Inequitable teacher turnover and performance-based appraisal: a global trend?

ABSTRACT

Student performance data is increasingly used to monitor and evaluate teachers. This study examines whether the turnover intentions – if teachers would change schools given the chance – of teachers in socioeconomically disadvantaged classrooms are moderated by teacher appraisal practices based on student academic performance data at the school system-level. Three-level hierarchical modelling in 46 education systems is conducted based on the Teaching and Learning International Survey (TALIS) from 2018. Results show that the effect of classroom socioeconomic composition on teacher turnover intentions increases as a function of performance-based teacher appraisal. However, this is only true when the appraisal is conducted by external authorities, and not so when conducted by the school management team. The models are then re-run by school consequences of appraisal such as dismissal, financial bonuses, or sanctions, and then by teacher characteristics, including gender, experience, and teaching subject. Experienced teachers in socioeconomically disadvantaged classrooms are more likely to change schools in school systems with more performance data-based teacher appraisal. These results underscore the potential pitfalls of performance data-based accountability systems for students in socioeconomically disadvantaged educational settings.

KEYWORDS:

• Accountability
• teacher appraisal
• educational equity
• teacher turnover

1 Introduction

Recent decades have seen a global push towards evermore educational accountability, where the perceived success of school leaders and teachers is increasingly tied to student performance data (i.e. how well students score on school-based or national and regional assessments) (OECD 2013a; Holloway, Sørensen, and Verger 2017; Smith and Kubacka 2017; Verger, Fontdevila, and Parcerisa 2019). As most schools are publicly funded, the argument for increases in accountability states that children – as well as the larger-society – are entitled to the ‘maximum-benefit’ of education (Brill et al. 2018). An overarching goal of the accountability discourse is the importance of continuously strengthening and improving education systems worldwide. While the use of student performance data for accountability purposes is now widespread, there is mounting concern that this practice puts undue stress on teachers and school leaders and distorts the democratic educational imperative (Ball 2003; Berryhill, Linney, and Fromewick 2014; Biesta 2007; Ball 2016; Jerrim and Sims 2021; Lee and Wong 2003; Valli and Buese 2007). Meanwhile, it has been shown that ‘teaching is increasingly a career of movement in and out’ (Skilbeck and Connell 2003, 32–33). This phenomenon has been observed in several countries and is a growing occurrence globally (Craig 2017). On the one hand, there is the issue of ‘macro’ mobility between the world of teaching and other professions, and on the other, there are ‘micro’ movements, concerning the mobility of teachers between schools (Vagi and Pivovarova 2017). Related to these latter ‘micro’ mobilities, an underexplored – particularly from a cross-national perspective – aspect of performance-based accountability is its relationship to the attrition of teachers in socioeconomically disadvantaged or ‘hard-to-staff’ settings. A wealth of research has demonstrated that schools with higher proportions of socioeconomically disadvantaged students deal with disproportionately higher turnover rates (Bonesrønning, Falch, and Strøm 2005; Feng 2009; Hanushek, Kain, and Rivkin 2004; Scafidi, Sjoquist, and Stinebrickner 2007), but less is known about the institutional correlates of such inequitable turnover.

The accountability drive continues despite mixed effects on student outcomes (Darling-Hammond 2004; Figlio and Loeb 2011; Hanushek and Raymond 2005; McNaughton, Lai, and Hsaio 2012; Ryan et al. 2017; van Geel et al. 2016; Wong, Cook, and Steiner 2010). It has been found that socioeconomic inequities between students widen under various accountability regimes (Strietholt et al. 2019). Historically, performance-driven accountability systems have neglected the issue of educational equity (Lee and Wong, 2003). Higher-achieving (and generally more socioeconomically advantaged) schools and students are rewarded while lower-achieving ones are punished. Darling-Hammond (2004) asks:

Will investments in better teaching, curriculum, and schooling follow the press for new standards? Or will standards and tests built upon a foundation of continued inequality simply certify student failure more visibly and reduce access to future education and employment?

(p. 1048).

The aim of the present study is to empirically address this problem and investigate whether increased student performance data-based accountability practices (in this case, formal teacher appraisal) is linked to increased micro mobilities of teachers between schools. Three-level generalized linear regressions are employed to test whether the relationship between classroom socioeconomic composition and teacher turnover intention (a random classroom-level slope) is moderated by school system-level accountability.

1.1 The accountability paradox

Educational reforms are increasingly focused on the notion of ‘standards’ (Ball 2003, 2016; OECD 2013a, 2013b; UNESCO 2015a; 2015b; UNESCO 2017), but accountability as a concept is hard to pin down (Bovens 2007, 2010). Bovens (2010) writes that while standards are in general seen as positive and normatively defined individual or organizational traits, they are to be distinguished from accountability mechanisms. This paper focuses on this latter definition of ‘accountability as a mechanism,’ an umbrella term for practices which aim to hold educational actors accountable for their performance.¹ In relation to the use of student performance data, the Organization for Economic Cooperation and Development (OECD 2013a) lists practices such as the use and publication of achievement data (as in the case of school league tables), the assessment of teacher practices, as well as the consequences of teacher appraisal as possible accountability mechanisms. There is some variation in the implementation of accountability and school improvement practices schools and school systems (Verger, Fontdevila, and Parcerisa 2019). Rivas and Sanchez (2022) discuss a ‘regulatory governance turn’ in Latin America, while Barbana, Dumay, and Dupriez (2020) highlight the lower-stakes accountability regimes more common in continental Europe. Despite these distinctions, ‘there is always a tension between using data for school development and instructional goals, and using data for accountability goals’ (Schidkamp 2019, 260) or ‘soft’ versus ‘strong’ or ‘neo-liberal’ forms of accountability where teachers are viewed in terms of their utilitarian value (Maroy and Voisin 2017).² There is also evidence of a convergence of test-based accountability reforms globally (Verger, Parcerisa, and Fontdevila 2018).

This paper focuses on a somewhat narrow dimension of accountability: the use of student performance-data to assess teachers and their teaching practices, and variations in the consequences of such assessments. Previous research suggests that such practices may have negative and unintended consequences for teachers as well as students at the lower end of the socioeconomic spectrum (Ball 2003; Darling-Hammond 2004; Jerrim and Sims 2021; Polesel, Rice, and Dulfer 2013). Due to the focus on performance, teaching practices may be constrained in order to ‘teach to the test’ and educational experiences consequently narrowed. In addition, the difficulty of raising student achievement is not constant across schools. While some policies account for the external factors which influence achievement and vary across schools, others do not and hold each school and teacher to a similar standard (Darling-Hammond 2004).

