Can LLMs Understand the Impact of Trauma? Costs and Benefits of LLMs Coding the Interviews of Firearm Violence Survivors Jessica H. Zhu, Shayla Stringfield, Vahe Zaprosyan, Michael Wagner, Michel Cukier, Joseph B. Richardson Jr. University of Maryland, College Park Correspondence: [email protected]
arXiv:2604.16132v1 [cs.CL] 17 Apr 2026
Abstract Firearm violence is a pressing public health issue, yet research into survivors’ lived experiences remains underfunded and difficult to scale. Qualitative research, including in-depth interviews, is a valuable tool for understanding the personal and societal consequences of community firearm violence and designing effective interventions. However, manually analyzing these narratives through thematic analysis and inductive coding is time-consuming and labor-intensive. Recent advancements in large language models (LLMs) have opened the door to automating this process, though concerns remain about whether these models can accurately and ethically capture the experiences of vulnerable populations. In this study, we assess the use of open-source LLMs to inductively code interviews with 21 Black men who have survived community firearm violence. Our results demonstrate that while some configurations of LLMs can identify important codes, overall relevance remains low and is highly sensitive to data processing. Furthermore, LLM guardrails lead to substantial narrative erasure. These findings highlight both the potential and limitations of LLM-assisted qualitative coding and underscore the ethical challenges of applying AI in research involving marginalized communities.
1
Introduction
Firearm violence remains a leading cause of death among children and Black men in the United States (Johns Hopkins). It disproportionately impacts people of color, and especially Black men. Understanding the lived experiences of firearm violence survivors through qualitative research (e.g., in-depth interviews, focus groups, participant observations) is necessary for developing and informing effective community violence interventions. Yet, since the legislation of the 1996 Dickey Amendment, firearm violence researchers have been restricted from NIH
and CDC federal funding (Zhu et al., 2024). While there was a reprieve under the Biden administration, when the CDC and NIH provided $25 million in funding for gun violence research, the Trump administration has reverted to severely restricting funding (Brownlee, 2026). The long-term impact of disinvestment in firearm violence research has impacted the field’s ability to integrate new technology, like large language models (LLMs). With the exponential growth of LLMs, qualitative coding platforms have integrated numerous AI-assisted services, from chatbots, to summarizations, to code suggestions, to complete automation, to assist along the qualitative coding process (Atlas.ti). These services are often only accessible at an additional premium (MAXQDA). Companies like OpenAI tout these LLM-powered AI automation and AI assistance tools as the starting point for the next big idea (OpenAI, 2025). However, studies have not yet demonstrated that LLMs can effectively characterize the lived experiences of minority communities, let alone those of the young Black men who are survivors of firearm violence. In this study, we explore the value of using LLMs to code the interviews of Black men who are survivors of community firearm related violence. We bridge a gap caused by the lack of research on automated qualitative coding on diverse datasets and the limited work exploring how machine learning (ML), specifically large language models (LLMs), can support firearm violence research. Through building a machine coding pipeline and comparing the results with human qualitative coders as ground truth, we discuss the challenges in effectively and ethically applying LLMs to research involving historically marginalized communities and firearm violence intervention.
2
Background
Qualitative methods are vital to capturing the lived experiences of individuals who have sustained violent injuries. They provide contextual insights into the social and environmental factors that contribute to violence. Qualitative coding techniques like thematic analysis enable the discovery of these insights in unstructured data. However, thematic analysis is inherently an iterative and reflective process (Ahmed et al., 2025). From re-reading texts, to identifying interesting concepts, and organizing them into themes, qualitative coding is a timeintensive, laborious process (Williams and Moser, 2019). This delays the publication of research findings and thereby also delays the implementation of science-backed approaches to reducing gun violence. Numerous studies have explored the benefits of machine learning supported qualitative coding. Beginning as early as 2008, researchers have used statistically defined natural language processing models, using word co-occurrence, latent semantic analysis, and token based topic models (Dam and Kaufmann, 2008; Sherin, 2013; Baumer et al., 2017; Crowston et al., 2012; Rodriguez and Storer, 2019; Lennon et al., 2021) to explore qualitative data, sometimes incorporating interactive human collaboration (Gao et al., 2024; Gebreegziabher et al., 2023). These approaches all demonstrated the potential for speedups in the qualitative coding process through the assistance of natural language processing and machine learning, especially for short texts (Feuston and Brubaker, 2021). However, these initial approaches still required various degrees of data labeling, rule development, parameter tuning, retraining, and additional human interpretation. The lack of transparency and potential for an overwhelming number of code recommendations was also found to detract from machine learning integration in qualitative coding (Rietz and Maedche, 2021; Lennon et al., 2021; Marathe and Toyama, 2018). With the advent of AI agents and generative LLMs, researchers and platforms are experimenting with using LLMs to fully automate the qualitative coding process with minimal human interaction. LLMs have been found to be helpful for deductive coding techniques (Ziems et al., 2024; Ranjit et al., 2025). LLMs have also found success in thematic analysis and inductive coding in scenarios like consumer product surveys (Dai
et al., 2023), online forums (Nagaraj Rao et al., 2025; Sharma and Wallace, 2025), quotes from social science studies (Parfenova et al., 2025), litigation documents (Zhong et al., 2025), and mental health interviews of healthcare professionals (Singh et al., 2024). All of the aforementioned studies except (Parfenova et al., 2025; Zhong et al., 2025), used closed-source models with over 100B parameters. These services and model size are both financially and computationally inaccessible for low-resourced, historically marginalized communities. In addition, these studies were primarily short form texts from online forums, surveys, or discussions in standard English, which is not representative of the long form interviews needed to understand and address the disproportionate impact of firearm violence on young Black men. This under-served population has been substantially biased against by ML and LLMs. There have been few ML applications for firearm violence intervention published despite the boom in ML research applied to other public health fields (Zhu et al., 2024). While LLMs have the potential to be beneficial to qualitative coding in social work research, they remain significantly limited by their potential for hallucinations and bias against BIPOC (Patton et al., 2023). LLMs have been found to generate covertly racist decisions against African American English (AAE). Reinforcement learning with human feedback often exacerbates implicit biases while decreasing explicit biases (Hofmann et al., 2024). Even guardrails, which are ostensibly supposed to protect users, have demonstrated biases against certain identities (Li et al., 2024). While there is debate as to the actual harms of implicit biases, research has found implicit biases to manifest as systemic racism (Galvan and Payne, 2024). To mitigate the propagation of systemic racism, researchers and developers must conduct a thorough analysis of LLMs on diverse datasets in partnership with communities before allowing LLMs to be a ubiquitous part of thematic analysis.
