Teach For America Teachers' Careers: Whether, When ... - CiteSeerX

0 downloads 152 Views 2MB Size Report
Educational Statistics' School and Staffing Survey have found that a better .... Because this is a study of teachers' ca
Teach For America Teachers’ Careers: Whether, When, and Why They Leave Low-Income Schools and the Teaching Profession

Morgaen L. Donaldson [email protected] [email protected] Harvard Graduate School of Education Project on the Next Generation of Teachers

Paper prepared for the 2008 annual meeting of the American Educational Research Association. New York, New York.

INTRODUCTION1 Low-income students fare poorly in U.S. schools. They score low on state tests, graduate from high school at depressed rates and attend college in small proportions (Murnane, 2007; U.S. Census Bureau, 2006). As they mature, the costs of their underperformance are borne by these students and society. Without the competencies and qualifications to obtain well-paid work, low-income individuals occupy low-wage, insecure jobs (Levy & Murnane, 2004) or no jobs at all (Finn, 2006). The government spends millions of taxpayers’ dollars to remediate those undereducated students who do make it to public colleges (Atwell, Lavin, Domina, & Levey, 2006) and, at the other end of the spectrum, support those who enroll in welfare in disproportionate numbers (Blank, 2007). And, of course, the country fails to maximize the potential of an entire segment of its population. High-quality teachers can make a tremendous difference in the achievement of low-income students (McCaffery, Lockwood, Koretz & Hamilton, 1996). As a result, there is growing consensus that attracting and retaining effective teachers for these students is vitally important (Education Trust, 2004). However, schools serving lowincome populations struggle to get and keep good teachers (Guin, 2004; Ingersoll, 2001). Furthermore, we know little about why this is so and what might be done to stem the departure of promising teachers. In particular, are teachers who are assigned more complicated teaching loads more likely to leave low-income schools and the profession? In this paper, I report the results of a large-scale study using discrete time survival analysis to explore turnover in a sample of 2029 Teach For America (TFA) teachers

1

I would like to thank the American Association of University Women and the Spencer Foundation for providing funding that supported this research. I would also like to thank Susan Moore Johnson, Richard Murnane, and John Willett for thoughtful guidance and ample support. I thank John Papay for insightful feedback. Lastly, I am most grateful to the Teach For America teachers who completed my survey.

2

working in low-income schools. Using responses of TFA teachers to an on-line questionnaire, I constructed a retrospective, teacher-period data set that allowed me to examine whether and when TFA teachers were most at risk of leaving their initial placement schools, transferring, and resigning from teaching for the first time. Additionally, I asked whether teachers who taught more complex assignments—multiple grades at the elementary level; multiple subjects at the secondary level; or out of field courses at the secondary level—were more likely to turn over. I found that, in general, TFA teachers who taught multiple assignments were more likely to leave their schools, transfer, and resign from teaching altogether. Math and social studies teachers lacking a major in those subjects were more likely to resign from the profession, while science teachers without a science major were less likely to leave teaching. BACKGROUND AND CONTEXT OF THE RESEARCH Teacher retention, generally, and new teacher retention, in particular, have garnered much attention in recent years (Guarino, Santibañez, Daley, & Brewer, 2004; Johnson, Berg, & Donaldson, 2005; National Commission on Teaching and America's Future, 2003). Educational leaders, policymakers, and researchers are concerned about teachers’ attrition from the profession overall and their migration away from schools serving lower-income students and into schools with higher-income students. Numerous studies demonstrate that attrition of new teachers is high (Hanushek, Kain, & Rivkin, 2004; Ingersoll, 2001; Luekens, Lyter, Fox, & Chandler, 2004; Murnane, Singer, Willett, Kemple, & Olsen, 1991). By some estimates, approximately 40% of teachers leave the profession within 5 years of starting to teach (Ingersoll, 2002) and 50% leave within 6 years (Kirby, Berends, & Naftel, 1999).

3

Estimates of migration also reveal a troubling picture. In many districts enrolling large numbers of low-income students, new teachers leave their schools at high rates. Nationwide, 15.2% of teachers at high-poverty schools leave their schools annually, compared to 10.5% of their counterparts in low-poverty settings (Ingersoll, 2001). Recently, 50% of Philadelphia’s novices departed that district within 3 years (Neild, Useem, Travers, & Lesnick, 2003). When promising teachers leave their schools, students and schools may suffer. Departing teachers are likely to be replaced by novices, leaving classes taught by a stream of first-year teachers who are, on average, less effective than their more experienced counterparts (Murnane & Phillips, 1981; Rockoff, 2004). When teachers leave, schools also lose the benefits of professional development and other resources they have invested in departing teachers (NCTAF, 2003). Moreover, routinely high levels of teacher turnover impede a school’s efforts to coordinate curriculum and communicate information about students from one year to the next. Lastly, the financial costs of teacher turnover are high. For example, the Boston Public Schools spent an estimated $3.3 million to replace 194 first-, second-, and third- year teachers in 2004-5 (Birkeland & Curtis, 2006). Noting these costs, districts and states have sought to reduce new teacher attrition and migration by instituting induction programs and other supports for novices (Smith & Ingersoll, 2004). Thus, there is a great need to understand why turnover—due to both attrition and migration--occurs in low-income schools and what educational leaders might do to decrease it. Little of the previous research on new teacher turnover has focused on whether the job that new teachers are assigned to perform is related to their probability of

4

turnover. In particular, almost no research has examined whether new teachers’ teaching assignments are associated with their likelihood of leaving their school or the profession. How might assignment be related to new teacher retention? Research uniformly suggests that new teachers struggle to perform their jobs (Huberman, 1990; Lortie, 1975; Veenman, 1984). Teaching is complex (Rowan, 1990) and learning to teach is a difficult undertaking. Regardless of the length or intensity of their preparation, most new teachers learn to teach on the job (Lortie, 1975). As such, new teachers may achieve competence and confidence faster if assigned a less complex portfolio of classes--a single grade or subject that matches their college major. Conversely, a more complex assignment may make new teachers’ already steep learning curve even more precipitous. Theoretically, the complexity of a new teacher’s assignment influences her retention such that: A complex assignment => low efficacy=> voluntary turnover In this way, teachers with more complex assignments experience less success and are more likely to leave. However, all new teachers may struggle, and new teachers with complex assignments may not feel appreciably less successful than those with single- or in-field assignments. According to this reasoning, they may not be more likely to leave their school or resign from the profession than those with less complex assignments. To date, few researchers have investigated the relationship between a teacher’s assignment and her retention. The research that does exist suggests that teaching assignment is related to turnover. Nationwide, studies based on National Center for Educational Statistics’ School and Staffing Survey have found that a better assignment is the number one reason teachers transfer (Luekens et al., 2004). Recently, 40% of

5

transferring teachers cited the desire for a better teaching assignment as their main motivation for moving (Luekens et al., 2004). Similarly, a comprehensive analysis of retention in Washington state revealed that teaching assignment was the most frequent reason teachers said they stayed in their school (Elfers, Plecki, & Knapp, 2006). However, these studies do not specify what it was about teachers’ assignments that prompted them to transfer or remain. Some research examines differences in teachers’ assignments and retention. Studies have shown that out-of-field teachers, those assigned to teach subjects outside of their college major or minor, frequently suffer low morale and job commitment (Ingersoll, 1999) and often transfer into assignments that better match their major or minor (Patterson, Roehrig, & Luff, 2003). This line of research has also explored the ways in which a teacher’s assignment is related to her retention in the profession. A teacher who is assigned to a class she feels unqualified to teach is 1.9 times more likely to say that she will leave teaching (MetLife, 2006). Novices with multiple grade or subject assignments are more likely to leave teaching than those with single grade or subject assignments (Johnson & Birkeland, 2003). Moreover, high school teachers, especially newer teachers, who teach a higher proportion of classes outside their certification area are more likely to quit teaching than those with fewer classes mismatched to their training (Mont & Rees, 1996). Although there is reason to believe that teaching multiple grades or subjects would cause a new teacher to struggle and consider leaving her school or the profession, some schools build their academic program around such assignments. Coalition of Essential Schools secondary schools, for example, often combine math and science or

6

English and social studies to offer a more integrated and, it is argued, relevant curriculum. Similarly, some elementary schools combine grades to allow teachers to better meet students’ developmental needs, which sometimes do not correspond with graded curricula. The fact remains, however, that these combinations likely increase the difficulty of a new teacher’s already challenging task of learning to teach. Educators’ views of out-of-field assignments are more uniformly negative. In fact, the No Child Left Behind Act explicitly forbids teachers from teaching out of field. Despite this prohibition, recent research suggests that out-of-field teaching continues to occur, especially in low-income schools (Peske & Haycock, 2006). Why Teach For America? To test whether complexity of teaching assignment was related to new teacher retention, I surveyed three entire cohorts of Teach For America teachers. Founded in 1990, TFA selects and places high-achieving individuals in low-income urban and rural classrooms. These teachers receive only five weeks of teacher preparation and formally agree to teach for only two years. However, some TFA teachers stay in teaching for more than two years (Skinner, 2005). Since 1990, TFA and programs like it have proliferated; TFA currently attracts large numbers of applicants from the nation’s most selective colleges (Azimi, 2007). I chose to conduct this study on TFA teachers because they are an important subpopulation of new teachers today. These teachers possess the few specific characteristics—high test scores and diplomas from selective colleges—known to make a difference in student achievement (Ehrenberg & Brewer, 1994; 1995). Furthermore, they instruct some of the nation’s poorest students. Schools need good information about what

7

retains TFA teachers--and others like them--through their 2-year obligation and what might prolong their stay. Conversely, schools need to know what hastens the departure of these teachers from their initial placement school and the profession. This study is the first to examine the retention of TFA teachers longitudinally and on a national scale. RESEARCH QUESTIONS In this study, I examined whether the retention of new teachers was related to their assignment complexity. More generally, I investigated the conditional probability (“risk”) that a new TFA teacher would: a) leave their initial school, b) transfer, or c) resign from teaching altogether in every year following their entry into teaching, given that they had not experienced the event of interest up to that point. My specific research questions were: 1: Do elementary-school teachers with multiple-grade assignments have a higher probability of leaving their school, transferring, or resigning from teaching than those with single-grade assignments? 2: Do secondary-school teachers with multiple-subject assignments have a higher probability of leaving their school, transferring, or resigning from teaching than those with single-subject assignments? 3: Do secondary-school teachers with out-of-field assignments have a higher probability of leaving their school, transferring, or resigning from teaching than those with in-field assignments? RESEARCH DESIGN My research design improves on prior research in several ways. As others have noted (e.g., Ingersoll, 2002), many studies of teacher retention have investigated predictors that are easily measurable, but not often policy-amenable. Much of the research has examined, for example, individual-level demographic predictors such as gender or race (e.g., Kirby, Berends, & Naftel, 1999). Research on institutional-level, policy-amenable 8

predictors has tended to investigate district-level factors such as salary, district size, or district racial composition (e.g., Gritz & Theobald, 1996). Few quantitative studies have examined the relationship between classroom-level variables and teacher turnover despite the fact that qualitative research suggests that new teachers’ career decisions are heavily influenced by their day-to-day experiences in classrooms (Johnson, 1990; Johnson and PNGT, 2004; Lortie, 1975). Second, many quantitative studies of teacher turnover have drawn on administrative datasets that fail to distinguish between voluntary and involuntary turnover. While a high turnover rate at urban schools may indicate that many teachers are choosing to leave, it may also indicate that they are being laid off or forced to transfer in large numbers. The policy response to these two scenarios would be quite different, cautioning researchers against drawing conclusions about teachers’ work preferences based on analyses of administrative data sets. Third, much prior research on teachers’ careers has been methodologically limited. Some of the most influential studies have focused on only one or two years of data (see, e.g., Ingersoll, 2001; 2002). This research divides teachers into “leavers” from the profession, “movers” to a new school, and “stayers” who remain in their current school. The problem with this approach is that it designates as a “stayer” someone who remains in her school from year 1 to 2, regardless of whether she moved previously. If we are concerned about the retention of teachers in low-income schools over time, we need to look at retention longitudinally, beginning when a teacher enters the profession and, implicitly, a low-income school.

9

Even when it has tracked teachers’ careers over more than two years, prior research has suffered an additional methodological problem. Researchers have tended to ignore “censored” teachers who had not experienced the event of interest (for example, had not left the profession) when data were collected. Instead, they focused on those who had experienced the event (e.g., the leavers), asking “when” they left teaching. Alternatively, they discarded the “when” question and asked “whether” teachers left during the observed time period (Singer & Willett, 2003). My study addresses these limitations. First, I investigated the relationship between a new teacher’s assignment, a potentially important classroom-level variable, and her retention. Second, I focused exclusively on teachers’ voluntary career decisions. Teachers who reported that they transferred or resigned from teaching involuntarily— 13.2% of all transfers and 2.4% of all resignations from the profession in this sample-were not classified as having experienced the outcome in question. Third, I tracked teacher retention over 4-6 years, and defined as “stayers” only those who do not leave the low-income school where they were initially placed. Lastly, I used survival analysis, which allowed me to incorporate those who had not experienced the outcome in question and answer the “when” question by estimating the unbiased “risk” that teachers would experience that outcome, given that they had not experienced it up to that point. Sample and Data Collection This study is based on a census of all teachers in the 2000, 2001, and 2002 TFA cohorts. Because this is a study of teachers’ careers, I followed teachers over a period long enough to observe their movement out of teachingi. Many studies of teachers’ careers indicate that attrition declines substantially after year 5 (Kirby et al., 1999;

10

Murnane et al., 1991; Stinebrickner, 2001). However, because this is a retrospective study, I chose relatively recent cohorts to reduce recall errors (Taris, 2000) and take advantage of TFA’s more reliable contact information for more recent cohorts. Thus, I chose cohort of teachers who would have accumulated four, five, or six years of teaching experience if they had taught continuously. Most data for this study came from an on-line survey that I distributed between January and March, 2007, to the census of teachers. The survey collected information on teachers’ individual characteristics (e.g., subject matter preparation and assignment; demographic information) and, where relevant, the timing of their first departure from their school and the teaching profession. I have incorporated TFA placement records, which specify the districts in which individuals were placed. I created the survey instrument, drawing on the School and Staffing Survey, questionnaires designed by Kardos (2004) and Liu (2004), research on new teacher retention, and literature question design (Dillman, 1978; Fowler, 1998, 2002; Payne, 1951) with a specific focus on reducing recall error (Sudman & Bradburn, 1982). I piloted the entire survey with 30 TFA teachers from cohorts immediately prior to 2000 or after 2002 and tested specific questions and the online survey process with 812 teachers demographically similar to TFA. Starting with a population of 3283, systematic data collection led to a final sample of 2029 and a response rate of 62%. Comparison of the sample and census yielded few significant differences.

11

Measures From the data, I constructed a retrospective, teacher-period dataset to capture important elements of the TFA teachers’ careers. Because of the challenges of respondent recall in retrospective research, most measures I collected were time-invariant (Ingersoll, 1999; Kelly, 2004; Taris, 2000). Outcomes This study included three related outcomes: a) first voluntary exit from one’s initial placement school by transfer or resignation from teaching (VEXITSCHL); b) first voluntary transfer from one’s school (VTRANSFER); and c) first voluntary resignation from the teaching profession (VEXIT). Each outcome is conditional, meaning that it depends on the individual not having experienced that outcome before. See appendix 1a for definitions of all variables. First voluntary exit from one’s initial school allows me to isolate the “stayers.” In this analysis, “stayers” are only those who do NOT voluntarily exit from their initial school; they remain in these schools over time. First voluntary transfer allows me to isolate the “movers,” who are those who transfer in a given year, contingent on their not having transferred or resigned from the profession before. First voluntary resignation from the profession permits me to specify the “leavers,” who are those who resign from teaching in a given year, contingent on their not having left before. Thus, some leavers have previously transferred. It was important for me to define “resign from the profession” this way to enable this measure to capture both those respondents whose first action is to resign from the profession entirely and

12

those individuals who first transferred and then resigned from teaching. This allows me to see how many TFA individuals remain in teaching today. Question Predictors Because I am interested in the differences between the retention of teachers with single- and multiple-assignments at a particular grade level, the predictor of interest for research questions 1 and 2 is the two-way interaction between a time-varying dummy variable that captures multiple assignment in year j (MULTIGRADE or MULTISUB) and a time-varying dummy variable that indicates whether the respondent taught at the specified school level in that same year (ELEM or MIDHS). The cumulative effect of teaching a multiple assignment at a particular grade level is the interaction effect plus the main effect representing multiple assignment. Multiple assignments were relatively common across the sample. In each year, 16 to 20% of all elementary teachers were assigned to teach more than one grade. From approximately one-third (35%) to one-half (50%)of all secondary teachers were assigned to teach more than one subject in each year. For the out-of-field analysis, I was also interested in differences in retention at a particular grade level. Thus, for research question 3, the predictor of interest is a threeway interaction between a time-varying dummy variable that captures subject assignment in year j (e.g., math_yr), a time-invariant dummy variable indicating whether the respondent majored in the college major corresponding to the subject assignment (e.g. MATH_maj), and a time-varying dummy variable indicating whether the respondent taught at the secondary level in year j (e.g. MIDHS).