Biesta (2007) traces the accountability movement in education to a wider global cultural shift which has its roots in the capitalist logic of the market, where states become providers and citizens consumers. Such shifts are seen through increasingly marketized education systems which are often characterized by standards, accountability and decentralization (Verger, Parcerisa, and Fontdevila 2018). In theory, performance-based accountability is positioned as a sort of constraint on school marketization and autonomy. According to Brill et al. (2018), ‘the rationale here is that when schools are granted more independence over the methods they use to achieve education outcomes, accountability increases in significance, as a more autonomous, school-led system depends even more heavily on a fair and effective accountability system’ (p. 2). In practice, however, such policies may end up compounding school market mechanisms as they are contingent upon the larger organizational context (Peters and van Nispen 1998). Ball (2003) writes:

Such developments are deeply paradoxical. On the one hand, they are frequently presented as a move away from ‘low-trust’, centralized, forms of employee control. Managerial responsibilities are delegated, initiative and problem-solving are highly valued. On the other hand, new forms of very immediate surveillance and self-monitoring are put in place; e.g. appraisal systems, target-setting, output comparisons.

(p. 219)

While most accept the need for school-improvement, few are satisfied with the ways in which it is carried out (Vanhoof and van Petegem 2007). Some argue that such ‘standards’-based reforms should be used primarily to strengthen curricula and inform better resource allocation and practice, while others argue that the problem stems from a fundamental ‘lack of effort’ on the part of educators, and that higher stakes consequences such as sanctions for those who fail to meet such standards are necessary (Darling-Hammond 2004; Maroy and Voisin 2017). The two perspectives can be summed up as the school-improvement perspective versus the accountability perspective and have also to do with ‘who determines the quality of education: the government or the school itself’ (Vanhoof and van Petegem 2007, 105). Traditionally, the accountability perspective was solely tied to external evaluations, and the school improvement perspective to school-based evaluations. However, school-based evaluations are becoming increasingly influenced by accountability, and the distinction between accountability (external evaluation) and school-improvement (internal evaluation) is no longer as clear. This is mirrored by the blurring between formative and summative teacher appraisal (Smith and Kubacka 2017). Two dimensions of accountability are therefore included in the present study: teacher appraisal conducted by external authorities or bodies (i.e. school inspectors, government authorities) and teacher appraisal conducted by the school management team.

1.2 Teacher turnover intentions and student socioeconomic composition

By now, the importance of teachers is well-accepted. This is especially true for socioeconomically disadvantaged students who have less access to at-home learning supports (Darling-Hammond 2000; Rivkin, Hanushek, and Kain 2005; Hattie 2003; Nilsen and Gustafsson 2016). Despite this, worldwide, the teaching profession faces challenges related to worsening working conditions, workload, and job stress, resulting in issues attracting and retaining individuals to the field, and even more so to socioeconomically disadvantaged settings (see for example Allen, Burgess, and Mayo 2018; Hanushek, Kain, and Rivkin 2004; Sass, Seal, and Martin 2011; Vagi and Pivovarova 2017; Valli and Buese 2007). Individuals who enter the profession with the goal of positively impacting society often leave feeling disillusioned (Perryman and Calvert 2020). This may be particularly true of novice teachers (Fantilli and Macdougall 2009; Hanushek and Rivkin 2010b, 2010a; Ingersoll 2001; Marinell et al. 2013), but research is scarce regarding other teacher characteristics associated with higher turnover rates. Issues related to attrition in low-socioeconomic status (SES) settings have been found in Sweden, Canada, the USA, Australia, Chile, and England (Allen, Burgess, and Mayo 2018; Carver-Thomas and Darling-Hammond 2017; Den; Elacqua, Hincapie, and Martinez 2019; Ingersoll 2002; Karsentil and Collin, 2013; Lindqvist et al., 2014; Loeb et al., 2005; Scafidi et al., 2007). A high teacher attrition rate hinders school climate and the success of students (Hanushek, Rivkin, and Schiman 2016). This may be because the school is forced to hire individuals with lower competency or qualification levels, the staff are not used to working alongside each other, or the school cannot fill the open spots and are forced to enlarge classes (Ingersoll 2001; Sorensen and Ladd 2020).

It therefore goes without saying that it is in the interest of education systems to keep teachers in hard-to-staff settings. In the context of this article, hard-to-staff signifies a socioeconomically disadvantaged context, and stems from previous work demonstrating issues of attrition in such settings (Bonesrønning, Falch, and Strøm 2005; Boyd et al. 2005; Darling-Hammond 1997; Hanushek et al., 2001; Scafidi, Sjoquist, and Stinebrickner 2007). This may be a school with a high proportion of socioeconomically disadvantaged (low-SES) students in more segregated school systems, or a classroom with a high proportion of low-SES students in case a system does not practice mixed ability grouping within schools. In both cases, the classroom setting is the most important. Allensworth, Ponisciak, and Mazzeo (2009) outline several main reasons teachers cite their dissatisfaction: principal effectiveness, dysfunctional administration, challenging students, low salary, and limited autonomy due which may be due to additional accountability practices. Sims and Allen (2018) note that often there are more job openings in such schools, and therefore teachers may take such positions with the intention to leave as soon as another opportunity comes along. In Turkey, for instance, teacher assignment is based on a seniority score, which is positively influenced by working in such schools (Özoğlu 2015). There is therefore likely a proportion of teachers who intend to leave such contexts regardless of school leadership or working conditions. As such, there may be a somewhat immovable link between classroom/school SES and teacher attrition across education systems. However, we know little about the institutional determinants of the remaining variation in teacher turnover.

1.2.1 Accountability and teacher turnover

While there is a host of literature discussing accountability and teacher appraisal practices (see for example Holloway, Sørensen, and Verger 2017), cross-national literature linking accountability to teacher turnover is scarce. There are several reasons that performance-based accountability may be tied to teacher turnover, but they are primarily due to increased workload, a lack of job autonomy and sense of purpose, and stress (Byrd-Blake et al. 2010; Feng, Figlio, and Sass 2018; Richards 2012; Ryan et al. 2017). Lower-performing schools already face greater administrative pressures in general (Browes 2021). A recent study by Jerrim and Sims (2021) found a link between performance data-driven accountability and teacher stress cross-nationally. von der Embse et al. (2016) find this link in four states in the USA. Using data from Florida, Feng, Figlio, and Sass (2018) find that schools which fell into a low-performing category experienced a jump in the probability of teacher turnover relative to schools that missed the cutoff. This was particularly true for the ‘best’³ teachers in the school. With data from Chile, Elacqua, Hincapie, and Martinez (2019) find that accountability practices (such as sanctions) result in higher teacher turnover rates in low-performing schools. In short, performance data-based accountability entails closer monitoring of the performance of teachers, a loss of job autonomy, longer hours and poor working conditions, and a school climate characterized by stress and ‘emotional contagion’ amongst the staff (Jerrim and Sims 2021). Other studies outline how accountability and teacher evaluation policies are often mismatched with teacher perspectives (La Londe 2017; Pizmony-Levy and Woolsey 2017). In one of the only cross-national studies on the topic, Smith and Holloway (2020) use OECD TALIS data from 2013 and find a link between decreased teacher satisfaction with their current school⁴ and school testing culture. This study follows a similar line of questioning but expands the analysis to 46 countries participating in TALIS 2018, with a special focus on turnover intentions specifically in socioeconomically disadvantaged contexts.