3
Methodology
In an effort to resolve one facet of concerns about LLM biases and effectiveness, we partnered with community violence intervention researchers to interrogate the effectiveness of LLMs in understanding the lived experiences of Black firearm violence survivors. We compare the effectiveness of language models in identifying codes in the interviews
of 21 firearm violence survivors with the codes identified by qualitative researchers (see Figure 1).
Figure 1: Our pipeline for analyzing interviews and creating codes
3.1
Data
We analyze the interviews of 21 Black/AfricanAmerican men that were previously collected between January 2013 and December 2015 with approval from the IRB (protocol number 343085-1). These subjects were identified after being admitted to a Level 2 trauma center in the DC metropolitan area for a violent injury. They were asked to be volunteers following their admission and were compensated with $50 per interview. Dr. Richardson, an experienced African American studies anthropologist and firearm violence researcher, led one on one interviews following a semi-structured protocol. Interviews lasted approximately 60 minutes and were conducted in a private room. The interviews were later manually transcribed and deidentified by two research assistants. Transcribers were undergraduate students and identified as white men and women. Once transcripts were cleaned, participant names were changed to pseudonyms prior to data analysis. The principal investigator kept a master spreadsheet with participant names and pseudonyms in a password protected folder. 3.2
Human Coding (HC)
Interviews were coded manually using qualitative software. The AI coding functionality in the software was not used by the team. Human coders followed grounded theory and used inductive thematic analysis coding techniques (Ahmed et al., 2025) to identify initial codes. They identified as Black/African-American women, with one undergraduate student and two research staff. One coder had more than two years of experience. Human coders were instructed to review the study protocol, which focused on specific experiences around high-risk sexual activity, previous trauma history, perceived community violence, and feelings of retaliation, prior to coding. They were
encouraged to highlight codes related to the initial focus, as well as any codes that were not a part of this initial objective. Once the human coders identified the initial codes, member check-in strategies were used to review the accuracy of the codes. Initial human annotated codes (HC codes) were then refined and merged into formal HC codes. Formal codes were reviewed by the principal investigator for accuracy prior to finalization. 3.3
Machine Coding Pipeline (MC)
We also built a machine coding pipeline using opensource models to evaluate the effectiveness of more accessible low-resourced LLMs (those that can be run on relatively affordable hardware) at automated qualitative coding. We first cleaned and processed the long form interviews into machine readable chunks before passing them to the LLMs. We experimented with open-source models running on university research servers to protect the sensitive interview data. Experiments were conducted on one 40GB A100 node, for 8B parameter LLMs; one 24GB L4 GPU, for 1B parameter LLMs; and a 16GB RAM laptop for all Sentence Transformer language models. Code is available at https: //github.com/jhzsquared/AIvsHumanCoding. 3.3.1
Data Processing
We removed auxiliary information from the transcripts (e.g., time stamps, additional commentary on background noise) and extracted text by speaker (interviewer or subject). Due to computational resource limitations when using larger models, we chunked the data into a maximum of 256 token length sections. We used 256 tokens to approximate the length of a paragraph’s worth of discussion on a given topic. We experimented with “paired chunks” where we sequentially separated interviews such that every line from an interviewer was first paired with the following subject’s turn and then split further only if it was too long. For “question chunks”, we extracted the questions from the original interview protocol. We then encoded the questions as well as every subject response turn. The embeddings from two Sentence Transformer models (“all-mpnet-base-v2” and “multiqa-MiniLM-L6-cos-v1” (Reimers and Gurevych, 2020)) were concatenated to form a final ensembled embedding. We then calculated the cosine similarity between the embedded questions and responses. Responses were assigned to the question (or an “other” category) where they had a maximum simi-
larity score, as long as they met a 20% similarity threshold. This threshold and the models were selected from observational checks of the 30 highest scoring matches from the question-response pairs after testing different thresholds and models. Responses were further split into at most 256 length token chunks. For the smallest LLM (1B parameters), we also experimented with passing each interview in their entirety (“full text”). We did not have sufficient computational capacity available to do the same for the larger, 8B parameter LLM. A breakdown of the data’s descriptive statistics is shown in Table 1. # of words per interview # of words per response # of response turns per interview # of paired chunks per interview # of question chunks per question Total # of questions in the protocol Total # of interviews
11503 (SD=4955) 34 (SD=64) 481 (SD=173) 249 (SD=92) 30 (SD=35) 28 21
Table 1: Average and standard deviation of various descriptive statistics of the corpus of semi-structured interviews
3.3.2 Code Generation For initial code generation, we used Llama-3.21B-Instruct (Llama 1B) and Llama-3.1-8B-Instruct (Llama 8B) (Grattafiori et al., 2024). We selected these models for their consistently high performance on standard benchmarks for instruction based tasks within the confines of our compute limitations and to evaluate the impact of larger model sizes and more flexible resources. To extract relevant themes from the preprocessed interview text, we passed zero-shot prompts while varying the identity the system should assume and the context of the queries (see Appendix A for additional prompt settings). We used the term “themes” rather than “codes”, as during initial tests, we found that the models were more helpful with the term “themes”. The word “codes” often resulted in output discussing coded language (i.e., slang) rather than qualitative codes. Figure 2 shows an example of one of the many prompts tested.
System: “You are an African American Studies anthropologist analyzing interviews to understand the experiences of gun violence survivors . Your response should be a numbered list with each item on a new line.” User: “List the themes observed in the following interview excerpt:" {INTERVIEW EXCERPT}” Figure 2: Example prompt – The first highlighted portion is the “identity” we have the system assume, and the second highlighted portion is the “context” that is provided.