13

Out-of-field assignments were quite widespread. In each of the six years of data collection, anywhere from 57 to 77% of math teachers, 16 to 31% of social studies teachers, and 38 to 50% of science teachers lacked a major in the field they were teaching. Out-of-field assignments were most prevalent in the first one or two years of respondents’ careers. I was also interested in the effect of time on the outcomes. In initial analyses, I used the most general specification of time: 6 dummy variables representing the up to 6 years in which respondents could have taught. Subsequently, I used more parsimonious representations of time to facilitate the three-way interaction of time with the multiple assignment interaction and the four-way interaction of time with the out-of-field interaction. Control variables I controlled for variables known to make a difference in teacher turnover: cohort (C1-C3), gender (female), race (black, latino, asian), age one entered teaching (agestartC), college major (e.g., MATH_maj), whether or not one is related to at least one teacher (family), and urbanicity of teaching assignment (RURAL). All of these are timeinvariant measures. I also included regional fixed effects to test the robustness of my findings. Data Analysis 1: Do elementary-school teachers with multiple-grade assignments have a higher probability of leaving their school than those with single-grade assignments? To answer this question, I used discrete time survival analysis (Singer & Willett, 1993; Willett & Singer, 1991), to fit the following no-intercept model: 1

14

p(VEXITSCHL=1) = ---------------------------------------------------1+e –[(!1T1ij + !2T2 ij +… !8T6 ij)+ (ß1ELEMij+ ß1MULTIGRADEij +ß3MULTIGRADE_ELEMij + "1 Z i)] where i represents the individual respondent, the ! parameters represent the risk a teacher will leave teaching for the first time in each year, ß1 represents the effect of ELEM on teachers’ conditional probability of leaving in year j; ß2 represents the effect of MULTIGRADE on teachers’ conditional probability of leaving in year j; ß3 represents the effect of MULTIGRADE_ELEM, the interaction term, on teachers’ conditional probability of leaving in year j; and "1 is a vector of parameters that represent the impact of time-invariant controls (Zi). From estimates of the ! parameters in this model, I recovered the fitted hazard probabilities that represent a teacher’s risk of leaving her school for the first time in each year. I also tested models that included interactions between the time variables and substantive question predictors. I answered the subsequent research questions by following this approach. As noted above, research question 3 required a three-way interaction between subject assignment, college major, and secondary level. Model-building strategy I built models by first determining a representation of time that was more parsimonious than the most general specification with six dummy variables. I then added the control variables. I subsequently added the two main effects (e.g., school level and multiple assignment) and the interaction effect (e.g. school level X multiple assignment) of interest. I then interacted this interaction and its components with each time variable to determine whether the effect of a complex assignment varied over time. I added all three- or four-way interactions with time and their components to the model at once and then dropped those that were not significant.

15

In building models, I examined -2 log likelihood goodness-of-fit statistics and compared differences on a chi-square distribution with the appropriate degrees of freedom. In adding the two-, three-, and four-way interactions that answer my research questions, I also evaluated their impact based on general linear hypotheses tests that determined whether individuals with less complex assignments (i.e. a single grade or subject or an in-field assignment) had a significantly higher probability of experiencing the outcome than those with more complex assignments in the time period specified. FINDINGS Baseline results: The Overall Sample of TFA teachers Figure 1 (see Appendix 2) depicts the fitted hazard function for exiting from the school in each yearii, provided that an individual has not exited in previous years, and the survivor function derived from these hazard probabilities. The fitted hazard function (the left panel) demonstrates that these teachers’ estimated, conditional risk of leaving their school is relatively low in year 1 (approximately 10%) but then increases rapidly. Their greatest risk of exiting their initial placement school is in years 2 and 3, when, in each year, approximately 50% of teachers who are still in their placement school are predicted to leave. This is consistent with the TFA program structure, which requires a 2-year commitment to the placement school and teaching profession. The teachers’ predicted risk of exit then declines. The cumulative effect of these exits is illustrated in the fitted survivor function (the right panel) in figure 1. Survivor functions depict the probability of not experiencing the event of interest (i.e. “surviving”) at the end of each time period. When the survivor function declines sharply—as it does between years 1 and 2—there is an elevated risk in

16

the hazard function. Thus, after two years, an estimated 44% remain in their school. Conversely, an estimated 56% of TFA teachers left their initial school within their first two years of teaching. By the end of 6 years, when data collection ceased, fewer that 10% of the original group of teachers are predicted to have remained in their initial school. Again, this is consistent with TFA’s emphasis on having their corps members teach in low-income schools for a short time. Respondents left their initial placement school by transferring to a new school or by resigning from the profession altogether. Figure 2 presents the fitted hazard function for transferring, provided that a person has not previously transferred or left the profession. I do not present an accompanying fitted survivor function because a person’s conditional probability of transfer depends not only on not having transferred previously but also on not having resigned from teaching. This leads to a biased survivor function. Figure 2 demonstrates that the predicted probability of transfer, conditional on not having transferred or left the profession, is low in year 1 but rises to an estimated 16% in year 2 and a high in year 3, at approximately 19%. This means that an estimated 16% of eligible teachers transferred out of their initial school in year 2 and an estimated 19% moved in year 3. We are also interested in the percentage of TFA teachers who remain in the teaching profession. Figure 3 contains the fitted hazard and survival functions with resignation from the profession as the outcome. Neither function suffers from bias because they depict the probability of resignation from the profession, provided a person has not yet left. In other words, an individual can transfer and still remain at risk for leaving the profession.

17

The fitted hazard function (the left panel) indicates that the fitted conditional risk of leaving the profession is low in year 1, rises to a high in year 2, and then generally declines thereafter. Approximately 35% of teachers who remained in teaching at the beginning of year 2 were estimated to have left by the beginning of year 3. The fact that year 2 is the riskiest time period for this sample is consistent with the 2-year commitment required by TFA. The fitted survivor function exhibits a steep decline in the probability of remaining in the profession in the first few years, reflecting the high risk of exit in years 2 and 3. Nonetheless, a relatively high proportion—an estimated 61% of the sample-remained in the teaching profession more than 2 years. 24% of the teachers are predicted to have remained in teaching beyond 6 years, when data collection ceased. Do these patterns differ for individuals with different teaching assignments? I now turn to the case of multiple grade assignments at the elementary level, multiple subject assignments at the secondary level, and out-of-field assignments in the upper grades.

What is the Effect of Multiple Grade Assignment on Teacher Turnover? Teachers with multiple grade assignments were significantly more likely to leave their school, transfer, and leave the profession in year 1 and in select, subsequent years. Figure 4 depicts the fitted hazard and survivor functions based on a model that controls for gender, race, college major, the presence of a teacher in one’s family, age of entry to the teaching profession, and urbanicity of placement. All covariates are held at the mean. These hazard and survival plots and those that follow depict the hazard and survival

18

probabilities for two prototypical individuals: one who always had a multiple assignment and one who always had a single assignment. I chose these two profiles to highlight the differences in their estimated risk of exit or transfer in each year. The fitted hazard function (the left panel) indicates that among elementary school teachers teaching multiple grades, the estimated probability of leaving one’s school in year 1 is 19.1%. By contrast, only 6.7% of single-grade teachers are predicted to exit in this year. Thus, the estimated odds that a multiple-grade teacher would exit were 3.29 times the estimated odds that a single-grade teacher would leave her school in this year. This difference is sizable and significant. As suggested by figure 4, multiple-grade teachers’ estimated hazard of exit from their school were not statistically or practically different from single-grade teachers’ risk in year 2. This likely reflects TFA’s expectation that corps members will teach in their assigned low-income school for at least two years. In years 3, 5, and 6, multiple-grade teachers had a higher fitted probability of leaving their school than single-grade teachers, but these differences were not statistically significant. In year 4, multi-grade teachers were predicted to be significantly less likely to leave their initial school. Among individuals who were still in their original school, the predicted probability that multiplegrade teachers would leave was 19.2% and the predicted risk that single-grade teachers would exit was 40.5%. The survivor function on the right in figure 4 reflects these differences in hazard probability for an elementary teacher who is assigned multiple-grades in each year compared to a person who is assigned single grades in each year. Multiple-grade teachers’ probability of remaining in their initial placement schools is considerably lower

19

in the first three years. A predicted 80.8% of multiple-grade teachers remain in their school for longer than one year, in contrast to almost 94% of single-grade teachers. An estimated 38.8% of multiple-grade teachers, compared with 45.2% of single-grade teachers, remain in their school beyond 2 years. Over time, the predicted gap between these two groups of teachers continues to narrow, perhaps reflecting TFA’s short-term emphasis. What accounts for multiple-grade elementary teachers’ elevated risk of leaving their school in year 1? It appears that a heightened risk of transfer in this year contributes heavily to this outcome. Figure 5 depicts the fitted hazard function with voluntary transfer as the dependent variable, controlling for all else. Figure 5 exhibits a stark difference between multiple-grade and single-grade teachers’ estimated probability of transfer in year 1. 14.1% of multiple-grade teachers still in their initial school are predicted to transfer in year 1. By contrast, the predicted risk of transfer for single-grade teachers is only 3.8%. In other words, the estimated odds that a multiple-grade teacher would transfer in year 1 are 4.20 times the odds that a single-grade teacher would move. In years 2 and 3, multiple-grade teachers who remain in their initial school have lower predicted probabilities of transfer than single-grade teachers who are still in their original school. It could be that those individuals who teach multiple grades in years 2 or 3 request such an assignment and, with their wish fulfilled, are thus more likely to stay in their school. It is less likely that respondents influenced their year 1 assignment, which makes it a less biased predictor of subsequent career decisions.

