Policy Research Working Paper 10597 Using Large Language Models for Qualitative Analysis can Introduce Serious Bias Julian Ashwin Aditya Chhabra Vijayendra Rao Development Economics Development Research Group November 2023 Policy Research Working Paper 10597 Abstract Large Language Models (LLMs) are quickly becoming that LLMs make in annotating interview transcripts are not ubiquitous, but the implications for social science research random with respect to the characteristics of the interview are not yet well understood. This paper asks whether LLMs subjects. Training simpler supervised models on high-qual- can help us analyse large-N qualitative data from open- ity human annotations with flexible coding leads to less ended interviews, with an application to transcripts of measurement error and bias than LLM annotations. There- interviews with displaced Rohingya people in Cox’s Bazaar, fore, given that some high quality annotations are necessary Bangladesh. The analysis finds that a great deal of caution in order to asses whether an LLM introduces bias, this is needed in using LLMs to annotate text as there is a risk paper argues that it is probably preferable to train a bespoke of introducing biases that can lead to misleading inferences. model on these annotations than it is to use an LLM for Here this refers to bias in the technical sense, that the errors annotation. This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at vrao@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Using Large Language Models for Qualitative Analysis can Introduce Serious Bias∗ Julian Ashwin Aditya Chhabra Vijayendra Rao† Maastricht University World Bank World Bank Keywords: Large Language Models, Qualitative Analysis, ChatGPT, Llama 2, Text as Data, Aspirations, Rohingya, Bangladesh JEL Codes: C63 C89 O1 1 Introduction Large Language Models (LLMs) are increasingly being used in social science research to, among other things, analyze and annotate text data (Gilardi et al., 2023). As LLMs become more accessible and popular we can expect that there will be a temptation to use them to analyze open-ended interview data such as those used by qualitative researchers (Small and Calarco, 2022) who follow an interpretative analytical approach. This relies on careful, nuanced, coding conducted by trained social scientists (Detering and Waters, 2018). Qualitative analysis of this kind lies at the core of fields like anthropology and sociology, and there is now a rapidly expanding literature on the use of Natural Language Processing (NLP) methods to analyze qualitative data in sociology (Bonikowski and Nelson, 2022), and qualitative analysis and NLP are also now being increasingly employed in more quantitative fields such as economics (Rao, 2023). Data generated from open-ended, in-depth, interviews is potentially very different from the benchmark datasets often used in the NLP literature to validate modelling approaches such as English language tweets and news, or product reviews. This is because qualitative research is often conducted in a ∗ The authors are grateful to the World Bank’s Knowledge for Change Program, and the World Bank-UNHCR Joint Data Center on Forced Displacement for financial support. Sudarshan Aittreya provided valuable research assistance for the project. † Corresponding author: vrao@worldbank.org manner in which the specific context matters for interpreting the data, and analyzed with codes that are "flexibly" developed that can be quite nuanced and complex. This is a particular problem in non-Western societies because LLMs have been shown to most resemble people from Western, Educated, Industrialized, Rich and Democratic (WEIRD) societies (Atari et al., 2023), and our example application falls into this category. We have interviews on a very specific topic (children’s aspirations) with a very specific population (Rohingya refugees and their hosts in Bangladesh) who are not well represented in the training data that LLMs are trained on (or in the data used in the NLP literature more broadly). We find that in such a context, using LLMs to annotate text is potentially dangerous. We test three different LLMs (ChatGPT and two versions of Meta’s Llama 2) and find that the prediction errors they make in annotation are not random with respect to the characteristics of the interview subject. This can lead to misleading conclusions in later analysis, as we shown in Figure 5. Statistical analysis based on LLM annotations can lead to estimated effects that are very different from those based on human expert annotations. It is therefore crucial to have some high quality expert annotations, even if it is just to assess whether the LLM is introducing bias or not. Given that some high quality annotations are needed to assess whether the LLM introduces bias, we argue that it is preferable to train a bespoke model on these annotations than it is to use an LLM. We show that iQual, a method we developed with others (Ashwin et al., 2022) to analyze large-N qualitative data by training supervised models on small human annotated samples, not only performs better than LLMs in terms of out-of-sample prediction accuracy but also introduces much less bias. LLMs can possibly assist this process by generating larger training sets (i.e. data augmentation, as proposed by Dai et al. (2023)) but we only find evidence of marginal benefits in a few cases. This suggests a potential way in which to reconcile the nuance and "reflexive" qualities of interpretative qualitative analysis with large representative samples. Crucially, we see LLMs and other NLP methods as assisting and extending traditional qualitative analysis, not replacing it. In order to create a coding tree that captures important and interesting variation across documents in a nuanced and context- aware manner, there is no substitute for a careful reading on at least a subset of those documents. Our application is based on open-ended interviews with displaced Rohingya people and their Bangladeshi hosts in Cox’s Bazaar, Bangladesh. These interviews focused on subjects aspirations and ambitions for their children (Callard, 2018) as well as their capacity to achieve those goals, i.e. their navigational capacity (Appadurai, 2004). They are analysed in detail in Ashwin et al. (2022), so we will not discuss the detail of data collection or related social science literature here. The substance of these interviews is not critical to the methodological contribution of this paper, but it is important to note that while "ambition" can be captured well by structured questions that yield quantitative data, aspirations and navigational capacity are subtle and complex concepts not easily defined are captured in structured surveys. It is precisely when dealing with these sorts of concepts that open- ended interviews and interpretative qualitative analysis is valuable. The complexity and nuance of the concepts may play a role in explaining the poor performance of LLMs in annotating interviews compared with other studies where the annotation tasks were substantially more straightforward, e.g. Mellon et al. (2022). Previous work has suggested that LLMs might outperform crowd-sourced human annotations (Gilardi et al., 2023), or even that a substantial proportion of workers on crowd-sourcing platform may be using LLMs in completing tasks (Veselovsky et al., 2023). Our results do not contradict these as for many annotation tasks LLMs may indeed perform very well and save researchers the expense and complication of crowd-sourcing. However, our results do suggest that researchers ought to be aware of the possibility of biases introduced by LLM annotation, particularly on data where a nuanced, contextual understanding of the documents is needed; LLMs, like other types of machine learning models, reflect the data they are trained on (Kearns and Roth, 2020) and many of the contexts in which qualitative analysis adds value require an understanding of communities and concepts that may not be adequately represented in this training data. The paper is structured as follows. The remainder of this Section discusses this paper’s contribution 2 in the context of related literature. Section 2 then very briefly introduces our dataset of annotated interview transcripts. Section 3 describes our approach to using LLMs for annotation (3.1) and the supervised NLP method introduced by Ashwin et al. (2022) which we refer to as iQual going forward (3.2). Section 4 then describes LLM-based out-of-sample performance in comparison to iQual (4.1) and then shows that LLMs introduce more bias and illustrates this could cause researchers to draw incorrect conclusions (4.2). Section 5 then concludes. 2 Data and Qualitative Analysis The interview transcripts, data collection and the qualitative coding process are explained in detail in Ashwin et al. (2022), so we restrict ourselves to a very brief description here. The population we sample are displaced Rohingyas based in the Cox’s Bazzar camp and local Bangladeshi residents. Along with a standard household survey including questions on demographics and economic conditions, the data include transcripts of 2,407 open-ended interviews with subjects on their aspirations for their eldest child. The interviews were conducted either in Bengali or in Rohingya which was then transcribed into Bengali, but we work with machine translations into English. The interviews take the form of an unstructured to-and-fro of question and answer (QA) pairs the interviewer and the subject. The interviews are on average 12.6 QA pairs long, with the average answer in each QA pair being 13.7 words long. Based on a close reading of a subset of transcripts, and following a "flexible coding" process (Detering and Waters, 2018), a coding tree was developed including 25 potentially overlapping categories, 19 of which we focus on in this paper. A full description of each code along with examples are shown in Appendix A. Following Callard (2018) the distinction between aspiration and ambition was adapted within the context and nature of "dreams" parents expressed for their children. For example, concrete and measurable dreams for child (e.g wishing a child would become a doctor, teacher, entrepreneur, or specific educational goals) was used as a definition for ambition while intangible, value oriented goals (e.g wishing the child to live with dignity or be a good human being) was classified as aspiration. Aspirations, were divided into "Religious" and "Secular" . Ambition was divided into seven major categories – Education (further sub-coded into High, Low, Neutral and Religious), Job Secular, Marriage, Entrepreneurship, Migration, Vocational Training, and No Ambition. While ambition and aspiration came up at any point in an interview, "capacity to aspire" or Navigational capacity was restricted to discussions of what parents were planning or able to do to fulfill dreams for their children. Navigational Capacity was coded into seven sub-codes – Low and High "Ability", Low and High "Budget", Low and High "Information Awareness", and Reliance on God. Of our sample of 2,407 interview transcripts, 789 are manually annotated by trained sociologists (co-authors on the Ashwin et al. (2022) paper) according to this coding structure. The annotations are defined at the level of QA pairs, allowing us to represent each annotation as a binary classification problem at the QA level. 3 Methods In this Section we first explain how we use LLMs to annotated our interview transcripts. We then briefly describe the iQual method which trains supervised models on our expert human annotations, as well as how we use LLMs for data augmentation in combination with iQual. We test three different LLMs- the closed-source ChatGPT (gpt-3.5-turbo) by OpenAI, as well as two open-source LLMs by Meta, the Llama-2 (13b) and its fine-tuned "chat" variant (Touvron et al., 2023).The base Llama-2 is pretrained on publicly available online data sources. The chat variant is then fine-tuned on publicly available instruction datasets and over 1 million human annotations. This fine-tuning is designed to make the model align with human preferences using techniques such as Reinforcement Learning with Human Feedback (RLHF). ChatGPT is also fine tuned using RLHFFor all three LLMs, our approach 3 Figure 1: LLM instructions example Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s future education and careers solely in the context of religion, without implying any monetary incentives. Specifically, your role is to discern whether the parent explicitly expresses a desire for their child to pursue a religious path. Assign the “Religious Aspirations” label only if the parent explicitly articulates aspirations for their child that include becoming a religious scholar, Hafiz, attending a madrassa, learning Arabic, Quran reading, Islamic covering, regular prayer, working in Islamic banks, or being a Islamic scholar. It is crucial not to assign this label if the child is already engaged in any of the mentioned activities, as current religious activities are not considered future aspirations. If there is no such direct reference to religious aspirations in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of religious aspirations was not addressed. Here are some examples: ### Input: Interviewer: What are your aspirations for your son’s future? Respondent: I hope that he will become a great Islamic scholar. I want him to learn Arabic and read the Quran regularly. ### Output: Interpretation: The parent explicitly expresses a desire for their child to pursue a path in religious education, specifically mentioning becoming an Islamic scholar, learning Arabic, and reading the Quran regularly. Therefore, this conversation aligns with the ‘Religious Aspirations’ label. Label: Religious Aspirations ### Input: Interviewer: Do you have any specific dreams for your daughter? Respondent: I want her to excel in her studies and become a successful engineer. ### Output: Interpretation: While the parent does express aspirations for their child’s future, these aspirations are not religious in nature. The parent hopes for their child to become an engineer, which doesn’t involve any religious activities as per the instruction. Therefore, this conversation should be labeled as ’Not Applicable’. Label: Not Applicable ### Input: Interviewer: What do you envision for your child’s future? Respondent: I would like my son to become a Hafiz. It’s a great honor in our religion. ### Output: Interpretation: The parent explicitly expresses a desire for their child to become a Hafiz, which is a religious aspiration as per the instruction. Therefore, this conversation aligns with the ‘Religious Aspiration’ label. Label: Religious Aspirations ### Input: Interviewer: What are your hopes for your children’s future careers? Respondent: I want them to have successful careers. I hope my daughter becomes a doctor and my son becomes a lawyer. ### Output: Interpretation: While the parent does express aspirations for their children’s future, these aspirations do not involve any religious activities or pursuits as per the instruction. Therefore, this conversation should be labeled as ‘Not Applicable’. Label: Not Applicable ### Input: Interviewer: What is your dream about your son who is studying in class one? Respondent: Still in Noorani. After reading Noorani, he will read Hafezi. By the grace of Allah, he will teach in a madrasa when he becomes a hafez. That is my dream. ### Output: (ChatGPT) Interpretation: The respondent expresses a clear religious aspiration for their child, which is to become a Hafiz and teach in a madrasa. There is no mention of any monetary incentives, so this response aligns with the criteria for "Religious Aspiration". Label: Religious Aspiration ### Output: (Llama-2-13b) Interpretation: The parent explicitly expresses a desire for their child to pursue a path in religious education, specifically mentioning becoming a Hafiz and teaching in a madrassa. Therefore, this conversation aligns with the "Religious Aspiration" label. Label: Religious Aspiration ### Output: (Llama-2-13b-chat) Interpretation: The parent explicitly expresses a desire for their child to pursue religious studies and become a Hafiz (one who has memorized the Quran). They also mention teaching in a madrasa, which is an Islamic school. Therefore, this conversation aligns with the "Religious Aspiration" label. Label: Religious Aspiration Note : This Figure shows example instructions given to the LLMs for the "Aspirations: Religious" code. These results are given to the LLM as a single prompt, but for clarity we have grouped the instructions into three parts here, as indicated by the colors. The first section (in green) describes the task and gives the definition of the code. The labelled examples (in blue) help the LLM understand the task at hand, often referred to as "few shot learning". For each example, we also provide a reasoning for the annotation (in red) so that the LLM is also asked to explain why it applies a certain label, which is known as "chain of thought" prompting. The second box shows an example of a QA pair to be annotated and the subsequent boxes show the responses of the three LLMs to this prompt. 