After the initial machine generated (MC) codes were generated, in mirroring the process of thematic analysis, we clustered the output codes into “formal codes”. We parsed the initial LLM output into individual codes, aggregated all results, and de-duplicated them. We then used BERTopic (Grootendorst, 2022) to cluster the initial codes into interpretable “formal” codes. We embedded the initial codes with “all-mpnet-base-v2”, the top performing pretrained Sentence Transformer model (Reimers and Gurevych, 2020). We used the topic model that maximized silhouette score after conducting grid-search over key hyperparameters. Parameter details can be found in Appendix B. We then prompted the originating LLM to generate formal MC code names using each topic model cluster’s representative keywords and its three most representative originating codes along with the LLM generated justification for those codes. See Appendix A for the prompt used to generate the formal codes. 3.3.3
Evaluation
To evaluate the quality of the machine codes, we introduce two metrics: Percent Captured and Percent Relevant. Percent Captured is the percentage of the formal human annotated codes with a machine code match. It reflects how well the machine automated coding pipeline extracts what the human annotators deemed to be important and could be considered a fuzzy recall score. Percent Relevant is the percentage of MC codes that match any HC codes (initial or formal). If there are many MC codes generated, the odds of a code matching with an HC code is higher, but the overall relevance of the output will be lower. While it could be considered analogous to a fuzzy precision score, Percent Relevant does not necessarily indicate that the ad-
ditional MC codes are hallucinations or not useful. These “irrelevant” MC codes could also be novel codes that the human annotators did not think of or did not deem important. To calculate these scores, we estimated the semantic similarity between the machine codes and human codes using the cosine similarity of their embeddings from “all-mpnet-base-v2” (Reimers and Gurevych, 2020). While human validation would be ideal, due to the intractable number of codes the LLM often generated and the numerous experiments completed, this was more time efficient. An MC code was considered to be a match with an HC code if its cosine similarity score was greater than 0.6. This threshold was determined through reviewing a sample of initial matches at varied thresholds. We evaluated both the initial and formal machine generated codes using these metrics.
than questions chunk and full text, outperforming them by 7% and 32% on average, respectively, per a Wilcoxon signed-rank test at .05 p-value. Question chunks had a statistically significant higher Percent Relevance than paired chunks, but only by a difference of 1%.
4
Figure 3: Percent Captured versus Percent Relevant of initial MC codes by data processing techniques using Llama 1B
Results
The human coders identified 41 initial codes and 11 formal codes across all 21 interviews. They spent an average of 12 hours each coding. Depending on the parameters, initial MC generation took 3.4 hours (SD = 3.3) on average over the 118 experiments that we ran. The large variation in time was primarily a factor of the data processing technique: the full interview experiments took 10 minutes on average, paired chunks took 7.3 hours on average, and question chunks took 2.5 hours on average. The formal code’s topic model pipeline took a trivial amount of time (less than 5 minutes on average). Statistics and scores from the best models are further summarized in Table 2. Their parameter settings are described in Appendix B 4.1
Initial MC Output
Prior to clustering, the MC pipeline generated orders of magnitude greater initial codes than the human coders. On average, 3072 (SD=2838) unique codes were identified from the initial LLM prompts. As few as 98 codes were generated using the full text data processing strategies and as many as 16000 for the paired chunk strategies. However, the full text strategy’s initial Percent Captured was never as high as the other data processing strategies though it had significantly higher relevance (see Figure 3). Full text output’s Percent Relevance was 11% and 12% higher on average than both question chunks and paired chunks, respectively. Paired chunks had significantly higher Percent Captured
Specifying the identity to be “an African American studies anthropologist” rather than “a Black anthropologist”, had a marginal, though statistically significant difference, of 0.4% in Percent Relevance. Specifying the context to be “the experiences of Black men as gun violence survivors” rather than just “the experiences of gun violence survivors” also resulted in a statistically significant 1% higher average Percent Captured, though there was no discernible impact on Percent Relevance. In comparing model size, Llama 8B outperformed Llama 1B by an average of 2.6% based on Percent Relevance, which was statistically significant per a Wilcoxon test at .05 p-value. Llama 8B did not outperform Llama 1B by Percent Captured, with Llama 1B actually outperforming (not statistically significantly) on average by 1.8%. We found that the best initial MC codes from either model were equally well aligned with the formal human codes, with 100% of the formal HC codes captured using both Llama 1B and Llama 8B (see Appendix C, Table 5 for the alignment of best and worst MC output to formal HC codes). However, the worst Llama 1B output had much lower scores than the worst Llama 8B experiment’s (Table 2). Overall, Percent Captured was as low as 27% (average of 71% and SD 18% across all experiments). Percent relevance was consistently low for all settings, with an average of 10% (SD = 5%)
across all experiments. The best and worst outputs all identified codes related to “Masculinity” and “Perceived community violence”, while only the best outputs also coded “Altercation leading to hospital visit”, “Behavior changes after incident”, “Reasoning for beef starting”, and “Sexual activity”. 4.2
Formal MC Output
After evaluating the Percent Captured and Percent Relevant of the initial MC output, we selected the best and worst results from each Llama model’s experiments for formal code generation. We passed these four sets of machine generated codes through the clustering pipeline. The specific prompts and data processing settings for these experiments are in Appendix C. Clustering the output greatly reduced the number of codes down to numbers close to the initial HC results (Table 2). While one may expect the quality of the codes to improve with aggregation, we found that clustering resulted in formal MC codes that were less aligned with the HC codes than the initial MC codes were. This is seen in the notable drop in Percent Captured between initial and formal codes in Table 2. This drop occurred regardless of the originating output’s scores. Clustering rebalanced the difference between the best Llama 8B and the worst Llama 8B’s output. It also neutralized the difference between the Best Llama 1B and the Best Llama 8B. After transforming the formal HC codes with each result’s pretrained topic model, we found that many of them were aligned with clusters where they lacked semantic similarity to the representative cluster name (i.e., the formal MC code). This indicates that in spite of their high silhouette scores, the clusters still encompass a wide breadth of codes. Only for a few human codes do their aligned clusters also have high semantic similarity. For example, for the formal HC code “Masculinity”, the best Llama 1B result’s topic model aligned it to the formal MC code “Dealing with masculinity and identity”, which was a semantic similarity match (per cosine similarity using “all-mpnet-base-v2”). However, the best Llama 8B result’s topic model aligned “Masculinity” to the formal MC code “Racial profiling and police bias”, even though it was most semantically similar to “Role models in masculinity”. Appendix C, Table 6 displays the alignment of all formal HC codes to clusters versus semantic match. Meanwhile, Llama 1B’s worst output fixated on non-standard English linguistic characteristics, like
the usage of the words “get” and “like”. None of its formal MC codes matched with HC codes based on semantic similarity. Clustering did substantially improve the overall relevance for all outputs (disregarding the worst Llama 1B since it had no matches after clustering). It reduced hundreds and thousands of initial codes to fewer than 60. Clustering also highlights potential new insights and relationships between codes. From the hierarchical structure that resulted from clustering the best Llama 8B’s codes (Figure 4), we can observe topics structured around employment and healthcare; social norms and relationships; descriptions of the source of injury; and varying facets of emotions. These interrelationships provide new insights into these narratives. We also examined the clusters that did not align with any formal HC codes to explore if the low Percent Relevance is a result of novel discoveries or hallucinations. We highlight some of the new codes from the best performing experiments in Table 3. With our MC pipeline, we are able to recall the interview excerpts that have led to each code. Upon reexamination of a sample of the most representative codes’ source data, we found that many of these new MC codes often hold a tenuous relationship with the source text, such that the LLM appears to have read between the lines. The new codes from Llama 8B were generally more justified by the representative source text than those from Llama 1B. For example, the code “Internet” from Llama 1B is partially justified, as they were discussing “World Star Hiphop”. While “Gang affiliation codes” was not justified at all since the representative source texts are primarily discussing sources of injury and fights. On the other hand, “Racial profiling and police bias” from Llama 8B is likely valid based on the discussions on unfair trespassing charges and differences in police treatment in neighborhoods. Meanwhile, “Public perception of the criminal justice system” is not clearly justified. Even though the interviews did discuss the subject’s experiences with the criminal justice system and how they perceive police, the representative texts do not clearly justify coding it as “public perception”. More experimentation and validation with humans in the loop is needed to determine if this discrepancy between Llama 1B and Llama 8B is a consistent trend.
