drug-discovery-prompts/upstream/mims-harvard-ToolUniverse/skills/tooluniverse-statistical-modeling/TOOLS_REFERENCE.md

title, task, lineage_type, upstream_source, upstream_sha, imported_at, prompt_class, upstream_changes, author, validated

title	task	lineage_type	upstream_source	upstream_sha	imported_at	prompt_class	upstream_changes	author	validated
Tools Reference: Statistical Modeling Skill		import	https://github.com/mims-harvard/ToolUniverse/blob/e2520a96/skills/tooluniverse-statistical-modeling/TOOLS_REFERENCE.md	e2520a96	2026-06-26	prompt	accepted	upstream	false

Tools Reference: Statistical Modeling Skill

Core Python Packages (Computational)

This skill is primarily computational. The statistical modeling is done with Python packages, not ToolUniverse API tools.

statsmodels (Primary statistical modeling)

Function	Purpose	Key Parameters
smf.ols(formula, data)	OLS linear regression	formula (R-style), data (DataFrame)
smf.logit(formula, data)	Binary logistic regression	formula, data
smf.mnlogit(formula, data)	Multinomial logistic	formula, data
smf.mixedlm(formula, data, groups)	Linear mixed-effects	formula, data, groups, re_formula
smf.gee(formula, groups, data, family)	GEE models	formula, groups, data, family
sm.OLS(y, X)	OLS (matrix interface)	y (array), X (array with constant)
sm.Logit(y, X)	Logistic (matrix)	y (binary), X (array)
sm.MNLogit(y, X)	Multinomial logistic (matrix)	y (coded), X (array)
OrderedModel(y, X, distr)	Ordinal logistic	y (codes), X (array), distr='logit'

statsmodels diagnostics

Function	Purpose	Returns
het_breuschpagan(resid, exog)	Heteroscedasticity test	(LM, p, F, Fp)
durbin_watson(resid)	Autocorrelation test	DW statistic
variance_inflation_factor(X, i)	Multicollinearity	VIF value
shapiro(resid)	Normality test	(W, p)

lifelines (Survival analysis)

Class/Function	Purpose	Key Parameters
CoxPHFitter()	Cox PH model	.fit(df, duration_col, event_col)
KaplanMeierFitter()	KM estimation	.fit(durations, event_observed)
logrank_test(T1, T2, E1, E2)	Log-rank test	durations and events for 2 groups
NelsonAalenFitter()	Cumulative hazard	.fit(durations, event_observed)

scipy.stats (Statistical tests)

Function	Purpose	Returns
ttest_ind(a, b)	Independent t-test	(t, p)
ttest_rel(a, b)	Paired t-test	(t, p)
mannwhitneyu(a, b)	Mann-Whitney U	(U, p)
chi2_contingency(table)	Chi-square test	(chi2, p, dof, expected)
fisher_exact(table)	Fisher's exact test	(OR, p)
f_oneway(*groups)	One-way ANOVA	(F, p)
kruskal(*groups)	Kruskal-Wallis	(H, p)
wilcoxon(a, b)	Wilcoxon signed-rank	(W, p)
shapiro(data)	Shapiro-Wilk normality	(W, p)

scikit-learn (Supplementary)

Class	Purpose	Key Methods
LogisticRegression(multi_class='multinomial')	Multinomial logistic	.fit(X, y), .predict_proba(X)
StandardScaler()	Feature scaling	.fit_transform(X)
LabelEncoder()	Label encoding	.fit_transform(y)

---

ToolUniverse Integration Tools (Data Retrieval)

These ToolUniverse tools can be used to retrieve data before modeling:

Clinical Trial Data

Tool	Parameters	Returns
ClinicalTrials_search_studies	action="search_studies", condition, intervention, limit	{total_count, studies}
get_clinical_trial_eligibility_criteria	nct_ids (array), eligibility_criteria="all"	[{NCT ID, eligibility_criteria}]

Drug Safety / Adverse Events

Tool	Parameters	Returns
FAERS_calculate_disproportionality	drug_name, adverse_event	{metrics: {PRR, ROR, IC}, signal_detection}
FAERS_stratify_by_demographics	drug_name, adverse_event, stratify_by	Stratified counts
FAERS_count_patient_reaction	medicinalproduct	[{term, count}]

Gene-Disease Evidence

Tool	Parameters	Returns
OpenTargets_target_disease_evidence	ensemblId, efoId	Evidence scores
OpenTargets_get_associated_targets_by_disease_efoId	efoId, size	{data: {disease: {associatedTargets}}}

Literature

Tool	Parameters	Returns
PubMed_search_articles	query, max_results	List of article dicts

---

Model Selection Decision Tree

Outcome Type?
|
|-- Continuous (numeric, wide range)
|   |-- Independent observations -> OLS (smf.ols)
|   |-- Clustered/repeated -> LMM (smf.mixedlm)
|
|-- Binary (0/1, yes/no)
|   |-- Independent observations -> Logistic (smf.logit)
|   |-- Clustered/repeated -> GEE Logistic (smf.gee)
|
|-- Ordinal (ordered categories, 3+ levels)
|   |-- Independent observations -> OrderedModel (distr='logit')
|   |-- If proportional odds violated -> Multinomial logistic
|
|-- Nominal (unordered categories, 3+ levels)
|   |-- -> Multinomial logistic (sm.MNLogit)
|
|-- Time-to-event (survival)
|   |-- With covariates -> Cox PH (CoxPHFitter)
|   |-- Descriptive / curves -> Kaplan-Meier (KaplanMeierFitter)
|   |-- Group comparison -> Log-rank test
|
|-- Count (integer, events)
|   |-- -> Poisson or Negative Binomial (smf.poisson / smf.negativebinomial)

---

Common Pitfalls

1. OrderedModel coefficient sign: In statsmodels OrderedModel, positive coefficient = higher odds of being in HIGHER category (same direction as R's polr with method="logistic")
2. Formula syntax: Use C(variable) for categorical variables in formulas, : for interaction, * for main effects + interaction
3. Reference levels: First level alphabetically is default reference. Use C(var, Treatment(reference='level')) to set explicitly
4. Convergence: OrderedModel may need method='bfgs' and maxiter=200+; logistic regression use maxiter=100 and disp=0
5. Missing data: statsmodels formula API drops NA rows automatically; matrix API does not
6. Odds ratio direction: OR > 1 = increased odds, OR < 1 = decreased odds, OR = 1 = no effect
7. Hazard ratio direction: HR > 1 = increased hazard (worse survival), HR < 1 = decreased hazard (better survival)