20

Turning to the question of retention in the profession, are multiple-grade teachers more likely than single-grade teachers to resign from teaching altogether? Figure 6 contains the fitted hazard and survivor functions with resignation from the profession as the outcome, controlling for all else. As suggested by the hazard function in Figure 6, multiple-grade teachers are predicted to be more likely than single-grade teachers to resign from the profession in year 1. In this year, 6% of multiple-grade teachers but just 2.8% of single-grade teachers are predicted to resign from teaching. Put another way, the estimated odds that a multiple-grade teacher would exit teaching in year 1 were 2.22 times the estimated odds that a single-grade teacher would leave. In all other years, however, multiple- and single-grade teachers’ predicted risk of exit were virtually the same. This is reflected in the fact that the lines representing each group in the hazard function nearly overlap in years 2-6. The same is true for the fitted survivor function; the fact that teachers with multiple-grade assignments in each year are predicted to remain in teaching at slightly lower rates than their single-grade counterparts is due almost entirely to the year 1 differences in exit probability. Among multiple-grade teachers, 62.3% are predicted to remain in teaching beyond year 2. 64.8% of singlegrade teachers are estimated to continue past this point in time. In summary, multiple-grade teachers are estimated to be more likely than singlegrade teachers to leave their school, transfer, and resign from the profession in year 1 and in particular other years. We now turn to the question of assignments at the secondary level. Are Teachers with Multiple Subject Assignments More Likely to Turn Over?

21

As with multiple-grade assignments, teachers assigned to teach multiple subjects at the secondary level in year 1 also were more likely to leave their schools and to leave the profession in that year. However, they differed from the elementary teachers in that they were not more likely to transfer. Multiple-subject secondary teachers’ higher estimated likelihood of exiting their school is captured in Figure 7 below. As shown in the fitted hazard function (the left panel), multiple-subject teachers’ predicted probability of exit from their school in year 1 is higher than that of single-subject teachers, controlling for all else. In this year, 15.0% of multiple-subject teachers are predicted to exit, in contrast to 9.8% of single-subject teachers. In other words, the estimated odds that multiple-subject teachers would leave their school in year 1 were 1.62 times the estimated odds that single-subject teachers would do so. Beyond year 1, however, there was no significant difference in the estimated, conditional probability that these two groups of teachers would exit their school. This is reflected in the nearly overlapping lines in the hazard function from years 2 through 6. The moderate difference in predicted risk in year 1 and the small differences thereafter are reflected in the fitted survivor function. 39.6% of single-subject teachers are predicted to remain in their school beyond year 2, compared to 36.4% of multiplesubject teachers. Overall, then, multiple- and single-subject teachers are estimated to leave their initial schools at comparable rates. Analysis of the effect of assignment at the secondary level on probability of transfer revealed no significant differences between teachers with multiple- and single-assignments. However, teachers with multiple assignments were

22

significantly more likely than their counterparts with single assignments to resign from the profession, as shown in Figure 8 below. In all time periods, teachers with multiple assignments are predicted to have a higher conditional probability of exit from the profession than their counterparts with single assignments, controlling for all else. For example, in year 1, 7.6% of multiplesubject teachers are estimated to leave the profession. 4.6% of single-subject teachers, by contrast, are predicted to resign. In this year, the estimated odds that multiple-subject teachers would resign from the profession are 1.72 times the estimated odds that singlesubject teachers would leave. Similarly, in years 3-6, teachers with multiple-subject assignments have higher predicted probabilities of leaving the profession. In year 3, the conditional, estimated probability that teachers with multiple assignments will resign is 29.3%; for teachers with single assignments, it is 24.4%. Multiple-subject teachers’ estimated odds of resigning are 1.28 times single-subject teachers’ odds for year 3. The estimated odds ratio for years 4, 5, and 6 are even larger: 1.44, 1.54, and 1.69, respectively. Multiple-subject teachers’ elevated estimated risk for resigning from teaching in each year leads to a sizable difference in the proportion of teachers in each group who remain in the profession, as shown in the survivor function. Although the difference between the estimated proportion of each group that remains in the profession beyond year 2 is relatively small (58.3% of multiple-subject teachers vs. 62.1% of single-subject teachers), by the end of data collection this gap has widened. 17.5% of teachers who have had a multi assignment in each year are predicted to remain in teaching beyond year

23

6; 26.1% of teachers who have had a single-subject assignment are estimated to continue to teach beyond this point, by contrast. In summary, respondents teaching multiple subjects were predicted to be more likely to resign from the teaching profession for the first time than those teaching single subjects. Moreover, they were more likely to leave their schools in year 1 than their single-subject counterparts. Are Individuals Who Teach Out-of-field More Likely to Leave or Transfer? Multiple-subject assignments at the secondary level are associated with an increased probability of leaving one’s school and resigning from the teaching profession. Individuals who teach multiple subjects are likely to teach out of field, which, following the National Center for Educational Statistics’ lead (Seastrom, Gruber, Henke, McGrath & Cohen, 2004)iii, I define as lacking a college major in the subject(s) they instruct. Are out-of-field assignments associated with increased risk for exit from one’s school, transfer, and resignation from the profession? The answer to this question is mixed. Math teachers without a major in mathematics, computer science, or engineering (Seastrom et al., 2004) were more likely to leave their school and resign from the profession than those with a major in these disciplines. Social studies without a corresponding major in history or social science (Seastrom et al., 2004) were more likely to resign from teaching than those with a related background. But science teachers without a science major were less likely to leave the profession than those with a major in physics, chemistry, natural, or life sciences (Seastrom et al., 2004). English teachers’ probability of exit, transfer, or resignation did not vary as a function of their major.

24

Math teachers fell most in line with my hypothesis that an out-of-field assignment would be associated with an increased risk of exit from one’s school and resignation from the profession. Math teachers lacking a math major were indeed more likely to leave their school for the first time, as shown in Figure 9 below. As shown in the left panel, in every year out-of-field math teachers were estimated to be more likely to leave their school than math teachers with a math major, controlling for all else. For example, in year 1, 14.2% of out-of-field math teachers were predicted to leave, as compared to 6.9% of in-field math teachers. Similarly, in year 2, 51.5% of out-of-field teachers were predicted to leave, as compared to 45.7% of their infield counterparts. Put in terms of odds ratios, the estimated odds that out-of-field math teachers would exit were 2.21 times those of in-field teachers in year 1, and 1.26 times those of their counterparts in year 2 and thereafter. Out-of-field math teachers’ elevated risk of exit, compared to that of in-field teachers, accumulates such that a substantially lower proportion of this group remains in their initial school, controlling for all covariates. In the right panel of Figure 10, we see that an estimated 41.6% of math teachers who have an out-of-field assignment in every year remain in their initial school beyond year 2; by comparison, an estimated 50.5% of in-field math teachers stay in their original school. This gap widens such that 4.9% of out-of-field teachers are predicted to remain beyond year 6, compared to 8.5% of in-field math teachers. Turning to voluntary transfer, there is a similar pattern in year 1 (see Figure 10). In year 1, out-of-field math teachers (defined here as those with a humanities, science, or foreign language major) are substantially more likely to transfer than their in-field

25

counterparts. Their estimated risk of transfer is 9.5% compared to only 3.0% for in-field math teachers. This translates into estimated odds that are 3.39 times those of in-field teachers’ estimated odds of transfer, controlling for all else. However, in years 2 through 6, out-of-field teachers who remain in their initial placement school are less likely to transfer than their in-field counterparts. In year 2, for example, 16.2% of out-of-field teachers who remain in their initial school are predicted to transfer, compared to 19.5% of in-field teachers. Turning to an analysis with resignation from the profession as the outcome (see Figure 11), we see that in every year out-of-field math teachers are more likely to leave teaching, given that they have not previously resigned from the profession. The left panel depicts the conditional probabilities that in-field and out-of-field math teachers will resign from the profession in each year, given that they have not previously done so and controlling for all else. In year 2, the estimated hazard of exit for in-field teachers is 30.6% while that for out-of-field teachers, defined here as those with a humanities, science, or foreign language major, is 39.0%. In each year, the estimated odds that an out-of-field math teacher will resign from the profession are 1.45 times those that an in-field teacher will leave. The right panel shows the cumulative effect of these differences. An estimated 66.8% of math teachers who are in-field in each year remain in the profession beyond year 2, as compared to an estimated 57.8% of the out-of-field math teachers, controlling for all else. This gap persists: 33.4% of individuals who teach math in each year and have a math major are estimated to stay in the profession beyond year 6, compared to only 22.1% of those who teach math in each year but do not possess a math major.

