You are here: Vanderbilt Biostatistics Wiki>Main Web>Projects>DivisOfGenIntMedAndPubHealth>PillCVD>PillCvdStatisticalAnalysis>PillCvdPreviousVersionOfStatAnalysis (09 Apr 2008, SvetlanaEden)EditAttach

- Number of
`preventable/ameliorable ADEs`

(see PILL-CVD definitions) per patient during the first 30 days after hospital discharge . - Number of
`potential ADEs`

(see PILL-CVD definitions) per patient during the first 30 days after hospital discharge .

- Number of Prescribed Medications
- Number of Medication Changes During Hospitalization
- Level of health literacy
- Investigation site
- Gender
- Race
- Age
- Primary language
- Social support
- Educational attainment
- Cognitive function

- Disease-Specific Quality of Life
- Disease Control
- Health Care Utilization

- Level of health literacy
- Cognitive function
- Educational attainment
- Social support
- Primary language
- Quality of life
- Disease control
- Gender
- Race
- Age
- Number of prescribed medications

- Number of
`preventable/ameliorable serious ADEs`

per subject - Number of
`non-preventable serious ADEs`

ADE per subject - Number of adverse events per subject (if collected)
- Number of medication changes during hospitalization
- Number of
`serious ADEs`

by severity

`preventable/ameliorable serious ADE`

is tested by introducing an interaction term (intervention with literacy level). - We use logistic regression to assess the association between the presence of
`preventable/ameliorable serious ADE`

and intervention. The model controls for gender the confounders listed under Primary Outcomes. - Logistic regression is used to assess the association between the presence of
`potential ADE`

and intervention. The model controls for the confounders listed under Primary Outcomes. - Poisson regression (or proportional odds ordinal logistic regression) is used to assess the association between the number of
`preventable/ameliorable ADEs`

and intervention. The model controls for the confounders listed under Primary Outcomes.

- The presence of at least one
`preventable`

serious medication error per patient during the first 30 days after hospital discharge. Note, serious medication error consists of three different errors. The presence of each error is considered separately. - Number of
`preventable`

serious medication errors per patient during the first 30 days after hospital discharge.

- Investigation site
- Gender
- Race
- Primary language
- Social support
- Educational attainment
- Cognitive function
- Level of health literacy
- Number of medications
- Study site

- Quality of life
- Disease control
- Gender
- Race
- Primary language
- Social support
- Educational attainment
- Cognitive function
- Level of health literacy
- Number of prescribed medications
- Number of
`preventable`

ADE per subject - Number of
`non-preventable`

ADE per subject - Number of adverse events per subject

- We use logistic regression to assess the association between the presence of each
`preventable`

ADE and intervention. The model controls for gender the confounders listed under Primary Outcomes. - Poisson regression (?) is used to to assess the association between the number of
`preventable`

ADE and intervention. The model controls for the confounders listed under Primary Outcomes. - The effect of health literacy level on number of (or at least one ?) ADE is tested by introducing an interaction term (intervention with literacy level) in Poisson (or logistic ?) regression model controlling for the confounders listed under Primary Outcomes.

The analytic plan is designed to quantify the efficacy of the pharmacist intervention compared to usual care in term of the key hypotheses. In addition, the analysis is designed to adjust for confounding due to incomplete randomization, clustering of effects within sites, loss to follow-up, and effect modification. Because the primary outcome is the presence of at least one serious medication error per patient, hypothesis 1 is a comparison of two proportions (intervention versus control) that will be tested with a chi square test. Next, univariable analyses will be performed for each candidate predictor using Fisher's exact test (for categorical predictors with small expected cell counts), chi square test(cor categorical predictors with large expected cell counts), or univariable logistic regression (for continuous predictors). Candidate predictors include age, race, primary language, social support, educational attainment, cognitive function, level of health literacy, number of medications, and study site. We will then control for potential confounders using multivariable logistic regression, including in the models all significant predictors of medication errors at a significance level of p < 0.1 from univariable testing. Any potential confounders that change the effect estimates for the intervention covariates by more that 10% will be retained as part of the final models. Each component of serious medication errors (i.e. preventable, ameliorable and potential ADEs) will be analyzed in a similar manner. For the secondary outcome of number of serious medication errors per patient, Poisson regression will be used, with the number of medications ordered at discharge as an obligatory covariate in the model. The effect of the intervention on other outcomes will be examined using t-tests (KCCQ, SAQ) or chi-square test (weight, blood pressure, health care utilization). To address hypothesis 2, subgroup analyses will be performed to determine if the intervention is more effective in patients with inadequate literacy skills (as measure by the s-TOFHLA) that in patients with marginal of adequate literacy skills. Effect modification will be determined in two ways: 1) using subgroup analyses in univariable and multivariable analyses (e.g., effect of interventions on patients of inadequate vs. those with marginal adequate literacy), and 2) using an interaction term in multivariable models (e.g., intervention*literacy as a variable in a model predicting serious medication errors). The first method is more iterpretable, while the second method provides a way to assess effect modification statistically, while controlling for multiple potential confounders. Because outcomes may be influenced by medical team and study site, we will adjust for these in our model as well. However, because randomization occurs at the patient level, these effects should not alter our statistical power. We will also perform subgroup analyses by site to identify any specific effects. All analyses will be performed according to the intention to treat principle. Two-sided p-values < 0.05 will be considered statistically significant.

Edit | Attach | Print version | History: r6 < r5 < r4 < r3 | Backlinks | View wiki text | Edit wiki text | More topic actions

Topic revision: r6 - 09 Apr 2008, SvetlanaEden

Copyright © 2013-2022 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.

Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback

Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback