1 of 41

DTSC-870 Project �

Casandra Clarke, Mohamed Bassimbo, �Alberlis Hernandez, Franciss Torres

2 of 41

Data Sources for Tennessee School Performance Analysis

This project integrates 93 separate Tennessee Department of Education data files spanning the 2018–2024 academic years and merges them at both the district and school levels. As shown on the right, Microsoft Access UNION ALL queries were used to stack year-specific tables into consolidated multi-year datasets for each domain, ensuring that original school-year records remained intact.

Once combined, each longitudinal table was aligned using shared identifiers (Year, District Number, and School Number), allowing records from different domains to be reliably joined. This approach ensured that every school–year observation remained distinct while enabling consistent tracking of individual schools across multiple academic years with comparable variables.

Extensive data cleaning was required to standardize school names and correct inconsistent identifiers. The final unified dataset provides a multi-dimensional view of each school-year record, combining academic outcomes, resource allocation, staffing conditions, and student demographics into a single analytic framework. This structure allows us to examine not only cross-sectional patterns, but also how school performance changes over time.

3 of 41

Hypothesis

Increasing per-pupil or per-school spending directly improves graduation rates and ACT scores
Schools with higher ACT scores and graduation rates are more likely to receive higher Tennessee Value-Added Assessment System (TVAAS) ratings
Schools in the top 30 and bottom 30 for a given outcome exhibit statistically significant and substantively large differences in related academic and demographic characteristics
Student demographics are strongly associated with school performance outcomes, but do not fully determine academic achievement or graduation rates

4 of 41

Task One: Design, Integrate, and Clean the School Performance Database��By Francis Torres

5 of 41

Data Frame Structure

This first stage of the database design integrates multiple datasets from the Tennessee Department of Education. As shown by the column inventory on the right, the integrated dataset includes a wide range of variables, including identifiers, staffing characteristics, financial measures, student outcomes, demographics, and school classifications.

�

Although each source dataset contributes its own specialized columns, all records share a common set of identifiers: Year, DistrictNumber, and SchoolNumber, which serve as the primary keys linking information across tables.

�

The relational diagram illustrates how each fact table joins back to the central school_profile table, which consolidates school identity, enrollment, demographic composition, and classification attributes. This was critical to preventing duplicate or mismatched records and for preserving a single, consistent representation of each school–year observation across the 2018–2024 period.

�

Following the establishment of this relational backbone, the second stage incorporates the financial_school and act_scores_school tables into the same schema. Like the other fact tables, these data connect to the central school_profile table through the shared Year–District–School identifiers, ensuring that financial and academic measures remain directly comparable across schools and over time.

�

6 of 41

Data Cleaning and Preprocessing

Raw Data

CSV / Excel files (2018–2024)

Standardization

Column names
School IDs
Year alignment

Cleaning & Harmonization

School names
School types
Duplicate removal

Transformation

Percent → numeric
Spending → $K / $M
Outcome encoding

Model-Ready Data

ACT models
TVAAS models
GAM / Ordinal inputs
Demographic Data Maps

To support a multi-year analysis of Tennessee high schools, we implemented a rigorous and consistent data-preprocessing pipeline across all datasets spanning the 2018–24 academic years. As illustrated in the preprocessing workflow diagram, raw school-level files were first standardized to ensure structural compatibility across years, including harmonized column names, corrected data types, and consistent identifiers for schools and districts. School names were cleaned to eliminate duplicate records, and each observation was uniquely defined by school and academic year.

�

As shown in the dataset size summary by year, this process preserved a large, stable sample of school records across most years, while also revealing substantial reductions in available data in later years due to partial reporting. This informed decisions about year inclusion and supported year-specific modeling rather than inappropriate pooling across incomplete periods.

�

School classifications were harmonized using official Tennessee Department of Education designations and subsequently collapsed into a condensed set of school types to reduce sparsity and improve model stability. Financial variables were rescaled to thousands or millions of dollars, and mixed percentage and decimal fields were standardized to a consistent numeric format to ensure comparability across variables and years.

