education-advanced-quantitative-modeling

# Education Advanced Quantitative Modeling ## Goal Select, specify, run, and report advanced quantitative models that match education research questions and data structures. ## Use After Use after: - `education-quantitative-data-cleaning` - `education-descriptive-statistics` - `education-inferential-statistics` - `education-validity-reliability-analysis` - `education-quantitative-study-design` - `education-learning-analytics-design` - `education-program-evaluation` Do not expose the skill name to users. Present it as "高级量化模型分析". ## Inputs - Research questions/hypotheses - Cleaned dataset - Variable roles and levels - Sample size and clustering structure - Timepoints if longitudinal/pre-post/panel - Measurement model if latent variables are used - Intervention/control group design if applicable ## Model Selection Guide | Research Need | Data Structure | Recommended Model | |---|---|---| | Students nested in classes/schools | Student + class/school IDs | Multilevel model / HLM | | Latent constructs measured by items | Item-level scale data | CFA / SEM | | Test indirect mechanism | IV -> mediator -> DV | Mediation model | | Test different effect by subgroup | Predictor x moderator | Moderation model | | Latent constructs + paths | Latent variables with relationships | SEM | | Item difficulty/discrimination | Test/item response data | IRT/Rasch | | Growth over time | 3+ timepoints | Growth curve / longitudinal mixed model | | Policy/intervention before-after with comparison group | Treatment/control + pre/post | DID | | Nonrandomized comparison | Observational treatment/control | Propensity score matching/weighting + regression | | Repeated measures within students/classes | Repeated observations | Mixed-effects model | ## Workflow 1. Confirm whether an advanced model is justified. 2. Identify unit and level: - item - timepoint - student - class - teacher - school - region 3. Check sample size/model complexity. 4. Select model family. 5. Specify variables and equations/path diagram. 6. Check assumptions and diagnostics. 7. Run baseline/simple model first. 8. Add complexity stepwise. 9. Interpret parameters with educational meaning. 10. Produce tables, diagrams, and reporting text. ## Tool Calls Dependency policy: inspect the current environment first. Do not install packages, clone repositories, or change the user's environment without explicit approval. Prefer a project-local Python virtual environment or R library, and record package versions used for reproducibility. ### R Packages ```r install.packages(c( "tidyverse", "lme4", "lmerTest", "performance", "broom.mixed", "lavaan", "semTools", "mediation", "interactions", "mirt", "TAM", "MatchIt", "WeightIt", "fixest", "did", "marginaleffects" )) ``` Multilevel model: ```r library(lme4) library(lmerTest) fit <- lmer(score_post ~ group + score_pre + (1 | class_id), data = data) summary(fit) performance::check_model(fit) ``` Random slope model: ```r fit <- lmer(score_post ~ group + score_pre + (1 + group | class_id), data = data) ``` SEM: ```r library(lavaan) model <- ' motivation =~ m1 + m2 + m3 + m4 engagement =~ e1 + e2 + e3 + e4 achievement ~ motivation + engagement engagement ~ motivation ' fit <- sem(model, data = data, estimator = "MLR", missing = "fiml") summary(fit, fit.measures = TRUE, standardized = TRUE) ``` Mediation: ```r model <- ' mediator ~ a * intervention outcome ~ b * mediator + cprime * intervention indirect := a * b total := cprime + (a * b) ' fit <- sem(model, data = data, se = "bootstrap", bootstrap = 5000) summary(fit, standardized = TRUE, ci = TRUE) ``` IRT: ```r library(mirt) fit <- mirt(item_data, 1, itemtype = "2PL") summary(fit) coef(fit, IRTpars = TRUE) ``` DID: ```r library(fixest) fit <- feols(outcome ~ treatment * post + covariates | school_id + time, data = data) summary(fit) ``` Propensity score matching: ```r library(MatchIt) m <- matchit(treatment ~ pre_score + gender + grade + ses, data = data, method = "nearest") matched <- match.data(m) ``` ### Python Packages ```bash pip install pandas numpy statsmodels linearmodels semopy pyirt scikit-learn causalinference ``` Mixed model: ```python import statsmodels.formula.api as smf fit = smf.mixedlm("score_post ~ group + score_pre", data=df, groups=df["class_id"]).fit() print(fit.summary()) ``` DID: ```python import statsmodels.formula.api as smf fit = smf.ols("outcome ~ treatment * post + covariates", data=df).fit(cov_type="cluster", cov_kwds={"groups": df["school_id"]}) print(fit.summary()) ``` ### Specialized Tools ```text Mplus: https://www.statmodel.com/ HLM: https://ssicentral.com/index.php/products/hlm-general/ Stata: https://www.stata.com/ AMOS: https://www.ibm.com/products/structural-equation-modeling-sem Winsteps: https://www.winsteps.com/ jamovi SEMLj module: https://semlj.github.io/ ``` ## Output Format ### 1. Model Choice Table | RQ/Hypothesis | Data Structure | Candidate Model | Recommended Model | Reason | |---|---|---|---|---| ### 2. Variable-Level Table | Variable | Role | Level | Type | Notes | |---|---|---|---|---| Levels: - item - timepoint - student - class - teacher - school - region ### 3. Model Specification | Component | Specification | |---|---| | Outcome | | | Fixed effects | | | Random effects | | | Latent variables | | | Covariates | | | Clustering | | | Estimator | | ### 4. Model Results Table | Parameter | Estimate | SE | Test Statistic | p | 95% CI | Interpretation | |---|---|---|---|---|---|---| ### 5. Diagnostics / Fit | Model Type | Diagnostics | |---|---| | Multilevel | ICC, random effects, residuals, convergence | | SEM | CFI, TLI, RMSEA, SRMR, standardized loadings | | IRT | item difficulty, discrimination, item fit | | DID | parallel trends, clustered SEs, robustness | | PSM | covariate balance, common support | ## Mermaid Templates ### Mediation ```mermaid flowchart LR X["Intervention"] --> M["Mediator"] M --> Y["Outcome"] X --> Y ``` ### Multilevel Data ```mermaid flowchart TD S["School"] --> C1["Class"] C1 --> ST1["Student"] C1 --> ST2["Student"] ``` ## Education-Specific Guidance - Education data are often nested; inspect ICC before ignoring class/school effects. - SEM requires theory-driven measurement and structural models. - DID requires credible parallel trends; do not use it just because data are pre/post. - IRT requires item-level response data and enough examinees/items. - Mediation with cross-sectional data should be interpreted cautiously. - Program evaluation with nonrandom assignment may require baseline controls or propensity methods. ## Quality Rules - Start with the simplest defensible model. - Do not use advanced models to disguise weak design. - Report model assumptions and diagnostics. - Avoid causal language without causal identification. - Keep model complexity proportional to sample size. - For nested education data, cluster-robust SEs or multilevel models may be needed. ## User-Facing Closure End by choosing the next modeling step: ```text 根据你的数据结构，最合适的高级模型是 [模型]，原因是 [简短原因]。接下来我可以先帮你写模型设定和变量层级表，或者直接生成 R/SPSS/Stata/Mplus 分析脚本。 ```