Are Early Absences Still Early Warnings?

Gwinnett County Public Schools (GCPS)

Map of Gwinnett County, Georgia

Image of GCPS demographics

Motivation

How should we screen for chronic absence risk at the beginning of the school year?

The primary motivation behind this paper is a very practical question: How should teams like this screen for chronic absence risk at the beginning of the school year?

Imagine you’re on one of these teams. It’s a week before school starts, and you want to compile a focus list of students most likely to be chronically absent without intervention. You want to rely on attendance data because it’s transparent, directly related to chronic absence, and monitorable throughout the year.

But you face two problems. First, many students are new to the district — no attendance history, so you’ll have to rely on what their attendance looks like in the first few weeks. Second, for returning students, you know who was chronically absent last year. But once school starts, you’re watching early attendance for everyone — some returning students may have new problems emerging, and some who struggled last year may not this year.

So maybe you just start using that first-month data for everyone? But how predictive is it? And if you already have a student’s attendance history, should you keep using that — or does a month of new data “overwrite” it?

What We Know About First-Month Absences

Screening for chronic absence risk

Olson (2014)

Within-year effects on end-of-year achievement

Gottfried & Kirksey (2017)
Keppens (2023)

Despite abundant research on chronic absenteeism, there’s very little to help with this specific problem of practice.

Early warning system design guides recommend monitoring for risk signs in the first 20-30 school days. But research on how predictive first-month absences actually are is quite limited. The most commonly cited work is Olson’s 2014 research brief, which demonstrated substantial increases in the relative risk of chronic absence among students missing 2-4 days in the first month. But relative risk only tells part of the story — we also need to know precision, how many students get flagged, and how many chronically absent students get identified early.

There’s also related work on the timing of absences within the year. Studies by Gottfried & Kirksey and later Keppens found that when absences occur matters for academic achievement — absences near end-of-year testing are particularly problematic. Importantly for us, Keppens’ month-by-month analysis found that first-month unexcused absences appear especially detrimental to test performance.

So there’s compelling evidence that first-month absences are meaningful. But all of these findings are based on pre-pandemic data.

Chronic Absence Remains High

Graph of chronic absence rates in GCPS and Georgia 2018 to 2025

Research Questions

Are early absences still early warnings?

RQ1 Are first-month absences useful for chronic absence screening?

RQ2 Are first-month absences associated with end-of-year test performance?

RQ1

Are first-month absences
useful for chronic absence screening?

METHOD

“First-Month” Absences

Absences occurring within the first 20 school days

Sampling Procedures

Two cohorts:
- 2018–19 (pre-pandemic)
- 2022–23 (post-pandemic)
K-12 students
Enrolled for at least the first 20 school days

Sample Characteristics

Analytical Approach

Evaluated transparent threshold rules

1+, 2+, 3+, etc., first-month days absent

Reviewed a variety of metrics

Classification efficiency metrics:
- Balanced Accuracy, Youden’s J (overall efficiency)
- F-scores (true positive efficiency)
Key metrics for operational decision-making:
- Sensitivity: Did we flag the kids who need help?
- Positive Predictive Value (PPV): Is this list worth acting on?
- Flag Rate: Can we realistically do this?

For the chronic absence screening, I evaluated threshold rules based on attendance data that resembles what attendance teams would actually use: flagging students if early absences are at least 1, 2, 3, and so on.

I computed confusion-matrix statistics at each threshold. Classification-efficiency metrics like balanced accuracy and F-scores because helped me identify a sensible cutpoint range. But with practice-oriented attendance risk screening in mind, three key metrics I specifically focused on were:

Sensitivity: Of students who end up chronically absent, how many did we catch early? Addressing the basic operational question of are we identifying the kids who need help?

Positive Predictive Value (or PPV): Of students we flag, how many actually become chronically absent? This gets at whether our focus list is worth acting on…when it flags a student as at-risk, is that true most of the time?

