README.md

---
language:
- en
tags:
- missing-data
- lost-to-follow-up
- retention
- breast-cancer
- sub-saharan-africa
- health-systems
license: cc-by-nc-4.0
pretty_name: SSA Breast Missing Data Patterns (Retention & Incomplete Tests)
task_categories:
- other
size_categories:
- 1K<n<10K
---

# SSA Breast Missing Data Patterns (Synthetic)

## Dataset summary

This module provides a **synthetic missing-data sandbox** for oncology care in African healthcare contexts, focusing on:

- **Realistic loss-to-follow-up (LTFU) and retention** patterns over 0–24 months.
- **Incomplete diagnostic and laboratory test results** (ordered vs completed vs available in records).
- **Non-random missingness** driven by facility type, distance, socioeconomic status (SES), and insurance.

The dataset is anchored in published evidence from:

- The **ABC-DO** sub-Saharan African breast cancer cohort (low LTFU with active tracing).
- Meta-analyses of **HIV ART retention** (60–70% retained at 2–3 years in routine care).
- Surveys of **breast cancer pathology services and management** (AORTIC, BMC Health Serv Res, JCO GO).
- Real-world challenges in SSA breast cancer care (BMJ Open 2021).

All records are **fully synthetic** and intended for methods development and teaching on missing data and retention, not for inference on real facilities.

## Cohort design

### Sample size and populations

- **Total N (baseline patients)**: 6,000.
- **Populations**:
  - `SSA_West`: 1,500
  - `SSA_East`: 1,500
  - `SSA_Central`: 1,000
  - `SSA_Southern`: 1,000
  - `AAW` (African American women): 1,000 (reference/high-resource context)

### Key baseline variables

- `sex`: predominantly `Female` (~96%), with a small proportion of `Male` to allow mixed-sex analyses.
- `age_years`: 18–90 (mean ~52, SD ~10).
- `facility_type`:
  - `Tertiary_urban`
  - `Regional_hospital`
  - `District_hospital`
- `urban_rural`:
  - `Urban`, `Periurban`, `Rural`.
- `distance_km` from facility:
  - Drawn from normal distributions by `urban_rural` (e.g., Urban mean ~5 km, Rural mean ~60 km).
- `ses` (socioeconomic status): `Low`, `Middle`, `High` with higher `Low` fractions in SSA cohorts.
- `insurance_status`: `None`, `National_insurance`, `Private`.

These variables drive **missingness mechanisms** (higher LTFU and test missingness with longer distance, low SES, and lack of insurance).

## Follow-up and retention

Patients are scheduled for visits at months:

- `visit_month`: [0, 3, 6, 9, 12, 18, 24].

Retention targets (probability of still in care at 12 and 24 months) vary by facility type, loosely anchored by ABC-DO and HIV ART literature:

- `Tertiary_urban`: 12m ≈ 0.85, 24m ≈ 0.75.
- `Regional_hospital`: 12m ≈ 0.75, 24m ≈ 0.60.
- `District_hospital`: 12m ≈ 0.65, 24m ≈ 0.50.

In the generator, a per-interval dropout probability is derived from these targets. This dropout risk is then **modulated by**:

- **Distance**: higher risk for patients living >75 km from the facility.
- **SES**: higher risk for `Low` vs `High` SES.
- **Insurance**: higher risk for `None` vs `National_insurance` or `Private`.

At each scheduled month, patients either:

- Remain in care and have a visit record (`visit_attended` True/False).
- Drop out (no further visits recorded).

Retention at 12 and 24 months is validated against the configuration.

## Test ordering and result availability (missingness)

### Baseline tests

For each patient at baseline, the following are simulated with facility-specific probabilities:

- `baseline_pathology_ordered`
- `baseline_pathology_result_available` (e.g., ER/PR/HER2 receptors)
- `baseline_cbc_ordered`
- `baseline_cbc_result_available`
- `baseline_imaging_ordered` (staging imaging)
- `baseline_imaging_result_available`

Approximate patterns by facility type:

- **Pathology receptors**:
  - High ordering and completion in `Tertiary_urban` (~90%+ completed).
  - Lower completion in `Regional_hospital` (~65–70%).
  - Substantial missingness in `District_hospital` (~40–50% completed).

- **CBC and imaging**:
  - More widely available, but still with gradients by facility and context.

### Follow-up labs (CBC)

At each **attended** follow-up visit (months >0):

- `cbc_ordered` is drawn from facility-specific probabilities.
- `cbc_result_available` is drawn conditional on ordering and reflects:
  - Higher completion in tertiary centres.
  - Lower completion in district hospitals.