�

Patterns of missing data were examined before modeling, as shown in the missing-value distribution across variables. Missingness was concentrated in specific outcome and accountability measures, particularly TVAAS-related variables, rather than uniformly across the dataset. Accordingly, missing data were handled selectively based on analytical relevance, preserving sample size while ensuring valid model inputs.

�

Outcome variables were prepared to align with modeling strategies, with TVAAS treated as an ordered categorical outcome and ACT scores and graduation rates treated as continuous measures. All datasets were then structured longitudinally, enabling year-specific modeling and cross-year comparisons.

�

7 of 41

Task Two: Explore Financial Trends in Per-Pupil and Per-School Expenditures��By Mohamed Bassimbo and Franciss Torres ��

8 of 41

9 of 41

10 of 41

11 of 41

12 of 41

13 of 41

To move beyond total spending, we use a Generalized Additive Mixed Model (GAMM), which allows flexible, non-linear relationships between financial inputs and enrollment while accounting for differences across school types.

Across all Tennessee school types, the GAMM reveals a consistent negative relationship between per-pupil spending and enrollment size: schools that spend more per student tend to be smaller, while larger schools tend to have lower per-pupil costs. This appears across Alternative, Charter, Magnet, Early College, Public High, Special Education, and Virtual schools, indicating a structural funding pattern rather than a school-type-specific effect.

The model’s smooth terms confirm that this relationship is highly significant. The two-dimensional smooth of school-level expenditures and nutrition spending is strongly significant (F ≈ 37, p < 2e-16), as is school-level nutrition spending (F ≈ 12, p < 2e-16). School type tells us how big a school usually is, but it doesn’t change how enrollment responds to per-pupil spending. The downward curve is nearly identical across all school types, reinforcing that per-pupil spending reflects scale and structure, not instructional effectiveness. If per-pupil spending mainly reflects school size and structure, the next question is whether it translates into improved student outcomes.

14 of 41

Task Three: Analyze the Relationship Between Student Outcomes and Per-Pupil Expenditures��By Francis Torres

15 of 41

So far, we’ve shown that per-pupil spending mainly reflects school size and structure rather than instructional differences. The next question is whether higher per-pupil spending actually translates into better academic outcomes. To test this, we use a Generalized Additive Mixed Model to examine the relationship between per-pupil spending and ACT scores, allowing for non-linear effects and differences across school types. In some school types, the relationship is slightly positive, in others it is flat or even negative, and in many cases the confidence bands overlap (Curves change direction, Patterns differ by school type but not systematically) We use a GAMM because it allows the relationship between spending and ACT scores to be non-linear, rather than forcing a straight-line assumption. Even with this model, spending does not emerge as a strong or reliable predictor of ACT performance. This suggests that per-pupil spending alone is not a sufficient explanation for differences in academic outcomes. In short, how much is spent per student matters far less than how effectively instruction is delivered

16 of 41

We see the same pattern when we move from test scores to longer-term outcomes. After finding no consistent relationship between per-pupil spending and ACT performance, we next examine whether spending is associated with graduation rates. “Again, we use a Generalized Additive Mixed Model to allow for non-linear effects and differences across school types. Some school types show slight positive trends, others are flat or even negative, and several curves change direction across the spending range. If per-pupil spending were a primary driver of student success, we would expect to see a consistent upward relationship across school types, and instead, the relationship is weak, inconsistent, and highly dependent on school context. Taken together with the ACT results, this suggests that financial inputs alone are insufficient to explain meaningful differences in student outcomes. So if money is not the primary driver, the next question is what explains differences in school performance

17 of 41

�

Task Four: Analyze the Relationship Between Student Outcomes and Tennessee Value-Added Assessment System (TVAAS) Ratings

By Francis Torres

18 of 41

TVAAS reflects instructional effectiveness across subjects and grade levels.