And the Flag Rate…the proportion of all students that we flagged. Because even a statistically ‘good’ rule can be operationally unfeasible if it flags every other student.

Analytical Approach (cont’d.)

Compared three beginning-of-year screening strategies

1 Prior-year chronic absence status only

Available for most students but not for new students

2 First-month absences only

Available for all students but likely less reliable

3 Hybrid
approach

Returning students:
Prior-year status

New students:
First-month absences

RESULTS: Key Screening Metrics for Operational Decision-Making

This slide demonstrates the inherent tradeoffs involved in selecting an “at-risk” threshold.

If we flag every student, we will catch every at-risk student, but our focus list will almost always be wrong.

As we move to the right, raising the threshold to 1+ days, 2+ days, we flag a smaller and smaller share of students. That would increase our capacity to provide targeted support. And the list we use would become more and more accurate, even as high as being right 90% of the time when a student is flagged if we go far enough. But we would catch fewer and fewer of all of the students who need support.

The classification efficiency metrics I reviewed showed that, depending on the statistic, somewhere between 1+ to 4+ is the most effective cutpoint.

On the low end of that, 1+ would identify almost every other student and we’d be wrong almost 70% of the time. Probably not going to be our top choice.

On the high end, we’d flag around 1 in 10 students–much more manageable–and when we flagged a student there’d be a 70% chance they’d actually need support–much more precise. But we would capture a much smaller share of all students needing support.

So let’s move on to the bigger practical question to help us determine the best cutpoint for our specific use case of screening at the beginning of the school year.

RESULTS: Beginning-of-Year Screening Strategy Comparison

Note: Based on post-pandemic sample (2022–23)

● Prior-year only
Sens 0.30 PPV 0.56 Flag Rate 0.08

● Hybrid 3+
Sens 0.49 PPV 0.57 Flag Rate 0.12

● Prior-year only
Sens 0.33 PPV 0.56 Flag Rate 0.07

● Hybrid 3+
Sens 0.43 PPV 0.57 Flag Rate 0.09

● Prior-year only
Sens 0.40 PPV 0.77 Flag Rate 0.13

● Hybrid 3+
Sens 0.47 PPV 0.77 Flag Rate 0.16

So now the big practical question — not just what threshold, but what strategy. Let me orient you. Orange is prior-year status only. Grey is first-month absences only. Blue is the hybrid approach. The threshold being used for first-month absences is shown at the top — we’re starting at 1+. X-axis is sensitivity, y-axis is PPV, and I’m showing percent flagged as well. Results are broken out by school level. The orange dot won’t move as I advance — prior-year status is static. It’s our baseline. The goal is for grey or blue to be to the right of and above orange — more at-risk students flagged without sacrificing precision. At 1+ and 2+, we’re not quite there yet. At 3+, the hybrid starts to deliver: precision is equal to or slightly better than prior-year only, while capturing roughly 20% more of the at-risk elementary population, 10% more at middle, and about 8% more in high school — flagging around 10 to 15% of students overall. Beyond 3+, we’re trading sensitivity for marginal precision gains. First-month absences alone never outperform the other strategies. The hybrid at 3+ is the sweet spot. One pattern worth flagging: performance is notably better in high school — much higher PPV at comparable sensitivity levels. That’s a meaningful signal for where supplemental information may be needed, especially in K–8.

RQ2

Are first-month absences
associated with end-of-year test performance?

Has the effect shifted post-pandemic?