The main hypothesis of this paper is that teachers working in classrooms with a high proportion of low-SES students face a more challenging uphill battle when it comes to influencing student achievement as compared to their peers working in more affluent or even mixed socioeconomic settings (Lazear 2001; Polesel, Rice, and Dulfer 2013). Therefore, if their evaluations are tied to student achievement data, they face a higher likelihood of stress and lower levels of autonomy over their profession, and therefore may be more likely to wish to work in a different school. Over and above this, I expand on previous research and investigate whether the consequences of appraisal interact with the classroom SES-turnover slope. Given that monetary sanctions and incentives are increasingly common consequences of appraisal in high-stakes accountability systems (OECD 2021), it is relevant to determine whether the relationship between accountability and teacher turnover intentions varies according to these consequences as past evidence on the importance of such incentives is mixed (Boyd et al. 2005; Falch 2011; Feng 2009; Hanushek, Kain, and Rivkin 2004; Jerrim and Sims 2021).

The study investigates the following research questions:

1. Are teachers’ odds of reporting that they wish to change schools positively related to the proportion of socioeconomically disadvantaged students in their target classroom?

2a. Is the marginal effect of classroom socioeconomic composition on teachers’ probability of wishing to change schools positively related to the level of achievement-data based accountability (teacher appraisal) in a school system?

2b. Does this relationship vary depending on who conducts the teacher appraisal?

3. Does the marginal effect of class SES composition on a teachers’ probability of changing schools vary as a function of teacher characteristics (gender, experience, subject-matter) or school consequences of assessment (dismissal, financial bonus, or material sanctions)?

2 Data and method

2.1 Data source

Data for this study come from the Teaching and Learning International Survey (TALIS) by the Organization for Economic Cooperation and Development (OECD) in 2018. The analysis incorporates between 40 and 46 systems (due to missingness on certain variables in the subgroup analyses). TALIS follows a two-stage stratified sampling procedure, and samples a nationally representative group of approximately 200 schools selected with a probability proportional to their size. Within each school, questionnaires are administered to the principal as well as a selection of approximately 20 teachers with equal probability based on the subjects taught at their school. Table 1 displays the descriptive statistics for system-level variables and number of teachers and schools sampled in each country. For descriptive statistics of all variables, see Table A1 in the Appendix.

Table 1. Descriptive statistics of system-level variables.

Code	Country		N Schools	N Teachers	External	SMT	HDI
36	AUS	Australia	233	3573	0.34	0.89	0.94
40	AUT	Austria	246	4255	0.41	0.43	0.91
56	BEL	Belgium	302	5257	0.30	0.52	0.92
76	BRA	Brazil	185	2447	0.43	0.83	0.76
100	BGR	Bulgaria	200	2862	0.73	0.72	0.82
152	CHL	Chile	179	1963	0.28	0.89	0.85
158	CTPE	Chinese Taipei	200	3835	0.10	0.78	0.92
170	COL	Colombia	154	2398	0.30	0.76	0.76
191	HRV	Croatia	188	3358	0.42	0.71	0.84
196	CYP	Cyprus	88	1611	0.39	0.66	0.87
203	CZE	Czech Republic	219	3447	0.29	0.94	0.89
208	DEN	Denmark	141	2001	0.34	0.87	0.93
233	EST	Estonia	195	3004	0.55	0.89	0.88
246	FIN	Finland	148	2851	0.14	0.42	0.93
250	FRA	France	176	3006	0.36	0.57	0.89
268	GEO	Georgia	192	3101	0.30	0.74	0.79
348	HUN	Hungary	189	3245	0.38	0.82	0.85
376	ISR	Israel	172	2627	0.49	0.91	0.91
380	ITA	Italy	191	3612	0.21	0.74	0.88
392	JPN	Japan	196	3555	0.25	0.70	0.92
398	KAZ	Kazakhstan	331	6566	0.52	0.84	0.82
410	KOR	South Korea	163	2931	0.17	0.49	0.91
428	LVA	Latvia	135	2315	0.54	0.97	0.85
440	LTU	Lithuania	195	3759	0.53	0.95	0.87
470	MLT	Malta	55	1656	0.48	0.80	0.89
484	MEX	Mexico	193	2926	0.43	0.65	0.77
528	NLD	Netherlands	116	1884	0.17	0.87	0.93
554	NZE	New Zealand	185	2257	0.21	0.88	0.92
578	NOR	Norway	185	4154	0.21	0.91	0.95
620	PRT	Portugal	200	3676	0.54	0.69	0.85
643	RUS	Russia	230	4011	0.69	0.95	0.82
682	SAU	Saudi Arabia	179	2744	0.57	0.52	0.86
702	SGP	Singapore	169	3280	0.07	0.96	0.94
703	SVK	Slovakia	176	3015	0.49	0.89	0.86
704	VNM	Vietnam	196	3825	0.36	0.74	0.70
705	SVN	Slovenia	132	2094	0.42	0.45	0.90
710	ZAF	South Africa	170	2046	0.48	0.86	0.71
724	ESP	Spain	399	7407	0.60	0.80	0.89
752	SWE	Sweden	180	2782	0.26	0.78	0.94
784	UAE	United Arab Emirates	521	8648	0.48	0.86	0.87
792	TUR	Turkey	196	3952	0.26	0.52	0.81
840	USA	USA	165	2560	0.20	0.75	0.92
926	ENG	England	149	2376	0.38	0.98	0.93
9134	CAN	Canada	122	1077	0.37	0.61	0.93
9642	ROM	Romania	199	3658	0.70	0.73	0.82
32001	ARG	Argentina	130	2099	1	1	0.83
156001	CHN	China	198	3976	0.31	0.92	0.76

Note. Several education systems are represented in place of the entire country (CAN = Alberta, CHN = Shanghai, and ARG = Buenos Aires). ‘External accountability,’ and ‘SMT accountability’ refer to the percentage of schools with each of these authorities responsible for using student achievement data for teacher appraisal. To control for differing national contexts, the HDI (Human Development Index, UNDP 2020a) is included as a system-level control variable. The HDI measure is computed based on a longevity/health dimension (life expectancy at birth), an educational dimension (expected years of schooling and mean years of schooling), and a standard of living dimensions (gross national income per capita). For more information as to how this scale is constructed, see UNDP (2020b).