Using ordinal logistic regression, we find that ACT scores strongly predict whether a school falls into the lowest or highest TVAAS categories, but offer limited separation among middle TVAAS levels. This suggests that while achievement helps distinguish clearly low- and high-performing schools, differences among average-performing schools are influenced by additional factors beyond test scores alone.

these results indicate that TVAAS offers a more informative lens for evaluating school performance than per-pupil spending, and this tells us that instructional effectiveness and student context matter more than financial inputs alone. That said, ACT scores capture academic achievement at a single point in time. To test whether this pattern extends to longer-term outcomes, we next examine graduation rate

19 of 41

Graduation rates, like ACT scores, strongly distinguish schools at the lowest and highest TVAAS levels but offer limited separation among middle categories. As graduation rates increase, the probability of a school being rated TVAAS level 5 rises, while the probability of being rated level 1 declines. In contrast, TVAAS levels 2, 3, and 4 overlap substantially across most of the graduation rate range, indicating that graduation rates alone cannot reliably differentiate between average-performing schools.

This finding aligns with the design of TVAAS, which measures student growth rather than proficiency or completion outcomes. Together, these findings suggest that instructional effectiveness, as captured by TVAAS, corresponds with achievement and completion outcomes at the extremes but is not fully explained by them. This reinforces the importance of incorporating demographic and contextual factors when evaluating differences among schools with similar outcome levels.

20 of 41

Up to this point, our analyses show achievement measures such as ACT scores and graduation rates meaningfully distinguish schools at the lowest and highest ends of the TVAAS scale.

To assess whether these relationships are consistent across different schools, we examine how TVAAS relates to outcomes across different school types. Across most school categories, including Charter, Magnet, Public High, and Virtual schools, low-performing TVAAS ratings are associated with lower ACT scores and graduation rates. As ACT scores and graduation rates increase, the probability of being rated TVAAS level 1 declines sharply, while the likelihood of being rated level 5 increases.

However, this separation weakens substantially among middle-performing schools. TVAAS levels 2, 3, and 4 overlap considerably across both ACT scores and graduation rates, indicating that instructional effectiveness alone does not fully explain variation among average-performing schools.

This suggests that while TVAAS captures meaningful differences in instructional impact at the extremes, additional demographic and contextual factors are likely necessary to explain performance differences within the middle range.

Taken together, these findings indicate that student outcomes are more aligned with instructional effectiveness than with how much is spent per pupil

21 of 41

Task Five: Compare Top 30 and Bottom 30 Schools Across Academic and Demographic Characteristics ��By Alberlis Hernandez��

22 of 41

Dropout Rates: Top 30 vs Bottom 30 Schools

Ranked schools by dropout rate�
Top 30 vs Bottom 30�
Independent t-tests�
Effect sizes

(Cohen’s d)

23 of 41

Statistical Significance of Dropout Differences

Extremely low p-values�
Large effect sizes

24 of 41

Dropout Rate & Demographics

Top 30: Higher White enrollment
Bottom 30: Higher African American enrollment

25 of 41

ACT Score Analysis

Ranked by ACT scores�
Top 30 vs Bottom 30�
T-tests & Cohen’s d

26 of 41

ACT Statistical Results

Very large mean differences�
All p-values < .001�
Cohen’s d ≈ 6–7

27 of 41

ACT Scores & Demographics

High ACT: More White & Asian students
Low ACT: More African American students

28 of 41

TVAAS Score Distribution by Enrollment

Composite scores (1–5)�
Weighted by enrollment�
Top 30 vs Bottom 30�
Chi-square tests

29 of 41

TVAAS Statistical Results

Significant in 3 of 4 years�
Moderate to strong effect sizes

30 of 41

Task Six: Examine the Relationship Between Student Demographics and School Performance Outcomes

By Casandra Clarke

31 of 41

Relationship Between Race and Graduation Outcomes (Top vs Bottom 30 Schools)

32 of 41

"Graduation Rate and Demographic Composition: Top vs Bottom 30 Schools (2018–2023)"

33 of 41

ACT Score & White %: Positive (r = 0.47)�ACT Score & African American %: Negative (r = –0.51)�ACT Score & Economically Disadvantaged %: Strong Negative (r = –0.76)�Grad Rate & White %: Strong Positive (r = 0.63)�

Summary: Schools with more white students and fewer economically disadvantaged students tend to have higher ACT scores and grad rates.