26

Social studies teachers followed a similar pattern, but only when resignation from the profession was the outcome. As shown in Figure 12, in each year, social studies teachers with a social studies major were significantly less likely to leave than those without a major, given that they had not left before and controlling for all covariates. In each year, the estimated odds that a social studies teacher lacking a corresponding major would leave the profession were 1.89 times the estimated odds that her counterpart with the appropriate major would resign, controlling for cohort, gender, race, age of entry, the presence of teachers in one’s family, and the urbanicity of teaching placement. For instance, in year 2, 47.4% of out-of-field social studies teachers (defined here as those with a science, foreign language, English, or arts major) were predicted to leave, as compared to 32.3% of in-field social studies teachers. Over time, these differences compounded, as shown in the right panel of Figure 12. Controlling for all else, 48.7% of individuals who taught social studies out of field each year were predicted to remain in the profession beyond year 2. In contrast, 64.9% of individuals who taught in field in each year were predicted to continue to teach beyond this point. Thus, for both math and social studies teachers, an in-field assignment in each year is associated with a higher likelihood of remaining in the profession in that year. Is this due to a more manageable teaching load, as my hypothesis would suggest? We now turn to the case of science teachers. As shown in figure 13, science teachers with a major in science are more likely than those without a major to resign from teaching in each year, provided they have not yet left and controlling for all else. In year 1, out-of-field (defined here as humanities,

27

math, and foreign language majors) science teachers’ estimated odds of resigning from the profession were .59 times the estimated odds that in-field science teachers would exit. Put another way, 10.2% of science teachers with a science major are predicted to leave in year 1, compared to 6.3% of science teachers with a humanities, math, or foreign language major. In each year thereafter, the differences in predicted, conditional exit probability were much smaller and not statistically significant. Thus, out-of-field math and social studies teachers are more likely to resign from the profession, but out-of-field science teachers are less likely to do so. DISCUSSION The findings of this study in general confirm my initial hypothesis that more complex assignments are associated with increased turnover among new teachers. In support of my hypothesis, I found that multiple-grade teachers were more likely than single-grade teachers to leave their school in most years and more likely to resign from the profession in year 1. I found that multiple-subject secondary teachers were more likely to leave their school in year 1 and resign from the profession in all years. Lastly, I found that out-offield math and social studies teachers were more likely to resign from the profession than in-field teachers of these subjects. However, contrary to my hypothesis, science teachers without a science major were less likely to leave than science teachers with a major. Across the sub-analyses, individuals’ year 1 assignment was most consistently related to their turnover. This connection between year 1 assignment and turnover lends credence to my hypothesis that assignment affects new teacher attrition and migration. Respondents who leave in year 1 fail to fulfill their obligation to TFA to remain in a lowincome school for two full years. People who knowingly break a commitment to which

28

they have agreed may be especially challenged by the circumstances in which they teach. Moreover, these teachers have the least influence over their assignment in year 1; thus, this is the year in which estimates of the relationship between assignment and turnover are least biased by individuals’ choice of which classes they teach. In years 2-6, multiple assignment was related to turnover in some cases and not in others. These mixed results are not surprising. After year 1, a teacher likely has more control over the classes assigned to her. If we observe that respondents teaching multiple grades or subjects have a higher probability of turnover in these years, it could be that they agree to teach a multiple assignment in a particular year, knowing that they will leave teaching at the end of that year. If we find that multiple-assignment teachers are more likely to stay, it could be that these teachers have achieved some degree of competence in the classroom and seek the new challenge posed by teaching a second grade or subject. In short, we would expect more mixed results as new teachers become experienced and gain more say in their teaching assignments. The one result that explicitly runs counter to my hypothesis is the finding that science teachers without a science major are less likely to leave than those with a science major. This finding reminds the reader that teachers may be both pushed out of the profession by a taxing teaching assignment (and other school- and job-related factors) and pulled out of teaching by alternative ways of earning money or spending their time. Many studies have found that science majors are more likely to resign from teaching than humanities majors (see, e.g., Kirby, Berends, & Naftel, 1999; Murnane et al., 1991) in part due to the higher salaries they could command in the science, technology, and engineering industries. Thus, the fact that in-field science teachers are more likely to

29

resign may simply reflect their access to a wider range of attractive, alternative occupations than those available to humanities or foreign language majors. The fact that science majors teaching science are more likely to resign than nonscience majors but math majors teaching math are less likely to do so than non-math majors is a bit puzzling because math majors arguably have the same range of occupational alternatives as science majors. Nonetheless, this is consistent with prior research that found science majors more likely to leave than math majors (Murnane et al., 1991). This finding, then, serves as an apt reminder of the observational nature of this study’s design. When we see that out-of-field teachers or those with multiple assignments are more likely to leave, it may not be due to their more complex teaching assignment, but other unmeasured factors. Threats to Validity Heeding this caution, it is important to consider threats to the internal and external validity of these findings. The main threat to internal validity is that something other than assignment is driving the elevated risk of turnover. My research design does not permit causal conclusions, as discussed, and several other factors may explain the relationship I observed. My design allowed me to rule out two threats to internal validity. One threat suffered by many studies of teacher turnover is the potential that the observed turnover is involuntary. In the case of my study, teachers with multiple-assignments may be more likely to be laid off or forced to transfer. I have been able to eliminate this threat by focusing only on voluntary exit and transfer. A second threat is that pre-entry differences

30

in teachers, rather than assignment or other workplace variables, cause different turnover outcomes. I have been able to minimize, though not eliminate, this threat by studying teachers in one program who have met the same selection criteria and completed the same preparation program. Beyond these threats, several remain. First, those people who assign TFA teachers to classes may identify certain teachers who are not likely to stay and place those teachers in multiple or out-of-field assignments. In this way, the probability of exiting or transferring drives the complex assignment, not the reverse. I intend to interview several individuals who make these decisions to better understand the factors they consider when assigning classes to TFA teachers. Second, the presence of a complex assignment may not, in and of itself, affect turnover. Instead, a complex assignment may be a proxy for poor working conditions. Disorganized, unruly, and under-resourced schools may be more likely to ask teachers to teach multiple assignments. The chaos in the hallways and absence of resources may provoke teacher turnover more than a complex assignment. Although most schools where TFA places teachers feature poor working conditions by definition, I have attempted to test this theory by fitting additional models with regional fixed effects. In essence, these models compare teachers with more and less complex assignments within regions, thus controlling for workplace conditions and other unmeasured characteristics that vary across region. This does not address the possibility that conditions vary by schools within a region. I have no mechanism by which to control for school-based differences within regions.

31

The main threat to external validity is the fact that TFA teachers differ markedly from generic new teachers. TFA teachers are graduates of selective colleges, sign up for an explicit two-year, cohort-based program, and receive very little preparation. In many ways, TFA teachers differ from “typical” new teachers today. However, these results are in line with earlier research on new teachers’ assignments (see, e.g. Mont & Rees, 1996).

IMPLICATIONS Although these results should be treated with caution, administrators should keep in mind that assigning a first-year teacher a more complex assignment may put her at risk of leaving the school. Although more research must be conducted on this topic, administrators who want to retain new teachers should probably make an effort to assign new teachers a single grade at the elementary level or a single subject well matched to the teacher’s college major at the secondary level. If program requirements or hiring timelines prevent this, administrators should observe new teachers with complex assignments carefully and offer support early if these teachers struggle. Overall, this study highlights an important area for further research. Teachers’ assignments have not been studied in any depth, yet they appear to be related to teachers’ career decisions. Qualitative investigations of how administrators assign teachers are needed. Such studies should examine assignment practices in large and small districts, in urban, suburban, and rural locales, and in bargaining and non-bargaining states. This would allow researchers to better understand whether and, if so, how practices differ by district size, urbanicity, and collective bargaining context. Such studies should track new

32

teachers with more- and less-complex assignments to understand how novices perceive such assignments and whether such assignments factor into their career decisions. Quantitative studies should examine in greater detail the prevalence, incidence, and impact of multiple and out-of-field assignments. How common are these? Who is more likely to teach a multiple assignment? What are the consequences of multiple assignments for student achievement? Has No Child Left Behind reduced out-of-field placements? When out-of-field teachers leave the profession, is it due more to a taxing assignment or an enticing opportunity outside education? Attracting and retaining good teachers for low-income children must be part of any effort to fracture the enduring link between poverty and low academic performance in the United States. This study aimed to aid these efforts by investigating whether teaching assignment is related to the retention of new Teach For America teachers. By teaching a single grade or subject for which they are well prepared, these teachers may meet and even exceed their two-year obligation to teach in low-income schools. In this way, they each contribute to the effort to provide low-income students a better education.

33

Table 1a: Parameter estimates, standard errors, and goodness-of-fit statistics from a series of hazard models in which assignment at the elementary level predicts the conditional probability that a teacher will leave her school, controlling for cohort, gender, race, age of entry to the profession, the presence of a teacher in one’s family, college major, and urbanicity of teaching placement (n=2029) Predictor C2 C3 T1 postT1 T4 T2to6 female black latino asian family agestartC SCITECHmaj HUMANmaj MATHmaj FLANGmaj RURAL elem_yr multi_grade multi_gradeelem T1_elem T4_elem T4_multi

Parameter estimate 1 2 Model Model 0.130 0.132 (0.091) (0.091) 0.156 0.156 (0.083)~ (0.083)~ -2.227 -2.178 (0.161)** (0.170)** 0.312 0.384 (0.192) (0.200)~ -0.454 -0.454 (0.133)** (0.133)** -0.097 -0.104 (0.046)* (0.046)* -0.150 -0.142 (0.076)* (0.077)~ -0.439 -0.427 (0.105)** (0.106)** -0.234 -0.229 (0.130)~ (0.130)~ -0.085 -0.074 (0.126) (0.126) -0.006 -0.005 (0.068) (0.068) -0.056 -0.058 (0.016)** (0.016)** 0.385 0.367 (0.127)** (0.129)** -0.024 -0.034 (0.107) (0.107) -0.174 -0.175 (0.121) (0.122) -0.246 -0.237 (0.143)~ (0.143)~ 0.074 0.076 (0.084) (0.085) -0.115 (0.084) -0.080 (0.093) 0.373 (0.162)*