METHOD

Outcome: Performance on end-of-year state Math & ELA tests

z-score based on state mean & SD

Sample

Same as Study 1 but further restricted to grade levels with end-of-year tests and prior achievement controls (Math: Grades 4-7; ELA: 4-8)

Academic Effects Modeling

For student \(i\) in school \(j\), achievement was modeled as:

\[ \begin{aligned} Y_{ij} &= \beta_0 + \\ &\quad \color{#d95f02}{\beta_1 (\text{DaysAbsent1stMonth}_{ij})} \\ &\quad \beta_2 (\text{Post}_{ij}) + \\ &\quad \color{#1f77b4}{\beta_3 (\text{DaysAbsent1stMonth}_{ij} \times \text{Post}_{ij})} + \\ &\quad \beta_4 (\text{PriorMA}_{ij}) + \beta_5 (\text{PriorELA}_{ij}) + \\ &\quad \beta_6 (\text{PriorChronicAbsence}_{ij}) + \\ &\quad \mathbf{X}_{ij}\boldsymbol{\gamma} + u_j + \varepsilon_{ij} \end{aligned} \]

where:

\(Y_{ij}\) is end-of-year math/ELA z-score for student \(i\) in school \(j\).
\(\text{DaysAbsent1stMonth}_{ij}\) is total days absent during the first 20 school days.
\(\text{Post}_{ij}\) is a binary indicator for the post-pandemic cohort.
\(\text{DaysAbsent1stMonth} \times \text{Post}\) tests whether the absence slope differs pre- vs. post-pandemic.
\(\text{PriorMA}_{ij}\) and \(\text{PriorELA}_{ij}\) are prior-year Math and ELA z-scores.
\(\text{PriorChronicAbsence}_{ij}\) is prior year chronic absence status.
\(\mathbf{X}_{ij}\) is a vector of student demographic covariates.
\(u_j \sim N(0, \tau^2)\) is a school-level random intercept.
\(\varepsilon_{ij} \sim N(0, \sigma^2)\) is the student-level residual.

RESULTS: 1st-Month Absences & End-of-Year Test Performance

Association with end-of-year achievement?
Has the effect shifted post-pandemic?

On the left we have the absence related conditional effects results for the ELA model in the top table…math in the bottom table.

Our first question, in orange up there, is whether there is an association with end-of-year achievement…after controlling for a variety of important confounds like prior achievement and prior absenteeism.

The first rows show the estimated effect of each additional day absent in the first month. I’m visualizing that effect on the chart to the right. That reference line is zero, so values below that indicate a negative average effect on test performance.

The dot is the estimated slope (from the table) and the bars are the 95% confidence interval around that estimate.

We can see that there are negative effects in ELA and Math. With a stronger effect in Math than in ELA. Both are statistically different from zero.

For context, if these were at the -0.07, around 3 absences would mean a roughly 0.2 SD reduction in test performance. Something certainly noticeable. For math, about a 0.1 SD reduction at 5 or more absences. So trivial to small effect at just a couple absences. A moderate effect once you get around 5+.

Our second question is whether that effect shifted post-pandemic. And I didn’t find evidence of that.

Here in blue we’re testing for that shift. We can see that the estimated effects for the post-pandemic group are slightly different. But those differences are trivial and not statistically different from the pre-pandemic estimate.

Implications for Policy and Practice

Early absences are still early warnings.
Improves beginning-of-year screening · Meaningful academic signal in math (at ≥5 first-month absences)

Hybrid 3+ is the recommended beginning-of-year screening strategy.
Prior-year status for returning students · First-month absences (≥3) for new students

Pre-pandemic findings still hold.
No shift in optimal threshold or academic effects · Existing guidance remains actionable

To sum everything up, I looked at whether first-month absences can reliably flag chronic absence risk at the beginning of the school year. I tested different screening strategies and assessed optimal cutpoints for total first-month absences. The key finding: A hybrid approach–using prior-year chronic absence status for returning students and first-month absences for new students–was the best screening strategy. And across levels–elementary, middle, and high–the optimal threshold was 3 or absences within the first 20 days for those new students. This “Hybrid 3+” strategy improved identification of at-risk students without sacrificing precision, while still flagging a relatively small portion of the student population.

I also looked at whether these first-month absences are associated with end-of-year test scores. And they are, especially in math – with the signal reaching meaningful levels around 5 or more absences in the first 20 days. So these early absences are not just an attendance flag. They may be an academic one too.

