• Download Summary

What was the project about?

Factors, such as personal traits, behaviors, or the environment, can affect a person’s risk of getting an illness. Doctors can use risk models, which account for these factors, to predict a person’s chance of getting an illness. The risk models group patients into different levels for certain illnesses, such as high risk or low risk.

Most risk models look at only a small number of factors, which affects how well the models can separate patients into different levels. Combining factors from different studies into a single risk model may improve how well the model works. Researchers can use statistical methods to combine data from different studies. But current methods don’t work when the studies look at different traits or other factors.

In this study, the research team developed a new method for combining data from studies that have information on different risk factors. The new method is called Generalized Meta-Analysis, or GENMETA.

What did the research team do?

The research team developed the GENMETA method to combine risk factor data from different studies. Study data could come from different sources, such as health records or surveys. Studies might also look at different risk factors.

To test GENMETA, the research team first used a computer program to create data that mimic real-world study data. The data had problems that often occur in real-world data. For example, some data may be missing from one study, or data may have been collected at different times. Even with these problems, GENMETA worked well. The team also developed an approach to detect such data problems.

Next, the research team used GENMETA to develop a risk model for breast cancer. They combined data from two studies. One study had many patients but only had data on a few risk factors. The other study had fewer patients but data on more risk factors. The team used the approach to detect data problems.

What were the results?

The risk model for breast cancer that the research team created using GENMETA successfully combined the data from the two different studies. The model used all the risk factors from the two studies.

The research team developed software for others to use GENMETA and shared it online for free.

What were the limits of the project?

GENMETA may not work if studies have different criteria about which patients can join the study. For example, studies may include patients in different age ranges or from different regions.

Future research could expand the method to handle patient differences across studies.

How can people use the results?

Researchers can use the new method to build risk models using data from different studies.

Background

Patient traits, lifestyle, and social and environmental conditions can affect a person’s risk for certain health conditions. Clinicians can use risk models to predict an individual patient’s chance of developing a condition. But to be useful, risk models must separate patients into distinct groups for which the harm-benefit ratio for specific interventions is clinically different. Most studies on risk factors for a given condition use a limited number of factors, which affects how well the model can separate patients into risk groups. Incorporating multiple risk factors from different studies into a single risk model may improve the generalizability and clinical utility of the model. Researchers can use meta-analysis methods to combine data from different studies, but existing methods do not work when the studies use different types of covariates.

Objective

To develop a new method for multivariate meta-analysis using summary-level information across studies with different covariates

Study Design

Methods and Results

The research team developed a new method, called Generalized Meta-Analysis (GENMETA), for estimating parameters of a multiple regression model through meta-analysis of studies with summary-level information that use different covariates. To verify the effectiveness of the new method, the team used a series of simulation studies to show that the method is robust even when underlying assumptions are violated. They also proposed a model diagnostic test to detect violations of model assumptions due to heterogeneity.

Next, the research team used GENMETA to develop models for predicting the risk of breast cancer by combining information from the Breast Prostate Colorectal Cancer Cohort study and the Breast Cancer Detection and Demonstration Project studies, which include different risk factors. The team first used the model diagnostic test to show that the underlying model assumption was unlikely to be violated. Then the team applied GENMETA to the two studies to carry out the meta-analysis. Results showed that the GENMETA method could produce estimates for all risk factors from the two different studies that were related to breast cancer.

The research team developed a software package to implement GENMETA and shared it online at no cost to users.

Limitations

The GENMETA method assumes that the populations across different studies are similar. This assumption may not hold if studies use different study designs, such as sampling methods.

Conclusions and Relevance

The new method can help researchers conduct meta-analyses using results from studies based on multiple, disparate data sources and with different covariates.

Future Research Needs

Future research could extend GENMETA to account for differences in study designs.