3 Model1 0.128 (0.091) 0.158 (0.083)~ -2.247 (0.207)** 0.218 (0.266) -0.612 (0.249)* -0.027 (0.082) -0.151 (0.077)~ -0.446 (0.106)** -0.233 (0.130)~ -0.072 (0.127) -0.003 (0.068) -0.057 (0.016)** 0.368 (0.130)** -0.034 (0.108) -0.191 (0.122) -0.238 (0.143)~ 0.081 (0.085) 0.229 (0.296) 0.178 (0.330) -0.607 (0.590) -0.559 (0.362) 0.491 (0.329) 0.245

4 Model 0.188 (0.092)* 0.200 (0.085)* -1.801 (0.251)** 0.641 (0.301)* -0.602 (0.251)* -0.002 (0.083) -0.158 (0.078)* -0.398 (0.111)** -0.181 (0.133) -0.039 (0.129) -0.003 (0.069) -0.057 (0.016)** 0.420 (0.131)** -0.006 (0.109) -0.181 (0.124) -0.217 (0.145) 0.283 (0.299) 0.162 (0.333) -0.583 (0.595) -0.586 (0.364) 0.491 (0.331) 0.264

34

T4_multielem T2to6_elem T2to6_multi T2to6_multielem T1_multi T1_multielem

(0.362) -1.753 (0.647)** -0.136 (0.112) -0.127 (0.123) 0.349 (0.212)~ -0.035 (0.390) 1.655 (0.672)*

(0.364) -1.852 (0.651)** -0.143 (0.113) -0.128 (0.123) 0.340 (0.214) -0.032 (0.392) 1.621 (0.677)* YES

5360.507

5311.107

Region Fixed Effects -2LL

5394.379

5388.610

Delta -2LL 0.7774(1) 5.7696(3) 28.103(9) 29.237(9) Comparison model 1 2 22 p value .3779 .1234 .0009 .0006 Observations 5044 5044 5044 5039 Standard errors in parentheses; ~ significant at 10%; * significant at 5%; ** significant at 1% 1

A model included postT1 interaction effects was fit, but these interactions were dropped because of collinearity with the T1 interaction effects. 2 model 2 fitted with regional fixed effects instead of the RURAL dummy (not shown)

Table 1b: General linear hypothesis tests for significant differences between single- and multi-assigned teachers in each year (based on Model 3) Year

Test

1

Ho: ßmulti_grade + ßmulti_gradeelem + ßT1_multi + ßT1_multielem =0 Ho: ßmulti_grade + ßmulti_gradeelem + 2* ßT2to6_multi + 2*ßT2to6_multielem = 0 Ho: ßmulti_grade + ßmulti_gradeelem + 3*ßT2to6_multi + 3*ßT2to6_multielem =0 Ho: ßmulti_grade + ßmulti_gradeelem + 4*ßT2to6_multi + 4*ßT2to6_multielem + ßT4_multi +ßT4_multielem =0 Ho: ßmulti_grade + ßmulti_gradeelem + 5*ßT2to6_multi + 5*ßT2to6_multielem =0 Ho: ßmulti_grade + ßmulti_gradeelem + 6*ßT2to6_multi + 6*ßT2to6_multielem =0

2 3 4 5 6

Chi square Results 23.430

p value

0.01

0.9347

1.84

0.1744

5.30

0.0213

2.43

0.1191

2.26

0.1326

.0000

35

Table 2a: Parameter estimates, standard errors, and goodness-of-fit statistics from a series of hazard models in which assignment at the elementary level predicts the conditional probability that a teacher will voluntarily transfer, controlling for cohort, gender, race, age of entry to the profession, the presence of a teacher in one’s family, college major, and urbanicity of teaching placement (n=2029)

Predictor C2 C3 TIMEC TIMEC2 TIMEC3 female black latino asian agestartC family SCITECHmaj HUMANmaj FLANGmaj MATHmaj RURAL elem_yr multi_grade multi_gradeelem TIMEC_elem TIMEC_multi TIMEC_multielem TIMEC2_elem TIMEC3_elem

Parameter Estimates (1) (2) Model Model 0.049 0.044 (0.122) (0.122) -0.004 -0.008 (0.111) (0.111) 1.905 2.012 (0.207)** (0.407)** -0.753 -0.756 (0.124)** (0.238)** 0.079 0.076 (0.020)** (0.037)* 0.263 0.256 (0.108)* (0.108)* -0.151 -0.182 (0.139) (0.140) -0.070 -0.073 (0.173) (0.173) -0.298 -0.306 (0.186) (0.186) 0.017 0.018 (0.018) (0.018) 0.032 0.038 (0.091) (0.092) 0.085 0.101 (0.167) (0.168) -0.261 -0.252 (0.137)~ (0.137)~ -0.007 -0.018 (0.183) (0.184) -0.127 -0.137 (0.168) (0.168) -0.179 -0.176 (0.117) (0.118) -0.081 -0.251 (0.114) (0.276) -0.028 0.022 (0.125) (0.288) 0.220 1.413 (0.214) (0.407)** 0.333 (0.545) 0.158 (0.578) -3.129 (1.012)** -0.118 (0.323) 0.007 (0.052)

(3) Model 0.062 (0.124) -0.012 (0.113) 2.034 (0.408)** -0.761 (0.239)** 0.077 (0.038)* 0.253 (0.109)* -0.185 (0.146) -0.108 (0.176) -0.320 (0.188)~ 0.019 (0.018) 0.035 (0.092) 0.128 (0.169) -0.210 (0.139) 0.057 (0.186) -0.131 (0.170) -0.222 (0.277) 0.060 (0.289) 1.436 (0.408)** 0.336 (0.547) 0.129 (0.580) -3.146 (1.011)** -0.125 (0.324) 0.008 (0.053)

36

TIMEC2_multi

-2.819 (0.224)**

-0.210 (0.336) 1.483 (0.648)* 0.044 (0.052) -0.194 (0.107)~ -2.930 (0.277)**

-0.181 (0.337) 1.474 (0.644)* 0.039 (0.052) -0.192 (0.106)~ -2.939 (0.335)**

3475.3905

3445.8701

3418.7326

TIMEC2_multielem TIMEC3_multi TIMEC3_multielem Constant Deviance-based hypothesis tests -2LL(df)

Delta -2LL 6.0938(3) 29.5204(9) 30.2616(9) Comparison model 1 11 P value .1071 .0005 .0004 Observations 5042 5042 5035 Standard errors in parentheses; ~ significant at 10%; * significant at 5%; ** significant at 1% 1 model 1 fitted with regional fixed effects instead of the RURAL dummy (not shown) Table 2b: GLH Tests for Significant Differences between Single- and Multi-Assigned Teachers in Each Year (based on Model 2) Year Test Chi p square value Results 1 Ho: ßmulti_grade + ßmulti_gradeelem = 0 25.08 .0000 2 Ho: ßmulti_grade + ßmulti_gradeelem + ßTIMEC_multi + 1.79 .1814 ßTIMEC_multielem + ßTIMEC2_multi + ßTIMEC2_multielem + ßTIMEC3_multi + ßTIMEC3_multielem=0 3

Ho: ßmulti_grade + ßmulti_gradeelem + 2*ßTIMEC_multi + 2* ßTIMEC_multielem + 4*ßTIMEC2_multi + 4*ßTIMEC2_multielem + 8*ßTIMEC3_multi + 8*ßTIMEC3_multielem=0

4.64

0.0312

4

Ho: ßmulti_grade + ßmulti_gradeelem + 3*ßTIMEC_multi + 3* ßTIMEC_multielem + 9*ßTIMEC2_multi + 9*ßTIMEC2_multielem + 27*ßTIMEC3_multi + 27*ßTIMEC3_multielem=0

.03

0.8639

5

Ho: ßmulti_grade + ßmulti_gradeelem + 4*ßTIMEC_multi + 4* ßTIMEC_multielem + 16*ßTIMEC2_multi + 16*ßTIMEC2_multielem + 64*ßTIMEC3_multi + 64*ßTIMEC3_multielem=0

.17

0.6798

6

Ho: ßmulti_grade + ßmulti_gradeelem + 5*ßTIMEC_multi + 5* ßTIMEC_multielem + 25*ßTIMEC2_multi + 25*ßTIMEC2_multielem + 125*ßTIMEC3_multi +125*ßTIMEC3_multielem=0

.03

0.8739

37

Table 3a: Parameter estimates, standard errors, and goodness-of-fit statistics from a series of hazard models in which assignment at the elementary level predicts the conditional probability that a teacher will voluntarily resign from teaching, controlling for cohort, gender, race, age of entry to the profession, the presence of a teacher in one’s family, college major, and urbanicity of teaching placement (n=2029)

Predictor C2 C3 T1 T5 lnTIME female black latino asian agestartC family SCITECHmaj HUMANmaj FLANGmaj MATHmaj RURAL elem_yr multi_grade multi_gradeelem T1_elem T1_multi T1_multielem T5_elem