Correlation Heatmap: Public Schools

34 of 41

ACT & White %: Very strong (r = 0.80)�ACT & African American %: Very strong negative (r = –0.77)�ACT & Economically Disadvantaged %: Extremely strong negative (r = –0.88)�Grad Rate & ACT: Very strong (r = 0.79)�

Summary: These schools show the clearest demographic divide , ACT and graduation outcomes are highly correlated with race and economic status.

Correlation Heatmap: Magnet School

35 of 41

	Average ACT Composite	grad_rate_state	african_american_pct	white_pct	hispanic_pct	economically_disadvantaged	limited_english_prof
Average ACT Composite	0	0	0	0	0.0822	0	0.077
grad_rate_state	0	0	0	0	0.5851	0	0.0385
african_american_pct	0	0	0	0	0	0	0.4019
white_pct	0	0	0	0	0.0012	0	0.6043
hispanic_pct	0.0822	0.5851	0	0.0012	0	0.015	0
economically_disadvantaged	0	0	0	0	0.015	0	0.1534
limited_english_prof	0.077	0.0385	0.4019	0.6043	0	0.1534	0

36 of 41

ACT & White %: Moderate (r = 0.43)�ACT & African American %: Moderate (r = –0.53)�ACT & Econ. Disadvantaged: Moderate (r = –0.51)�

Summary: Patterns are similar to public schools, but weaker student composition still matters, but not as sharply.

Correlation Heatmap: Charter School

37 of 41

Very weak correlations overall�Only notable: ACT & Econ. Disadvantaged (r = –0.76)�

Summary: Demographics don’t predict much in virtual schools possibly due to smaller or more varied populations.

Correlation Heatmap: Virtual

38 of 41

"Average Student Outcomes and Demographics by School Type (2018–2023)"

Public High Schools:

2nd highest ACT (18.95) and grad rate (91.48%)
Predominantly White (70.56%) and lower disadvantage (28.41%)

Charter & Magnet Schools:

Serve highly disadvantaged and diverse students (over 60% Black, high econ. disadvantage)
Lower ACT scores (17.44, 17.3) and grad rates (~85%)

Alternative Schools:

Lowest ACT scores (15.23) and graduation rate (77.88%)
Highest % EL students (7.14%) and high economic disadvantage (37.37%)

Early College:

Highest ACT scores (22.4) and graduation rate (95.8%), Lowest % economically disadvantaged (26.6%) and EL students (2.4%)

39 of 41

Conclusions

Increases in per-pupil or per-school spending were not significantly associated with higher graduation rates or ACT scores. While spending increased over time, financial spending alone does not directly drive academic performance.
Schools with higher ACT scores and graduation rates were substantially more likely to receive higher TVAAS ratings.
Schools in the top 30 and bottom 30 for a given outcome exhibited statistically significant and substantively large differences in academic performance and selected demographic characteristics.
Student demographics are strongly correlated with school performance outcomes, but variation within demographic groups indicates that demographics alone do not fully explain academic achievement or graduation rates

40 of 41

Limitations

Spending data may strengthen trends, but it doesn't explain them

It doesn't explain how funds are used

Generalized Additive Mixed Models are based on school-level data and identify associations rather than causal effects. While these models captured significant nonlinear trends, their low R² values indicate that school-level predictors explain only part of the variation in student outcomes.

Aggregation Bias & Lack of Disaggregation

Statewide averages hide dramatically different stories across districts

No demographic breakdown (Urban vs Rural, High Poverty vs Affluent)

41 of 41

Future Work

Build Predictive modeling

Create Models to predict future ACT scores, Graduation rates, or TAAVS ratings

Incorporate additional variables (attendance, teacher quality, course access).
Explore long-term trends beyond 2018–2023.
Evaluate the policies over time

Whether new policies shift the outcome
Did certain districts benefit more than others?

Add Geographic Analysis

See where low or high-performing districts cluster
Which regions have the most economic disadvantage