2.6 Statistical analyses
If antidepressant use was reported, and plasma concentration was undetectable, the respective concentration was set at half of the LLQ. For amitriptyline and nortriptyline, where also semi-quantitative concentrations below LLQ were available, half of the lowest detectable concentration was taken. Because the number of participants in the various individual antidepressant groups were too low for the analyses, the individual antidepressants subclasses needed to be pooled into two groups: SSRIs and TCAs. Although venlafaxine is a selective serotonin-norepinephrine reuptake inhibitor (SNRI), we analyzed venlafaxine combined with the SSRIs. Since the dose of venlafaxine in our cohort did not exceed 150 milligram (mg), SNRIs mainly inhibit reuptake of serotonin and thus closely resemble SSRIs19.
Relationship between drug exposure and fall risk were analyzed with plasma concentrations expressed on both a continuous and categorical scale. If a participant used two or more antidepressants concomitantly, only the antidepressant with the highest concentration was included in the analyses. First, the continuous plasma concentration levels of the individual antidepressants were analyzed. After that, we standardized the concentrations of the different antidepressants by creating z-scores to be able to combine the continuous concentrations. Second, for each antidepressant the median of the plasma concentration was calculated and the concentrations were reclassified into a binary category (below (which was set as reference) or above the median). Third, to explore the influence of the highest plasma concentrations, the antidepressant plasma concentrations were divided into four categories. The reference category contained the concentrations below the LLQ. After that, the concentrations above LLQ were equally divided into tertiles. For the TCAs, the remaining number of participants per category was too low. Thus the reference category contained the lowest tertile and the concentrations below LLQ and the middle and highest tertile were combined. The same was applied for the SSRI follow-up concentrations. Also, delta concentrations (concentration at follow-up minus concentration at baseline) were calculated, reflecting concentration changes over time. These were reclassified into a binary category, negative (decrease or equal concentration over time) or positive (concentration increase over time)
Baseline characteristics were compared between fallers and non-fallers in antidepressant users. Differences between groups were tested using chi-square test, t-test or a Mann-Whitney U test (categorical and continuous non-normally and normally distributed data, respectively). To study the association between the antidepressant plasma concentration at baseline and falls during follow-up, Cox proportional hazard models were used to calculate hazard ratios (HRs). First, we built a model adjusted for age and gender (model 1). Second, the following covariates were selected as potential confounders based on a Directed Acyclic Graph (DAG) and included in the models if they changed the HR of the association by ≥10% (model 2): region, BMI, smoking, alcohol, pain, depressive symptoms, MMSE, number of medication and estimated glomerular filtration rate (eGFR). If covariates could not be added, due to the lack of 10 fall events per covariate 20, only the unadjusted model is presented. For follow-up visit and delta concentrations, logistic regression models were used to calculate odds ratios (ORs) for the association between concentration and fall risk prior to follow-up visit. The same models as described for the Cox proportional hazard models were applied.
To assess whether antidepressant plasma concentrations were correlated with dosage, we calculated Pearson correlation coefficients. Since there were no signs of multicollinearity (correlation coefficient <0.5) between plasma concentration and dosage, as a possible marker for example for disease severity, was added to model 2 to test whether dosage could be a confounder.
P-values <0.05 were considered statistically significant. Statistical analyses were performed in SPSS (version 28.0, IBM, Armonk, NY, USA).