This yields **visit-level missingness** that depends on both visit attendance and facility/test capacity.

### Non-random missingness

Missingness is deliberately **not MCAR**:

- Baseline pathology results are more often missing in:
  - `District_hospital` and `Regional_hospital` than `Tertiary_urban`.
  - `Low` SES vs `High` SES.
- Follow-up CBC results are more often missing among:
  - Patients with long travel distances.
  - Those without insurance.

The validation script checks for **higher missingness in Low vs High SES** for both baseline pathology and follow-up CBC.

## Files and schema

### Baseline table

Files:

- `missing_data_baseline.parquet`
- `missing_data_baseline.csv`

Columns (per patient):

- Identifiers and demographics:
  - `sample_id`
  - `population`
  - `region`
  - `is_SSA`
  - `sex`
  - `age_years`
- Access and facility characteristics:
  - `facility_type`
  - `urban_rural`
  - `distance_km`
  - `ses`
  - `insurance_status`
- Baseline tests (ordered and result availability):
  - `baseline_pathology_ordered`
  - `baseline_pathology_result_available`
  - `baseline_cbc_ordered`
  - `baseline_cbc_result_available`
  - `baseline_imaging_ordered`
  - `baseline_imaging_result_available`

### Visit-level table

Files:

- `missing_data_visits.parquet`
- `missing_data_visits.csv`

Columns (per scheduled time point while in care):

- Identifiers and baseline covariates:
  - `sample_id`
  - `population`
  - `facility_type`
  - `urban_rural`
  - `ses`
  - `insurance_status`
  - `distance_km`
- Visit information:
  - `visit_month` (0, 3, 6, 9, 12, 18, 24)
  - `visit_attended` (True/False)
- Follow-up CBC:
  - `cbc_ordered`
  - `cbc_result_available`

Patients with early loss to follow-up have **shorter visit histories**, so the visit table is an **unbalanced panel** that mimics real program data.

## Generation

The dataset is generated with:

- `missing_data_patterns/scripts/generate_missing_data.py`

using configuration:

- `missing_data_patterns/configs/missing_data_config.yaml`

and literature inventory:

- `missing_data_patterns/docs/LITERATURE_INVENTORY.csv`

Key steps:

1. **Baseline cohort**: populations, sex, age, facility type, urban/rural, distance, SES, insurance.
2. **Baseline tests**: pathology receptors, CBC, and imaging, with ordering/completion probabilities by facility type.
3. **Visits and retention**: scheduled visits from 0 to 24 months, with facility-specific dropout probabilities tuned to match retention targets and modified by distance/SES/insurance.
4. **Follow-up CBC**: ordering and result availability for each attended visit, by facility type.

## Validation

Validation is performed with:

- `missing_data_patterns/scripts/validate_missing_data.py`

and summarized in:

- `missing_data_patterns/output/validation_report.md`

Checks include:

- **C01–C02**: Baseline sample size and population counts vs configuration.
- **C03**: Retention at 12 and 24 months by facility vs configured targets.
- **C04**: Baseline pathology receptor result availability vs expected rates.
- **C05**: Follow-up CBC result availability vs expected rates for attended visits.
- **C06–C07**: Non-random missingness by SES (Low vs High) for baseline pathology and follow-up CBC.
- **C08**: Overall missingness in key baseline and visit variables.

## Intended use

This dataset is intended for:

- Developing and benchmarking **missing-data methods** (imputation, inverse probability weighting, joint models).
- Exploring **selection bias** introduced by LTFU and incomplete tests.
- Teaching about:
  - How health-system factors (facility, distance, SES, insurance) shape missing data.
  - Differences between MCAR, MAR, and MNAR mechanisms in realistic African oncology settings.

It is **not intended** for:

- Estimating real-world retention or test completion at specific facilities.
- Evaluating individual centres or countries.
- Clinical decision-making.

## Ethical considerations

- All data are **synthetic** and derived from literature-informed parameter ranges.
- Facility and population labels are generic and **must not** be interpreted as real institutions.
- The goal is to enable more robust and equitable analyses under realistic data limitations in African healthcare settings.

## License

- License: **CC BY-NC 4.0**.
- Free for non-commercial research, method development, and education with attribution.

## Citation

If you use this dataset, please cite:

> Electric Sheep Africa. "SSA Breast Missing Data Patterns (Retention & Incomplete Tests, Synthetic)." Hugging Face Datasets.

and relevant literature on retention and pathology services in sub-Saharan Africa (e.g., Foerster et al. 2020, Rosen & Fox 2007, Adesina et al. 2020, Joko-Fru et al. 2021).