Parameter 1 Model 0.108 (0.089) 0.168 (0.082)* -3.165 (0.158)** 0.303 (0.160)~ -1.064 (0.120)** -0.327 (0.074)** -0.247 (0.107)* -0.208 (0.133) 0.265 (0.122)* -0.095 (0.018)** -0.031 (0.067) 0.251 (0.123)* 0.248 (0.107)* -0.112 (0.142) -0.058 (0.118) 0.229 (0.081)**

Estimates 2 Model 0.106 (0.089) 0.167 (0.082)* -3.174 (0.159)** 0.305 (0.160)~ -1.074 (0.121)** -0.316 (0.074)** -0.238 (0.108)* -0.204 (0.133) 0.270 (0.123)* -0.095 (0.018)** -0.029 (0.067) 0.227 (0.125)~ 0.249 (0.107)* -0.111 (0.142) -0.065 (0.119) 0.221 (0.082)** -0.092 (0.083) -0.009 (0.091) 0.097 (0.154)

3 Model 0.107 (0.089) 0.169 (0.082)* -3.183 (0.299)** 0.407 (0.293) -1.113 (0.226)** -0.320 (0.074)** -0.244 (0.108)* -0.202 (0.133) 0.270 (0.123)* -0.096 (0.018)** -0.029 (0.067) 0.228 (0.125)~ 0.248 (0.107)* -0.113 (0.143) -0.068 (0.119) 0.222 (0.082)** -0.272 (0.292) 0.056 (0.335) 0.473 (0.584) -0.230 (0.406) 0.194 (0.428) 0.074 (0.742) -0.484 (0.403)

4 Model 0.105 (0.089) 0.168 (0.082)* -3.143 (0.238)** 0.304 (0.160)~ -1.066 (0.121)** -0.318 (0.074)** -0.244 (0.108)* -0.206 (0.133) 0.269 (0.123)* -0.095 (0.018)** -0.029 (0.067) 0.227 (0.125)~ 0.249 (0.107)* -0.109 (0.142) -0.066 (0.119) 0.221 (0.082)** -0.056 (0.087) -0.045 (0.097) 0.059 (0.162) -0.447 (0.297) 0.296 (0.283) 0.489 (0.486)

5 Model 0.148 (0.090)~ 0.209 (0.083)* -3.132 (0.239)** 0.301 (0.160)~ -1.021 (0.122)** -0.334 (0.075)** -0.178 (0.112) -0.143 (0.136) 0.316 (0.125)* -0.095 (0.018)** -0.020 (0.068) 0.255 (0.126)* 0.248 (0.108)* -0.133 (0.144) -0.078 (0.120) -0.028 (0.089) -0.090 (0.098) 0.037 (0.164) -0.457 (0.298) 0.295 (0.284) 0.475 (0.488)

38

T5_multi

0.059 (0.187)

0.103 (0.194)

0.336 (0.422) -0.050 (0.709) 0.252 (0.295) -0.128 (0.338) -0.382 (0.566) 0.135 (0.265)

5800.3248

5798.7258

5785.2339

5790.5149

5730.008

Comparison model Delta -2LL(df)

2 13.4919(9)

2

7.8806(1)

1 1.599(3)

21 8.2292 (3)

p value Observations

.005 6341

.660 6341

.142 6341

T5_multielem lnTIME_elem lnTIME_multi lnTIME_multielem Constant

0.095 (0.195)

0.518 (0.234)* YES

Regional fixed effects Deviance-based hypothesis tests -2LL(df)

8.2109(3) .042 6341

.042 6336

Standard errors in parentheses~ significant at 10%; * significant at 5%; ** significant at 1% 1 model 2 fitted with regional fixed effects instead of the RURAL dummy (not shown) Table 3b: GLH Tests for Significant Differences between Single- and Multi-Assigned Teachers in Each Year (based on Model 4) Year Test Chi square p value Results 1 Ho: ßmulti_grade + ßmulti_gradeelem + ßT1_multi + ßT1_multielem 4.57 .0325 =0 2-6 Ho: ßmulti_grade + ßmulti_gradeelem = 0 0.01 .9197 Results of equivalent GLH tests based on Model 5, with regional fixed effects, were the same.

39

Table 4a: Parameter estimates, standard errors, and goodness-of-fit statistics from a series of hazard models in which assignment at the secondary level predicts the conditional probability that a teacher will voluntarily leave the school, controlling for cohort, gender, race, age of entry to the profession, the presence of a teacher in one’s family, college major, and urbanicity of teaching placement (n=2029) Parameter Estimates Predictor (1) (2) (3) (4) (5) (6) C2 0.130 0.136 0.126 0.126 0.132 0.194 (0.091) (0.091) (0.091) (0.091) (0.091) (0.092)* C3 0.156 0.166 0.159 0.159 0.161 0.204 (0.083)~ (0.083)* (0.083)~ (0.083)~ (0.083)~ (0.085)* T1 -2.227 -2.525 -2.833 -2.833 -2.837 -2.638 (0.161)** (0.273)** (1.040)** (1.040)** (1.040)** (1.105)* postT1 0.312 0.008 -1.021 -1.021 0.023 0.392 (0.192) (0.297) (0.751) (0.751) (0.302) (0.335) T4 -0.454 -0.455 -1.801 -1.801 -0.454 -0.447 (0.133)** (0.133)** (0.723)* (0.723)* (0.133)** (0.133)** T2to6 -0.097 -0.091 0.311 0.311 -0.091 -0.069 (0.046)* (0.046)~ (0.232) (0.232) (0.046)~ (0.047) female -0.150 -0.151 -0.152 -0.152 -0.151 -0.159 (0.076)* (0.077)* (0.077)* (0.077)* (0.077)* (0.078)* black -0.439 -0.441 -0.439 -0.439 -0.440 -0.389 (0.105)** (0.106)** (0.106)** (0.106)** (0.106)** (0.111)** latino -0.234 -0.242 -0.249 -0.249 -0.242 -0.193 (0.130)~ (0.130)~ (0.130)~ (0.130)~ (0.130)~ (0.133) asian -0.085 -0.084 -0.082 -0.082 -0.080 -0.046 (0.126) (0.126) (0.126) (0.126) (0.126) (0.129) agestartC -0.056 -0.056 -0.056 -0.056 -0.056 -0.055 (0.016)** (0.016)** (0.016)** (0.016)** (0.016)** (0.016)** family -0.006 -0.002 -0.000 -0.000 -0.002 -0.001 (0.068) (0.068) (0.068) (0.068) (0.068) (0.069) RURAL 0.074 0.088 0.094 0.094 0.087 (0.084) (0.085) (0.085) (0.085) (0.085) SCITECHmaj 0.385 0.397 0.401 0.401 0.401 0.453 (0.127)** (0.129)** (0.130)** (0.130)** (0.129)** (0.131)** HUMANmaj -0.024 -0.039 -0.037 -0.037 -0.034 -0.006 (0.107) (0.107) (0.108) (0.108) (0.108) (0.109) FLANGmaj -0.246 -0.259 -0.264 -0.264 -0.254 -0.236 (0.143)~ (0.143)~ (0.143)~ (0.143)~ (0.143)~ (0.144) MATHmaj -0.174 -0.168 -0.175 -0.175 -0.168 -0.157 (0.121) (0.122) (0.122) (0.122) (0.122) (0.124) midhs_yr 0.244 0.509 1.382 0.237 0.278 (0.226) (1.038) (0.764)~ (0.234) (0.237) multi_sub 0.306 1.471 1.471 0.309 0.398 (0.225) (0.761)~ (0.761)~ (0.232) (0.236)~ multi_submidhs -0.176 -1.738 -1.738 -0.263 -0.350 (0.244) (0.837)* (0.837)* (0.256) (0.260) T1_multi_sub -0.944 -0.944 0.217 0.394 (1.286) (1.286) (1.062) (1.117) T1_multimidhs 1.639 1.639 0.165 -0.002 (1.347) (1.347) (1.087) (1.141) T1_midhs -0.873 0.272 0.453 (1.289) (1.064) (1.118) T4_multi_sub 1.603 1.603 (0.752)* (0.752)* T4_multimidhs -2.101 -2.101 (0.845)* (0.845)*

40

T4_midhs

1.426 (0.754)~ -0.455 (0.243)~ 0.592 (0.276)* -0.440 (0.244)~ 0.873 (1.289)

T2to6_multi_sub T2to6_multimidhs T2to6_midhs postT1_midhs Regional fixed effects Deviance-based hypothesis tests -2LL Change in -2LL

1.426 (0.754)~ -0.455 (0.243)~ 0.592 (0.276)* -0.440 (0.244)~ YES

5394.3792

5390.7836

5377.7163

5377.7163

5386.6357

5335.5663

3.5956(3)

13.0673(9)

13.0673(9)

4.1479(3)

4.480(3)

Comparison model 1 2 2 2 p value .3806 .1596 .1596 .2459 Observations 5044 5044 5044 5044 5044 ~ significant at 10%; * significant at 5%; ** significant at 1%; Standard errors in parentheses 1 model 2 with regional fixed effects instead of RURAL dummy, not shown

21 .2141 5039

Table 4b: GLH Tests for Significant Differences between Single- and Multi-Assigned Teachers in Each Year (based on Model 5) Year Test Chi square p Results value 1 Ho: ßmulti_sub + ßmulti_submidhs + ßT1_multi_sub + 4.34 .0373 ßT1_multisubmidhs = 0 2-6 Ho: ßmulti_sub + ßmulti_submidhs=0 0.17 0.6777

