Model | Analysis Sample | Retest Correction Method** | Additional Specifications to Basic Model* | Strengths | Weaknesses |
---|---|---|---|---|---|
1: No correction | SAGES surgery sample (delirium-positive and delirium-negative) N = 560 | No correction for retest effect | Main effect: delirium group Interactions: time × delirium group Outcome: raw GCP | -Does not make manipulations to observed data -Does not require a control group | -Difficult to separate retest effects from effects due to delirium/surgery |
2: Mean difference correction | SAGES surgery sample (delirium-positive and delirium-negative) N = 560 | Step 1: Calculate mean retest effect in NSC group: \( R{(NSC)}_t=\frac{1}{N}\sum \limits_i GCP{(NSC)}_{it}- GCP{(NSC)}_{i0}; \) t ∈ (1, 2, 6), i ∈ 1, 2, …, N) Step 2: Calculate retest-corrected scores in surgical group: GCPRCit = GCPit − R(NSC)t; where R(NSC)6 = RSC(NSC)12 = RSC(NSC)18 | Main effect: delirium group Interactions: time × delirium group Outcome: mean difference corrected GCP | -Straightforward, constant (within occasion) transformation for all people -Correction is uncorrelated with any other variable at each time point | -Potential variability due to the fact that precision of retest correction estimation is not accounted for |
3: Predicted difference correction | SAGES surgery sample (delirium-positive and delirium-negative), N = 560 | Step 1: Estimate linear regression equations in NSC group expressing the dependence of GCP score at follow-up on baseline GCP, centered at the overall mean of the GCP at baseline in the NSC group; The regression is estimated separately for each observation time point (follow-up month) t ∈ (1,2,6): \( \kern0.75em GCP{(NSC)}_{it}={b}_{0t}+{b}_{1t}\left( GCP{(NSC)}_{i0}-{\overline{GCP(NSC)}}_0\right)+{e}_t \) Step 2: Compute expected GCP performance given the NSC group regression equations: \( E{(GCP)}_{it}={b}_{0t}+{b}_{1t}\left({GCP}_{i0}-{\overline{GCP(NSC)}}_0\right) \) Step 3: Calculate retest effect in surgical (SRG) group: R(SRG)it = GCPit − E(GCP)it; t ∈ (1, 2, 6) Step 4: Calculate retest-corrected scores in surgical group: GCPRCit = GCPit − R(SRG)t; t ∈ (1, 2, 6,12,18) | Main effect: delirium group Interactions: time × delirium group Outcome: predicted difference corrected GCP | -Retest effects will be more appropriately modeled if a baseline variable is known to predict retest effects | -The degree of correction could differ by group if predictors of the retest effect differ by group -Inconsistent literature on variables predicting retest effect -Variability due to precision of retest correction estimation is not accounted for in the model |
4: Model-based correction | SAGES surgery sample (delirium-positive and delirium-negative), N = 560 & non-surgical comparison (NSC) sample, N = 118 | Raw GCP is utilized for both the NSC and surgical groups, but this method differs from Model 1 (no correction) in that NSC data is modeled as the comparison group for both the delirium+ and delirium- groups. Relative differences between the delirium+ and delirium- groups are then calculated with post hoc tests. | Main effect: NSC, delirium group Interactions: time × group Outcome: raw GCP | -Standard errors are appropriately conservative -Can model differences compared to another surgical group and a NSC sample | -Model estimates are reported in reference to the NSC group -Potential analyses are restricted to variables observable in the NSC group -Maintaining a NSC group is expensive and may introduce additional bias over time due to drop out or latent differences between the NSC and surgical groups. |