4 to prompting remains consistent. 3.1 Annotation with LLMs We follow several well-established practices to improve the effectiveness of LLMs in annotating our interview transcripts. We provide a prompt that includes precise directives for the LLM, and employ "few-shot learning" (Brown et al., 2020) as well as "chain of thought" prompting (Wei et al., 2022), as explained below. For each code, we created detailed textual instructions, similar to those one would give to human annotators. These instructions include enough details to ensure that, in principle, the model is fully aware of the specific standards and definitions required for coding transcripts. Each code and each question-answer pair are annotated by the LLM independently. By incorporating both few-shot learning and CoTP, we are in line with best practices and give the LLMs a good chance at annotating accurately. Previous work has shown that using these techniques can help LLM out- perform crowd workers in text annotation tasks (Gilardi et al., 2023). The choice of prompt given to an LLM when giving it a certain task can make a substantial difference to its performance. We give the models a thorough briefing of what each code represents and how to identify its presence in a conversation. This includes the context, certain specific terms or activities that could be indicators, and the need to distinguish between current circumstances and future aspirations, as shown in the green text of Figure 1. These instructions provide a benchmark for the model to understand the coding system and thereby infer the respective codes from the interview transcripts. Instructions for each of the codes are shown in Appendix A. Few-shot learning and chain of thought prompting (CoTP) are two powerful techniques that can be combined to improve the performance and interpretability of LLMs. Few-shot learning provides examples of a task to the model, which helps guide its behavior and understanding of the task at hand. We provide the LLM with four examples that follow the detailed instructions, as shown by the blue text in Figure 1, to demonstrate correct behavior to the model. These examples are chosen to be instructive of the how the code should be applied and are similar to the examples one would use to explain a code in traditional qualitative analysis. We also apply chain of thought prompting (CoTP) in these examples to nudge the model to generate an interpretation of the transcript and articulate its line of reasoning before assignment of the final code. It is beneficial in complex tasks where reasoning and interpretation play crucial roles, such as our coding task. The underlying idea is that by having the model outline its thinking process, we can encourage it to reason more deeply and accurately, while also producing outputs that are more interpretable and trustworthy. For our task, we have used both few-shot learning and CoTP by asking the model not only to provide a label for each transcript, but also to give an interpretation explaining why it chose that label. An example of a full prompt for the ’Religious Aspirations’ code are shown in Figure 1, with the instructions and few shot examples for all codes shown in Appendix A. 3.2 Training supervised models on interpretative annotations (iQual) An alternative to using LLMs to annotate large corpora of text documents would be to create high quality annotations on a smaller sub-sample and then training supervised models to predict these annotations on the remainder of the documents. We thus train a separate classifier for each code on a numerical representation of the text at the QA level. As discussed in Appendix B, there are many options for both the classifier we can use here (e.g. random forest, logistic regression, neural networks, SVM), as well as how to represent the text numerically (e.g. tf-idf ngram vectors, sentence embeddings, translations or transliterations). Using k-fold cross-validation we select the best performing model, the text representation and a variety of hyperparameters, so that the approach which performs best in out-of-sample prediction is selected. In each case, we hold out a test set of 200 interviews in order to assess out-of-sample performance. Details about this methodology are provided in Ashwin et al. 5 (2022), and it is implementable in an open source Python package.1 The crucial intuition though is simply that we use a subset of high quality expert annotations to train a small bespoke model for each code. These models rely only on the annotated training data, unlike the pre-trained LLMs which are trained on huge quantities of text from a huge range of contexts. Rather than asking LLMs to directly annotate text, another potential use for them is for data augmentation in combination with a supervised model, such as those described above. Data augmentation is a common technique in machine learning to generate more variation in a training set while preserving the important signals. For example, when training a model on a labelled dataset of images of animals, one might generate extra variation in the training data by rotating the images by 90 degrees or transforming them into a mirror image of themselves. The idea is to generate more training observations where the noise in the data is different but the signals are the same. A good example of this from the natural language processing literature is back-translation, where text is translated into a different language and then back into the original, so that the exact phrasing and style of the text is different but the meaning is the same (Edunov et al., 2018). Using LLMs for data augmentation has been found to increasing prediction performance in some contexts, so we follow the approach set out in Dai et al. (2023) as an additional experiment here. The example prompts and further details on the augmentation are shown in Appendix A.1. We thus test two different versions of iQual: first, training supervised models on the human annotations without the use of LLMs, and second, training the model on data augmented by the LLMs to generate more variation in the text while preserving the meaning. 4 Results We assess the performance of LLMs in our annotation tasks along two dimensions. Firstly, we assess how accurate of the LLMs predictions relative to our expert human annotations, finding that performance is poor relative to our simpler supervised models. Secondly, and more importantly, we investigate whether the annotations provided by LLMs or iQual introduce bias. We here mean bias in the technical sense that the prediction errors which the models make are not random. 1 https://github.com/worldbank/iQual 6 Figure 2: Out-of-sample prediction performance of different methods Aspiration: Religious Aspiration: Secular Ambition: No Ambition Ambition: Job Secular Ambition: Vocational Training Ambition: Entrepreneur Model Ambition: Education High ChatGPT Ambition: Education Neutral Llama−2 13B chat Ambition: Education Low Ambition: Education Religious Llama−2 13B Ambition: Marriage iQual + ChatGPT aug Ambition: Migration iQual Capacity: Reliance On God Capacity: Ability High Random Capacity: Ability Low Capacity: Budget High Capacity: Budget Low Capacity: Awareness Information High Capacity: Awareness Information Low 0.00 0.25 0.50 0.75 Out−of−sample F1 score Note : This Figure compares the out-of-sample prediction performance of LLM and supervised approaches, compared to the expert human annotations. Each code is shown along the vertical axis, and the test set F1 scores are shown on the horizontal axis. The F1 score that would be achieved by random guessing is shown as the black triangle and all models perform better than this. The performance of each model for each code is shown as a separate point with the color and shape of the point denoting the model. Averaging the F1 scores across all codes, iQual performs best with 0.542, followed by iQual + ChatGPT aug (0.541), ChatGPT (0.414), Llama-2 13B (0.290) and finally Llama-2B chat (0.274). 4.1 Out-of-sample Performance Given that the interview transcripts are annotated with a series of binary variables at the QA level, we can assess LLM prediction accuracy with the out-of-sample F1 score for each code. We compare the performance of each LLM to the supervised models trained on annotated data, with and without augmentation. Figure 2 shows the results on the comparative performance of the different annotation approaches, as measured by the test set F1 score.2 Given that many of the codes are very sparse, a useful comparison is the F1 score that random guessing would achieve, which is shown as black triangles. All models across all codes do better than random (i.e. have a higher F1 score than that which random guessing would achieve). In all but one case (Capacity: Awareness Information High) ChatGPT is the best performing LLM. However, in all but one case (Capacity: Budget Low) all LLMs perform worse than iQual in terms of these F1 scores. When used for augmentation, ChatGPT does improve performance slightly in some cases, but it slightly worsens performance just as often. If we measure performance in terms of accuracy (i.e. the proportion of observations that are correctly classified) rather than F1 score we get the same results. iQual achieves accuracy of 0.969. In contrast, ChatGPT only achieves 0.909, Llama-2 13B 0.854 abd Llama-2 13B chat 0.851. These results are of course specific to our context, and a different annotation structure on a different set of text data may lead to different results. However, in our case it is clear that LLMs generate less accurate annotations than training much smaller models on a subset of human annotations does. 2 The F1 score is the harmonic mean of the precision and recall, where precision is the number of true positive divided by the sum of true positives and false positive, and recall is the number of true positive results divided by the sum of true positives and false negatives. It thus symmetrically represents both type 1 and type 2 errors. 7 4.2 Bias If the annotations generated by LLMs are inaccurate, this is not necessarily a hugely consequential problem. If the mistakes they make are random, with a large enough sample we should still be able to come to correct conclusions. However, if the mistakes are not random, then using LLM annotations can lead to completely incorrect conclusions. In other words, if the LLMs errors are biased, then relying on these annotations could lead researchers to identify relationships in the data that are purely a result of these algorithmic biases rather than reality. We look at two ways in which the predicted annotations could be biased. Firstly, and most straightforwardly, we show that LLMs over-predict annotations that are very sparse (i.e. there are many more false positives than false negatives). Secondly, we show that in many cases LLM prediction errors are systematically associated with characteristics of the interview subject (e.g. refugee status, gender, education). Figure 3 shows the degree of over-prediction across different annotations. Each model is shown as a separate panel and the bars show the degree of over-prediction as a percentage of all answers. All three LLMs we tested systematically over-predict most of the annotations. This is a problem in itself, as we might be interested in the prevalence of a particular concept, but it is especially problematic if we want to compare the prevalence of different annotations. For example, if we wished to compare the prevalence of secular and religious aspirations in our sample, using the annotations provided by ChatGPT would lead us to very misleading conclusions. While ChatGPT over-predicts both the "Aspirations: Secular" and "Aspirations: Religious" codes, as can be seen from the uppermost two rows of Figure 3, "Aspirations: Secular" is over-predicted much more frequently than "Aspirations: Religious"; in the expert human annotations "Aspirations: Secular" appears around 1.2 times more frequently than "Aspirations: Religious", while in the ChatGPT annotations "Aspirations: Secular" appears around 3 times more frequently than "Aspirations: Religious". Figure 3: LLMs systematically over predict annotations ChatGPT Llama−2 13B chat Llama−2 13B iQual + ChatGPT aug iQual Aspiration: Religious Aspiration: Secular Ambition: No Ambition Ambition: Job Secular Ambition: Vocational Training Ambition: Entrepreneur Model Ambition: Education High Ambition: Education Neutral ChatGPT Ambition: Education Low Llama−2 13B chat Ambition: Education Religious Ambition: Marriage Llama−2 13B Ambition: Migration iQual + ChatGPT aug Capacity: Reliance On God iQual Capacity: Ability High Capacity: Ability Low Capacity: Budget High Capacity: Budget Low Capacity: Awareness Information High Capacity: Awareness Information Low 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 Over−prediction (percentage of answers) Note : This Figure shows the average percentage of answers in which each model over or under predicts each annotation. Each model is shown as a separate panel, with each code shown along the vertical axis and the percentage of answers in which there is an net over-prediction is shown on the horizontal axis. A score of 50% thus means that half of all observations are a false positive. If the value is positive, then the model assigns the annotation too frequently while if it is negative then the model doesn’t assign the annotation frequently enough. The LLM models systematically over-predict most of the annotations. Of perhaps even greater concern than over-prediction we find that the LLM’s predictions are systematically biased with respect to the interview subjects’ characteristics (e.g. refugee status, demographics, education and income). To test whether prediction errors are systematically related to subject characteristics, we regress prediction errors for each model on a range of subject characteristics. We then calculate the F statistic of this regression, which tells us whether there is some statistically significant relationship between the prediction errors and subject characteristics (e.g. a model might 8 over-predict a certain code for men but under-predict for women). Figure 4: LLM models fail bias test much more regularly that iQual ChatGPT Llama−2 13B chat Llama−2 13B iQual + ChatGPT aug iQual Aspiration: Religious Aspiration: Secular Ambition: No Ambition Ambition: Job Secular Ambition: Vocational Training Ambition: Entrepreneur Ambition: Education High Ambition: Education Neutral p−value Annotation Ambition: Education Low <1% Ambition: Education Religious Ambition: Marriage <5% Ambition: Migration >5% Capacity: Reliance On God Capacity: Ability High Capacity: Ability Low Capacity: Budget High Capacity: Budget Low Capacity: Awareness Information High Capacity: Awareness Information Low −1 0 1 2 −1 0 1 2 −1 0 1 2 −1 0 1 2 −1 0 1 2 log F statistic Note : This Figure shows the result of an F-test for a statistical association between the prediction errors of each model with the characteristics of the interview subject. Each model is shown as a separate panel, with each code shown along the vertical axis. The log F statistic of this test is shown along the horizontal axis with the color of the points indicating the statistical significance of the test statistic. The subject characteristics include refugee status; age and sex of eldest child; age, education and sex of interview subject, "refugee"; total number of children; household assets and income; and history of trauma experience. The LLM models display a bias much more frequently that the supervised models. The full results for each regression are shown in Appendix C Figure 4 shows these F statistics that test whether the prediction errors of each annotation approach are systematically related to the interview subjects’ characteristics, with the full regression in each case reported in Appendix C. The higher the (log) F statistic is, the stronger the evidence of bias. The color of the points indicates the level of statistical significance and each model is shown as a separate panel. We see that for iQual in the left-most panel, there is evidence of bias in only one of the 19 codes. While we should be cautious in interpreting results with this code, there is not much cause for concern. However, for the LLMs we find strong evidence of bias in many of the codes. This tells us that the prediction errors the LLMs make are not random and conducting analysis on the basis of its predictions is likely to result in misleading interpretations. The F tests shown in Figure 4 tell us that there is some statistical association between prediction errors and subject characteristics. We can see concrete examples of how this can lead to misleading conclusions in Figure 5. This Figure shows estimated coefficients for regressions of the prevalence of an annotation in an interview on dummy variables for the subjects’ refugee status and the gender of their eldest child. So if the coefficient on refugee status is positive than this code appears more in interviews with refugees than in interviews with hosts. Six of the codes are shown as separate panels and the results based on the annotations of the five different models are shown for each coefficient, following the same color scheme as Figures 2 and 3, but with the coefficient based on only the expert annotations shown in black. The differences in the estimates across annotation methods are because the errors they make are not random with respect to refugee status and the gender of the eldest child. 