41

Table 5a: Parameter estimates, standard errors, and goodness-of-fit statistics from a series of hazard models in which assignment at the secondary level predicts the conditional probability that a teacher will voluntarily resign from teaching, controlling for cohort, gender, race, age of entry to the profession, the presence of a teacher in one’s family, college major, and urbanicity of teaching placement (n=2029) Parameter Estimate Predictor (1) (2) (3) Final (4) (5) Model C2 0.108 0.112 0.117 0.112 0.158 (0.089) (0.089) (0.089) (0.089) (0.090)~ C3 0.168 0.179 0.182 0.177 0.220 (0.082)* (0.082)* (0.082)* (0.082)* (0.084)** T1 -3.165 -3.164 -4.062 -3.019 -4.129 (0.158)** (0.159)** (1.296)** (1.055)** (1.334)** T5 0.303 0.302 1.133 0.302 1.161 (0.160)~ (0.160)~ (0.618)~ (0.160)~ (0.622)~ lnTIME -1.064 -1.053 -2.054 -1.049 -1.932 (0.120)** (0.121)** (0.710)** (0.121)** (0.718)** female -0.327 -0.320 -0.326 -0.321 -0.344 (0.074)** (0.074)** (0.074)** (0.074)** (0.075)** black -0.247 -0.248 -0.252 -0.247 -0.185 (0.107)* (0.108)* (0.108)* (0.108)* (0.113) latino -0.208 -0.218 -0.216 -0.219 -0.149 (0.133) (0.133) (0.134) (0.133)~ (0.136) asian 0.265 0.260 0.252 0.262 0.308 (0.122)* (0.122)* (0.123)* (0.122)* (0.125)* agestartC -0.095 -0.095 -0.095 -0.095 -0.095 (0.018)** (0.018)** (0.018)** (0.018)** (0.018)** family -0.031 -0.025 -0.023 -0.026 -0.017 (0.067) (0.067) (0.067) (0.067) (0.068) SCITECHmaj 0.251 0.266 0.266 0.267 0.288 (0.123)* (0.125)* (0.125)* (0.125)* (0.126)* HUMANmaj 0.248 0.232 0.236 0.234 0.237 (0.107)* (0.107)* (0.107)* (0.107)* (0.108)* FLANGmaj -0.112 -0.125 -0.126 -0.122 -0.152 (0.142) (0.142) (0.143) (0.142) (0.144) MATHmaj -0.058 -0.044 -0.046 -0.044 -0.064 (0.118) (0.119) (0.119) (0.119) (0.120) RURAL 0.229 0.247 0.244 0.246 (0.081)** (0.082)** (0.082)** (0.082)** midhs_yr 0.001 -0.858 -0.000 -0.787 (0.188) (0.810) (0.191) (0.822) multi_sub 0.076 -1.077 0.108 -0.970 (0.187) (0.806) (0.190) (0.817) multi_submidhs 0.181 0.890 0.103 0.765 (0.208) (0.875) (0.214) (0.886) T1_midhs 0.779 -0.080 0.847 (1.323) (1.063) (1.361) T1_multi_sub 0.732 -0.453 0.800 (1.321) (1.064) (1.359) T1_multimidhs -0.004 0.784 -0.076 (1.389) (1.100) (1.425) T5_midhs -0.605 -0.621 (0.669) (0.672) T5_multi_sub -1.354 -1.409 (0.680)* (0.684)* T5_multimidhs 1.337 1.363

42

0.059 (0.187)

-0.037 (0.265)

(0.802)~ 0.804 (0.739) 1.180 (0.734) -0.784 (0.809) 0.996 (0.797)

5800.3248

5792.3623

5776.9042

5785.9355

5717.2118

1 7.9625

2 15.4581

2 6.4268

21 15.5961

0.047 6340

.079 6340

.093 6340

.076 6335

lnTIME_midhs lnTIME_multisub lnTIME_multisubmidhs Constant

-0.047 (0.266)

Regional fixed effects Deviance-based hypothesis tests -2LL(df) Comparison model Delta -2LL(df) p value Observations

6341

(0.806)~ 0.720 (0.746) 1.101 (0.742) -0.715 (0.817) 1.344 (0.817)~ YES

~ significant at 10%; * significant at 5%; ** significant at 1%; Standard errors in parentheses 1 model 2 fitted with regional fixed effects instead of the RURAL dummy (not shown) Table 5b: GLH Tests for Significant Differences between Single- and Multi-Assigned Teachers in Each Year (based on Model 3) Year Test Chi square p Results value 1 Ho: ßmulti_sub + ßmulti_submidhs + ßT1_multi_sub + ßT1_multisubmidhs 4.29 .0382 =0 2 Ho: ßmulti_sub + ßmulti_submidhs+ .69314718* ßlnTIME_multisub + 0.40 0.5257 .69314718 *ßlnTIME_multisubmidhs=0 3 Ho: ßmulti_sub + ßmulti_submidhs+ 1.0986123 * ßlnTIME_multisub + 4.67 0.0308 1.0986123 *ßlnTIME_multisubmidhs=0 4 5 6

Ho: ßmulti_sub + ßmulti_submidhs+ 1.7917595* ßlnTIME_multisub + 1.7917595 *ßlnTIME_multisubmidhs=0 Ho: ßmulti_sub + ßmulti_submidhs+ ßT5_multi_sub +ßT5_multimidh +1.609* ßlnTIME_multisub + 1.609*ßlnTIME_multisubmidhs=0 Ho: ßmulti_sub + ßmulti_submidhs+ 1.7917595* ßlnTIME_multisub + 1.7917595 *ßlnTIME_multisubmidhs=0

2.96

0.0856

1.54

0.2147

2.96 0.0856

Results of equivalent GLH tests based on Model 11, with regional fixed effects, were the same.

43

Table 6a: Parameter estimates, standard errors, and goodness-of-fit statistics from a series of hazard models in which in-field vs. out-of-field math assignment at the secondary level predicts the conditional probability that a teacher will leave her school, controlling for cohort, gender, race, age of entry to the profession, the presence of a teacher in one’s family, college major, and urbanicity of teaching placement (n=2029) Parameter Estimates Predictor C2 C3 T1 postT1 T4 T2to6 female black latino asian family agestartC HUMANmaj SCITECHmaj FLANGmaj MATHmaj RURAL midhs_yr math_yr MATHmajXplacement math_yrXmidhs MATHXmidhs MATHmajXplacementXmidhs

(1) vexitschl 0.130 (0.091) 0.154 (0.083)~ -2.234 (0.166)** 0.306 (0.196) -0.454 (0.133)** -0.098 (0.046)* -0.141 (0.076)~ -0.439 (0.106)** -0.232 (0.130)~ -0.093 (0.126) -0.005 (0.068) -0.055 (0.016)** -0.033 (0.108) 0.392 (0.129)** -0.251 (0.143)~ -0.219 (0.165) 0.072 (0.084) 0.006 (0.071)

(2) Model 0.129 (0.091) 0.154 (0.083)~ -2.161 (0.184)** 0.403 (0.216)~ -0.435 (0.135)** -0.109 (0.047)* -0.137 (0.076)~ -0.442 (0.106)** -0.241 (0.131)~ -0.097 (0.127) -0.010 (0.068) -0.056 (0.016)** -0.047 (0.109) 0.362 (0.132)** -0.265 (0.143)~ -0.022 (0.332) 0.067 (0.084) -0.024 (0.083) -0.067 (0.083) -0.206 (0.373)

(3) Model 0.118 (0.091) 0.142 (0.083)~ -2.026 (0.196)** 0.567 (0.231)* -0.394 (0.136)** -0.128 (0.048)** -0.131 (0.076)~ -0.446 (0.107)** -0.252 (0.131)~ -0.104 (0.127) -0.012 (0.068) -0.056 (0.016)** -0.042 (0.109) 0.365 (0.132)** -0.253 (0.143)~ 0.717 (0.812) 0.068 (0.084) -0.178 (0.116) -0.234 (0.118)* -0.920 (1.035) 0.335 (0.165)* -0.872 (0.879) 0.717 (1.113)

(4) Model 0.115 (0.091) 0.141 (0.083)~ -2.084 (0.221)** 0.562 (0.232)* -0.393 (0.137)** -0.127 (0.048)** -0.128 (0.076)~ -0.443 (0.106)** -0.253 (0.131)~ -0.106 (0.127) -0.011 (0.068) -0.057 (0.016)** -0.038 (0.109) 0.364 (0.132)** -0.251 (0.144)~ 0.388 (0.886) 0.066 (0.085) -0.168 (0.117) -0.212 (0.122)~ -1.444 (1.217) 0.184 (0.178) -0.273 (0.966) 1.108 (1.301)

44

T1_MATH T1_MATHmajXplacement T1_math_yr T1_MATHmajXplacementXmidhs T1_MATHXmidhs T1_math_yrXmidhs -2LL(df)

5394.6768

5365.3054

5359.7694

Comparison model 1 2 Change in -2LL 29.3714 5.536 p-value