README.md

---
configs:
- config_name: filtered_incidence
  data_files:
  - split: train
    path: "data/incidence_train.parquet"
  - split: test
    path: "data/incidence_validation.parquet" 
- config_name: filtered_termination
  data_files:
  - split: train
    path: "data/termination_train.parquet"
  - split: test
    path: "data/termination_validation.parquet"

license: apache-2.0
tags:
- Actuarial Science
- LTC Products
- Underwriting
- Insurance
- Categorical Variable
- Panel Data
- SOA(LTC)
size_categories:
- 100M<n<1B
language:
- en
task_categories:
- feature-extraction
- summarization
---


# Project 1 Proposal of the Long Term Care(LTC) Aggregated Dataset 

KAO, HSUAN-CHEN(Justin)  <br> 
NetID: hk310

## Dataset Details

The long-term care aggregated dataset, essential for conducting experience studies, is an extensive and valuable compilation of variables central to the analysis and prediction of long-term care (LTC) insurance products. This dataset integrates two critical files: one detailing claim incidence and the other capturing policy terminations. This merger is crucial for valuation purposes, enabling a holistic view of the insurance lifecycle.

The dataset features a broad spectrum of variables, from demographic information such as `Gender`, `Issue_Age_Bucket`, and `Marital_Status`, to more nuanced policy-specific details including `Premium_Class`, `Underwriting_Type`, and `Coverage_Type_Bucket`. Additionally, the termination component enriches the dataset with variables like `Claim_Type`, `Region`, `Diagnosis_Category`, `Claim_Duration`, `Exposure`, `Benefit_Expiry`, `Deaths`, `Recovery`, `Terminations`, and `Others_Terminations`. These elements offer insights into the reasons for policy cessation, whether due to the policyholder's death, recovery from the condition leading to the claim, or other factors leading to the discontinuation of coverage.

By combining the incidence and termination datasets, insurers gain a comprehensive understanding of both the initiation and conclusion of LTC insurance policies. This complete perspective is vital for actuaries and analysts to assess risk, set appropriate reserves, design tailored products, and determine pricing strategies that reflect the true cost of providing LTC coverage. It also aids in regulatory compliance and ensures that insurance products are both financially viable for the provider and beneficial for the consumer. Through careful analysis of this aggregated data, insurers can predict trends, modify underwriting practices, and make informed decisions to manage their portfolios effectively.


### Dataset Sources

<!-- Provide the basic links for the dataset. -->
Please copy and paste the URL; do not click the link directly. The SOA has blocked hyperlinks from external websites.
- **Repository:** SOA Actuarial Practice Modules [ https://www.soa.org/sections/long-term-care/long-term-care-resources/ ]
- **Paper Link:** Long‐Term Care Intercompany Experience
Study [ https://www.soa.org/4a31a6/globalassets/assets/files/resources/experience-studies/2020/2000-2016-ltc-aggregate-database-report.pdf ]
- **Data Process:** Data_Cleaning_Process.ipynb [ https://github.com/justinkao44/STA663_Project_1 ]


## Uses

The long-term care (LTC) aggregated dataset is intended to serve as a foundational tool for actuaries and data scientists aiming to conduct comprehensive experience studies within the insurance sector. Experience studies are essential for understanding the past performance and projecting the future trends of insurance products, and this dataset provides the necessary detailed information to perform such analyses.

Traditional Framework(Used mostly in the insurance companies now): <br>
One of the primary uses of this dataset is to apply the chain ladder method, a traditional actuarial technique used to predict future claim costs and the reserve amounts needed to cover these potential claims. The chain ladder method relies on the assumption that historical claims development can predict future claims development, making it possible to estimate the reserves required for incurred but not reported (IBNR) claims as well as incurred but not enough reserved (IBNER) claims.

Bayesian Framework(Innovative methods recommended): <br>
The dataset can be utilized within a Bayesian framework to enhance the predictive modeling process. A Bayesian approach allows for the incorporation of prior knowledge or expert opinion into the statistical models, updating these beliefs with data from the dataset to generate a posterior distribution of the expected claims. This method is particularly useful when dealing with complex systems or when the available data is sparse or contains a high level of uncertainty.

1. Regarding this LTC aggregated dataset, Bayesian hierarchical models can be applied to account for multiple levels of variability, such as between different policyholders, across various regions, and over time. These models can also help in understanding the effects of policy features and policyholder characteristics on the likelihood and timing of claims, providing a deeper insight into risk factors.

2. Another significant use of the dataset in the Bayesian framework is the development of predictive distributions for various risk metrics. These could include the probability of claim terminations due to death or recovery, the expected number of claims within certain diagnosis categories, or the expected claim durations.