9 Figure 5: Examples of misleading conclusions when using LLM Aspiration: Secular Capacity: Awareness Information High Capacity: Awareness Information Low Refugee status Male eldest child Model Expert annotations iQual 50 25 0 5 −0 0 25 0 5 0 4 3 2 1 00 01 00 02 05 00 02 05 .0 .0 .0 .0 .0 .0 .0 0. 0. −0 −0 −0 −0 0. 0. 0. 0. 0. 0. −0 −0 iQual + ChatGPT aug Ambition: Education High Ambition: Education Neutral Ambition: Education Low ChatGPT Llama−2 13B Llama−2 13B chat Refugee status Male eldest child 50 25 0 5 0 2 00 02 04 06 1 00 01 02 00 02 05 .0 .0 .0 .0 0. 0. 0. 0. 0. 0. 0. −0 −0 0. 0. 0. −0 −0 Coefficient Note : This Figures shows the estimated coefficients for regressions of the prevalence of a code in an interview on dummy variables indicating the subjects’ refugee status and the gender of their eldest child. Codes are shown as separate panels, with the error bars represent 95% confidence intervals and color indicates which approach was used to generate the annotations. The coefficient for a regression estimated on just the Coefficients for all codes are shown in Figure 7 We see in Figure 5 that in many cases the coefficient based on the ChatGPT predicted annotations (in pink) is very different from that based on the true human annotations (in black). For example, for "Ambition: Education Low" in the bottom right panel, the coefficient on refugee status is positive and significant using the expert annotations but negative and insignificant on the ChatGPT annotations. The coefficient on whether the eldest child is male meanwhile is negative and insignificant on the human annotations but positive and significant using the ChatGPT annotations. This is just one example, but we can see here how relying on the LLM annotations can lead to potentially dangerous misunderstandings. For example, based on the ChatGPT annotations we might conclude that subjects are more likely to have low educational ambitions for their male children, while in the expert annotations there is no evidence for that. In fact once we increase the sample size using iQual we find a marginally significant effect of the opposite sign. We can also note here that the coefficients using iQual are not different from those using just the expert annotations, but have much smaller standard errors. Using supervised models to scale up expert human annotations thus increases precision while not introducing bias, as argued in Ashwin et al. (2022). Using ChatGPT to extend the sample size through data augmentation does not appear to introduce additional bias, although it does not have a substantial benefit either. Given that some expert annotations will be necessary in order to identify whether LLM (or crowd sourced) annotations are biased, this suggests that training smaller bespoke models on these annotations may be more reliable than relying on LLMs to annotate large samples. 5 Discussion LLMs are trained on a wide range of text and consequently may not be suited for nuanced and context-specific tasks. First, they may introduce systematic biases when used to annotated text. In our example, we see that the errors that LLMs make in annotations (compared to expert human 10 annotations) are not random. Second, LLMs over-predict many of our codes. We can think of this as the LLM bringing the "pre-conceptions" it has learned from its training data to the annotation task. Consequently, LLMs are probably not suited for coding most qualitative data that requires nuanced and contextual analysis. This sort of analysis has traditionally been the province of anthropology and sociology but is increasingly being used by more quantitative fields such as economics and political science. To analyze large-N qualitative data, such as those obtained from open-ended in-depth interviews, we show that a coding scheme based on a close-reading of transcripts by experts in qualitative analysis with a sub-sample of the full corpus of data is vital for interpretation and analysis. Firstly, high quality annotations are necessary in order to assess whether the LLM is introducing bias in its annotations. Secondly, these high quality annotations can then be used as a training set for smaller bespoke models. These bespoke models may be able to leverage LLMs through data augmentation, but importantly they are trained on context-specific data so researchers have better control, and an overview, of the information that is used. We suspect that these limitations will continue even as LLMs improve, and we encourage researchers using LLMs for annotation tasks to be aware of and check for bias. References Appadurai, A. (2004), ‘The capacity to aspire: Culture and the terms of recognition’, Culture and Public Action, ed. Vijayendra Rao and Michael Walton, Stanford, California: Stanford University Press pp. 59–84. Ashwin, J., Rao, V., Biradavolu, M., Chhabra, A., Haque, A., Krishnan, N. and Khan, A. (2022), ‘A method to scale-up interpretative qualitative analysis, with an application to aspirations in cox’s bazaar, bangladesh’. Atari, M., Xue, M. J., Park, P. S., Blasi, D. and Henrich, J. (2023), ‘Which humans?’, https://doi.org/10.31234/osf.io/5b26t . Bonikowski, B. and Nelson, L. K. (2022), ‘From ends to means: The promise of computational text analysis for theoretically driven sociological research’, Sociological Methods & Research 51(4), 1469– 1483. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A. et al. (2020), ‘Language models are few-shot learners’, Advances in neural information processing systems 33, 1877–1901. Callard, A. (2018), Aspiration: The agency of becoming, Oxford University Press. Dai, H., Liu, Z., Liao, W., Huang, X., Wu, Z., Zhao, L., Liu, W., Liu, N., Li, S., Zhu, D. et al. (2023), ‘Chataug: Leveraging chatgpt for text data augmentation’, arXiv preprint arXiv:2302.13007 . Detering, N. M. and Waters, M. (2018), ‘Flexible coding of in-depth interviews: A twenty-first century approach’, Sociological Methods and Research 50(2)), 708–738. Edunov, S., Ott, M., Auli, M. and Grangier, D. (2018), ‘Understanding back-translation at scale’, arXiv preprint arXiv:1808.09381 . Gilardi, F., Alizadeh, M. and Kubli, M. (2023), ‘Chatgpt outperforms crowd-workers for text- annotation tasks’, Proceedings of the National Academy of Sciences 120(30), e2305016120. Kearns, M. and Roth, A. (2020), The Ethical Algorithm, Oxford University Press. 11 Mellon, J., Bailey, J., Scott, R., Breckwoldt, J. and Miori, M. (2022), ‘Does gpt-3 know what the most important issue is? using large language models to code open-text social survey responses at scale’, Using Large Language Models to Code Open-Text Social Survey Responses At Scale (December 22, 2022) . Rao, V. (2023), Can economics become more reflexive? exploring the potential of mixed-methods, in ‘Handbook on the Economics of Discrimination and Affirmative Action, A. Deshpande Editor’, Springer. Small, M. L. and Calarco, J. M. (2022), Qualitative Literacy: A Guide to Evaluating Ethnographic and Interview Research, University of California Press. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E. and Lample, G. (2023), ‘Llama: Open and efficient foundation language models’. Veselovsky, V., Ribeiro, M. H. and West, R. (2023), ‘Artificial artificial artificial intelligence: Crowd workers widely use large language models for text production tasks’, arXiv preprint arXiv:2306.07899 . Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D. et al. (2022), ‘Chain-of-thought prompting elicits reasoning in large language models’, Advances in Neural Information Processing Systems 35, 24824–24837. 12 A Prompts for Classification 13 High Ability System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their demonstrated capacity to support their children’s education and future aspirations. Specifically, your role is to discern whether the parent explicitly expresses any steps they have taken that go above and beyond their capacity to ensure a better future for their child. Assign the “High Ability” label only if the parent explicitly articulates measures that they’ve taken that could be considered extraordinary efforts to secure their child’s future. This can include, but is not limited to, investments in education, personal sacrifices, or unique strategies. Keep in mind that financial capacity is not the sole determinant of “High Ability”; parents can still be considered high ability even if they have a limited budget. If there is no such direct reference to high ability in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of high ability was not addressed. Few-shot examples Input: Input: Interviewer: How are you supporting your child’s Interviewer: What are you doing to ensure your ambition to become a scientist? daughter’s future? Respondent: I’ve been saving up to send him to a Respondent: I have been working two jobs to save up science summer camp every year since he was 10. I for her college education. I also spend my free time also spend time with him doing experiments from his helping her with her homework and encouraging her science kit every weekend. to participate in extracurricular activities that will enhance her college application. Output: Interpretation: The parent is making significant Output: efforts to support their child’s ambition, including Interpretation: The parent is making personal saving money for summer camps and dedicating sacrifices, such as working multiple jobs and personal time for science experiments. This dedicating personal time to support their child’s demonstrates a high ability to support their child’s education. This shows a high ability to secure their education and future aspirations. child’s future. Label: High Ability Label: High Ability Input: Input: Interviewer: How do you plan to help your son Interviewer: What are you doing to help your achieve his dreams? daughter become a doctor? Respondent: I want him to be happy and successful. Respondent: I am doing my best to support her. I I will support him in whatever he chooses to do. always tell her to study hard. Output: Output: Interpretation: While the parent expresses a desire to Interpretation: Although the parent expresses a desire support their child’s dreams, they do not specify any to support their child’s ambition, they do not specify extraordinary measures they are taking to ensure a any extraordinary measures they are taking to ensure better future for their child. Therefore, this dialogue a better future for their child. Therefore, this dialogue does not demonstrate high ability. does not demonstrate high ability. Label: Not Applicable Label: Not Applicable 14 Low Ability System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their expressed capacity to support their children’s education and future aspirations. Specifically, your role is to discern whether the parent explicitly indicates any limitations or constraints in their means to aid their child’s future. Assign the “Low Ability” label only if the parent explicitly states that they have limited or no means (e.g., food, income, resources) to support their child’s education and future aspirations. If there is no such direct reference to low ability in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of low ability was not addressed. Few-shot examples Input: Input: Interviewer: What are your aspirations for your Interviewer: Do you think your children will be able child’s education? to pursue their dreams? Respondent: I wish I could give them the best Respondent: They are very ambitious and I’m trying education possible, but I barely make enough to feed my best to support them. them. Output: Output: Interpretation: Although the parent mentions their Interpretation: The parent expresses a direct effort to support their child’s dreams, there is no limitation in their ability to support their child’s explicit reference to the limitations or constraints in education due to lack of sufficient income. This aligns their means to do so. Hence, the “Low Ability” label with the criteria for the “Low Ability” label. is not applicable. Label: Low Ability Label: Not Applicable Input: Input: Interviewer: How are you planning to support your Interviewer: Are you able to provide the resources child’s dream of becoming a doctor? your child needs for his studies? Respondent: I would love to, but we are struggling Respondent: Yes, we have managed to provide them to make ends meet. I can’t afford to pay for their with the necessary books and materials. higher education. Output: Output: Interpretation: The parent confirms their ability to Interpretation: The parent directly states their provide resources for their child’s education. There inability to support their child’s future aspirations due is no explicit reference to limitations or constraints in to financial constraints. This meets the criteria for the their means to support their child’s future. Therefore, “Low Ability” label. the “Low Ability” label is not applicable. Label: Low Ability Label: Not Applicable 15 High Awareness System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their awareness and understanding of the means to achieve their aspirations for their children’s futures. Specifically, your role is to discern whether the parent demonstrates a comprehensive and logical understanding or has the capacity to describe the means of reaching their dreams for their child. Assign the “High Awareness” label only if the parent demonstrates in their statements a clear understanding and awareness of the processes, resources, and strategies necessary to achieve their aspirations for their child’s future. The assignment of this label is inferential, based on your analysis of their comprehension, rather than explicit statements. If there is no evidence of high awareness in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of high awareness was not addressed. Few-shot examples Input: Interviewer: What do you envision for your son’s future? Respondent: I want my son to become a doctor. I’ve Input: already started saving for his medical school fees Interviewer: How are you preparing your child for and we spend extra time every day working on his their future career? science and math homework. I also plan to get him Respondent: I believe in my son’s capabilities. He’s into extracurricular activities that will enhance his smart and he’ll figure it out. application for medical school. Output: Output: Interpretation: The respondent does not provide any Interpretation: The respondent demonstrates a clear specific strategies, resources, or processes to prepare understanding of the steps necessary to achieve their their child for their future career. They express faith aspiration for their child’s future, which involves in their child’s abilities but do not demonstrate a high him becoming a doctor. They have financial plans level of awareness about the steps necessary to achieve in place, focus on relevant academic subjects, and their career goals. have considered the importance of extracurricular Label: Not Applicable activities. This indicates a high level of awareness about the process. Label: High Awareness Input: Interviewer: What are your dreams for your daughter’s future? Input: Respondent: My dream is for her to become a Interviewer: What are your thoughts about your successful engineer. We are ensuring she excels in daughter’s future? her math and science classes, and we’ve enrolled Respondent: I want her to be independent and happy. her in a robotics club to give her early exposure to engineering concepts. We’re also exploring Output: scholarship opportunities for engineering programs. Interpretation: While the respondent has a general aspiration for their child’s future, they do not provide Output: any specific strategies, resources, or processes to Interpretation: The respondent demonstrates a achieve this aspiration. Therefore, it does not clear understanding and awareness of the processes, demonstrate a high level of awareness about the steps resources, and strategies necessary to achieve their necessary to achieve this goal. aspiration for their child’s future. They have identified Label: Not Applicable relevant academic subjects, extracurricular activities, and potential financial resources, indicating