Skip to Content
Interactive Textbook on Clinical Symptom Research Logo


Home Button

Statistical Models for Prognostication
Author Bio
Introduction
Currently selected section: Predictions: Statistical Models
Insight: Statistical Models
Ingredients: Statistical Models
Theoretical Aspects
Central Concepts
Regression Models
Problems: Regression
Practical Advice
Example 1
Example 2




Chapter 8: Statistical Models for Prognostication: Predictions from Statistical Models
        

Analysis of Clinical Trials

Statistical models may be beneficial in the analysis of randomized clinical trials because of:

  • Correction for imbalance between randomized groups
    This issue is, however, under debate among investigators.

    • PRO: Imbalance in prognostic factors may be considered to cause a bias in the interpretation of a differrence between randomized patient groups. Randomization only guarantees comparability between groups in the long run, that is, with infinite sample sizes. In practice, imbalance may arise by pure chance. This imbalance can be corrected by a statistical model. Note that the model should generally be pre-specified before starting the analysis. For example, see the GUSTO-III trial (1997).
    • CON: Imbalance can never be fully corrected for. There may be imbalance for one known factor in the trial, the factor of age, for example, but there may well be imbalance on other unknown factors. Randomization a priori ensures balance in both known and unknown baseline characteristics. A posteriori, chance may have caused imbalance in both known and unknown characteristics, while only imbalance in known characteristics can be corrected for.
  • Estimation of a more individualized treatment effect
    Covariable adjustment is rarely needed for the purpose of adjustment for imbalances in randomized trials. But covariable adjustment reduces residual variation and hence lowers standard errors of treatment effect estimates. In nonlinear models such as the logistic regression or survival analysis, covariable adjustment is the only way to obtain unbiased patient-specific treatment effect estimates. Without adjusting for important prognostic factors, crude treatment effect estimates in randomized trials are biased towards the null.

    An analysis that does not consider existing heterogeneity among patients may hence be considered mis-specified. Important individual prognostic characteristics should be included in the analysis to account for heterogeneity as much as possible. The resulting estimate of the treatment effect may be interpreted as more individualized.

  • A higher statistical power for detection of a true treatment effect In linear regression analysis, the standard error of the treatment effect decreases when individual prognostic characteristics are taken into account. In contrast, in non-linear models the standard error of the treatment effect increases when individual prognostic characteristics are taken into account. However, the change in treatment effect is larger than the increase in standard error, causing a net increase in power.
In non-randomized studies, correction for imbalance in predictors is also an important application of statistical models. However, differences between groups may exist that cannot be corrected for, e.g. if not all predictor variables are known (Kunz and Oxman, 1998).

 

QUESTION 2.3

In randomized clinical trials, balance in prognosis between randomized groups is:

Selection ABy definition found if the randomization procedure was honestly followed.
Selection BExpected beforehand, but may not be found after realization of the trial.

QUESTION 2.4

Which analysis has the highest power to estimate the treatment effect in a randomized controlled trial, where strict balance is found in well-known prognostic factors?

Selection AAn unadjusted analysis of the treatment effect
Selection BAn adjusted analysis of the treatment effect, where prognostic factors are taken into account

Previous Page