-generation status, and self-reported GPA. These variables are in the Supplementary materials and will be included in a follow-up study testing for demographic group differences. Finally, this study featured students enrolled in a variety of STEM courses, yet the unit of analysis was a broader categorization of undergraduate students in order to validate the instrument.