a high level of awareness about the process. Label: High Awareness 16 Low Awareness System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their awareness and understanding of the means to achieve their aspirations for their children’s futures. Specifically, your role is to discern whether the parent demonstrates a lack of knowledge, uncertainty, or confusion regarding how to help their child achieve their aspirations. Assign the “Low Awareness” label only if the parent demonstrates in their statements a lack of understanding, cluelessness, or uncertainty about the processes, resources, and strategies necessary to achieve their aspirations for their child’s future. The assignment of this label is inferential, based on your analysis of their comprehension, rather than explicit statements. If there is no evidence of low awareness in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of low awareness was not addressed. Few-shot examples Input: Input: Interviewer: What are your plans to ensure your Interviewer: What are your dreams for your children’s daughter becomes a doctor, which you mentioned is future? her dream? Respondent: I just want them to be happy and Respondent: I don’t really know. I suppose she will successful. I’m not sure how to make that happen, have to study a lot. though. Output: Output: Interpretation: The respondent shows uncertainty Interpretation: Although the respondent expresses about the specific steps and resources required to aspirations for their children’s future, they help their daughter achieve her dream of becoming acknowledge their uncertainty about how to a doctor. This indicates a lack of awareness about the help them achieve success. This suggests a lack processes involved, hence the “Low Awareness” label of awareness about the necessary strategies and is assigned. resources, warranting the “Low Awareness” label. Label: Low Awareness Label: Low Awareness Input: Interviewer: How do you plan to help your children Input: achieve their dreams? Interviewer: How are you supporting your son’s Respondent: I believe education is the key. I’m doing ambition to be an engineer? my best to save money for their schooling and to Respondent: I’m not sure how to help him. I didn’t provide them with a conducive learning environment receive much education myself. at home. Output: Output: Interpretation: The respondent admits to not Interpretation: The respondent demonstrates an knowing how to assist their son in achieving his understanding of the importance of education in ambition of becoming an engineer. This demonstrates achieving their children’s dreams and has a plan to a lack of understanding about the necessary steps and save money for their schooling. This indicates an resources, thus the “Low Awareness” label is assigned. awareness of the necessary resources and strategies, Label: Low Awareness so the “Low Awareness” label is not applicable. Label: Not Applicable 17 High Budget System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their financial capacity to support their children’s education and future aspirations. Specifically, your role is to discern whether the parent explicitly demonstrates having substantial financial resources, either through an ability to save or spend on significant items (e.g., setting up a shop) or services like private tuition or coaching centers. Assign the “High Budget” label only if the parent explicitly articulates possessing or having access to substantial financial resources that they are able or willing to utilize for their child’s education and future aspirations. If there is no such direct reference to high budget in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of high budget was not addressed. It is important to note that the presence of a parent-child relationship in the conversation, or the discussion of the child’s future unrelated to high budget, does not automatically warrant the “High Budget” label. Only discussions that directly and explicitly relate to the parent’s financial capacity in supporting their child’s future should receive this label. Few-shot examples Input: Input: Interviewer: What are your plans for your daughter’s Interviewer: How are you helping your children with education? their studies? Respondent: I am planning to send her to a private Respondent: We are managing somehow. I try to school. I have also arranged for a private tutor to teach them myself when I get time after work. help her with her studies at home. Output: Output: Interpretation: The respondent doesn’t mention Interpretation: The respondent explicitly mentions any significant financial investment towards their that they have the financial capacity to send their children’s education, indicating that the topic of high child to a private school and afford a private tutor, budget is not addressed. indicating a high budget. Label: Not Applicable Label: High Budget Input: Input: Interviewer: How do you support your son’s dream Interviewer: What are your future plans for your to become a scientist? child’s education? Respondent: I have enrolled him in a science coaching Respondent: I want to send him to a good college, center and I am also planning to buy him a personal but I’m not sure if I can afford it. computer for his research work. Output: Output: Interpretation: The respondent mentions a desire Interpretation: The respondent mentions that they to support their child’s education but also expresses are able to afford enrollment in a coaching center and uncertainty about their financial capacity to do so, a personal computer for their child, indicating a high indicating that the topic of high budget is not budget. addressed. Label: High Budget Label: Not Applicable 18 Low Budget System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their financial capacity to support their children’s education and future aspirations. Specifically, your role is to discern whether the parent explicitly indicates a limited financial capacity, directly related to a lack of funds, regardless of their willingness to allocate a budget for their child’s education. Assign the “Low Budget” label only if the parent explicitly articulates a scarcity or lack of financial resources that they have available for their child’s education and future aspirations. If there is no such direct reference to low budget in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of low budget was not addressed. Few-shot examples Input: Input: Interviewer: What are your dreams for your children’s Interviewer: How do you plan to support your child’s future? education? Respondent: I want them to get a good education Respondent: We’re doing our best to save money, and have a successful career, but I’m not sure how but it’s tough. We might have to consider taking out we’ll afford it. We’re barely making ends meet as it is. a loan. Output: Output: Interpretation: The respondent explicitly expresses a Interpretation: The respondent mentions the concern about their financial capacity to support their difficulty in saving money and the consideration child’s education. The respondent is uncertain about of a loan, which suggests a struggle with financial how to afford the costs of their children’s education, constraints. This indicates a lack of financial which indicates a lack of financial resources. resources to support their child’s education. Label: Low Budget Label: Low Budget Input: Input: Interviewer: What are your plans for your child’s Interviewer: How do you see your child’s future? education? Respondent: I see my child becoming a doctor. We’re Respondent: We’ve been exploring different schools doing all we can to support this dream. and educational programs. We want the best for our child. Output: Interpretation: The respondent expresses a hopeful Output: vision for their child’s future and their commitment Interpretation: The respondent talks about exploring to support it. However, there is no explicit mention different education options for their child, indicating a of financial constraints or limitations in their ability proactive approach. There is no explicit mention of a to support their child’s education. lack of financial resources or constraints in supporting Label: Not Applicable their child’s education. Label: Not Applicable 19 High Education System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their educational aspirations for their children’s futures. Specifically, your role is to discern whether the parent explicitly expresses an intention or desire for their child to achieve a high level of education, either secular or religious, regardless of whether the level is explicitly specified. Assign the “High Education” label only if the parent explicitly articulates a desire or intention for their child to pursue a high level of education in the future. If there is no such direct reference to high education in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of high education was not addressed. Few-shot examples Input: Input: Interviewer: What are your aspirations for your Interviewer: Do you have any specific goals for your children’s future? children’s education? Respondent: I have always dreamt of them becoming Respondent: Yes, I want them to study hard and successful in their respective fields. I want my eldest excel in their academics. I hope they can attend son to pursue medicine and my daughter to become university and get a good degree. an engineer. Output: Output: Interpretation: The respondent explicitly states Interpretation: The respondent clearly articulates a a desire for their children to attend university desire for their children to pursue careers in fields and achieve a degree, which is considered high that require high education, such as medicine and education. Therefore, the dialogue aligns with the engineering. Therefore, the dialogue aligns with the “High Education” label. “High Education” label. Label: High Education Label: High Education Input: Input: Interviewer: What are your plans for your children’s Interviewer: What are your dreams for your children’s education? future? Respondent: I just want them to be happy and Respondent: I want them to be good people, to be healthy. Whatever they choose to do in life, I will respectful and kind. I hope they find happiness in support them. whatever they choose to do. Output: Output: Interpretation: Although the respondent expresses Interpretation: While the respondent expresses a supportive attitude towards their children’s future aspirations for their children’s character and plans, there is no explicit mention or intention of high happiness, there is no explicit mention of a desire for education. Therefore, the dialogue aligns with the high education. Therefore, the dialogue aligns with “Not Applicable” label. the “Not Applicable” label. Label: Not Applicable Label: Not Applicable 20 Low Education System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their educational aspirations for their children’s futures. Specifically, your role is to discern whether the parent explicitly expresses an intention or desire for their child to achieve a low level of education, either secular or religious, equivalent to or lower than higher secondary education. Do not assign this label if the child is currently studying below lower or secondary education, or if the scale of education is unspecified. Assign the “Low Education” label only if the parent explicitly articulates a desire or intention for their child to pursue a level of education that is equivalent to or lower than higher secondary education in the future. If there is no such direct reference to low education in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of low education was not addressed. Few-shot examples Input: Input: Interviewer: What are your aspirations for your son’s Interviewer: Do you have any educational goals for education? your child? Respondent: I want him to finish his schooling up to Respondent: Yes, I want my child to pursue a degree the 10th grade and then start working. in medicine. Output: Output: Interpretation: The respondent explicitly states that Interpretation: The respondent expresses a desire for they want their child to complete education up to the their child to pursue a degree in medicine, which 10th grade, which is equivalent to or lower than higher is a higher level of education than higher secondary secondary education. Therefore, the dialogue should education. Therefore, the dialogue should be assigned be assigned the “Low Education” label. the “Not Applicable” label. Label: Low Education Label: Not Applicable Input: Input: Interviewer: What do you wish for your daughter’s Interviewer: What are your hopes for your son’s future? future? Respondent: I hope she completes her education till Respondent: I hope he becomes a successful grade 8 and then gets married. businessman. Output: Output: Interpretation: The respondent expresses a desire for Interpretation: The respondent expresses a hope for their daughter to finish education up to the 8th grade, their son’s future but does not specify a level of which is lower than higher secondary education. This education. Therefore, the dialogue should be assigned aligns with the “Low Education” label. the “Not Applicable” label. Label: Low Education Label: Not Applicable 21 Neutral Education System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their educational aspirations for their children’s futures. Specifically, your role is to discern whether the parent expresses a vague or general desire for their child’s education, accompanied by a future job or religious education ambition, without implying a high or low level of education. Assign the “Neutral Education” label only if the parent articulates a general or ambiguous desire for their child to study in the future, accompanied by the mention of a future job or religious education ambition, without specifically indicating a high or low level of education. If there is no such direct reference to neutral education in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of neutral education was not addressed. Few-shot examples Input: Input: Interviewer: What are your aspirations for your Interviewer: What are your expectations for your daughter’s future? child’s education? Respondent: My wish is for her to receive a good Respondent: I want my child to be an engineer. education and perhaps find a good job in the future. Output: Output: Interpretation: The respondent expresses a specific Interpretation: The respondent expresses a general career ambition for their child, which implies a high desire for their child to receive education and hints level of education. This does not align with the at a future job ambition. However, there is no specific “Neutral Education” label. mention of a high or low level of education. Label: Not Applicable Label: Neutral Education Input: Input: Interviewer: How do you envision your child’s future? Interviewer: What are your dreams for your son’s Respondent: I just want them to be happy and future? healthy. If they choose to pursue education or a Respondent: I want him to be successful in whatever career, that’s up to them. he chooses to do, be it in his education or career. Output: Output: Interpretation: The respondent expresses a general Interpretation: The respondent expresses a general wish for their child’s happiness and health, but doesn’t desire for their child’s success in education and career specify any educational or career ambitions. This does but doesn’t specify a level of education. not align with the “Neutral Education” label. Label: Neutral Education Label: Not Applicable 22 Religious Education System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations for their children’s future education with a particular focus on religious education. Specifically, your role is to discern whether the parent explicitly expresses an intention or desire for their child to pursue religious education or become a religious figure in the future. Do not assign this label if the child is already engaged in any of the mentioned activities, as current religious education is not considered a future ambition. Assign the “Religious Education” label only if the parent explicitly articulates a desire or intention for their child to engage in religious education or assume a religious role in the future. Few-shot examples Input: Input: Interviewer: What are your aspirations for your Interviewer: How do you envision your son’s future? youngest son? Respondent: I want him to be a scholar of our Respondent: I want him to become a doctor, but I religion. He should learn and teach our holy also want him to get a religious education. I have scriptures. plans to enroll him in a Madrasa when he’s a bit older. Output: Output: Interpretation: The parent expresses a clear intention Interpretation: The parent explicitly expresses a for their child to engage in religious education and desire for their child to pursue religious education in even become a religious figure