## Dataset Structure and Instances

### Incidence Dataset Structure Description

1. LTC Claim Incidence
- **Group_Indicator**: Categorical variable indicating the group to which the policyholder belongs.
- **Gender**: Categorical variable indicating the gender of the policyholder.
- **Issue_Age_Bucket**: Categorical variable denoting the age range when the policy was issued.
- **Incurred_Age_Bucket**: Categorical variable representing the age range when the claim was incurred.
- **Issue_Year_Bucket**: Categorical variable indicating the year range when the policy was issued.
- **Policy_Year**: Categorical variable indicating the specific year of the policy.
- **Marital_Status**: Categorical variable indicating the marital status of the policyholder.
- **Premium_Class**: Categorical variable indicating the classification of the premium.
- **Underwriting_Type**: Categorical variable indicating the type of underwriting applied to the policy.
- **Coverage_Type_Bucket**: Categorical variable indicating the coverage type category.
- **Tax_Qualification_Status**: Categorical variable indicating the tax qualification status of the policy.
- **Inflation_Rider**: Categorical variable indicating whether an inflation protection rider is attached to the policy.
- **Rate_Increase_Flag**: Categorical variable indicating if there has been a rate increase on the policy.
- **Restoration_of_Benefits**: Categorical variable indicating whether benefits have been restored.
- **NH_Orig_Daily_Ben_Bucket**: Categorical variable indicating the original daily benefit amount for nursing home care.
- **ALF_Orig_Daily_Ben_Bucket**: Categorical variable indicating the original daily benefit amount for assisted living facilities.
- **HHC_Orig_Daily_Ben_Bucket**: Categorical variable indicating the original daily benefit amount for home health care.
- **NH_Ben_Period_Bucket**: Categorical variable indicating the benefit period for nursing home care.
- **ALF_Ben_Period_Bucket**: Categorical variable indicating the benefit period for assisted living facilities.
- **HHC_Ben_Period_Bucket**: Categorical variable indicating the benefit period for home health care.
- **NH_EP_Bucket**: Categorical variable indicating the elimination period for nursing home care.
- **ALF_EP_Bucket**: Categorical variable indicating the elimination period for assisted living facilities.
- **HHC_EP_Bucket**: Categorical variable indicating the elimination period for home health care.
- **Region**: Categorical variable indicating the geographical region of the policy.
- **Active_Exposure**: Numeric variable indicating the active exposure amount.
- **Total_Exposure**: Numeric variable indicating the total exposure amount.
- **Claim_Count**: Numeric variable indicating the count of claims.
- **Count_NH**: Numeric variable indicating the count of nursing home claims.
- **Count_ALF**: Numeric variable indicating the count of assisted living facility claims.
- **Count_HHC**: Numeric variable indicating the count of home health care claims.
- **Count_Unk**: Numeric variable indicating the count of claims with unknown categorization.

 
2.  filtered_incidence instances

```
{'Group_Indicator': 'Individual',
 'Gender': 'Female',
 'Issue_Age_Bucket': '50-54',
 'Incurred_Age_Bucket': '60-64',
 'Issue_Year_Bucket': '2003-2005',
 'Policy_Year': ' 7-9 years',
 'Marital_Status': 'Single',
 'Premium_Class': 'Preferred',
 'Underwriting_Type': 'Other',
 'Coverage_Type_Bucket': 'Comprehensive',
 'Tax_Qualification_Status': 'Tax-qualified',
 'Inflation_Rider': 'Inflation protection',
 'Rate_Increase_Flag': 'Yes',
 'Restoration_of_Benefits': 'No',
 'NH_Orig_Daily_Ben_Bucket': ' < 100',
 'ALF_Orig_Daily_Ben_Bucket': ' < 100',
 'HHC_Orig_Daily_Ben_Bucket': ' < 100',
 'NH_Ben_Period_Bucket': 'Unknown',
 'ALF_Ben_Period_Bucket': 'Unknown',
 'HHC_Ben_Period_Bucket': 'Unknown',
 'NH_EP_Bucket': '  90',
 'ALF_EP_Bucket': '  90',
 'HHC_EP_Bucket': '   0',
 'Region': '04: West',
 'Active_Exposure': 19.583332,
 'Total_Exposure': 19.583332,
 'Claim_Count': 0,
 'Count_NH': 0,
 'Count_ALF': 0,
 'Count_HHC': 0,
 'Count_Unk': 0}
```

### Termination Dataset Structure Description

1. LTC Claim Termination

- **Gender**: Categorical variable indicating the gender of the policyholder.
- **Incurred_Age_Bucket**: Categorical variable denoting the age range when the claim was incurred.
- **Incurred_Year_Bucket**: Categorical variable indicating the year range when the claim was incurred.
- **Claim_Type**: Categorical variable indicating the type of claim made.
- **Region**: Categorical variable indicating the geographical region of the policy.
- **Diagnosis_Category**: Categorical variable providing the category of diagnosis related to the claim.
- **Claim_Duration**: Categorical variable representing the duration that the claim has been active.
- **Exposure**: Numeric variable indicating the measure of risk that the insurer has been exposed to for the policy.
- **Deaths**: Numeric variable indicating the number of deaths among the policyholders.
- **Recovery**: Numeric variable indicating the number of policyholders who have recovered.
- **Terminations**: Numeric variable indicating the number of policy terminations.
- **Benefit_Expiry**: Numeric variable indicating the number of terminations due to benefits reaching their expiration date.
- **Others_Terminations**: Numeric variable indicating the number of terminations due to other unspecified reasons.

