Skip to main content

Table 4 Advantages and limitations to methods for modeling repeated biomarkers of exposure in association with a binary, non-time-varying outcome

From: Statistical methods for modeling repeated measures of maternal environmental exposure biomarkers during pregnancy in association with preterm birth

 

Advantages

Limitations

Method

- Simple implementation

- Collinearity in longitudinal phthalate measures can cause instable effect estimates and inflated variance estimates

Multiple logistic regression model

- Jointly account for longitudinal phthalate measures in one model

- Requires time points to be uniform

  

- Only the subjects with complete data are used

  

- Difficult interpretation

Method

- Simple implementation

- No straightforward way to combine results from multiple regression models to assess aggregate effect of phthalate levels on preterm birth

Parallel cross-sectional logistic regression models

- Subjects with incomplete data can be retained

- Control for family-wise error rate using Bonferroni correction may be too conservative

 

- Simple interpretation

 

Method

- Simple implementation

- Difficult to handle time-varying covariates

Model using mean exposure across visits as a summary

- Simple way to account for and summarize longitudinal phthalate measures

- Limited if data are unbalanced and/or not missing at random

 

- Straightforward interpretation

- Trends of phthalate measures relevant to the outcome may be missed

 

- Improved power when exposure has poor stability over time and exposure levels themselves are most relevant to the outcome

 

Method

- Simple implementation

- May be inappropriate when maximum concentrations are indicative of recent rather than acute exposure

Model Using maximum exposure value across visits as summary

- Straightforward interpretation

- Deposition of time-varying covariates is questionable

 

- Powerful when the association is not driven by the longitudinal trend and/or average level but rather an acute instance of phthalate exposure

 

Method

- Flexible modeling of exposure pattern over time in Stage 1

 

Two stage mixed effects model

- Examines effect of characteristics carried from Stage 1 in Stage 2

- Uncertainty from Stage 1 is not incorporated in Stage 2 which may lead to biased results

 

- Naturally accounts for between subject heterogeneity

- May not be useful when phthalate levels are unstable over time

Method

- Accounts for longitudinal nature of exposure

- Not temporally logical

Generalized additive mixed model to contrast exposure trajectories

- Trends of exposure can be depicted parametrically or non-parametrically for each group

- Risk cannot be estimated

Method

- Allows risk estimation based on cluster identity

- Requires dataset to be balanced and complete

Gaussian mixture model by clustering the exposure values

- Characteristics of each cluster well-depicted by a multivariate Gaussian distribution

- Requires longitudinal phthalate measures to follow a multivariate Gaussian distribution

 

- Direct interpretation

- Subtle characteristics cannot be captured by the first two moments

  

- Computationally expensive

Method

- Accounts for longitudinal nature of the exposure and time-varying covariates

- May be underpowered if trends of phthalate levels are unimportant

Functional clustering model

- Allows risk estimation based on cluster identity

- Trends are unreliable if the data are sparse with (few time points for each subject)

 

- Does not require exposure to be balance and complete

 
 

- Direct interpretation

 

Method

- Accounts for longitudinal nature of the exposure and time-varying covariates

- Difficult interpretation

Functional logistic regression model

- Does not require exposure to be balanced and complete

- Trends are unreliable if the data are sparse (few time points for each subject)

 

- Longitudinal information is entirely retained in FPC scores

- Choice of number of principal and number of basis function via BIC is ad-hoc