in the future. This the future, alongside becoming a doctor. This aligns aligns with the “Religious Education” label. with the “Religious Education” label. Label: Religious Education Label: Religious Education Input: Input: Interviewer: What are your dreams for your Interviewer: What are your plans for your children’s daughter’s future? education? Respondent: I hope she becomes a successful Respondent: I want them to study hard and go to engineer. I want her to have a good education and a university. A good education is the key to a better life. stable career. Output: Output: Interpretation: The parent talks about their child’s Interpretation: While the parent has clear aspirations education, but there is no explicit reference to for their child’s future, there is no explicit mention religious education or the child assuming a religious of religious education or the desire for the child to role in the future. Therefore, the “Religious take on a religious role. Therefore, the “Religious Education” label is not applicable. Education” label is not applicable. Label: Not Applicable Label: Not Applicable 23 Entrepreneur System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s futures. Specifically, your role is to discern whether the parent explicitly expresses any intention or wishes related to their child’s potential future entrepreneurship. Assign the “Entrepreneur” label only if the parent explicitly articulates a desire or intention for their child to pursue a non-wage enterprise job or start and run their own business, regardless of whether the business type is specified. If there is no such direct reference to entrepreneurship in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of entrepreneurship was not addressed. Few-shot examples Input: Input: Interviewer: What are your dreams for your children’s Interviewer: What are your aspirations for your future? daughter’s future? Respondent: I want them to be happy and healthy. Respondent: I want her to be independent and I hope they get a good education and find jobs that successful. I am thinking of setting up a small they love. business for her, maybe a bakery or a boutique. Output: Output: Interpretation: While the respondent talks about Interpretation: The respondent explicitly expresses a their children’s future, there is no direct reference to desire for their child to run a business in the future. entrepreneurship or starting a business. Therefore, This aligns with the definition of the “Entrepreneur” this does not meet the criteria for the “Entrepreneur” label. label. Label: Entrepreneur Label: Not Applicable Input: Input: Interviewer: What do you envision for your child’s Interviewer: What do you hope your son will do in future? the future? Respondent: I want my child to be a doctor. I believe Respondent: I want him to be a successful person. I that is a noble profession and I want him to serve the am encouraging him to start his own tech firm. community. Output: Output: Interpretation: The respondent has directly expressed Interpretation: The respondent has expressed a their desire for their child to start and run their own specific desire for their child to become a doctor, which business in the future, which fits the criteria for the is distinct from entrepreneurship. This does not meet “Entrepreneur” label. the criteria for the “Entrepreneur” label. Label: Entrepreneur Label: Not Applicable 24 Secular Job System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s future careers in secular jobs. Specifically, your role is to discern whether the parent explicitly expresses a desire or intention related to their child pursuing a specific secular job, occupation, or work type, such as a doctor, government job, NGO job, or teacher in a non-religious school. Assign the “Secular Job” label only if the parent explicitly articulates a desire or intention for their child to pursue a particular secular career in the future. If there is no such direct reference to a secular job in the parent’s statement, or if there is a reference to a religious job instead, assign the “Not Applicable” label to indicate that the topic of secular jobs was not addressed. Few-shot examples Input: Input: Interviewer: What are your aspirations for your son’s Interviewer: What are your hopes for your daughter’s future? future? Respondent: I just want him to be happy and Respondent: I hope she becomes a successful lawyer. healthy. Whatever career he chooses, I will support She has always been good at debating and I believe him. she would excel in that field. Output: Output: Interpretation: While the parent expresses a general Interpretation: The parent explicitly expresses a desire for their child’s well-being and happiness, there desire for their child to pursue a career as a lawyer, is no explicit mention of a specific secular job or which is a secular job. Therefore, the “Secular Job” career. Therefore, the “Not Applicable” label should label should be assigned. be assigned. Label: Secular Job Label: Not Applicable Input: Input: Interviewer: Do you have any specific career in mind Interviewer: What are your dreams for your child’s for your daughter? future? Respondent: Yes, I would like her to pursue a career Respondent: My dream is for my child to be a good in engineering. She has always been fascinated by person and contribute positively to society. how things work. Output: Output: Interpretation: While the parent expresses a general Interpretation: The parent explicitly expresses a desire for their child’s future, there is no explicit desire for their child to pursue a career in engineering, mention of a specific secular job or career. Therefore, which is a secular job. Therefore, the “Secular Job” the “Not Applicable” label should be assigned. label should be assigned. Label: Not Applicable Label: Secular Job 25 Marriage System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s futures. Specifically, your role is to discern whether the parent explicitly expresses any intention or wishes related to their child’s potential future marriage. Assign the “Marriage” label only if the parent articulates a desire or intention for their child to get married in the future. If there is no such reference to marriage in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of marriage was not addressed. Few-shot examples Input: Input: Interviewer: What are your aspirations for your son, Interviewer: What do you hope for your eldest son, Ahmed? Omar? Respondent: My son is very smart, I hope he will get Respondent: Omar is a good boy. I hope he will find a good job in the future. And when he is settled, I a good wife and start a family of his own. would like him to marry a good woman. Output: Output: Interpretation: The respondent explicitly mentions Interpretation: The respondent explicitly mentions their hope for their son to find a good wife and start their desire for their son to get married in the future, a family, which directly relates to the child’s potential after getting a good job. This directly relates to future marriage. Therefore, the label “Marriage” the child’s potential future marriage, so the label should be assigned. “Marriage” should be assigned. Label: Marriage Label: Marriage Input: Input: Interviewer: Can you tell me about your dreams for Interviewer: What are your dreams for your children? your daughter’s future? Respondent: I want them to be happy and successful Respondent: I want my daughter to become a in their respective fields. I have always encouraged successful lawyer. She’s always had a passion for them to pursue their passions. justice. Output: Output: Interpretation: The respondent talks about their Interpretation: The respondent mentions their aspirations for their children’s happiness and success, aspirations for their daughter’s future, but it is related but there is no explicit mention of marriage. to her career, not her marital status. Since there Therefore, the label “Not Applicable” should be is no explicit mention of marriage, the label “Not assigned. Applicable” should be assigned. Label: Not Applicable Label: Not Applicable 26 Migration System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s futures. Specifically, your role is to discern whether the parent explicitly expresses any intention or wishes related to their child’s potential future migration. Assign the “Migration” label only if the parent explicitly articulates a desire or intention for their child to migrate in the future for reasons such as work, studying, or resettling. If there is no such direct reference to migration in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of migration was not addressed. Few-shot examples Input: Input: Interviewer: What are your thoughts on your Interviewer: Have you considered options for your daughter’s future after she completes her studies? daughter’s higher education? Respondent: I am hoping to send her to a foreign Respondent: Yes, I am considering sending her university for her post-graduation studies. I believe abroad for her Master’s degree. I think she will it will provide her with better opportunities. benefit from the exposure and opportunities there. Output: Output: Interpretation: The parent is explicitly expressing a Interpretation: The parent is explicitly expressing a desire for their child to migrate for further studies. desire for their child to migrate for higher education. This aligns with the “Migration” label as the parent This aligns with the “Migration” label as the parent is directly referring to the child’s potential future is directly referring to the child’s potential future migration. migration. Label: Migration Label: Migration Input: Input: Interviewer: What are your aspirations for your son’s Interviewer: What are your plans for your son after future? he completes his high school? Respondent: I want him to have a good education Respondent: I want him to enrol in a good university and a successful career. I hope he becomes a here and study engineering. We have excellent respected doctor. universities in our country. Output: Output: Interpretation: The parent is discussing their child’s Interpretation: The parent is discussing their child’s future but there is no explicit mention of migration. future but there is no explicit mention of migration. The parent is expressing aspirations related to the The parent is expressing aspirations related to the child’s career but not necessarily in the context of child’s education but it is specifically mentioned to moving abroad. Hence, this does not align with the be within their own country. Hence, this does not “Migration” label. align with the “Migration” label. Label: Not Applicable Label: Not Applicable 27 No Ambition System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations for their children’s futures. Specifically, your role is to discern whether the parent expresses helplessness, unwillingness, or a lack of dreams or plans for their child’s future. Assign the “No Ambition” label only if the parent explicitly conveys a lack of ambitions or plans for their child’s future or if they demonstrate a sense of helplessness or unwillingness concerning their child’s future. If there is no such direct expression of a lack of ambition in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of ambition was not addressed. Few-shot examples Input: Input: Interviewer: What are your hopes for your daughter’s Interviewer: What are your plans for your child’s future? education? Respondent: What future? We are just surviving. Respondent: I can’t afford to send him to school. She will just have to get married and that’s it. He’ll have to work like the rest of us. Output: Output: Interpretation: The respondent expresses a sense of Interpretation: The respondent expresses a sense helplessness regarding their daughter’s future and of helplessness and lack of plans for their child’s does not express any dreams or plans for her future education, indicating a lack of ambition for the child’s beyond getting married. This aligns with the “No future. This aligns with the “No Ambition” label. Ambition” label. Label: No Ambition Label: No Ambition Input: Input: Interviewer: What are your dreams for your son’s Interviewer: How do you see your child’s future? future? Respondent: I hope he will be able to get a good Respondent: I want him to become a doctor and education and have a successful career. serve the community. Output: Output: Interpretation: The respondent expresses hopes Interpretation: The respondent clearly expresses a for their child’s future, indicating that they have dream for their child’s future, indicating that they ambitions for the child. This does not align with the have ambitions for the child. This does not align with “No Ambition” label. the “No Ambition” label. Label: Not Applicable Label: Not Applicable 28 Reliance on God System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their reliance on divine intervention or faith in God to support their children’s education and future aspirations. Specifically, your role is to discern whether the parent explicitly expresses full or partial reliance on God to fulfill their dreams for their children or if they currently rely on God’s will in their lives. Assign the “Reliance on God” label only if the parent explicitly articulates a full or partial dependence on God to accomplish their aspirations for their children’s future or if they demonstrate a current reliance on divine will. If there is no such direct reference to reliance on God in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of reliance on divine intervention was not addressed. Few-shot examples Input: Input: Interviewer: What are your hopes for your children’s Interviewer: Do you have any dreams for your future? daughter’s future? Respondent: I hope they will do well in their studies Respondent: Yes, I dream that she will become a and become successful. But ultimately, it’s all up to doctor. But we leave everything in God’s hands. If God’s will. it’s His will, it will happen. Output: Output: Interpretation: The respondent clearly states their Interpretation: The respondent expresses a dream for hopes for their children’s future, but also expresses a their daughter’s future but also mentions that the reliance on God’s will to determine the outcome. This fulfillment of this dream is dependent on God’s will. demonstrates a partial reliance on divine intervention This shows a reliance on divine intervention for their for their children’s future. child’s future. Label: Reliance on God Label: Reliance on God Input: Input: Interviewer: How do you plan to support your Interviewer: What are your plans for your son’s children’s education? education? Respondent: I work hard every day to provide for Respondent: I just want him to do his best. The rest their needs. I believe in the power of hard work and is up to him. dedication. Output: Output: Interpretation: The respondent expresses a desire for Interpretation: The respondent talks about their son to do well in his education, but does not supporting their children’s education through their mention any reliance on God or divine intervention. own efforts and does not mention any reliance on Therefore, the “Reliance on God” label does not apply God or divine intervention. Therefore, the “Reliance in this case. on God” label does not apply in this case. Label: Not Applicable Label: Not Applicable 29 Religious Aspirations System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s future education and careers solely in the context of religion, without implying any monetary incentives. Specifically, your role is to discern whether the parent explicitly expresses a desire for their child to pursue a religious path. Assign the “Religious Aspiration” label only if the parent explicitly articulates aspirations for their child that include becoming a religious scholar, Hafiz, attending a madrassa, learning Arabic, Quran reading, Islamic covering, regular prayer, working in Islamic banks, or being a great Islamic scholar. It is crucial not to assign this label if the child is already engaged in any of the mentioned activities, as current religious activities are not considered future aspirations. If there is no such direct reference to religious aspirations in the parent’s statement, or if there is a reference to secular aspirations, assign the “Not Applicable” label to indicate that the topic of religious aspirations was not addressed. Few-shot examples Input: Interviewer: What are your aspirations for your son’s Input: future? Interviewer: What do you envision for your child’s Respondent: I hope that he will become a great future? Islamic scholar. I want him to learn Arabic and read Respondent: I would like my son to become a Hafiz. the Quran regularly. It’s a great honor in our religion. Output: Output: Interpretation: The parent explicitly expresses a Interpretation: The parent explicitly expresses a desire for their child to pursue a path in religious desire for their