[Paper Review] Case-control studies for rare diseases: improved estimation of several risks and of feature dependences

Nanny Wermuth, Giovanni M. Marchetti|arXiv (Cornell University)|Mar 8, 2012

Gene expression and cancer classification29 references3 citations

TL;DR

This paper proposes a novel case-control analysis method that improves risk estimation for rare diseases by separately modeling dependence structures in cases and controls, then comparing them to identify key feature combinations and refine odds-ratio estimates. By leveraging distinct regressor dependencies in each group, the approach enhances model accuracy and reveals disease-specific risk patterns beyond standard logistic regression.

ABSTRACT

To capture the dependences of a disease on several risk factors, a challenge is to combine model-based estimation with evidence-based arguments. Standard case-control methods allow estimation of the dependences of a rare disease on several regressors via logistic regressions. For case-control studies, the sampling design leads to samples from two different populations and for the set of regressors in every logistic regression, these samples are then mixed and taken as given observations. But, it is the differences in independence structures of regressors for cases and for controls that can improve logistic regression estimates and guide us to the important feature dependences that are specific to the diseased. A case-control study on laryngeal cancer is used as illustration.

Motivation & Objective

• Address the limitation of standard logistic regression in case-control studies, which treats mixed samples from cases and controls as a single population, potentially distorting dependence structures.
• Overcome the risk of model-based estimates being extrapolated beyond data support by integrating evidence-based data summaries with modeling.
• Identify critical feature combinations that distinguish cases from controls by exploiting differences in regressor dependence structures between the two groups.
• Improve estimation of odds-ratios and risk profiles for rare diseases by using separate graphical models for cases and controls.
• Supplement traditional logistic regression with direct comparisons of dependence structures to detect effects significant in one group but not the other.

Proposed method

• Define categorical variables to create cross-classifications with manageable cell counts, ensuring comparability between cases and controls on key features.
• Fit separate regression graphs (using concentration graphs) to the regressor variables in the case group and the control group independently.
• Use the cliques from the separate graphs to derive well-fitting logit regression models for each group, capturing distinct dependence structures.
• Estimate expected counts for each cell in the cross-classification using models fitted separately to cases and controls, based on the well-fitting graphs.
• Combine observed and estimated counts from both groups to compute mixed-count odds-ratios and assess differential feature accumulation.
• Apply goodness-of-fit tests based on saturated models derived from initial data processing to validate the separate models for cases and controls.

Experimental results

Research questions

• RQ1 How can dependence structures among risk factors differ between cases and controls in a case-control study of a rare disease?
• RQ2 What improvements in odds-ratio estimation can be achieved by modeling cases and controls separately rather than pooling them?
• RQ3 Which combinations of risk factors show significant differences in their joint distribution between cases and controls, indicating disease-specific associations?
• RQ4 Can separate graphical models for cases and controls reveal important feature dependencies that are obscured in pooled logistic regression?
• RQ5 To what extent can evidence-based data summaries from separate case and control analyses improve model interpretation and risk profile identification?

Key findings

• Separate modeling of regressor dependence structures in cases and controls revealed distinct patterns not visible in pooled logistic regression, particularly for high-risk combinations of tobacco and alcohol use.
• For the cell with levels (L=0, V=0, C=0, R=0, A=0, E=0), the estimated odds-ratio was 4.7, derived from estimated counts of 23.79 (controls) and 2.85 (cases), indicating a strong imbalance in risk factor distribution.
• The estimated count for the (L=1, V=0, C=0, R=0, A=0, E=0) cell was 3.29 in cases and 19.55 in controls, suggesting that even with low exposure levels, the disease group had higher relative risk in certain subgroups.
• The method successfully identified that certain feature combinations, such as high tobacco and alcohol exposure, were more prevalent in cases than in controls, with estimated counts showing consistent deviations from expected under independence.
• The use of separate models for cases and controls led to more accurate and interpretable risk estimates, particularly in sparse data regions where pooled models may fail.
• The approach allowed detection of statistically significant effects in one group (e.g., controls) that were not apparent in the pooled analysis, highlighting the importance of group-specific dependence structures.