Time Spent (hrs) # of Initial Codes # of Formal Codes Silhouette Score % Captured Initial Formal % Relevant Initial Formal
HC 35 41 11 N/A
Best L1B 1.45 2997 57 .694
Worst L1B .16 184 15 .785
Best L8B 5.75 6968 45 .68
Worst L8B .63 690 54 .73
N/A N/A
100% 36%
27% 0%
100% 36%
64% 45%
N/A N/A
9.0% 28%
6.5% 0%
13.8% 28%
19.7% 36%
Table 2: Statistics from the best and worst Llama 1B and Llama 8B experiments. Model settings are described in Appendix C.
Figure 4: The relationship between the top 20 most common codes from the best performing Llama 8B experiment
4.3
LLM Refusals
Although numerous codes were generated, we found that on average, 44% (SD = 25%) of each prompt request refused to generate output. Refusals also occurred during clustering naming. The primary reasons the LLMs used to justify refusal are displayed in Figure 5. These were determined through a regular expression of keywords related to each category. These keywords are listed in Appendix B. The primary justification LLMs provided for refusal was the descriptions of the survivor’s firearm violence experiences, which were often deemed too graphic. In some cases, the LLMs refused to “discuss content that promotes or glorifies violence”. There were cases where they also refused “to sexualize or objectify survivors of firearm violence”. The LLMs deemed the descriptions
and questions on the subject’s experiences were too graphic, even though we explicitly stated that we are requesting themes in our capacity as researchers. In addition, explicit and AAE language in the interviews, especially the use of the n* word, almost guaranteed refusal. This is despite the text not being derogatory. Discussions regarding sexual activity and race were also often refused, which is problematic as these topics were focal points of the original study. These interviews research not only the factors leading to traumatic events, but also if young Black men at high risk for violent victimization may also be engaged in high risk sexual behavior. Due to LLM refusals, much of the narrative discussing these research questions is ignored. It is important to note that we did not find a corre-
Llama 1B Continuation Checkup Perpetrator-Victim Dynamics Frustration and Anger Perpetrators Gang Affiliation Codes: Blood Gangs Internet
Valid N N N N N N M
Llama 8B Stigma and Shame in Vulnerable Populations Social Commentary and Moral Judgment Mental Health and Wellbeing of Black Men Racial Profiling and Police Bias Geographic Location and Sense of Place Public Perception of the Criminal Justice System Self-Awareness and Personal Growth
Valid M M Y Y M M Y
Table 3: New machine generated formal codes from the best performing experiments (Y= sufficient data to support it; M=data may support it; N=insufficient data to support it)
codes for human analysis, defeating the time saving goals of automation. But with the inclusion of a second clustering step, the machine pipeline generated a reasonable (fewer than 100) number of human readable codes. Yet these “formal” MC codes appeared to have lower validity overall than the initially generated codes, as well as a greater likelihood of highlighting hallucinated, misrepresented, and/or stereotypical codes, as shown by our Percent Captured and Percent Relevant scores.
Figure 5: Distribution of justifications for LLM refusals. Note: Justifications may be aligned with multiple topic areas (e.g., firearm violence would count for both gun and violence).
lation between the percent refused per experiment and the number of outputs generated, nor with our quality metrics: Percent Relevant and Percent Captured (Pearson’s correlation coefficient of < |.2|). This demonstrates that the formal MC codes we successfully captured were substantially present in the data that was not ignored. It also leads to the question of what additional, informative codes exist within the 44% of data that was erased.
5
Discussion
These results show that while a low-resourced fully automated pipeline has the potential to extract codes, the results are inconsistent. The best MC experiments effectively extracted the formal human codes with minimal tuning and additional background. However, there were significant differences in the quality of the initial MC codes with minimal changes to the data processing techniques, core prompt, and identity. The MC pipeline also initially generated an intractable number (on average 3000) of unique
These LLMs also were not difference-aware (Wang et al., 2025). Their guardrails were biased (Li et al., 2024). Our results indicate that not only are LLMs biased against AAE, but they are also biased against the experiences of firearm violence survivors, especially the traumas they have survived. These biases and misinformed guardrails resulted in narrative erasure of the interviews, where up to 65% of some subjects’ experiences were ignored. While Llama is known for being more safety minded against controversial prompts (DontPlanToEnd), these findings demonstrate significant concerns about the efficacy of LLMs that are marketed to assist all qualitative research, to include marginalized and historically under-served populations, like those impacted by firearm violence. Though human coders come with their own biases, unlike our automated pipeline, they were able to understand AAE and not fixate on colloquialisms. While there is an emotional burden to reading traumatic experiences, they successfully extracted codes relevant to the study protocol and were not stopped by the depictions of violent injuries pertinent to this study. However, they needed over 10 times as many hours to code, which demonstrates the significant potential for time savings if models are accessible, reliable, and ethically aligned. Given the resource and especially computational
constraints that many minority communities and positive impact organizations face, it is promising that Llama 1B, which required on average 6GB of VRAM, nears Llama 8B’s performance, which required 34GB of VRAM to run our pipeline. However, Llama 1B’s performance was less reliable and more impacted by fixations on colloquialisms characteristic of AAE. Regardless of the model, automated pipelines will still require human experts to validate codes. Luckily, as demonstrated by our novel code investigation, the validation time can be reduced through effective data tracking strategies combined with more interpretable clustering techniques like BERTopic. Nevertheless, significantly more adjustments will still be needed for reliable, unbiased inductive coding.