child to become a Hafiz, which education, specifically mentioning becoming an is a religious aspiration as per the instruction. Islamic scholar, learning Arabic, and reading the Therefore, this conversation aligns with the “Religious Quran regularly. Therefore, this conversation aligns Aspiration” label. with the “Religious Aspiration” label. Label: Religious Aspiration Label: Religious Aspiration Input: Input: Interviewer: Do you have any specific dreams for Interviewer: What are your hopes for your children’s your daughter? future careers? Respondent: I want her to excel in her studies and Respondent: I want them to have successful careers. become a successful engineer. I hope my daughter becomes a doctor and my son becomes a lawyer. Output: Interpretation: While the parent does express Output: aspirations for their child’s future, these aspirations Interpretation: While the parent does express are not religious in nature. The parent hopes for aspirations for their children’s future, these their child to become an engineer, which doesn’t aspirations do not involve any religious activities involve any religious activities as per the instruction. or pursuits as per the instruction. Therefore, this Therefore, this conversation should be labeled as “Not conversation should be labeled as “Not Applicable”. Applicable”. Label: Not Applicable Label: Not Applicable 30 Secular Aspirations System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s future education, career, and personal development without religious context. Specifically, your role is to discern whether the parent explicitly expresses a desire for their child to develop positive character traits, moralistic values, or general well-being. Assign the “Secular Aspiration” label only if the parent explicitly articulates aspirations for their child that include becoming a good person, being responsible, having good manners, living a healthy life, achieving professional success, or having a good family. If there is no such direct reference to secular aspirations in the parent’s statement, or if there is a reference to religious aspirations, assign the “Not Applicable” label to indicate that the topic of secular aspirations was not addressed. Few-shot examples Input: Input: Interviewer: Can you share your aspirations for your Interviewer: What are your dreams for your children? children? Respondent: My dream for my children is that Respondent: I want them to be kind, respectful and they grow up to be responsible, hardworking, and educated individuals. I hope they lead a healthy life successful in their careers. I also hope they have a and contribute positively to society. happy family life. Output: Output: Interpretation: The respondent explicitly expresses Interpretation: The respondent clearly articulates a desire for their children to develop positive traits secular aspirations for their children, which include such as kindness and respectfulness. They also becoming responsible and hardworking individuals, mention the importance of education and leading a achieving professional success, and having a good healthy life, which are secular aspirations for their family life. Therefore, the dialogue should be assigned children. Therefore, the dialogue should be assigned the “Secular Aspiration” label. the “Secular Aspiration” label. Label: Secular Aspiration Label: Secular Aspiration Input: Input: Interviewer: What kind of person would you like your Interviewer: What kind of future do you envision for child to become and how do you want them to live your children? their life? Respondent: I just hope they’re happy, whatever Respondent: I hope they will follow our religious they choose to do. teachings, pray regularly, and become a great Islamic scholar. Output: Interpretation: While the respondent expresses a Output: desire for their children’s happiness, they do not Interpretation: The respondent discusses faith-based explicitly state any secular aspirations such as aspirations for their child, such as following religious developing positive character traits or achieving teachings, praying regularly, and becoming a great professional success. Therefore, the dialogue should Islamic scholar. Therefore, the dialogue should be be assigned the “Not Applicable” label. assigned the “Not Applicable” label. Label: Not Applicable Label: Not Applicable 31 Vocational Training System prompt Your task involves interpreting and analyzing transcripts of interviews conducted with parents, focusing on their aspirations concerning their children’s future vocational training. Specifically, your role is to discern whether the parent explicitly expresses a desire or intention related to their child pursuing vocational training or learning a specific skill in the context of ambition, such as tailoring, handicrafts, or sewing. The label should only be assigned if training, teaching, or learning is mentioned in the context of ambition. Assign the “Vocational Training” label only if the parent explicitly articulates a desire or intention for their child to undergo vocational training or learn a particular skill in the future as part of their ambition. If there is no such direct reference to vocational training in the parent’s statement, assign the “Not Applicable” label to indicate that the topic of vocational training was not addressed. Few-shot examples Input: Input: Interviewer: What are your aspirations for your Interviewer: How do you envision your child’s future? youngest daughter? Respondent: Well, I believe in letting him choose his Respondent: I want her to learn sewing. It’s a skill own path. But I do encourage him to learn carpentry. that can help her in the future. It’s a useful skill to have. Output: Output: Interpretation: The parent explicitly mentions a Interpretation: The parent expresses a desire for their desire for their daughter to learn sewing, a vocational child to learn carpentry, a vocational skill. Therefore, skill. Therefore, the dialogue should be labeled as the dialogue should be labeled as ’Vocational ’Vocational Training’. Training’. Label: Vocational Training Label: Vocational Training Input: Input: Interviewer: Do you have any specific plans for your Interviewer: What are your dreams for your children? son’s future? Respondent: I just want them to be happy and Respondent: I hope he continues his studies and healthy. That’s all that matters. becomes a doctor one day. Output: Output: Interpretation: The parent expresses a general wish Interpretation: Although the parent expresses a for their children’s well-being, but there is no explicit desire for their child’s future, there is no mention mention of vocational training or learning a specific of vocational training or learning a specific skill. skill. Therefore, the dialogue should be labeled as Therefore, the dialogue should be labeled as ’Not ’Not Applicable’. Applicable’. Label: Not Applicable Label: Not Applicable 32 A.1 LLM prompts for Data Augmentation We directed the LLM with the following “system prompt”: Revise a given dialogue between an interviewer and respondent while maintaining the core meaning. Additionally, you may alter names, places, and other identifying information while preserving the core meaning of the dialogue. Create a single revision of the dialogue in the following format: Interviewer: (question/statement by the interviewer) Respondent: (answer by the respondent) B iQual Appendix This Appendix summarises the key features of the iQual approach to supervised NLP for qualitative analysis. The method is set out in more detail in Ashwin et al. (2022) and available through a Python package at https://github.com/worldbank/iQual. B.1 Summary of Approach The overall goal of our iQual is to use our subset of annotated interviews to provide reliable annotations for the remainder of the sample. Broadly, we do this by training a series of classifier models on our annotated set and then using this model to predict annotations for the unannotated set. We train and predict for each annotations separately, so the model for Religious Aspiration will be trained and make its predictions separately from the model for Secular Aspiration. As the qualitative annotations are defined at the level of question-answer pairs (QA), we can represent each annotation as a binary classification problem at the QA level. Figure 6, also shown in Ashwin et al. (2022), illustrates our overall methodology for a single annotation. On the left hand side we see a "human" sample of size Nh , in which interviews include both text w and annotations y , and a "machine" sample in which interviews include only the text. h As annotations are defined at the QA (question-answer pair) level, so we represent wi,s as the sth h QA in interview i in the human sample, with yi,s being the binary annotation on that QA. In other h words, if the annotation Religious Aspiration, yi,s will be equal to one if that QA has been annotated as displaying religious aspirations, and will be zero otherwise. 33 Figure 6: Methodology Aggregate to interview level Human (# ) # !# = (!,! # , … , !,$ ) ⋮ Interview 1 # # !# = (% ! ,! , … , % ! ,$ ) # # !,! !,! Measurement errors ⋮ .# # − # # !,$ !,$ Aggregate to interview level ⋮ .!# = ( # /!,! , … , # /!,$ ) ⋮ Interview # .!# = ( # # /% ! ,! , … , / % ! ,$ ) Assess bias, efficiency and # # % ! ,! % ! ,! interpretability ⋮ # # % ! ,$ % ! ,$ Enhanced sample Train classifier .& 4 = # , # # 2 /',( = (',( ; ) Machine (& ) Substantive analysis Interview 1 & !,! ⋮ Aggregate to interview level & !,$ Predict annotations .!& = ( & /!,! , … , & /!,$ ) & & 2 /',( = (',( ; ) ⋮ ⋮ . & & & /%" ,! , … , ! = ( /%" ,$ ) Interview & & %" ,! ⋮ & %" ,$ h h We then train some classifier f () parameterised by θ to predict yi,s based on the QA text wi,s . As we will discuss below, there are many options for both the classifier we can use here, as well as how to represent the text numerically. A key point here is that the text representation must by full unsupervised - i.e. we do not use any information about y or any further information about the interview subject when creating a numerical representation of the text. The text representation, classifier and a variety of hyperparameters are chosen using k-fold cross-validation. Given this trained classifier we can then predict annotations a the QA level for our unannotated "machine" sample. This m gives us the predicted annotations y ˆi,s . We then aggregate the QA level annotations to the interview level using aggregation function g (). The choice of this aggregation function is at least in part a substantive question that depends on the research question. For example, if we take the mean value of y across QA pairs for each interview this gives us a measure of the intensity with which this concept comes up. On the other hand, if we take the maximum value across the interview this gives us a measure of interviews in which this concept comes up at least once. We perform this aggregation for both the observed human annotations Y h , the "in-sample" predicted human annotations Y ˆ h and the "out-of-sample" predicted machine annotations ˆ . The predicted annotations for the human sample can then by used to assess the measurement Y m errors introduced by the model. For the quantification of measurement errors, we also make use of bootstrapping. The observed human annotations and machine annotations are then combined to give an enhanced sample Y ˜ . Once we have verified that the enhancement does indeed add value, we proceed with substantive analysis. B.2 Choices There are many possible options for the numerical representation of the text representation w, the classifier f () and the aggregation function g (). We fix the aggregation function as taking the mean across QA pairs and we use cross-validation to select the text representation and the classifier. As we train the classifier for each annotation independently, this allows for the fact that a different 34 classification model of text representation may be optimal for different annotations. THe following gives an exhaustive list of the text representations, models and hyperparameters that are selected over during cross validation. In addition to these options we also cross-validate whether the question is included along with the answer in the text representation. 35 Table 1: Statistical methods for text vectorization Method Name Description Hyperparameters Hyperparameters (Used) (Options) TfidfVectorizer TfidfVectorizer is a method for converting text into • ngram_range: The • max_features: The numerical representations, range of n-grams to maximum number of specifically term frequency- consider when creating the words to keep in the inverse document frequency vocabulary. vocabulary based on word (TF-IDF) vectors. It counts frequency. [1000, 10000] • min_df : The minimum the frequency of words in number of documents a • ngram_range: The lower a document and down- word must be in to be and upper boundary of weights the importance included in the vocabulary. the range of n-values for of commonly used words. different word n-grams to This can be useful for text • max_df : The maximum be extracted. { (1,1), classification tasks, as it number of documents a (1,2), (1,3) } allows the classifier to focus word can be in to be on the words that are most included in the vocabulary. distinctive to a particular • max_features: The document. maximum number of words to keep in the vocabulary, based on word frequency. • use_idf : A boolean flag indicating whether to use the inverse-document- frequency weighting. • norm: The type of normalization to apply to the vectors. • smooth_idf : A boolean flag indicating whether to smooth the idf values. • sublinear_tf : A boolean flag indicating whether to apply sublinear scaling to the term frequency. CountVectorizer CountVectorizer is a method for converting text into • ngram_range: The • max_features: The numerical representations, range of n-grams to maximum number of specifically a sparse matrix consider when creating the words to keep in the of word counts. It counts vocabulary. vocabulary, based on word the frequency of words in frequency. [1000, 10000] • min_df : The minimum a document and does not number of documents a • ngram_range: The lower down-weight the importance word must be in to be and upper boundary of of commonly used words. included in the vocabulary. the range of n-values for This can be useful for different word n-grams to text classification tasks, as • max_df : The maximum be extracted. { (1,1), it allows the classifier to number of documents a (1,2), (1,3) } consider all words equally, word can be in to be rather than down-weighting included in the vocabulary. • binary: Whether to use the importance of commonly binary or frequency counts. • max_features: The used words. maximum number of {True, False} words to keep in the vocabulary, based on word frequency. • binary: A boolean flag indicating whether to create binary vectors, with 0/1 values indicating the presence/absence of a word in a document. 36 Table 2: Pre-trained embeddings for text vectorization Model Name Dimensions Description all-mpnet-base-v2 768 This a pre-trained language understanding model that combines the advantages of masked language modeling (MLM) and permuted language modeling (PLM) to address the limitations of both methods. It leverages the dependency among predicted tokens through PLM and takes auxiliary position information as input to make the model see a full sentence, reducing the position discrepancy between pre-training and fine-tuning. This model was pre-trained on a large-scale dataset and generates a vector of 768 dimensions. all-roberta-large-v1 1024 This is a pre-trained language understanding model with a vector representation of 1024 dimensions. It was developed as an improvement upon the BERT model and was trained using the masked language modeling (MLM) objective. It has achieved strong performance on natural language processing tasks and can be fine- tuned on labeled datasets for specific tasks such as classification or language translation. average_word_embeddings 300 This is a method for converting text into numerical representations, _glove.6B.300d specifically word embeddings. It uses a pre-trained GloVe model to generate 300-dimensional vector representations for each word in a document, and then averages these vectors to create a single representation for the entire document. This can be useful for text classification tasks, as it allows the classifier to consider the semantic relationships between words, rather than just their frequencies. distiluse-base-multilingual- 512 This is a pre-trained language understanding model that maps text cased-v2 into a 512-dimensional vector representation. It is a smaller and faster version of the popular transformer model, BERT, and has been trained on a large multilingual dataset, allowing it to process text in multiple languages. It has also been cased, meaning it can distinguish between upper and lower case letters. This model is useful for natural language processing tasks such as language translation and text classification, and can be fine-tuned on labeled datasets for specific tasks. 