RQ2b

Author

Calum Davey 

library(data.table)
Warning: package 'data.table' was built under R version 4.4.3

Abstract

1. Background

1.1 Evidence on teacher effectiveness growth

  • VA growth in 1-5 years, then v slow change (Podolsky 2016)
  • How big? 0.1 SD in pupil attainment
  • For context, 25th and 75th percentile teachers will differ by about 0.25 SD (Slater 2012)
  • Kraft & Papay 2012

–> A new teacher does get better with experience but does not typically improve so much as to close the gap with the top performers solely through experience. For example, if an average first-year teacher is – hypothetically – at the 40th percentile of effectiveness initially, after 5 years they might move to the 50th or 60th percentile with that ~0.1 SD gain

1.2 Estimation challenges

  • Cross-sectional: who are the most effective teachers now, in terms of experience?
  • FE model of VA by experience (multiple years required) (Kraft 2015, Ladd 2017, Jackson 2016)
  • Poisson/Cox model of time-to-leaving by experience
  • Teacher quality decomposition (Wiswall 2013)

1.3 Research questions

  1. How does our measure of teaching quality change over time and relate to teacher retention? This is about several things:
    1. Are more experienced teachers, in the MATs, more or less effective then less effective teachers?
    2. How much does teachers’ VA change during their career?
    3. Is time-to-leaving teaching associated with VA (initial, current, or trend)?
    4. How does answering 2 & 3 help explain 1?

1.4 Contribution

  • Evidence in England
  • Student fixed-effects with within-year test scores in other subjects
  • Close attention to survivor bias, with competing risks framing

2. Methods

  • Using data from one MAT, MAT 2 and/or MAT 3
d  <- fread("../../TED/MAT3/01_MAT3_hr_clean.csv")
  • Estimation methods
  • What will the TVA be? Will this be conditional, ie accounting for child and school-level factors (isolating the teacher level?). For example, will we have an estimate of TVA which takes into account allocation to classes, so will address differences between classes and also any associations that we have observed regarding under which circumstances teachers can have the highest VA draw from their overall average?

3. Results

3.1 Description of the data

d[, crs_date := as.Date(cont_relevant_service_date, format = c("%d/%m/%Y"))]
d[, date := as.Date(as.character(academic_year), format = "%Y")]
d[, duration := as.numeric(d$date - d$crs_date)/365.25]

# distribution of years of experience 
hist(d$duration)

summary(d$duration)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
 -1.374   2.198   6.196   8.190  11.794  55.280 
# balance of the panel 
table(d$academic_year)

2023 2024 2025 
1954 1953 1853 
d[, .(rowCount = .N), by = staff_id][, (freq = .N), by = rowCount][order(-rowCount)]
   rowCount    V1
      <int> <int>
1:       41     1
2:        3  1219
3:        2   626
4:        1   810
# data available from first 5 years of teaching 
d[duration < 6, .(rowCount = .N), by = date]
         date rowCount
       <Date>    <int>
1: 2023-11-12      981
2: 2024-11-12      984
3: 2025-11-12      889
# number of teachers who started in first year 
d[duration < 4, .(rowCount = .N), by = staff_id][rowCount == 3, (freq = .N)]
[1] 198
# plot of cohort 
dates <- d[, .(date = max(date), crs_date = min(crs_date)), by = staff_id][order(-crs_date)]
plot(x = as.Date("2025-11-12"), y = 1, xlim=c(min(dates$crs_date), max(dates$date)), 
     ylim = c(0,nrow(dates)), type = 'n', xlab = 'Date', ylab = '')
for (i in 1:nrow(dates)){
  lines(x = dates[i, 2:3], y = c(i, i), lwd = .01, col = "#8D10FF")
}
abline(v=unique(dates$date), col="red", lwd=0.8)

# plot of the observed data, ordered by date starting in teaching
plot(x = as.Date("2025-11-12"), y = 1, xlim=c(min(dates$date), max(dates$date)), 
     ylim = c(0,nrow(dates)), type = 'n', xlab = 'Date', ylab = '')
for (i in 1:nrow(dates)){
  lines(x = dates[i, 2:3], y = c(i, i), lwd = .001, col = "#8D10FF")
}
abline(v=unique(dates$date), col="red", lwd=0.8)