TALIS provides teacher weights that are inversely proportional to the probability of selecting a given teacher within a school to the TALIS sample. However, there are no senate weights provided by TALIS, which are sometimes necessary in cross-national analysis to ensure that estimates are not biased by a differing number of teachers sampled per education system. Therefore, both teacher and school senate weights are computed.⁵

2.2 Variables

For descriptive statistics of all variables, see Table A1 in the Appendix. The dependent variable is a question asked to the teachers in the teacher questionnaire about their desire to change schools (see Figure 1). Here, ‘agree’ and ‘strongly agree’ are coded as 1, while ‘strongly disagree’ and ‘disagree’ are coded as 0.

Figure 1. Turnover intentions in the 2018 TALIS Teacher Questionnaire.

Teachers were also asked about the socioeconomic composition of their target classrooms, asking them to indicate the proportion of socioeconomically disadvantaged students, ranging from 1 = none, 2 = 1% to 10%, 3 = 11% − 30%, 4 = 31% − 60%, 5 = more than 60%. The focus is on the classroom level so as not to exclude teachers working in socioeconomically heterogeneous schools, but where within-school teacher sorting may occur. Teacher-level control variables include experience (coded 1 if a teacher has less than 5 years of experience), subject matter specialization (coded 1 if a teacher does not have the relevant education for the subject matter they currently teach across mathematics, science, humanities, and other subjects), job satisfaction⁶ (scale fulfills metric invariance across countries), self-efficacy (scale fulfills metric invariance across countries), and gender (coded 1 as female).

The explanatory variable of interest comes from the principal questionnaire and is therefore at the school level before aggregation. Principals were asked, ‘Who uses the following types of information as part of the formal appraisal of teachers’ work in this school?’ (OECD 2019, 12). This study focuses on answers related to student-achievement data, such as students’ external results (e.g. national test scores) and school-based and classroom-based results (e.g. performance results, project results, test scores). For the present paper, only school management team (SMT) or external individuals or bodies were included.⁷ Each of these categories was coded separately as 1 for each school. Each was then aggregated to the country level, with school weights applied.

Principals were also asked to indicate the frequency with which any of the following occur in the school, following a formal teacher appraisal, material sanctions such as reduced annual increases in pay are imposed, salary increase or payment of a financial bonus, or a dismissal or non-renewal of contract. Frequency-related responses include ‘never,’ ‘sometimes,’ ‘most of the time,’ or ‘always.’ The variables for sanctions, financial bonus and dismissal were each coded as 1 if the principal did not select ‘never.’

Control variables include school type (coded as 1 if private school), school location (coded as 1 if school is in a rural or remote location or a small town), and the proportion of students from socioeconomically disadvantaged homes (1 = none, 2 = 1% to 10%, 3 = 11% − 30%, 4 = 31% − 60%, 5 = more than 60%) For more information on these variables, see OECD (2019, 8). Last, the hierarchical models control for institutional settings via the Human Development Index (HDI, UNDP 2020a).

2.3 Method

Mplus (Muthén and Muthén 1998–2018) and RStudio were used for analyses. Mplus automatically employs either full information maximum likelihood (FIML) or multiple imputation (in the case of Bayesian analysis) to deal with missing values. In RStudio, multivariate imputation by chained equation (the ‘mice’ package, van Buuren and Groothuis-Oudshoorn 2011) was conducted to account for missing variables into a newly imputed complete dataset. As auxiliary variables, the imputation model included dummy variable for each education system, school socioeconomic composition, location, type, as well as teacher characteristics including age, gender, experience, subject matter, self-efficacy, job-satisfaction and social-utility value.

For the first research question, for each country the relationship between classroom socioeconomic composition and a teacher’s odds of wishing to change schools is examined. For this, individual logistic regression models are run in RStudio with the ‘glm’ function, clustering the standard errors by school (Abadie et al. 2017) using the ‘ClubSandwich’ package (Pustejovsky 2021) and the heteroscedasticity consistent ‘CR0’ variance estimator to account for the nested data structure. The model is as follows:

(1) $\mathit{Log}\left(T_{\mathit{ij}}\mathit{odds}\right)=\mathit{\alpha }+\mathit{\beta }.\mathit{ClassroomSE}{S}_{\mathit{ij}}+\mathit{\gamma }.\mathit{Teacher}.\mathit{control}{s}_{\mathit{ij}}+\mathit{\delta }.\mathit{School}.\mathit{control}{s}_j$(1)

Where $T_{\mathit{ij}}$ refers to whether the teacher i in school j wishes to change schools and coefficient $\mathit{\beta }$.$\mathit{ClassroomSES}$ refers to the socioeconomic composition of a classroom, $\mathit{\gamma }$ the teacher controls, and δ the school controls. Estimates are reported in odds ratios alongside the 95% confidence intervals. The odds ratios minus 1 divided by their standard errors (the t-value) is taken to plot the estimates alongside the country-level accountability measures.

For the final two research questions, three-level generalized regressions are estimated in Mplus (type = ‘THREELEVEL RANDOM’) following a cumulative model building procedure (Raudenbush and Bryk 2002; Sommet and Morselli 2017). The intra-class correlation coefficient of the dependent variable is 0.05. The model of interest estimates whether the relationship between wishing to change schools and classroom SES (the random slope) varies as a function of school-system level accountability practices:

(2) $T_{\mathit{ijk}}=\mathit{\alpha }+\mathit{\beta }.\mathit{ClassroomSE}{S}_{\mathit{ijk}}+\mathit{\gamma }.\mathit{Teacher}.\mathit{control}{s}_{\mathit{ijk}}+\mathit{\delta }.\mathit{School}.\mathit{control}{s}_{\mathit{jk}}+\mathit{\theta }.\mathit{HD}{I}_k+\mathit{\pi }.\mathit{Accoun}{t}_k+(\mathit{\sigma }.\mathit{ClassroomSE}{S}_{\mathit{ijk}}\times \mathit{Accoun}{t}_k)+v_{\mathit{jk}}+u_{1\mathit{k}}$(2)

Where the probability ($T_{\mathit{ijk}}$) that a teacher i in school j in country k wishes to change schools is estimated as a function of the overall likelihood that teachers wish to change schools across countries (α), classroom SES ($\mathit{\beta })$, teacher-level controls ($\mathit{\gamma })$, school level controls (δ), country HDI (θ), and country-level accountability (π). The model includes a cross-level interaction (σ), a random effect at the school level ($v_{\mathit{jk}})$ and a random slope and intercept at the country-level ($u_{1\mathit{k}}).$ As the system level variables represent the percentage of schools where teacher appraisal based on achievement-data is conducted externally or by the school management team, there is no need to center. Random slope models are quite useless if effects move in opposite directions, resulting in a null effect. Within-country two-level models are also estimated for this reason, and these results can be found Figure B in the Appendix. In a final step, the model is estimated again for subsamples of the teacher population: by gender, novice status, subject matter taught (mathematics/science against all other subjects), and whether their appraisal has consequences such as dismissal, financial bonus, or sanctions.