37 Table 3: Classifier Options I Method Description Hyperparameters (Options) Hyperparameters (Used) LogisticRegression This is a linear classifier that uses a logistic • C: The inverse of the • penalty: The type of function to predict regularization strength, with regularization to use: L1 or the probability of a higher values indicating less L2. sample belonging to regularization. • C: Inverse of regularization a particular class. It • penalty: The type of strength. [0.00002, 10000] is commonly used for regularization to use, either binary classification L1 or L2. tasks, but can also • fit_intercept: A boolean flag be used for multi- indicating whether to fit an class classification by intercept term. implementing a one- • tol: The tolerance for stopping versus-rest approach. criteria. • intercept_scaling: The scaling of the intercept term, if it is being fitted. • class_weight: The class weights to use for unbalanced classes. • max_iter: The maximum number of iterations for the optimization algorithm. SGDClassifier This is a linear classifier that uses stochastic • loss: The loss function to use, • loss: The loss function to use. gradient descent to with options such as "hinge", ("modified_huber") learn the parameters of "log", "modified_huber", • penalty: The type of the model. The modified "squared_hinge", and regularization to use: L1 or huber loss function is a "perceptron". L2. smooth approximation • penalty: The type of • learning_rate: The learning rate of the hinge loss, which regularization to use, with schedule to use. ("optimal") is commonly used for options such as L1, L2, • alpha: The constant that linear classification "elasticnet", and "none". multiplies the regularization tasks. • alpha: The regularization term. [0.00002, 1000] strength, with higher values indicating stronger regularization. • l1_ratio: The proportion of L1 regularization to use in the elasticnet penalty. • tol: The tolerance for the stopping criteria. • learning_rate: The learning rate for the optimization algorithm, with options such as "constant", "optimal", and "invscaling". • eta0: The initial learning rate for the "constant" and "invscaling" learning rate schedules. • power_t: The exponent for the "invscaling" learning rate schedule. 38 Table 4: Classifier Options II Method Description Hyperparameters (Options) Hyperparameters (Used) RandomForestClassifier This is an ensemble classifier that uses • n_estimators: The number of • n_estimators: The number of multiple decision trees decision trees in the forest. trees in the forest. [100, 1000] to make predictions. • criterion: The function to • max_depth: The maximum It randomly selects a measure the quality of a split, depth of the tree. [10, 100] subset of features to with options such as "gini" and consider at each split "entropy". in the tree, which helps • max_depth: The maximum to reduce overfitting depth of the decision tree. and improve the • min_samples_split: The generalization of the minimum number of samples model. required to split an internal node. • min_samples_leaf: The minimum number of samples required to be at a leaf node. • min_weight_fraction_leaf: The minimum weighted fraction of the sum total of weights required to be at a leaf node. • max_features: The number of features to consider when looking for the best split. • max_leaf_nodes: The maximum number of leaf nodes in the tree. • min_impurity_decrease: The minimum decrease in impurity required to split the node. • bootstrap: A boolean flag indicating whether to use bootstrap samples when building the trees. • oob_score: A boolean flag indicating whether to use out- of-bag samples to estimate the generalization error. DecisionTreeClassifier This is a classifier that uses a tree structure to • criterion: The function to • max_depth: The maximum make decisions based on measure the quality of a split, depth of the tree. [5, 100] the features of a sample. with options such as "gini" and • min_impurity_decrease: A node At each node in the tree, "entropy". will be split if this split induces a the classifier considers a • splitter: The strategy to use decrease of the impurity greater single feature and splits when searching for a split, with than or equal to this value. the data based on the options such as "best" and [0.00002,10000] value of that feature. "random". The final decision is • max_depth: The maximum made based on the path depth of the tree. taken through the tree. • min_samples_split: The minimum number of samples required to split an internal node. • min_samples_leaf: The minimum number of samples required to be at a leaf node. • min_weight_fraction_leaf: The minimum weighted fraction of the sum total of weights required to be at a leaf node. • max_features: The number of features to consider when looking for the best split. • max_leaf_nodes: The maximum number of leaf nodes in the tree. • min_impurity_decrease: The minimum decrease in impurity required to split the node. 39 Table 5: Classifier Options III Method Description Hyperparameters (Options) Hyperparameters (Used) MLPClassifier This is a classifier that uses a neural network • hidden_layer_sizes: The • hidden_layer_sizes: The ith with multiple layers number of neurons in each element represents the number of to make predictions. hidden layer. neurons in the ith hidden layer. It is commonly used • activation: The activation [(100,), (100, 100), (100, 100, for classification function to use, with options 100)] tasks and can handle such as "identity", "logistic", • activation: Activation function both continuous and "tanh", and "relu". for the hidden layer. ("tanh", categorical data. The • solver: The algorithm to use for "relu") number of layers and the optimization, with options such • alpha: L2 penalty (regularization number of units in each as "lbfgs", "sgd", and "adam". term) parameter. [0.01, 1] layer can be adjusted to • alpha: The regularization fit the complexity of the strength, with higher task. values indicating stronger regularization. • batch_size: The number of samples to use in each iteration of the optimization algorithm. • learning_rate: The learning rate for the optimization algorithm, with options such as "constant", "invscaling", and "adaptive". • learning_rate_init: The initial learning rate for the "constant" and "invscaling" learning rate schedules. • power_t: The exponent for the "invscaling" learning rate schedule. • max_iter: The maximum number of iterations to run the optimization algorithm. • shuffle: A boolean flag indicating whether to shuffle the training data before each epoch. • tol: The tolerance for the stopping criteria. • warm_start: A boolean flag indicating whether to reuse the solution of the previous call to fit. • momentum: The momentum for the optimization algorithm. • nesterovs_momentum: A boolean flag indicating whether to use Nesterov’s momentum. • early_stopping: A boolean flag indicating whether to use early stopping to terminate the optimization early. • validation_fraction: The fraction of the training data to use as validation data for early stopping. • beta_1: The beta 1 parameter for the Adam optimization algorithm. 40 Table 6: Classifier Options III Method Description Hyperparameters (Options) Hyperparameters (Used) KNeighborsClassifier This is a non-parametric classifier that uses the • n_neighbors: The number of • n_neighbors: Number of K nearest neighbors neighbors to use when making a neighbors to use by default for of a sample to make prediction. kneighbors queries. [10,10000] a prediction. It is • weights: The weight function to • weights: weight function used commonly used for use when making a prediction, in prediction. ("uniform", classification tasks with options such as "uniform" "distance") and can handle and "distance". both continuous and • algorithm: The algorithm to use categorical data. The for finding the nearest neighbors, number of neighbors with options such as "brute" and to consider (K) is a "kd_tree". hyperparameter that • leaf_size: The number of points can be adjusted to fit at which to switch to a brute the complexity of the force search for the nearest task. neighbors. • p: The power parameter for the Minkowski distance metric. • metric: The distance metric to use, with options such as "euclidean", "manhattan", and "minkowski". • metric_params: Additional parameters for the distance metric. SVC This is a classifier that uses a support vector • C: The regularization strength, • C: Penalty parameter C of the machine (SVM) to find with higher values indicating error term. [0.00001, -00] the optimal hyperplane stronger regularization. to separate the different • kernel: The kernel to use for the classes. It is commonly decision function, with options used for classification such as "linear", "poly", "rbf", tasks and can handle "sigmoid", and "precomputed". both continuous and • degree: The degree of the categorical data. The polynomial kernel. kernel function used to • gamma: The kernel coefficient project the data into for the rbf, poly, and sigmoid a higher dimensional kernels. space can be adjusted to • coef0: The independent term fit the complexity of the in the polynomial and sigmoid task. kernels. • shrinking: A boolean flag indicating whether to use the shrinking heuristic. • probability: A boolean flag indicating whether to enable probability estimates. • tol: The tolerance for the stopping criteria. • class_weight: The class weights to use for unbalanced classes. • verbose: The level of verbosity in the output. • decision_function_shape: The shape of the decision function, with options such as "ovo" and "ovr". 41 C Bias tests Table 7: Bias tests for Aspiration: Religious Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.0347∗∗∗ -0.0716∗∗∗ -0.0139 -0.0032 −1.55 × 10−5 (0.0132) (0.0185) (0.0126) (0.0104) (0.0107) data_roundR3 0.0059 -0.0206 -0.0233∗ 0.0116 0.0248∗∗ (0.0127) (0.0178) (0.0121) (0.0100) (0.0103) refugee -0.0101 -0.0167 -0.0005 0.0039 -0.0044 (0.0083) (0.0116) (0.0079) (0.0065) (0.0067) num_child 0.0007 -0.0008 0.0008 -0.0003 -0.0018 (0.0021) (0.0029) (0.0020) (0.0017) (0.0017) hh_head_sex -0.0076 0.0115 0.0003 -0.0091 -0.0009 (0.0075) (0.0105) (0.0072) (0.0059) (0.0061) hh_head_age 0.0004 0.0006 -0.0004 0.0004 0.0004 (0.0003) (0.0005) (0.0003) (0.0003) (0.0003) parent_eduyears 1.25 × 10−5 -0.0020∗ -0.0004 -0.0004 1.67 × 10−5 (0.0008) (0.0012) (0.0008) (0.0007) (0.0007) parent_reledu -0.0088 -0.0231 -0.0039 0.0073 0.0063 (0.0142) (0.0199) (0.0136) (0.0112) (0.0115) eld_sex -0.0052 -0.0113 -0.0012 -0.0074 -0.0087∗ (0.0058) (0.0081) (0.0055) (0.0046) (0.0047) eld_age −9.15 × 10−5 0.0003 0.0007 -0.0003 -0.0006∗ (0.0005) (0.0007) (0.0004) (0.0004) (0.0004) hh_asset_index -0.0022 -0.0041 0.0002 -0.0016 -0.0015 (0.0024) (0.0033) (0.0023) (0.0019) (0.0019) hh_income 0.0008 0.0027 0.0010 -0.0009 -0.0006 (0.0017) (0.0024) (0.0016) (0.0014) (0.0014) int_trauma_exp 0.0019 0.0010 0.0023∗ -0.0004 0.0012 (0.0012) (0.0017) (0.0012) (0.0010) (0.0010) Fit statistics Observations 667 667 667 667 667 R2 0.01430 0.02935 0.01555 0.02448 0.02312 Adjusted R2 -0.00378 0.01154 -0.00252 0.00659 0.00520 F-test 0.79090 1.6477 0.86071 1.3679 1.2898 F-test, p-value 0.66022 0.07445 0.58739 0.17634 0.21956 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 42 Table 8: Bias tests for Aspiration: Secular Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.1790∗∗∗ -0.0602∗∗ -0.0812∗∗∗ -0.0247 -0.0113 (0.0301) (0.0243) (0.0242) (0.0152) (0.0137) data_roundR3 0.0877∗∗∗ 0.0308 -0.0205 0.0064 0.0075 (0.0289) (0.0233) (0.0233) (0.0146) (0.0132) refugee 0.0243 0.0054 0.0166 0.0137 0.0052 (0.0188) (0.0152) (0.0152) (0.0095) (0.0086) num_child 0.0052 0.0054 0.0013 0.0039 0.0012 (0.0048) (0.0039) (0.0038) (0.0024) (0.0022) hh_head_sex 0.0130 0.0018 -0.0039 -0.0026 -0.0008 (0.0171) (0.0138) (0.0138) (0.0086) (0.0078) hh_head_age -0.0006 8.01 × 10−5 3.67 × 10−5 0.0002 −4.79 × 10−6 (0.0007) (0.0006) (0.0006) (0.0004) (0.0003) parent_eduyears -0.0039∗∗ -0.0030∗ -0.0012 -0.0018∗ -0.0004 (0.0019) (0.0015) (0.0015) (0.0010) (0.0009) parent_reledu 0.0294 0.0085 -0.0036 -0.0015 0.0127 (0.0323) (0.0261) (0.0261) (0.0163) (0.0147) eld_sex 0.0043 -0.0086 0.0100 -0.0008 -0.0012 (0.0132) (0.0106) (0.0106) (0.0067) (0.0060) eld_age −9.1 × 10−5 -0.0006 0.0004 −6.99 × 10−5 -0.0001 (0.0011) (0.0009) (0.0009) (0.0005) (0.0005) hh_asset_index -0.0026 -0.0024 -0.0008 0.0046∗ 0.0028 (0.0054) (0.0044) (0.0044) (0.0027) (0.0025) hh_income -0.0022 0.0012 0.0036 -0.0008 -0.0008 (0.0039) (0.0032) (0.0032) (0.0020) (0.0018) int_trauma_exp -0.0021 -0.0010 -0.0012 0.0003 −7.47 × 10−5 (0.0028) (0.0023) (0.0023) (0.0014) (0.0013) Fit statistics Observations 667 667 667 667 667 R2 0.09448 0.02157 0.01093 0.01966 0.00522 Adjusted R2 0.07786 0.00362 -0.00721 0.00167 -0.01303 F-test 5.6861 1.2017 0.60246 1.0930 0.28618 F-test, p-value 2.3 × 10−9 0.27754 0.84103 0.36292 0.99143 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 43 Table 9: Bias tests for Ambition: No Ambition Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.1563∗∗∗ -0.4332∗∗∗ -0.5035∗∗∗ -0.0085 -0.0087 (0.0219) (0.0336) (0.0352) (0.0075) (0.0081) data_roundR3 0.0406∗ 0.0864∗∗∗ -0.0430 -0.0044 -0.0061 (0.0210) (0.0323) (0.0338) (0.0072) (0.0078) refugee 0.0385∗∗∗ 0.0284 0.0613∗∗∗ 0.0136∗∗∗ 0.0150∗∗∗ (0.0137) (0.0210) (0.0221) (0.0047) (0.0051) num_child -0.0032 0.0014 -0.0087 -0.0002 -0.0008 (0.0035) (0.0053) (0.0056) (0.0012) (0.0013) hh_head_sex -0.0436∗∗∗ -0.0442∗∗ -0.0607∗∗∗ -0.0054 -0.0107∗∗ (0.0124) (0.0191) (0.0200) (0.0043) (0.0046) hh_head_age 3.79 × 10−5 -0.0012 -0.0011 0.0002 0.0002 (0.0005) (0.0008) (0.0009) (0.0002) (0.0002) parent_eduyears 0.0061∗∗∗ 0.0052∗∗ 0.0085∗∗∗ -0.0001 -0.0004 (0.0014) (0.0021) (0.0022) (0.0005) (0.0005) parent_reledu 0.0138 0.0115 0.0124 -0.0104 -0.0161∗ (0.0235) (0.0362) (0.0379) (0.0081) (0.0087) eld_sex -0.0030 0.0005 -0.0112 -0.0021 9.85 × 10−5 (0.0096) (0.0147) (0.0154) (0.0033) (0.0036) eld_age -0.0004 -0.0011 −3.39 × 10−6 7.4 × 10−5 7.2 × 10−5 (0.0008) (0.0012) (0.0012) (0.0003) (0.0003) hh_asset_index 0.0042 0.0111∗ 0.0148∗∗ -0.0010 -0.0012 (0.0039) (0.0060) (0.0063) (0.0014) (0.0015) hh_income 0.0011 0.0025 -0.0009 0.0004 0.0006 (0.0028) (0.0044) (0.0046) (0.0010) (0.0011) int_trauma_exp 0.0004 -0.0002 -0.0028 -0.0002 -0.0009 (0.0021) (0.0032) (0.0033) (0.0007) (0.0008) Fit statistics Observations 667 667 667 667 667 R2 0.08578 0.06986 0.08313 0.04273 0.05054 Adjusted R2 0.06900 0.05280 0.06630 0.02516 0.03312 F-test 5.1135 4.0935 4.9410 2.4325 2.9013 F-test, p-value 3.3 × 10−8 3.49 × 10−6 7.31 × 10−8 0.00426 0.00063 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 44 Table 10: Bias tests for Ambition: Job Secular Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.0379∗∗ -0.1023∗∗∗ -0.1063∗∗∗ -0.0030 -0.0041 (0.0178) (0.0229) (0.0254) (0.0165) (0.0159) data_roundR3 0.0177 0.0197 0.0031 -0.0024 -0.0190 (0.0171) (0.0220) (0.0244) (0.0159) (0.0152) refugee 0.0260∗∗ 0.0205 0.0135 0.0078 0.0071 (0.0112) (0.0143) (0.0159) (0.0103) (0.0099) num_child 0.0061∗∗ 0.0056 0.0053 0.0036 0.0031 (0.0028) (0.0036) (0.0040) (0.0026) (0.0025) hh_head_sex 0.0002 0.0121 0.0085 -0.0055 -0.0021 (0.0101) (0.0130) (0.0144) (0.0094) (0.0090) hh_head_age 0.0008∗ 0.0009 0.0003 −5.25 × 10−5 -0.0003 (0.0004) (0.0006) (0.0006) (0.0004) (0.0004) parent_eduyears -0.0010 -0.0009 -0.0007 −5.75 × 10−5 0.0001 (0.0011) (0.0014) (0.0016) (0.0010) (0.0010) parent_reledu -0.0202 -0.0249 -0.0732∗∗∗ -0.0137 -0.0120 (0.0192) (0.0246) (0.0273) (0.0178) (0.0171) eld_sex 0.0022 -0.0004 -0.0010 -0.0095 -0.0106 (0.0078) (0.0100) (0.0111) (0.0072) (0.0070) eld_age -0.0009 -0.0004 -0.0002 −3.47 × 10−5 0.0006 (0.0006) (0.0008) (0.0009) (0.0006) (0.0006) hh_asset_index -0.0010 0.0002 0.0012 0.0009 0.0003 (0.0032) (0.0041) (0.0046) (0.0030) (0.0029) hh_income 0.0004 0.0034 0.0029 -0.0011 -0.0018 (0.0023) (0.0030) (0.0033) (0.0022) (0.0021) int_trauma_exp -0.0024 -0.0008 -0.0050∗∗ -0.0005 -0.0007 (0.0017) (0.0022) (0.0024) (0.0016) (0.0015) Fit statistics Observations 667 667 667 667 667 R2 0.03887 0.01983 0.02149 0.01108 0.01631 Adjusted R2 0.02123 0.00185 0.00353 -0.00706 -0.00174 F-test 2.2039 1.1029 1.1967 0.61066 0.90381 F-test, p-value 0.01031 0.35453 0.28112 0.83416 0.54279 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 45 Table 11: Bias tests for Ambition: Vocational Training Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) 0.0016 -0.1262∗∗∗ -0.0516∗∗ 0.0120∗∗ 0.0104∗ (0.0051) (0.0233) (0.0207) (0.0058) (0.0055) data_roundR3 -0.0040 -0.0118 -0.0328∗ -0.0085 -0.0063 (0.0049) (0.0223) (0.0199) (0.0056) (0.0053) refugee -0.0021 -0.0014 0.0028 0.0012 -0.0019 (0.0032) (0.0146) (0.0130) (0.0036) (0.0035) num_child -0.0010 -0.0023 -0.0073∗∗ -0.0001 0.0004 (0.0008) (0.0037) (0.0033) (0.0009) (0.0009) hh_head_sex -0.0028 0.0245∗ 0.0033 -0.0014 0.0015 (0.0029) (0.0132) (0.0118) (0.0033) (0.0032) hh_head_age 