And the overarching question was whether the pandemic has resulted in any shifts that should result in policy and practice shifts. While higher prevalence rates have naturally made screening a little more efficient in some respects, the main take-away is no: our post-pandemic analysis showed that the ideal threshold pre- and post-pandemic was the same and that academic effects were statistically no different. School districts should be able to continue using pre-pandemic research on chronic absence with confidence.

Limitations
Scope
Single school district.

Sample
Students enrolled at least first 20 days.

Design
Adjusted associations, not causal estimates.

Future Directions
Screening timing
When does current-year attendance alone become sufficient?

K-8 screening improvement
Can additional administrative signals improve early identification?

Generalizability. Results based on data from a single school district; contextual differences (e.g., policies, transportation, demographics) may mean different results elsewhere.
Sample restrictions. Sample restricted to students enrolled through Day 20. Very early mobility may be underrepresented, and this does not include analyses of new students enrolling after school begins.
Non-causal. Screening analyses were designed specifically to inform early warning indicator design, not to explain causal relationships. Achievement models estimated adjusted associations, not causal effects.

Two future directions follow directly from this work. One is a timing question: At what point in the school year does current-year attendance become as predictive as prior-year chronic absence for returning students? That’s a very actionable design question for early warning systems. Second, especially for K–8, there may be value in predictive approaches that incorporate additional administrative signals—while still being transparent and acceptable for school use.

Thank you!

Matthew Lovelace

Director of Research · Gwinnett County Public Schools

Matthew.Lovelace@gcpsk12.org

References

Gottfried, M. A., & Kirksey, J. J. (2017). “When” students miss school: The role of timing of absenteeism on students’ test performance. Educational Researcher, 46(3), 119–130.

Keppens, G. (2023). School absenteeism and academic achievement: Does the timing of the absence matter? Learning and Instruction, 86, 101769.

Olson, L. S. (2014). Why september matters: Improving student attendance. Policy brief. Baltimore Education Research Consortium.

RESULTS: Classification Performance Metrics

This additional slide is the ‘where do cutpoints seem to want to be’ picture. Across cohorts, you see a highly similar pattern: metrics that evenly weight sensitivity and specificity (that’s the top 2 figures) they tend to suggest 1+ early absences as the optimal cutpoint. But for attendance intervention, the cost of over-indentfication is real: staff time, family contact capacity and impact, and the risk of diluting the quality of support provided. Which is why I paid particular attention to F-scores, which are for situations where more emphasis is placed on correct classification of positive cases. It’s more aligned with the practical work of early intervention teams, who would generally want to avoid false positives and generating an unmanageable focus list of students. F1 (equally weighting PPV and Sensitivity) indicates 2+ as ideal. F0.5, weighs precision more heavily, suggest a slightly higher threshold band of 3 or more total absences in the first 20 days.

Are Early Absences Still Early Warnings?

Gwinnett County Public Schools (GCPS)

Motivation

What We Know About First-Month Absences

Screening for chronic absence risk

Within-year effects on end-of-year achievement

Chronic Absence Remains High

Research Questions

Are first-month absencesuseful for chronic absence screening?

METHOD

“First-Month” Absences

Sampling Procedures

Sample Characteristics

Analytical Approach

Evaluated transparent threshold rules

Reviewed a variety of metrics

Analytical Approach (cont’d.)

RESULTS: Key Screening Metrics for Operational Decision-Making

RESULTS: Beginning-of-Year Screening Strategy Comparison

Are first-month absencesassociated with end-of-year test performance?

METHOD

Outcome: Performance on end-of-year state Math & ELA tests

Sample

Academic Effects Modeling

RESULTS: 1st-Month Absences & End-of-Year Test Performance

Implications for Policy and Practice

Thank you!

References

RESULTS: Classification Performance Metrics

Are first-month absences
useful for chronic absence screening?

Are first-month absences
associated with end-of-year test performance?