2. filtered_termination instances

```
{'Gender': 'Male',
 'Incurred_Age_Bucket': '75-79',
 'Incurred_Year_Bucket': '2001-2002',
 'Claim_Type': 'HCC',
 'Region': '03: South',
 'Diagnosis_Category': '08: Other',
 'Claim_Duration': '56',
 'Exposure': '3',
 'Deaths': '0',
 'Recovery': '0',
 'Terminations': '0',
 'Benefit_Expiry': '0',
 'Others_Terminations': '0'}
```


## Dataset Creation

### Curation Rationale

The creation of this dataset stems from the increased popularity of long-term care (LTC) products amid rising longevity rates. It aims to provide insights into the trends and patterns of the claim filing process. The expansion of the LTC product line necessitates that insurance companies establish more accurate reserves to ensure financial health and sustainability. This dataset is pivotal in laying the groundwork for such analysis.

From my experience as a data analysis and administration intern at RGA, I understand the difficulties that arise from inadequate reserve practices and the lack of precise predictions in claim filings. These challenges underscore the need for a robust dataset that allows for comprehensive research into these critical areas. By leveraging this dataset, we can enhance the precision of actuarial predictions and reserve estimations, thereby contributing to the stability and efficiency of LTC insurance operations.
### Source Data

The dataset has been meticulously compiled by the Society of Actuaries, drawing from a wealth of information provided by various insurance companies within the industry. This collaboration ensures a rich and diverse collection of data, encompassing a wide range of long-term care insurance policies, claims, and termination details.


#### Who are the source data producers?

- Society of Actuaries(SOA)
- American Academy of Actuaries(AAA)
- Canadian Institute of Actuaries(CIA)


#### Personal and Sensitive Information

In aligning with HIPAA regulations to safeguard personal and sensitive information, I have taken significant steps to ensure the dataset is devoid of data that could be considered private. As a member of the Society of Actuaries and having access to a wealth of personal data, I was keenly aware of the necessity to adhere strictly to HIPAA's guidelines, including the Privacy Rule, Security Rule, and Breach Notification Rule. Therefore, I personally undertook the anonymization process, meticulously eliminating or modifying any information that could potentially lead to the identification of individuals, whether directly or indirectly. This includes removing details such as addresses, names, and any demographic or health-related information that could compromise an individual's privacy.

My goal in the anonymization process was to minimize the loss of valuable information to maintain the dataset's utility for research and analytical purposes. This careful balance ensures that the data can be safely used by vetted users, facilitating the advancement of understanding and management within the long-term care insurance sector, without infringing on individual privacy rights.

I encourage all users of this dataset to notify me directly if they encounter any data or information that could potentially lead to privacy concerns. My commitment to privacy and data protection is paramount, and I am dedicated to taking prompt action to address and rectify any issues that may arise.


## Bias, Risks, and Limitations

### Bias

- **Selection Bias**: The dataset primarily sourced from insurance companies participating with the Society of Actuaries may not represent all types of LTC insurance providers, particularly smaller or regional companies that might have different claim patterns or policyholder demographics.
- **Reporting Bias**: There might be inconsistencies in how data is reported across different companies, affecting the uniformity and comparability of the information. For instance, the threshold for defining a claim or the categorization of terminations could vary, leading to potential biases in analysis.
- **Survivorship Bias**: The data might inherently focus more on policies that have led to claims or terminations, overlooking the ones that are still active without incidents. This could skew the understanding of risk and the overall policyholder experience.

### Risks

- **Privacy Risk**: Despite the efforts to anonymize the data, there is always a risk, however minimal, that individuals could be re-identified through a combination of variables, especially in smaller, more homogeneous datasets.
- **Misinterpretation**: Without a deep understanding of the intricacies of insurance data and the specific context of LTC products, there's a risk that the data could be misinterpreted, leading to inaccurate conclusions or policy decisions.
- **Missing Data Risk**: Given that some data points are missing, analyses could be biased if the missing data is not random. This can lead to inaccurate conclusions.

### Limitations

- **Temporal Relevance**: The dataset, while comprehensive, captures data up to a certain point in time. The rapidly evolving nature of the LTC insurance market and changing demographics might limit the relevance of historical data for predicting future trends.
- **Complexity of LTC Products**: LTC insurance products are complex, and the dataset might not fully capture the nuances of different policy features, benefit triggers, and underwriting practices, which could affect the analysis.

### Recommendations

[TBD]



## Dataset Card Contact

[justinkao.44@duke.edu]