While there is debate about the usefulness of comparing educational systems which necessarily have a host of unobservable cultural and demographic differences (Nóvoa and Yariv-Marshal, 2003; Strietholt et al. 2019), educational policies often lack sufficient variability within countries, thereby complicating their assessment (Strietholt et al. 2014). Unobservable selection effects at school- and student-levels can also be accounted for when aggregating institutional variables to the country level, such as if a certain type of school is targeted for external teacher appraisal. Nevertheless, although the models control for HDI, it must be stressed that the results are correlational, and that other system-level unobserved variables may be present.

3 Results

3.1 Inequitable teacher turnover intentions within countries

The first research question estimates the odds of a teacher wishing to change to a different school as a function of classroom socioeconomic composition for each country, with characteristics such as location and school type under control. Figure 2 illustrates this relationship.

Figure 2. Teacher turnover intentions and classroom socioeconomic composition .

Note. The x-axis displays the within-country odds ratios (and 95% CIs) of a teacher wishing to change to a different school per unit increase on the classroom SES disadvantage scale of their target classroom. Standard errors clustered by school. Teacher weights applied.

In Figure 2, countries are ranked by the size of their odds ratios. An odds ratio of 1 implies that a teacher is just as likely to wish to change schools at all levels of SES classroom composition (i.e. the null hypothesis). An odds ratio of 1.10 (such as in Japan or Portugal) signifies that the odds of a teacher wishing to change schools increase by 10 percentage points for every unit increase on the class SES disadvantage scale. By and large, a majority of education systems save for 11 (Georgia, Colombia, Sweden, Mexico, Italy, Japan, Brazil, Denmark, Norway, Croatia and Lithuania) display statistically significant odds ratios above 1. The educational systems with negligible relations between socioeconomic composition and teacher turnover intentions do not have any immediately obvious unifying characteristics, highlighting the fact that the determinants of teacher mobility are complex and not straightforwardly predicted.

The standard errors account for the hierarchical sample design by using a heteroskedasticity consistent variance estimator and clustering the standard errors at the school level. It should be highlighted that the estimates are displayed alongside very large confidence intervals and therefore a high degree of uncertainty. Nevertheless, the results confirm findings from national studies and demonstrate that socioeconomically motivated teacher turnover intentions are indeed a problem in most educational systems. The median odds ratio in the sample was 1.22 (Austria). Georgia (.95) and Colombia (1.02) displayed the smallest increases in odds as a function of each unit increase in classroom SES, while Romania (1.56), Latvia (1.51) and Canada (Alberta) (1.55) displayed the largest increases in odds as a function of classroom socioeconomic composition.

3.2 School system-level accountability as a moderating factor

Based on the above findings, it is clear that few education systems (out of the pool of 46) have managed to avoid the problem of teacher turnover intent altogether. Is this relationship stronger in school systems with more performance-based teacher appraisal? Figure 3 displays the cross-national relationship between classroom SES-teacher turnover intent and the percentage of schools which practice teacher appraisal based on performance data.

Figure 3. Cross-national relationships between accountability and classroom SES-based turnover intentions.

Note. The y-axis displays the within-country odds ratio – 1 divided by the standard error. The x-axis displays the proportion of schools where teacher appraisal is conducted by either external bodies or the school management team.

At first glance, it appears that the main focus of this paper (the link between performance data-based teacher appraisal and teacher turnover intentions in low-SES contexts) is only confirmed in the cases where teacher appraisal is carried out externally (by school inspectors or government representatives, for example). Education systems show considerable differences in within-country variability depending on who conducts the teacher appraisal. In most education systems, a majority of schools appraise teachers according to performance data via the school management team. There are a few exceptions to this, including Finland, Austria, Slovenia, and South Korea, in which less than half of schools practice this type of appraisal. In contrast, much fewer educational systems practice performance data-based appraisal via external actors. Argentina (Buenos Aires), Bulgaria, Romania and Russia display the most widespread use. Although selection effects exist at the school level (for instance, if certain teachers are monitored for unobservable reasons endogenously related to teacher turnover intentions), these are removed accountability practices are aggregated to the system-level and turnover intention rates across countries are the subject of analysis. From Figure 3, I conclude that the turnover-classroom SES relationship is positively related to the system level of external accountability. Inequitable turnover intentions are more frequent in countries where teacher practices are monitored externally using student performance data.

Three-level regressions then model this relationship empirically. Figure 4 displays the estimates from this model where turnover intentions are regressed on classroom socioeconomic composition (at the classroom level), controlling for the Human Development Index at the system-level. Figure 4 shows that for each added level of socioeconomic disadvantage, the odds of wishing to change to a different school increases. However, it does not seem to matter whether there are more than 30 percent of socioeconomically disadvantaged students in the classroom or more than 60 percent, as the odds ratios are not significantly different at the 95% level.

Figure 4. Three-level generalized linear estimates of teachers turnover intentions on classroom socioeconomic composition.

Note. Model controls for HDI.

The findings so far demonstrate a clear link between turnover intentions and classroom socioeconomic composition that is found in a majority of the within-country logistic regressions and is also generalizable across the countries in the TALIS sample. Does this relationship hold when additional controls are added into the model? Moving to the successive hierarchical model results conducted in Mplus, Table 2 shows the cumulative models and estimates for research questions 2a and 2b, which ask primarily whether the relationship between classroom SES and teacher turnover intent differs as a function of school accountability after including teacher level controls such as gender, experience, subject specialization, job satisfaction and self-efficacy, and school level controls such as school location, type, and socioeconomic composition. The point estimates represent the median of the posterior probability distributions for each parameter.

Table 2. Estimates from successive three-level generalized models.

