Socially desirable reporting to survey questions can lead to measurement error that is specific to questions, data collection modes, interviewers, and respondents. Social desirability scales commonly used in the literature, such as the Marlowe-Crowne Scale, assume that the respondent-level component is stable across all sensitive survey questions. To relax this assumption, we use mixed Rasch models (MRMs), a measurement tool that incorporates both respondent and question information to measure the respondent's socially desirable response behavior. MRMs are a mixture of latent class and item response theory (IRT) models. IRT models are applied to scales consisting of multiple items measuring a latent construct. One of the latent item characteristics modeled is “item difficulty,” which refers to the extent to which an item taps into the construct. When applied to a desirable/undesirable question, item difficulty also captures the perceived sensitivity of the item. Thus, more desirable items have a higher probability and less desirable items have a lower probability of being endorsed, though they both tap into the same level of the construct. The respondent's perception of the item's sensitivity could be affected by his or her characteristics, such as the need for social conformity. Such respondent-level variability in the perceived sensitivity of the item could lead to differences in the item difficulty estimates across respondents. We explore the use of MRMs to capture such respondent-level variability through latent classes. We hypothesize that these latent classes from MRMs capture respondents with different needs for social conformity and in turn different socially desirable response behavior. We confirm the nature of these latent classes by testing their association with objective survey context measures, such as interview privacy and interview mode. We apply these models to desirable, undesirable, and relatively neutral item scales from the Lebanon World Mental Health Survey.