Assessment of neurological and psychiatric disorders like depression are unusual from a speech processing perspective, in that speakers can be prompted or instructed in what they should say (e.g. as part of a clinical assessment). Despite prior speech-based depression studies that have used a variety of speech elicitation methods, there has been little evaluation of the best elicitation mode. One approach to understand this better is to analyze an existing database from the perspective of articulation effort, word affect, and linguistic complexity measures as proxies for depression sub-symptoms (e.g. psychomotor retardation, negative stimulus suppression, cognitive impairment). Here a novel measure for quantifying articulation effort is introduced, and when applied experimentally to the DAIC corpus shows promise for identifying speech data that are more discriminative of depression. Interestingly, experiment results demonstrate that by selecting speech with higher articulation effort, linguistic complexity, or word-based arousal/valence, improvements in acoustic speech-based feature depression classification performance can be achieved, serving as a guide for future elicitation design.