	Dependent variable: ‘I would like to change to a different school if that were possible’ = 1
	(1)		(2)		(3)
	Coefficient	SD	Coefficient	SD	Coefficient	SD
Classroom level
Classroom SES	0.137***	0.011
Novice teacher	0.223***	0.012	0.186***	0.013	0.197***	0.014
Unspecialized teacher	0.100***	0.011	0.066***	0.012	0.079***	0.014
Gender	0.065***	0.011	-0.011	0.012	0.007	0.014
Job satisfaction	-0.234***	0.006	-0.307***	0.006	-0.286***	0.009
Self-efficacy	0.166***	0.008	0.037***	0.005	0.056***	0.008
School level
School SES	0.022***	0.007	0.061***	0.009	0.059***	0.005
Private school	0.120***	0.020	0.135***	0.023	0.131***	0.018
Rural school	0.050***	0.013	0.056***	0.013	0.053***	0.013
System level
HDI	-0.541	0.824	-0.855	0.903	-0.886	0.932
Accountability (External)	-0.336	0.353	-0.543	0.352	-0.512	0.355
Accountability (SMT)	-0.361	0.353	-0.313	0.409	-0.271	0.377
Cross-level interactions
Classroom SES × Accountability (External)					0.105*	0.058
Classroom SES × Accountability (SMT)					0.040	0.061
Classroom SES × HDI					0.246*	0.155
Variance components
School slope			0.004***	0.001	0.003***	0.001
Country slope			0.001***	0.001	0.001***	0.001
N
Teacher	130979
School	8814
System	46

*** = p < .000, ** = p < .01, * = p < .05.

Model 1 (Table 2) shows that the effect of one unit increase in the classroom socioeconomic composition scale equates to about a 0.14 probability increase in turnover intentions averaged across all countries in the sample. In addition, novice teachers are significantly more likely to wish to change schools, as are teachers without a specialization in the subject matter they currently teach. Teachers with higher levels of self-efficacy were also more likely to wish to change schools, and there was an inverse relationship between job satisfaction and a teacher’s turnover intent. Moving to the school level, in line with the main hypothesis of this paper, teachers were more likely to report that they wish to change schools if a school had a higher proportion of low-SES students, and if they worked in a rural school. These results suggest that classroom socioeconomic composition matters alongside school characteristics.

Model 2 (Table 2) shows the same three-level model, this time with the addition of a random slope (the relationship between classroom SES and teacher turnover intent). The variance of the random slope in model 2 varies significantly across education systems and therefore the estimation proceeds to the next step. In the final model (model 3) in Table 2, the cross-level interactions are introduced, where the level one slope is regressed on the level three predictor variables of those responsible for teacher appraisal based on student-achievement data and HDI. The results confirm the findings depicted in Figure 3. In other words, they confirm that the effect of classroom SES on teacher turnover intent is stronger in school systems where teacher appraisal based on student achievement data is conducted externally. The coefficient represents a median probability increase of about 0.11 increasewith each classroom SES disadvantage category. Similar to the plot in Figure 3, there was no change in the marginal effect in relation to teacher appraisal based on student achievement data conducted by the school management team. However, the relationship between classroom SES and teacher turnover intent was also stronger in higher-HDI countries. . Here, the median probability increase in the slope as a function of HDI was about 0.25. Last, the variance of the slope remains statistically significant in Model 3, which suggests that factors not captured by the model are relevant for explaining the slope variation.

To further investigate the null results for accountability conducted by the school management team, within-country models are estimated for this same research question (with school-level accountability). These results are in Figure B in the Appendix, and confirm that the relationship is both positive and negative and therefore no general cross-national pattern can be assumed.

3.3 Teacher characteristics and school consequences of appraisal

The three-level hierarchical generalized models are then re-run for different sub-groups of the data, by school-level consequences of formal teacher appraisal (dismissal, salary changes, and sanctions following formal teacher appraisal, see Table 3), as well as by certain teacher characteristics (gender, experience, and subject area, see Table 4). Only the estimates of the random slope regressed on teacher appraisal actors and HDI are presented here. For full cumulative models for each subgroup see Tables C1 and C2 in the Appendix. Table 3 shows teachers who work in schools where formal teacher appraisal may have serious consequences for their careers. In each case, the estimates are positioned in contrast to those where such consequences are not a part of school practice. The relationship between the random slope and school accountability practices only varies in a few cases. There is a positive association between external teacher appraisal and the slope in schools where dismissal is not a consequence of teacher appraisal, especially in higher HDI countries. There is no difference in the relationship between accountability and the slope between teachers in schools with financial bonuses or sanctions. Table 4 shows the model run for sub-groups based on teacher gender, experience, and subject matter. The positive moderating effect of external teacher appraisal on the random slope was larger for male teachers than for female teachers. Teachers with more than five years of experience were also more likely to report wishing to leave the school in school systems with more external teacher appraisal. Finally, there was no significant difference between teachers of mathematics and science teachers and teachers of other subjects.

Table 3. Estimates from three-level random slope models by school consequence subgroup.

	DV – Teacher turnover intent – Class SES random slope
	(1)				(2)				(3)
	Dismissal		No dismissal		Financial bonus		No financial bonus		Sanctions		No sanctions
	Coefficient	SD	Coefficient	SD	Coefficient	SD	Coefficient	SD	Coefficient	SD	Coefficient	SD
Teacher appraisal (External)	0.089	0.069	0.134*	0.085	0.109	0.145	0.106	0.084	0.131	0.126	0.085	.059
Teacher appraisal (SMT)	0.062	0.077	0.003	0.086	0.026	0.150	0.030	0.068	-0.054	0.142	0.051	.059
HDI	1.11	0.200	0.427*	0.252	0.060	0.336	0.336	0.224	0.043	0.287	0.203	.117
Teacher controls	Yes		Yes		Yes		Yes		Yes		Yes
School controls	Yes		Yes		Yes		Yes		Yes		Yes
HDI	Yes		Yes		Yes		Yes		Yes		Yes
N
Teacher	59,907		54,612		14,114		96,086		17,941		87,786
School	3985		3689		4679		7790		1250		5910
Country	46		46		41		41		40		41

*** = p < .000, ** = p < .01, * = p < .05.

Table 4. Estimates from three-level random slope models by teacher characteristic subgroup.

	DV – Teacher turnover intent – Class SES random slope
	(1)				(2)				(3)
	Female		Male		Novice		Experienced		Math/sci		Other
	Coefficient	SD	Coefficient	SD	Coefficient	SD	Coefficient	SD	Coefficient	SD	Coefficient	SD
Teacher appraisal (External)	0.086*	0.066	0.101*	0.060	0.022	0.100	0.132*	0.057	0.107	0.079	0.090	0.067
Teacher appraisal (SMT)	0.028	0.069	-0.030	0.090	0.074	0.100	0.028	0.059	0.008	0.083	0.049	.069
HDI	0.263	0.162	0.190	0.179	0.458*	0.281	0.226	0.173	0.574**	0.228	0.065	.193
Teacher controls	Yes		Yes		Yes		Yes		Yes		Yes
School controls	Yes		Yes		Yes		Yes		Yes		Yes
HDI	Yes		Yes		Yes		Yes		Yes		Yes
N
Teacher	90583		40,396		23,122		107,857		42,717		88,262
School	8585		8164		7255		8772		8579		8694
Country	46		46		46		46		46		46