6
Conclusion
Our research demonstrates that an automated pipeline for coding long-form interviews still has significant limitations. Through implementing open-source pre-trained instruction tuned models, our automated pipeline demonstrated high Percent Captured using as few as 12 GPU hours. However, more low-resourced methods are needed to support qualitative coding of non-short form, survey based, standard English data. We cannot rely on AI qualitative coding tools to understand the lived experiences of diverse and vulnerable communities, without significant improvements to LLMs. Though an automated coding pipeline may appear faster, it cannot yet match the consistent quality of human coding. The time spent evaluating machine generated codes rapidly devalues any initial time savings. If one further factors in the time to tune and train an AI pipeline, the time gains are drastically reduced. We currently cannot use AI assistants and LLMs to code these narratives. There remains a significant risk of narrative erasure as well as nonrepresentative and hallucinated results. Methods are needed that can efficiently increase coding quality and reliability, and decrease the risks of misrepresentation and implicit biases for low-resourced settings. As our research shows, larger models may not even offer a substantial improvement in performance, despite their significant increase in computational, financial, and environmental burden on an already underfunded research community. Therefore, we must be deliberate in building lowresourced AI tools that represent minority commu-
nities, like the young Black men disproportionately impacted by firearm injuries.
Limitations This study draws from subjects in the Washington, DC area. As such, they represent only one of many AAE regional patterns. In addition, interviews were manually transcribed. While transcriptions were further reviewed by researchers fluent in AAE, there is still a potential for human error. Our prompting strategy was also limited and did not include the study protocol as we wanted to minimize the potential for the LLM to generate themes from the protocol rather than the interview. Including the study protocol would otherwise lead to a more informed context, assuming the LLM has sufficient capacity for longer inputs. We minimized the potential discrepancies in human versus machine codes, by asking the human coders to code items not relevant to the initial study objective as well. Further human validation of our evaluation and data processing pipeline would be beneficial prior to generalizations. However, through our limited samples, we believe our thresholds to be sufficiently reliable for this study. We also recognize that further examination of LLMs in other model families, or of larger sizes with quantization would be beneficial. We used full precision Llama 3.2 1B and 3.1 8B due to resource limitations, the widespread usage of Llama, and to avoid additional variations from quantization.
Ethical Considerations Aside from the potential costs of LLM automated coding already discussed (e.g., narrative erasure, misrepresentation, and stereotyping), there is a risk in using these LLM generated output code to generalize the experiences of gun violence survivors. While we do share the final formal codes from MC, further validation of the MC pipeline is needed for these codes to be a reliable source of qualitative analysis. Data will not be released to protect interview subjects. To further ensure privacy, we have manually validated that none of the output codes are identifiable.
Acknowledgments Thank you to Hannah Balconoff and Bailey Skeeter for their work coding the interviews and the undergraduate researchers who supported transcriptions. The authors would also like to acknowledge the
University of Maryland supercomputing resources (https://hpcc.umd.edu) made available for conducting the research reported in this paper as well as the support of various University of Maryland College Park grants, to include their Summer Research Fellowship, School of Public Health’s Prevention Research Center Seed Grant Program, Consortium for Race, Gender, and Ethnicity: Faculty Seed Grants for Developing Qualitative Work, College of Behavioral and Social Sciences Dean’s Research Initiative, and Research and Scholarship Award.
References Sirwan Khalid Ahmed, Ribwar Arsalan Mohammed, Abdulqadir J. Nashwan, Radhwan Hussein Ibrahim, Araz Qadir Abdalla, Barzan Mohammed M. Ameen, and Renas Mohammed Khdhir. 2025. Using thematic analysis in qualitative research. Journal of Medicine, Surgery, and Public Health, 6:100198.
in human-AI collaboration for qualitative analysis. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW2):1–25. Manuel J. Galvan and B. Keith Payne. 2024. Implicit bias as a cognitive manifestation of systemic racism. Daedalus, 153(1):106–122. Jie Gao, Yuchen Guo, Gionnieve Lim, Tianqin Zhang, Zheng Zhang, Toby Jia-Jun Li, and Simon Tangi Perrault. 2024. Collabcoder: A lower-barrier, rigorous workflow for inductive collaborative qualitative analysis with large language models. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, New York, NY, USA. Association for Computing Machinery. Simret Araya Gebreegziabher, Zheng Zhang, Xiaohang Tang, Yihao Meng, Elena L. Glassman, and Toby Jia-Jun Li. 2023. PaTAT: Human-AI collaborative qualitative coding with explainable interactive rule synthesis. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM.
Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Atlas.ti. Quickly gain insights with AI-powered Abhinav Pandey, Abhishek Kadian, Ahmad Altools. https://atlasti.com/trainings/ Dahle, Aiesha Letman, Akhil Mathur, Alan Schelquickly-gain-insights-with-ai-powered-tools. ten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Accessed: 2025-12-20. Goyal, Anthony Hartshorn, Aobo Yang, Archi MiEric P. S. Baumer, David Mimno, Shion Guha, Emily tra, Archie Sravankumar, Artem Korenev, Arthur Quan, and Geri K. Gay. 2017. Comparing grounded Hinsvark, and 542 others. 2024. The Llama 3 herd theory and topic modeling: Extreme divergence or of models. Preprint, arXiv:2407.21783. unlikely convergence? Journal of the Association Maarten Grootendorst. 2022. BERTopic: neural topic for Information Science and Technology, 68(6):1397– modeling with a class-based TF-IDF procedure. 1410. arXiv preprint arXiv:2203.05794. Chip Brownlee. 2026. Trump has slashed federal Valentin Hofmann, Pratyusha Ria Kalluri, Dan Jurafsky, funding for gun violence prevention. but how much and Sharese King. 2024. AI generates covertly racist exactly? https://www.thetrace.org/2026/01/ decisions about people based on their dialect. Nature, trump-public-safety-gun-violence-funding/. 633(8028):147–154. Accessed: 2026-04-15. Kevin Crowston, Eileen E. Allen, and Robert Heckman. 2012. Using natural language processing technology for qualitative data analysis. International Journal of Social Research Methodology, 15(6):523–543. Shih-Chieh Dai, Aiping Xiong, and Lun-Wei Ku. 2023. LLM-in-the-loop: leveraging large language model for thematic analysis. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9993–10001. Association for Computational Linguistics. Gregory Dam and Stefan Kaufmann. 2008. Computer assessment of interview data using latent semantic analysis. Behavior Research Methods, 40(1):8–20. DontPlanToEnd. UGI leaderboard: Uncensored general intelligence. https://huggingface.co/ spaces/DontPlanToEnd/UGI-Leaderboard. Accessed: 2025-12-20. Jessica L. Feuston and Jed R. Brubaker. 2021. Putting tools in their place: The role of time and perspective
Johns Hopkins. 2024. New report highlights U.S. 2022 gun-related deaths: Firearms remain leading cause of death for children and teens, and disproportionately affect people of color. https://publichealth.jhu.edu/ 2024/guns-remain-leading-cause-of-deathfor-children-and-teens. Accessed: 2025-1220. Robert P Lennon, Robbie Fraleigh, Lauren J Van Scoy, Aparna Keshaviah, Xindi C Hu, Bethany L Snyder, Erin L Miller, William A Calo, Aleksandra E Zgierska, and Christopher Griffin. 2021. Developing and testing an automated qualitative assistant (AQUA) to support qualitative analysis. Family Medicine and Community Health, 9(Suppl 1):e001287. Victoria R Li, Yida Chen, and Naomi Saphra. 2024. ChatGPT doesn’t trust chargers fans: Guardrail sensitivity in context. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6327–6345, Miami, Florida, USA. Association for Computational Linguistics.
Megh Marathe and Kentaro Toyama. 2018. Semiautomated coding for qualitative research. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. ACM.
Bruce Sherin. 2013. A computational study of commonsense science: An exploration in the automated analysis of clinical interview data. Journal of the Learning Sciences, 22(4):600–638.
MAXQDA. AI assist: AI-powered qualitative data analysis with MAXQDA. https://www.maxqda.com/ products/ai-assist. Accessed: 2025-12-20.
Satpreet Harcharan Singh, Kevin Jiang, Kanchan Bhasin, Ashutosh Sabharwal, Nidal Moukaddam, and Ankit Patel. 2024. RACER: An llm-powered methodology for scalable analysis of semi-structured mental health interviews. In Proceedings of the 1st Workshop on NLP for Science (NLP4Science), pages 73–98. Association for Computational Linguistics.
Varun Nagaraj Rao, Eesha Agarwal, Samantha Dalal, Dana Calacci, and Andrés Monroy-Hernández. 2025. QuaLLM: An llm-based framework to extract quantitative insights from online forums. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 1355–1369. Association for Computational Linguistics. OpenAI. 2025. ChatGPT | the intelligence age. https: //www.maxqda.com/products/ai-assist. Accessed: 2025-12-20. Angelina Parfenova, Andreas Marfurt, Jürgen Pfeffer, and Alexander Denzler. 2025. Text annotation via inductive coding: Comparing human experts to llms in qualitative data analysis. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 6456–6469. Association for Computational Linguistics. Desmond Upton Patton, Aviv Y. Landau, and Siva Mathiyazhagan. 2023. ChatGPT for social work science: Ethical challenges and opportunities. Journal of the Society for Social Work and Research, 14(3):553–562. Jaspreet Ranjit, Hyundong J. Cho, Claire J. Smerdon, Yoonsoo Nam, Myles Phung, Jonathan May, John R. Blosnich, and Swabha Swayamdipta. 2025. Uncovering intervention opportunities for suicide prevention with language model assistants. Nils Reimers and Iryna Gurevych. 2020. Making monolingual sentence embeddings multilingual using knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics. Tim Rietz and Alexander Maedche. 2021. Cody: An AIbased system to semi-automate coding for qualitative research. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM. Maria Y. Rodriguez and Heather Storer. 2019. A computational social science perspective on qualitative data exploration: Using topic models for the descriptive analysis of social media data. Journal of Technology in Human Services, 38(1):54–86. Ansh Sharma and James R Wallace. 2025. DeTAILS: Deep thematic analysis with iterative llm support. In Proceedings of the 7th ACM Conference on Conversational User Interfaces, CUI ’25, pages 1–7. ACM.
Angelina Wang, Michelle Phan, Daniel E. Ho, and Sanmi Koyejo. 2025. Fairness through difference awareness: Measuring Desired group discrimination in LLMs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6867–6893, Vienna, Austria. Association for Computational Linguistics. Michael Williams and Tami Moser. 2019. The art of coding and thematic exploration in qualitative research. International Management Review, 15:45. Mian Zhong, Pristina Wang, and Anjalie Field. 2025. Hicode: Hierarchical inductive coding with llms. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 31048–31066. Association for Computational Linguistics. Jessica Zhu, Michel Cukier, and Jr Richardson, Joseph. 2024. Nutrition facts, drug facts, and model facts: putting AI ethics into practice in gun violence research. Journal of the American Medical Informatics Association, 31(10):2414–2421. Caleb Ziems, William Held, Omar Shaikh, Jiaao Chen, Zhehao Zhang, and Diyi Yang. 2024. Can large language models transform computational social science? Computational Linguistics, 50(1):237–291.