*** = p < .000, ** = p < .01, * = p < .05.

4 Discussion

Educational scholars of different backgrounds contest the role of performance-based accountability practices such as the formal appraisal of teacher effectiveness based on student test scores, but such practices continue to become more widespread. While some report on and advocate for the benefits of accountability in the school system for student learning (Figlio and Loeb 2011; Hanushek and Raymond 2005), others relay the potential risks of an unchecked focus on performance, particularly for teachers (Ball 2003; Darling-Hammond 2004; Ryan et al. 2017; Valli and Buese 2007). Increased accountability may lead to increased workload and a diminished sense of professional well-being and autonomy (Valli and Buese 2007). Despite this debate, student achievement-based accountability practices are becoming more widely practiced on a global scale. Jerrim and Sims (2021) recently found a modest link between school system-level accountability and the degree of teacher and principal stress cross-nationally, and Smith and Holloway (2020) find a link between performance data-based teacher appraisal and job satisfaction. This study expanded upon this line of questioning and confirmed that higher levels of performance data-based accountability are also linked to teachers’ intentions of changing schools, particularly if they are working in lower-SES environments. There is also an interest as to whether this effect varied across subgroups based on different school consequences of the appraisal and different teacher characteristics. Ultimately, this study is a first step in the cross-national investigation of the policy-level correlates of teacher attrition and turnover.

The results from the three-level models show that the marginal effect of classroom SES on a teacher’s turnover intent increases as a function of system-level accountability. However, this was only found for external appraisal.⁸ There was no difference in the association in school systems with higher levels of achievement-based teacher appraisal conducted by the school management team. In general, this finding also supports a clearer distinction between internal and external accountability as outlined previously by Vanhoof and van Petegem (2007). External monitoring may not be achieving its intended consequence when it comes to appraisal. The first implication, therefore, is to limit appraisal to those with knowledge of the particular school context. Past work has suggested that principals may play a mediating role between school goals and external monitoring (Qian and Walker 2019). As such, teacher appraisal conducted by the school management team (which often involves the principal) may have a more ‘sincere’ and constructive quality (Qian and Walker 2019). Furthermore, Ehren, Paterson and Baxter (2019) outline several ways in which low levels of trust intertwine with accountability policies. First, a lack of expertise on the part of external monitors, second, a misunderstanding and lack of support for why educational actors fail to meet targets, and third, a lack of a shared sense of purpose and view regarding the accountability exercise.

There was also a difference in the marginal effect of classroom SES on teacher turnover intent (the slope) as a function of HDI, or the level of economic and social development in a country. This signifies that teachers in low-SES classrooms were more likely to report wishing to change schools if they lived in more economically developed countries. Past work from Chile has pointed out that in some cases, teachers in lower-SES settings are more likely to stay put throughout their careers (Meckes and Bascope 2012). This may be due to a lack of job opportunities in more remote (and often lower-income) settings, but it could also be – as mentioned earlier – that a greater proportion of teachers in higher-income countries enter the profession with intentions to change to more desirable educational settings later in their careers.

Regarding the analyses across subgroups, there was a mostly null pattern of results, with a few exceptions. The anti-compensatory relationship between accountability and teacher attrition was stronger for teachers in schools where dismissal was not a consequence of appraisal. There were no significant results regarding either financial bonus or material sanctions. These results were somewhat surprising, as the possibility of dismissal may entail a more punitive, high-stakes (and stressful) working environment. In such a case, we would expect to see a stronger relationship between the slope and system level accountability in schools where dismissal or material sanctions were imposed. However, Jerrim and Sims (2021) similarly found that facing dismissal did not predict a stronger relationship between accountability and teacher stress cross-nationally. It could be that schools less likely to fire teachers as a consequence of their appraisal have greater issues related to teacher retention and hiring in the first place as compared to schools where firing is common practice. In this way, there may be some unobserved characteristics associated with the variables related to school consequences. There may also be an important mediating role of the perceived pressure of accountability mechanisms and past work has found that schools experience pressure differently (Verger, Ferrer-Esteban, and Parcerisa 2021). By breaking down each subgroup by school consequence, associations between turnover intentions and other school characteristics across contexts can be observed. For instance, in Table C2 Model 2a (Appendix), the association between turnover intentions and school socioeconomic composition and location (rural or remote) remains despite the presence of financial bonus as a possible consequence of appraisal, indicating these are not sufficient to eliminate such potential turnover determinants. Last, female teachers are slightly less likely to report wishing to change schools than male teachers when sanctions are used as a consequence of appraisal.

A noteworthy finding in the teacher characteristic subgroup analyses demonstrated that experienced teachers were more likely to report wishing to change schools in higher accountability systems. This finding was also surprising, given that it goes against some past evidence on general patterns of teacher turnover (Fantilli and Macdougall 2009; Hanushek and Rivkin 2010b, 2010a; Ingersoll 2001; Marinell et al. 2013; Ryan et al. 2017), although these data largely come from a single country (the USA). As Hanushek and Rivkin (2010b) point out, if more stringent accountability practices in socioeconomically disadvantaged schools tend to push out the least effective teachers, it could indeed be a positive thing for students if schools are able to replace them with better teachers. However, our findings are more in line with those of Feng, Figlio, and Sass (2018) who find that accountability pushes out more experienced teachers. Given the link between teacher competence and experience after the first few years on the job (Rice 2003), this finding has negative implications for socioeconomically disadvantaged students in higher-accountability systems.

4.1 Limitations

There are several limitations to this analysis which constrain the findings. First, the findings may be vulnerable to selection effects at the system-level. For instance, school-systems with more challenging teaching conditions or a lower-quality teacher workforce may be more ‘in need’ of and therefore likely to implement external accountability checks. However, this is somewhat unlikely as the models control for HDI, but not every possible confounding factor can be accounted for related to the teacher job market and micro mobility. There was also a considerably higher amount of variation regarding the external accountability measure at the system level as compared to the school management team accountability measure.

It is again worth stressing that the findings are correlational. As such, no causal claims can be made of the present analysis or of the focused upon research questions. However, it does suggest a relationship between the two which warrants further investigation by researchers and policymakers alike. Although the results suggest that teacher attrition issues may be more prevalent in systems where their appraisal is conducted based on student-achievement data externally, the within-country results show that this is not the case for all education systems. In these models (Figure B in Appendix), lower-level selection effects may be present, as schools with ‘school management teams’ may be more or less attractive to teachers than those with external accountability monitors for other unobserved reasons. Nationally-focused research will better answer such remaining questions.

5 Conclusions

While more research will need to replicate the findings with other data sources and variables, this study does have several preliminary implications for educational policymakers. Although the heterogeneity seen in the within-country analyses show that it is difficult to generalize without caveat the relationship between accountability and teacher turnover intentions across education systems, the pattern for external appraisal suggests a global trend. As pointed out by past work (Ball 2003; Jerrim and Sims 2021; Lee and Wong 2003), the long-run effects of increasing external accountability may not be obviously detrimental to students via clear-cut achievement changes but may hurt certain students on a larger scale via inequitable learning opportunities through issues with teacher attrition.