A
Prompts
The full list of prompts, identities, and context experimented with are as follows. A.1
Identities
• an anthropologist • an African American Studies anthropologist • a Black anthropologist A.2
Contexts
• the experiences of gun violence survivors • the experiences of Black men as gun violence survivors
A.3
Prompts
Prompts by data processing technique: For paired chunks: • base theme: “user”: “What themes are observed in the following interview excerpt? Your response should be a numbered list with each item on a new line. {INTERVIEW}” • base_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the themes observed in the following interview excerpt: {INTERVIEW}” • cot_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: "List the themes observed in the following interview excerpt. Provide quotes from the interview that demonstrate the themes. {INTERVIEW} ” • “base_c”: “system”: “You are {IDENTITY} applying inductive coding techniques to understand {CONTEXT} from interview data. Your response should be a numbered list with each item on a new line.” “user”: “List the codes observed in the following interview excerpt: {INTERVIEW}” • “novel_cot_t”: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the novel themes observed in the following interview excerpt. Provide quotes from the interview that demonstrate the themes. {INTERVIEW}" For question chunks: • base_theme “user”: “What themes are observed in the following interview responses to the question: {QUESTION}? Your response should be a numbered list with each item on a new line. {INTERVIEW}”
• base_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the key themes observed in the following interview responses to the question: {QUESTION}:\n {INTERVIEW}” • cot_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the key themes observed in the following interview responses to the question: {QUESTION}. Provide quotes from the interview that demonstrate the themes. {INTERVIEW}” • base_c: “system”: “You are {IDENTITY} applying inductive coding techniques to understand {CONTEXT} from interview data. Your response should be a numbered list with each item on a new line.” “user”: “List the codes observed in the following interview responses to the question: {QUESTION}. Responses: {INTERVIEW}” • novel_cot_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the novel themes observed in the following interview responses. Provide quotes from the interview that demonstrate the themes. {INTERVIEW}” For full text: • base_theme: “user”: “What themes are observed in the following interview? Your response should be a numbered list with each item on a new line. {INTERVIEW}” • base_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the themes observed in the following interview: {INTERVIEW}”
• cot_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the themes observed in the following interview. Provide quotes from the interview that demonstrate the themes. {INTERVIEW}”
To analyze the source of refusals, we relied on regular expressions. We did experiment with topic modeling the refusal sentences as well, but the sentences were too noisy for well structured and interpretable clusters. The terms used in the regular expressions for each refusal category are as follows: • illegal: illegal, criminal, crime
• base_c: “system”: “You are {IDENTITY} applying inductive coding techniques to understand {CONTEXT} from interview data. Your response should be a numbered list with each item on a new line.” “user”: “List the codes observed in the following interview: {INTERVIEW}"
• violence: violent, violence, war, brutality • guns: firearm, shot, gun, shooting • explicit: explicit, n-word, profane, profanity, obscenity, nigga • stereotypes: hate, speech, derogatory, stereotype, slur, discriminatory, discriminate, stigma
• novel_cot_t: “system”: “You are {IDENTITY} analyzing interviews to understand {CONTEXT}. Your response should be a numbered list with each item on a new line.” “user”: “List the novel themes observed in the following interview. Provide quotes from the interview that demonstrate the themes. {INTERVIEW}” To get the topic name: “system”: “You are an assistant that extracts high-level topics from texts. Only return the topic name.” “user”: This is a list of texts where each collection of texts describe a topic. After each collection of texts, the name of the topic they represent is mentioned as a short-highly-descriptive title. — Topic: Sample texts from this topic: - DOCUMENTS Keywords: KEYWORDS Topic name:” The DOCUMENTS are the originating code and justification (as relevant). The KEYWORDS are outputs of BERTopic.
B
Pipeline Parameters
We used a temperature of 0.6 and top_p of 0.9 for both Llama 3.1 8B (L8B) and 3.2 1B (L1B) Instruct models to support more creative but still reliable output. The prompt settings and topic modeling parameters we found to have the best and worst codes are shown in Table 4
• mental_health: mental, suicide, crisis, selfdestructive • graphic: dangerous, graphic, disturbing, harmful • sex: sex, condom, HIV, AIDS, std • drugs: drug, marijuana, substance abuse, weed • gender: women, gender, man, men, woman, female, male • race: black, African, racial • minors: child abuse, child, minor, children • privacy: identify, personal, individual, medical • prison: jail, prison, justice system, incarceration • misc: all sentences that did not fall into one of these categories
C
Initial HC and formal MC output
We share the initial human annotated codes as well as the formal output codes from the best and worst Llama 1B and Llama 8B pipeline results to allow further comparison of the output. We have manually removed duplicates, topic name refusals, and “null” topic. Duplicated topic names are marked with “(number of duplicates)”. Given the systemic errors in pipeline results, the MC outputs should
Data processing Prompt Identity
Context
n_neighbors n_components min cluster size
Best L1B base_c Question chunks African American Studies anthropologist the experiences of gun violence survivors 15 5 15
Worst L1B base_c Full text
Best L8B base_c Paired chunks
Worst L8B cot_t Question chunks Black anthropol- Black anthropol- African Ameriogist ogist can studies anthropologist the experiences the experiences the experiences of gun violence of Black men of Black men survivors as gun violence as gun violence survivors survivors 5 20 5 5 5 2 5 40 5
Table 4: Pipeline parameters resulting in best and worst LLM codes
not be directly used to draw conclusions on the lived experiences of firearm violence survivors. Initial human annotated codes: • Use of deescalation tactics • Feelings of spirituality from incident • ACEs • Influence of neighborhood factors on self peer • Trauma recidivism • Influence of neighborhood factors on self older adults
• Sexuality • Rooted prejudice • Street Codes • Reasoning for problems escalating or "beefs" starting • Support System • Reason for hospital visit/ Diagnosis at hospital • Videos of incident • Reaction to the injury • Probation or parole
• Thoughts on racism
• Putting in work definition
• Demographic information
• Previous incarceration history
• Thoughts of incident being life changing moment
• Perceived community violence
• Daily routine
• Masculinity
• Symptoms of traumatic stress
• Misconception of medical care
• Current PTSD
• Influences on the youth
• Substance use
• Altercations with police
• Perpetuation of violence through the system
• Health care coverage
• Social media use
• Altercation leading up to hospital visit
• Religion
• Frequency of healthcare
• Sexual activity
• Carrying illegally for protection
• Gang involvement (not neighborhood)
HC Best L1B Aftermath of vio- Trauma and injury lent injury
Best L8B Worst L1B Trauma and physical aftermath
Altercation leading to hospital visit Behavior changes after incident
Physical altercation
Physical altercation
Worst L8B Prior experiences with violence and injury
Emotional response to the incident
Pre-incident vs. Post-incident behavior Feelings on retalia- Desire for retalia- Struggle with retalition tion ation Influences on the Youth Youth influence youth
Immediate desire for retaliation Youth are seen as role models for their peers Masculinity Masculinity Masculinity Masculinity Defining masculinity Perceived commu- Neighborhood vio- Community vio- Gun vio- Racialized violence nity violence lence lence lence and victimhood Previous incarcera- Ex-offenders Past incarceration Multiple stints in tion history experience prison Reasoning for beef Beef (conflict, dis- Beef starting pute, or conflict) Sexual activity Sexual activity Leisure activities Substance use Substance use Substance use Substance Substance use as a abuse way to escape or avoid problems
Table 5: Initial machine codes that are most semantically similar to the human codes. Darkened cells indicate there was no machine match
• Feelings on retaliation
• Sense of Disconnection
• Lack of praise for those doing good
• Post-racial Society
• Struggles for black men in society
• Perpetrator-Victim Dynamics
Best Llama-3.2-1B Instruct formal codes (2 topics refused output, 1 null): • Neighborhood • Physical Discomfort • Relationships
• Systemic Inequality • Stability • Violence (3) • Older
• Perpetrators
• Work
• Substance Use
• Identity
• Fear
• Protection
• Lack
• Confusion
• Affirmation
• Safety
• Good
• Care
• Self-preservation
• Resilience and Coping Mechanisms
• Police
• Spiritual
• Self-defense (2)
• Self-improvement Past
• Gang Affiliation Codes: Blood Gangs
Best Llama 3.1-8B Instruct formal codes:
• Security
• Societal Norms and Expectations
• Anger
• Relationship with Authority
• Self-safety
• Violence and Aggression
• Gun Violence
• Career Development and Stability
• Self-desire
• Resilience and Survival
• Internet
• Danger and Risk in Vulnerable Contexts
• Avoidance
• Avoidance Tactics in Conversation
• Defensiveness
• Childhood and Coming of Age
• Situations
• Role Models in Masculinity
• Transporters
• Racial Profiling and Police Bias
• Boys
• Family Dynamics and Structure
• Last
• Healthcare Access Barriers
• Dealing with Masculinity and Identity
• Fear and Anxiety
• Continuation
• Emotional Vulnerability
• De-escalation
• Emotional Expression and Regulation
• Social Relationships
• Gun Violence
• Trauma
• Geographic Location and Sense of Place
• Moment
• Time Perception and Transition
• Family Dynamics
• Lack of Guidance and Structure for Young Black Men
• Perceived Threat of Death • Checkup • Nonverbal Communication • Perceived Societal Expectations • Trust in Authority (2) • Change • Education • Frustration and Anger • Empathy
• Public Perception of the Criminal Justice System • Relationship Dynamics • Gun Violence and Victimization • Social Support Systems • Self-Blame and Accountability • Substance Use and Abuse • Uncertainty and Ambiguity in Life Experiences
• Validation
• Peer Influence on Substance Abuse
• Faith and Spirituality
• Youth’s Relationship with Authority Figures
• Stigma and Shame in Vulnerable Populations
• Personal Growth and Self-Improvement
• Autonomy and Self-Determination in Masculinity
• Systemic Inequality and Social Unrest
• Mental Health and Wellbeing of Black Men • Contextualizing Gun Violence
• Gun Culture and Ownership Practices • Rediscovery of Purpose and Meaning
• Social Commentary and Moral Judgment
• Sense of Belonging to a Close-Knit Community
• Prioritization of Needs Over Health
• Survival and Self-Preservation
• Avoidance of Disclosure
• Physical Assault
• Street Culture and Identity
• Substance Abuse Among Young People
• Conflict Resolution Strategies
• The Influence of Peer Relationships on Youth
• Impact of Gun Violence on Survivors
• Resilience in the Face of Adversity
• Self-Awareness and Personal Growth
• Physical and Emotional Recovery After Trauma
Worst Llama 3.2-1B Instruct formal codes (3 refused and 2 null): • Health: High Blood Pressure
• Emotional Regulation and Self-Control • Positive Role Models in the Home and Community
• Relationship • Get
• Impact of Past Experiences on Behavior and Life Choices
• Like
• Prioritizing Family Stability
• Gunshot
• Social Isolation and Limited Resources
• Privilege
• Masculinity Socialization
• Racial Identity
• Normalization of Violence in Communities
Worst Llama 3.1-8B Instruct formal codes:
• Youth’s Attitudes Towards Violence
• Job Insecurity and Limited Opportunities
• Delayed Understanding and Acceptance
• Trauma and Mental Health Consequences
• Personal Growth and Self-Improvement
• Struggling Young Adults’ Motivations and Aspirations
• Youth Vulnerability to Gang Involvement
• Defining Masculinity
• Reentry and Reincarceration Challenges
• Racism in a Post-Racial Society
• Family Strains and Communication Breakdowns
• Personal Agency and Control
• Importance of Guidance and Mentorship
• Psychological Aftermath of Trauma
• Limited Familiarity with Internet Technology
• Lack of Consequences for Violence in Youth
• Increased Vigilance and Caution in Daily Life
• Uncertainty and Ambivalence • Conflict De-escalation • Regret Over Missed Opportunities in Life • The Search for Identity and Belonging • Prioritizing Neighborhood Safety • Loss of Autonomy and Control • Substance Abuse as a Coping Mechanism • Gun Violence and Trauma in Communities • Faith and Spiritual Growth • Macho Identity • Side Effects of Opioids • Relationships and Support Systems • Negative Influences on Young People • Struggling with Substance Addiction • Dependence on Others for Support • Limited Mobility and Stability • Limited Job Opportunities
Dealing with masculinity and identity
Sexual activity Substance use
Reasoning for beef starting
Lack
Substance use
Dealing with masculinity and identity Gun violence
Role models in masculinity
Relationship
Relationship
Get
Uncertainty and Ambiguity in life experiences
Impact of gun violence on survivors Substance and abuse
use
Like
Career develop- Communityment and stabil- specific apity proaches to addressing gun violence Relationship dyGet namics
Racial profiling and police bias
Physical trauma and injury Role models in masculinity
Semantic Match Physical trauma and injury
Worst L1B Cluster Match Get Semantic Match
Table 6: The formal machine codes that the human codes align best with. Darkened cells indicate there was no match.
Previous incar- Empathy ceration history
Perceived community violence
Masculinity
Societal norms and expectations
Altercation leading to hospital visit Behavior changes after incident Feelings on retaliation Influences on the youth Societal norms and expectations
Best L8B Cluster Match Societal norms and expectations
Best L1B Cluster Match Semantic Match Aftermath of vi- Family dynam- Trauma olent injury ics
HC
a
Gun violence and trauma in communities
Family strains and communication breakdowns
Racism in post-racial society
Resilience in the face of adversity
Worst L8B Cluster Match Emotional regulation and selfcontrol Regret over missed opportunities in life Peer influence on substance abuse
Struggling with substance addiction
Normalization of violence in communities
Negative influences on young people Defining masculinity
Semantic Match Psychological aftermath of trauma