Monitoring teacher and school effectiveness via testing and performance data is repeatedly touted as a type of equalizer with regards to ‘quality assurance’ in schools (OECD 2021). In practice, such approaches likely further entrench educational inequities in many cases given that raising student achievement is much more difficult when students have lower levels of home educational resources and if accountability mechanisms are uncoupled from the local educational context. This study has provided new evidence that such regimes are more likely to push teachers out of more challenging educational settings. Policymakers may need to rethink the use of student performance results in the appraisal of teacher performance in socioeconomically disadvantaged schools and classrooms in favour of other forms of feedback. Moreover, appraisal and feedback should sooner come from those with more contextual school knowledge – such as the school principal or other teachers.

Disclosure statement

No potential conflict of interest was reported by the author.

Additional information

Notes on contributors

Leah Natasha Glassow

Leah N Glassow is a lecturer and researcher at the Department of Education and Special Education at the University of Gothenburg. Her main research interests are in inequalities in teacher sorting and mobility, educational policy, and school equity.

Notes

1. The term ‘standards’-based reforms also refers to this type of accountability in the context of this study.

2. For more on accountability typologies, see Maroy and Voisin (2017).

3. Measured by value-added.

4. Turnover intentions is part of this scale.

5. Although these weights are computed, their use in Mplus is restricted as Mplus employs a Bayesian estimator for three-level models with categorical outcomes. In order to determine that our estimates are not biased by differing sample sizes within countries, a system-level variable is added indicating sample size and the slope regressed on this new variable. The slope is unrelated to the sample size in each country.

6. TALIS distinguishes between job satisfaction at the current school and job satisfaction with the profession. To control for whether teachers have a general feeling with the profession overall, we include the job satisfaction with the profession as a control.

7. Educational systems displayed almost zero within-country variability when appraisal was conducted by the principal.

8. Supplemental analyses (Appendix B) shows that this conclusion is not without caveat.

Appendix

Table A1. Descriptive statistics.

	Mean	SD	Min	Max
Classroom SES	2.29	1.05	1	5
Novice teacher	.18	–	0	1
Unspecialized	.25	–	0	1
Female teacher	.69	–	0	1
Job satisfaction	12.06	2.00	3.27	16.48
Self-efficacy	12.67	2.00	.67	19.22
Turnover intent	.21	–	0	1
School SES	2.53	1.01	1	5
Private school	.34	–	0	1
Rural school	.18	–	0	1
Accountability (External)	.77	.155	.07	1
Accountability (SMT)	.40	.173	.42	1
HDI	.86	.055	.71	.95

Table C1. Teacher characteristics.

	(1a)		(1b)		(2a)		(2b)		(3a)		(3b)
	Novice		Experienced		Math/sci		Other		Female		Male
	Coef.	SD	Coef.	SD	Coef.	SD	Coef.	SD	Coef.	SD	Coef.	SD
Teacher level
Novice					.221***	.021	.186***	.014	.176***	.016	.191***	.022
Gender	−.020	.023	−.008	.013	−.037*	.017	−.007	.013
Job satisfaction	−.324*	.009	−.308***	.004	−.319***	.004	−.305***	.007	−.320***	.007	−.289***	.004
Self-efficacy	.054*	.009	.031*	.004	.030***	.004	.039***	.005	.043***	.006	.028***	.004
School level
Composition	.033**	.011	.061***	.007	.058***	.010	.053***	.007	.055***	.008	.063***	.009
Location	.094***	.025	.043***	.013	.056***	.018	.051***	.014	.057***	.016	.055***	.022
Type	.029	.034	.153***	.018	.146***	.025	.129***	.019	.114***	.021	.159***	.026
System level
Classroom SES
× Teacher appraisal (External)	.022	.100	.132*	.057	.107	.079	.090	.067	.086*	.066	.101*	.060
× Teacher appraisal (SMT)	.074	.100	.028	.059	.008	.083	.049	.069	.028	.069	−.030	.090
× HDI	.458*	.281	.226	.173	.574**	.228	.065	.193	.263	.162	.190	.179

*** = p < .000, ** = p < .01, * = p < .05.

Table C2. School consequences of appraisal.

	(1a)		(1b)		(2a)		(2b)		(3a)		(3b)
	Dismissal		No dismissal		Financial bonus		No financial bonus		Sanctions		No sanctions
	Coef.	SD	Coef.	SD	Coef.	SD	Coef.	SD	Coef.	SD	Coef.	SD
Teacher level
Novice	.213***	.019	.159***	.016	.150***	.019	.195**	.016	.164***	.031	.185***	.015
Gender	.016	.019	.02	.013	−.018	.018	.008	.013	−.046*	.028	−.008	.014
Job satisfaction	−.301***	.006	−.331***	.009	−.311***	.007	−.321***	.004	−.307***	.006	−.300***	.006
Self-efficacy	.036***	.006	.038***	.005	.039***	.006	.028***	.003	.037***	.005	.039***	.007
School level
Composition	.063***	.010	.056***	.010	.059***	.013	.058***	.008	.091***	.017	.049***	.009
Location	.097***	.019	.020	.015	.104**	.038	.073***	.015	.106**	.040	.055***	.016
Type	.082***	.035	.091***	.040	.068	.049	.171***	.022	.224***	.045	.049***	.009
System level
Classroom SES
× Teacher appraisal (External)	.089	.069	.134*	.085	.109	.145	.106	.084	.131	.126	.085	.059
× Teacher appraisal (SMT)	.062	.077	.003	.086	.026	.150	.030	.068	−.054	.142	.051	.059
× HDI	.107	.200	.427*	.252	.060	.336	.223	.224	.043	.287	.203	.117

*** = p < .000, ** = p < .01, * = p < .05.

Figure B.

Note. Figure represents the coefficients for the interaction term classroom SES * achievement-based teacher appraisal (a dummy) regressed on teacher turnover intent. Statistically significant estimates shown in blue. Argentina and Norway are excluded due to no within-system variation in certain cases. Figure B shows the estimates from the within-country two-level logistic models. In these models, control variables include teacher gender, experience, self-efficacy, job satisfaction, school location, type, and socioeconomic composition. Positive estimates indicate an anti-compensatory general moderating effect of achievement-based teacher appraisal on SES-based turnover. The estimates are depicted as odds ratios minus one, and therefore signify marginal increases or decreases in the odds of wishing to change schools in percentage points. This is a random-intercepts model only, as in the majority of education systems, the teacher mobility intentions-classroom SES slope did not vary as a function of accountability at the school level.