9.34 × 10−5 6.68 × 10−5 -0.0004 -0.0002 -0.0001 (0.0001) (0.0006) (0.0005) (0.0001) (0.0001) parent_eduyears -0.0004 -0.0038∗∗∗ -0.0016 -0.0007∗ -0.0007∗∗ (0.0003) (0.0015) (0.0013) (0.0004) (0.0003) parent_reledu 0.0026 -0.0235 -0.0186 0.0071 0.0043 (0.0054) (0.0250) (0.0223) (0.0062) (0.0060) eld_sex 0.0026 0.0113 0.0038 0.0034 0.0009 (0.0022) (0.0102) (0.0091) (0.0025) (0.0024) eld_age −1.58 × 10−5 0.0009 0.0008 0.0001 7.07 × 10−5 (0.0002) (0.0008) (0.0007) (0.0002) (0.0002) hh_asset_index 3.77 × 10−5 -0.0010 -0.0036 0.0009 0.0002 (0.0009) (0.0042) (0.0037) (0.0010) (0.0010) hh_income 0.0002 0.0016 0.0016 0.0005 0.0005 (0.0007) (0.0030) (0.0027) (0.0008) (0.0007) int_trauma_exp -0.0001 -0.0010 -0.0007 -0.0006 -0.0002 (0.0005) (0.0022) (0.0019) (0.0005) (0.0005) Fit statistics Observations 667 667 667 667 667 R2 0.01689 0.02743 0.01987 0.02429 0.01802 Adjusted R2 -0.00115 0.00958 0.00189 0.00638 6.82 × 10−6 F-test 0.93641 1.5368 1.1049 1.3565 1.0004 F-test, p-value 0.50962 0.10618 0.35279 0.18220 0.44678 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 46 Table 12: Bias tests for Ambition: Entrepreneur Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.0273∗∗ 0.0001 -0.0034 -0.0136∗ -0.0092 (0.0119) (0.0112) (0.0096) (0.0070) (0.0073) data_roundR3 -0.0035 -0.0025 0.0032 0.0014 0.0058 (0.0114) (0.0108) (0.0092) (0.0068) (0.0070) refugee -0.0112 -0.0105 -0.0016 -0.0057 -0.0015 (0.0075) (0.0070) (0.0060) (0.0044) (0.0046) num_child 0.0002 0.0017 0.0008 0.0024∗∗ 0.0019∗ (0.0019) (0.0018) (0.0015) (0.0011) (0.0012) hh_head_sex -0.0186∗∗∗ 0.0074 0.0138∗∗ 0.0004 0.0036 (0.0068) (0.0064) (0.0055) (0.0040) (0.0041) hh_head_age -0.0003 0.0005∗ 0.0004 0.0001 4.05 × 10−5 (0.0003) (0.0003) (0.0002) (0.0002) (0.0002) parent_eduyears 0.0007 -0.0006 -0.0003 0.0007 0.0003 (0.0007) (0.0007) (0.0006) (0.0004) (0.0005) parent_reledu 0.0047 -0.0092 -0.0183∗ -0.0005 0.0003 (0.0128) (0.0121) (0.0104) (0.0076) (0.0078) eld_sex 0.0111∗∗ -0.0086∗ -0.0111∗∗∗ 0.0007 -0.0018 (0.0052) (0.0049) (0.0042) (0.0031) (0.0032) eld_age 0.0003 -0.0003 -0.0003 −2.76 × 10−5 -0.0001 (0.0004) (0.0004) (0.0003) (0.0002) (0.0003) hh_asset_index -0.0025 0.0004 0.0016 -0.0024∗ -0.0015 (0.0021) (0.0020) (0.0017) (0.0013) (0.0013) hh_income 0.0003 -0.0007 −6.87 × 10−5 0.0006 0.0013 (0.0016) (0.0015) (0.0013) (0.0009) (0.0009) int_trauma_exp 0.0024∗∗ 0.0001 -0.0004 6.71 × 10−5 -0.0004 (0.0011) (0.0011) (0.0009) (0.0007) (0.0007) Fit statistics Observations 667 667 667 667 667 R2 0.02655 0.02647 0.02927 0.01755 0.01095 Adjusted R2 0.00869 0.00861 0.01146 -0.00047 -0.00719 F-test 1.4864 1.4819 1.6436 0.97374 0.60359 F-test, p-value 0.12408 0.12581 0.07548 0.47254 0.84009 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 47 Table 13: Bias tests for Ambition: Education High Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.1557∗∗∗ -0.2259∗∗∗ -0.3257∗∗∗ 0.0112 0.0268 (0.0233) (0.0294) (0.0298) (0.0172) (0.0171) data_roundR3 0.0246 0.0196 -0.0146 -0.0020 -0.0192 (0.0224) (0.0283) (0.0286) (0.0165) (0.0164) refugee 0.0162 -0.0075 0.0063 0.0095 0.0119 (0.0146) (0.0184) (0.0186) (0.0108) (0.0107) num_child 0.0056 -0.0043 0.0096∗∗ 0.0009 -0.0005 (0.0037) (0.0047) (0.0047) (0.0027) (0.0027) hh_head_sex 0.0154 0.0404∗∗ 0.0077 0.0128 0.0211∗∗ (0.0133) (0.0167) (0.0169) (0.0098) (0.0097) hh_head_age 0.0007 0.0005 0.0010 -0.0004 -0.0007∗ (0.0006) (0.0007) (0.0007) (0.0004) (0.0004) parent_eduyears -0.0036∗∗ -0.0044∗∗ -0.0007 -0.0008 -0.0010 (0.0015) (0.0018) (0.0019) (0.0011) (0.0011) parent_reledu -0.0241 0.0226 0.0018 -0.0001 0.0163 (0.0251) (0.0317) (0.0320) (0.0185) (0.0184) eld_sex 0.0074 0.0143 0.0337∗∗∗ 0.0023 0.0025 (0.0102) (0.0129) (0.0131) (0.0075) (0.0075) eld_age −5.05 × 10−5 0.0002 0.0007 -0.0003 0.0003 (0.0008) (0.0010) (0.0011) (0.0006) (0.0006) hh_asset_index 0.0027 -0.0039 -0.0009 0.0067∗∗ 0.0054∗ (0.0042) (0.0053) (0.0054) (0.0031) (0.0031) hh_income 0.0017 0.0065∗ 0.0061 0.0006 0.0006 (0.0030) (0.0038) (0.0039) (0.0022) (0.0022) int_trauma_exp 0.0013 0.0007 0.0022 0.0009 -0.0020 (0.0022) (0.0028) (0.0028) (0.0016) (0.0016) Fit statistics Observations 667 667 667 667 667 R2 0.03744 0.03659 0.03663 0.01590 0.02704 Adjusted R2 0.01978 0.01891 0.01896 -0.00215 0.00919 F-test 2.1198 2.0700 2.0724 0.88080 1.5148 F-test, p-value 0.01415 0.01702 0.01687 0.56652 0.11370 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 48 Table 14: Bias tests for Ambition: Education Neutral Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.3200∗∗∗ -0.3254∗∗∗ -0.1237∗∗∗ -0.0377∗ -0.0144 (0.0331) (0.0350) (0.0354) (0.0223) (0.0220) data_roundR3 0.0043 0.0721∗∗ -0.0076 0.0498∗∗ 0.0312 (0.0318) (0.0336) (0.0340) (0.0214) (0.0211) refugee 0.0227 0.0118 0.0608∗∗∗ 0.0268∗ 0.0181 (0.0207) (0.0219) (0.0222) (0.0140) (0.0138) num_child 0.0039 0.0021 -0.0010 -0.0070∗∗ -0.0060∗ (0.0053) (0.0056) (0.0056) (0.0035) (0.0035) hh_head_sex 0.0383∗∗ 0.0027 0.0084 0.0155 0.0166 (0.0188) (0.0199) (0.0201) (0.0127) (0.0125) hh_head_age 0.0007 0.0017∗∗ -0.0004 −7.1 × 10−6 −8.23 × 10−5 (0.0008) (0.0009) (0.0009) (0.0005) (0.0005) parent_eduyears -0.0020 -0.0077∗∗∗ -0.0022 -0.0014 -0.0012 (0.0021) (0.0022) (0.0022) (0.0014) (0.0014) parent_reledu 0.0487 -0.0836∗∗ 0.0042 0.0187 0.0229 (0.0356) (0.0377) (0.0381) (0.0240) (0.0236) eld_sex 0.0318∗∗ -0.0033 -0.0124 0.0022 -0.0044 (0.0145) (0.0154) (0.0155) (0.0098) (0.0096) eld_age 0.0005 -0.0023∗ -0.0015 -0.0006 -0.0003 (0.0012) (0.0012) (0.0013) (0.0008) (0.0008) hh_asset_index -0.0079 -0.0116∗ 0.0076 0.0074∗ 0.0048 (0.0060) (0.0063) (0.0064) (0.0040) (0.0040) hh_income 0.0031 0.0029 0.0036 -0.0012 0.0014 (0.0043) (0.0046) (0.0046) (0.0029) (0.0029) int_trauma_exp -0.0048 -0.0039 -0.0062∗ 0.0020 0.0011 (0.0031) (0.0033) (0.0033) (0.0021) (0.0021) Fit statistics Observations 667 667 667 667 667 R2 0.04520 0.05661 0.03922 0.04675 0.02593 Adjusted R2 0.02768 0.03930 0.02159 0.02926 0.00806 F-test 2.5800 3.2705 2.2248 2.6727 1.4507 F-test, p-value 0.00236 0.00013 0.00952 0.00162 0.13825 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 49 Table 15: Bias tests for Ambition: Education Low Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.0171 -0.1612∗∗∗ -0.0657∗∗∗ 0.0112 0.0150∗ (0.0118) (0.0224) (0.0188) (0.0089) (0.0090) data_roundR3 -0.0123 0.0068 0.0050 -0.0032 -0.0062 (0.0113) (0.0215) (0.0181) (0.0086) (0.0086) refugee 0.0157∗∗ 0.0230 0.0022 -0.0042 -0.0117∗∗ (0.0074) (0.0140) (0.0118) (0.0056) (0.0056) num_child -0.0016 -0.0014 0.0036 0.0010 0.0004 (0.0019) (0.0035) (0.0030) (0.0014) (0.0014) hh_head_sex -0.0075 -0.0048 -0.0066 -0.0038 -0.0077 (0.0067) (0.0127) (0.0107) (0.0051) (0.0051) hh_head_age -0.0003 3.83 × 10−6 -0.0002 -0.0002 -0.0002 (0.0003) (0.0005) (0.0005) (0.0002) (0.0002) parent_eduyears -0.0002 0.0004 -0.0014 0.0003 -0.0007 (0.0007) (0.0014) (0.0012) (0.0006) (0.0006) parent_reledu -0.0051 -0.0289 -0.0347∗ -0.0024 -0.0063 (0.0126) (0.0240) (0.0203) (0.0096) (0.0097) eld_sex 0.0173∗∗∗ 0.0011 0.0048 0.0048 0.0032 (0.0052) (0.0098) (0.0083) (0.0039) (0.0039) eld_age 0.0008∗ 0.0003 -0.0007 −9.99 × 10−5 2.42 × 10−5 (0.0004) (0.0008) (0.0007) (0.0003) (0.0003) hh_asset_index 0.0006 -0.0021 -0.0009 -0.0011 -0.0014 (0.0021) (0.0040) (0.0034) (0.0016) (0.0016) hh_income 0.0002 -0.0001 0.0010 0.0002 -0.0003 (0.0015) (0.0029) (0.0025) (0.0012) (0.0012) int_trauma_exp 0.0003 −8.98 × 10−5 0.0006 0.0008 0.0014 (0.0011) (0.0021) (0.0018) (0.0008) (0.0008) Fit statistics Observations 667 667 667 667 667 R2 0.03226 0.01698 0.01891 0.01307 0.01801 Adjusted R2 0.01450 -0.00106 0.00091 -0.00504 −3.96 × 10−6 F-test 1.8168 0.94150 1.0507 0.72175 0.99978 F-test, p-value 0.04214 0.50451 0.40003 0.73085 0.44735 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 50 Table 16: Bias tests for Ambition: Education Religious Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.0076 -0.0266∗ -0.0251∗ 0.0147 0.0256∗∗ (0.0135) (0.0146) (0.0141) (0.0107) (0.0109) data_roundR3 -0.0209 -0.0424∗∗∗ -0.0260∗ -0.0037 -0.0028 (0.0129) (0.0140) (0.0136) (0.0102) (0.0105) refugee 0.0096 0.0154∗ 0.0031 0.0080 0.0075 (0.0084) (0.0091) (0.0089) (0.0067) (0.0069) num_child -0.0017 -0.0007 -0.0021 -0.0012 -0.0030∗ (0.0021) (0.0023) (0.0022) (0.0017) (0.0017) hh_head_sex -0.0055 0.0094 -0.0105 -0.0032 -0.0045 (0.0077) (0.0083) (0.0080) (0.0061) (0.0062) hh_head_age 1.68 × 10−6 0.0001 0.0003 3.15 × 10−6 −3.36 × 10−5 (0.0003) (0.0004) (0.0003) (0.0003) (0.0003) parent_eduyears -0.0012 -0.0005 -0.0006 -0.0005 -0.0012∗ (0.0008) (0.0009) (0.0009) (0.0007) (0.0007) parent_reledu 0.0048 -0.0313∗∗ -0.0153 0.0127 0.0157 (0.0145) (0.0157) (0.0152) (0.0115) (0.0118) eld_sex 0.0067 0.0032 0.0053 −4.65 × 10−5 -0.0019 (0.0059) (0.0064) (0.0062) (0.0047) (0.0048) eld_age 0.0002 0.0007 0.0004 -0.0002 -0.0003 (0.0005) (0.0005) (0.0005) (0.0004) (0.0004) hh_asset_index 0.0039 0.0070∗∗∗ 0.0015 0.0025 0.0037∗ (0.0024) (0.0026) (0.0025) (0.0019) (0.0020) hh_income 0.0002 -0.0010 -0.0003 -0.0015 -0.0010 (0.0018) (0.0019) (0.0018) (0.0014) (0.0014) int_trauma_exp 0.0014 −9.22 × 10−5 0.0022∗ 9.23 × 10−5 0.0009 (0.0013) (0.0014) (0.0013) (0.0010) (0.0010) Fit statistics Observations 667 667 667 667 667 R2 0.02348 0.05358 0.02626 0.01853 0.03037 Adjusted R2 0.00556 0.03621 0.00839 0.00052 0.01258 F-test 1.3103 3.0853 1.4698 1.0289 1.7073 F-test, p-value 0.20751 0.00029 0.13050 0.41997 0.06115 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 51 Table 17: Bias tests for Ambition: Marriage Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) 0.0031 0.0204 0.0462∗∗∗ 0.0029 -0.0052 (0.0091) (0.0126) (0.0153) (0.0082) (0.0078) data_roundR3 0.0036 -0.0231∗ -0.0479∗∗∗ 0.0106 0.0136∗ (0.0087) (0.0121) (0.0147) (0.0078) (0.0075) refugee -0.0091 -0.0068 -0.0032 -0.0008 -0.0002 (0.0057) (0.0079) (0.0096) (0.0051) (0.0049) num_child -0.0002 0.0005 0.0036 -0.0029∗∗ -0.0028∗∗ (0.0014) (0.0020) (0.0024) (0.0013) (0.0012) hh_head_sex 0.0003 -0.0018 0.0013 0.0006 −4.3 × 10−5 (0.0052) (0.0072) (0.0087) (0.0046) (0.0044) hh_head_age 4.94 × 10−5 −2.9 × 10−5 -0.0004 0.0003 0.0004∗ (0.0002) (0.0003) (0.0004) (0.0002) (0.0002) parent_eduyears -0.0007 -0.0025∗∗∗ -0.0006 -0.0004 -0.0003 (0.0006) (0.0008) (0.0010) (0.0005) (0.0005) parent_reledu -0.0162∗ -0.0172 -0.0178 -0.0103 -0.0093 (0.0098) (0.0136) (0.0164) (0.0088) (0.0084) eld_sex -0.0010 0.0077 0.0256∗∗∗ -0.0072∗∗ -0.0079∗∗ (0.0040) (0.0055) (0.0067) (0.0036) (0.0034) eld_age -0.0005∗ 0.0001 0.0005 -0.0004 -0.0004 (0.0003) (0.0004) (0.0005) (0.0003) (0.0003) hh_asset_index -0.0008 -0.0034 -0.0061∗∗ -0.0014 -0.0023∗ (0.0016) (0.0023) (0.0028) (0.0015) (0.0014) hh_income 0.0016 0.0026 0.0025 0.0010 0.0009 (0.0012) (0.0016) (0.0020) (0.0011) (0.0010) int_trauma_exp 0.0003 -0.0015 -0.0007 −9.1 × 10−5 0.0002 (0.0009) (0.0012) (0.0014) (0.0008) (0.0007) Fit statistics Observations 667 667 667 667 667 R2 0.02664 0.05043 0.08818 0.02193 0.03117 Adjusted R2 0.00878 0.03301 0.07144 0.00398 0.01339 F-test 1.4918 2.8945 5.2702 1.2217 1.7535 F-test, p-value 0.12203 0.00065 1.6 × 10−8 0.26352 0.05235 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 52 Table 18: Bias tests for Ambition: Migration Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) 0.0020 -0.0100 0.0057 0.0024 0.0082 (0.0081) (0.0091) (0.0080) (0.0057) (0.0067) data_roundR3 0.0044 -0.0111 -0.0099 -0.0051 -0.0076 (0.0078) (0.0088) (0.0077) (0.0054) (0.0064) refugee -0.0099∗ 0.0009 0.0031 -0.0001 -0.0009 (0.0051) (0.0057) (0.0050) (0.0036) (0.0042) num_child 0.0009 -0.0004 0.0018 4.24 × 10−6 0.0006 (0.0013) (0.0014) (0.0013) (0.0009) (0.0011) hh_head_sex -0.0080∗ 0.0050 8.42 × 10−5 -0.0011 -0.0021 (0.0046) (0.0052) (0.0045) (0.0032) (0.0038) hh_head_age -0.0004∗∗ -0.0003 −1.95 × 10−5 3.56 × 10−6 -0.0001 (0.0002) (0.0002) (0.0002) (0.0001) (0.0002) parent_eduyears −8.93 × 10−5 0.0003 0.0003 -0.0002 -0.0007∗ (0.0005) (0.0006) (0.0005) (0.0004) (0.0004) parent_reledu 0.0062 0.0088 0.0081 0.0054 0.0071 (0.0088) (0.0098) (0.0086) (0.0061) (0.0072) eld_sex 0.0067∗ 0.0014 -0.0043 0.0003 0.0007 (0.0036) (0.0040) (0.0035) (0.0025) (0.0029) eld_age 5.58 × 10−5 0.0004 7.59 × 10−5 6.47 × 10−5 0.0001 (0.0003) (0.0003) (0.0003) (0.0002) (0.0002) hh_asset_index -0.0014 -0.0001 -0.0004 1.38 × 10−5 0.0009 (0.0015) (0.0016) (0.0014) (0.0010) (0.0012) hh_income 0.0003 -0.0002 -0.0002 −2.67 × 10−5 -0.0008 (0.0011) (0.0012) (0.0010) (0.0007) (0.0009) int_trauma_exp -0.0009 -0.0004 −9.94 × 10−5 0.0005 0.0004 (0.0008) (0.0009) (0.0008) (0.0005) (0.0006) Fit statistics Observations 667 667 667 667 667 R2 0.03138 0.00724 0.02165 0.00823 0.01693 Adjusted R2 0.01361 -0.01098 0.00370 -0.00997 -0.00111 F-test 1.7658 0.39750 1.2059 0.45233 0.93836 F-test, p-value 0.05019 0.96466 0.27459 0.94137 0.50766 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 53 Table 19: Bias tests for Capacity: Reliance On God Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.0735∗∗∗ -0.1134∗∗∗ -0.0993∗∗∗ 0.0008 0.0066 (0.0158) (0.0217) (0.0198) (0.0120) (0.0117) data_roundR3 -0.0198 -0.0501∗∗ -0.0186 -0.0073 -0.0080 (0.0151) (0.0208) (0.0190) (0.0115) (0.0112) refugee 0.0215∗∗ 0.0242∗ 0.0325∗∗∗ 0.0103 0.0070 (0.0099) (0.0136) (0.0124) (0.0075) (0.0073) num_child 0.0004 0.0014 0.0004 0.0012 -0.0006 (0.0025) (0.0034) (0.0031) (0.0019) (0.0019) hh_head_sex -0.0002 -0.0085 -0.0230∗∗ 0.0097 0.0062 (0.0090) (0.0123) (0.0112) (0.0068) (0.0067) hh_head_age −6.25 × 10−5 -0.0003 -0.0003 -0.0005∗ -0.0004 (0.0004) (0.0005) (0.0005) (0.0003) (0.0003) parent_eduyears -0.0010 -0.0015 -0.0008 -0.0005 -0.0009 (0.0010) (0.0014) (0.0012) (0.0007) (0.0007) parent_reledu -0.0275 -0.0125 -0.0085 -0.0181 -0.0071 (0.0170) (0.0233) (0.0213) (0.0129) (0.0126) eld_sex 0.0074 0.0095 0.0077 -0.0012 -0.0020 (0.0069) (0.0095) (0.0087) (0.0052) (0.0051) eld_age 0.0007 0.0018∗∗ 0.0007 0.0002 0.0003 (0.0006) (0.0008) (0.0007) (0.0004) (0.0004) hh_asset_index 0.0066∗∗ 0.0082∗∗ 0.0075∗∗ 0.0042∗∗ 0.0037∗ (0.0028) (0.0039) (0.0036) (0.0022) (0.0021) hh_income 0.0006 -0.0018 0.0002 0.0014 0.0018 (0.0021) (0.0028) (0.0026) (0.0016) (0.0015) int_trauma_exp 0.0013 -0.0007 0.0007 -0.0002 0.0004 (0.0015) (0.0020) (0.0019) (0.0011) (0.0011) Fit statistics Observations 667 667 667 667 667 R2 0.02232 0.02407 0.02126 0.01891 0.01410 Adjusted R2 0.00438 0.00617 0.00331 0.00091 -0.00398 F-test 1.2443 1.3443 1.1841 1.0507 0.77974 F-test, p-value 0.24830 0.18863 0.29035 0.40002 0.67179 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 54 Table 20: Bias tests for Capacity: Ability High Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.1597∗∗∗ -0.1068∗∗∗ -0.0103 -0.0190 -0.0156 (0.0253) (0.0215) (0.0167) (0.0135) (0.0138) data_roundR3 0.0807∗∗∗ 0.0153 -0.0145 0.0050 0.0120 (0.0243) (0.0206) (0.0160) (0.0129) (0.0132) refugee -0.0042 -0.0002 0.0038 -0.0008 0.0021 (0.0159) (0.0134) (0.0105) (0.0085) (0.0086) num_child 0.0071∗ 0.0048 0.0002 0.0009 0.0014 (0.0040) (0.0034) (0.0027) (0.0021) (0.0022) hh_head_sex 0.0045 0.0061 -0.0218∗∗ 0.0001 -0.0043 (0.0144) (0.0122) (0.0095) (0.0077) (0.0078) hh_head_age −3.34 × 10−5 0.0003 -0.0006 0.0001 -0.0002 (0.0006) (0.0005) (0.0004) (0.0003) (0.0003) parent_eduyears -0.0024 -0.0010 0.0017 0.0005 -0.0002 (0.0016) (0.0013) (0.0010) (0.0008) (0.0009) parent_reledu -0.0200 -0.0648∗∗∗ 0.0138 0.0102 0.0130 (0.0272) (0.0231) (0.0180) (0.0145) (0.0148) eld_sex 0.0175 -0.0014 0.0117 0.0046 0.0061 (0.0111) (0.0094) (0.0073) (0.0059) (0.0060) eld_age -0.0003 -0.0002 0.0005 5.98 × 10−5 0.0001 (0.0009) (0.0008) (0.0006) (0.0005) (0.0005) hh_asset_index 0.0013 -0.0035 -0.0008 0.0015 0.0038 (0.0046) (0.0039) (0.0030) (0.0024) (0.0025) hh_income 0.0026 0.0016 0.0018 0.0021 2.23 × 10−5 (0.0033) (0.0028) (0.0022) (0.0018) (0.0018) int_trauma_exp 0.0017 0.0016 0.0011 0.0018 0.0014 (0.0024) (0.0020) (0.0016) (0.0013) (0.0013) Fit statistics Observations 667 667 667 667 667 R2 0.08009 0.02463 0.01936 0.01452 0.01975 Adjusted R2 0.06321 0.00673 0.00137 -0.00356 0.00177 F-test 4.7448 1.3763 1.0762 0.80300 1.0982 F-test, p-value 1.8 × 10−7 0.17214 0.37742 0.64763 0.35848 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 55 Table 21: Bias tests for Capacity: Ability Low Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.1072∗∗∗ -0.1025∗∗∗ -0.0482∗∗∗ -0.0090 -0.0170 (0.0214) (0.0221) (0.0180) (0.0140) (0.0138) data_roundR3 0.0694∗∗∗ 0.0658∗∗∗ 0.0352∗∗ 0.0001 0.0058 (0.0205) (0.0212) (0.0173) (0.0134) (0.0133) refugee 0.0103 -0.0070 -0.0106 0.0018 0.0014 (0.0134) (0.0138) (0.0113) (0.0088) (0.0087) num_child -0.0024 0.0012 0.0025 0.0043∗ 0.0036 (0.0034) (0.0035) (0.0029) (0.0022) (0.0022) hh_head_sex -0.0369∗∗∗ -0.0229∗ -0.0154 -0.0049 -0.0024 (0.0122) (0.0126) (0.0102) (0.0080) (0.0079) hh_head_age 6.31 × 10−5 0.0002 0.0003 −3.59 × 10−5 0.0002 (0.0005) (0.0005) (0.0004) (0.0003) (0.0003) parent_eduyears 0.0027∗∗ 0.0025∗ 0.0015 -0.0009 -0.0004 (0.0013) (0.0014) (0.0011) (0.0009) (0.0009) parent_reledu -0.0160 -0.0109 -0.0292 -0.0159 -0.0373∗∗ (0.0230) (0.0238) (0.0194) (0.0151) (0.0149) eld_sex 0.0096 0.0097 -0.0016 0.0029 -0.0011 (0.0094) (0.0097) (0.0079) (0.0061) (0.0061) eld_age 7.46 × 10−5 -0.0005 -0.0004 9.65 × 10−5 −8.11 × 10−5 (0.0008) (0.0008) (0.0006) (0.0005) (0.0005) hh_asset_index -0.0027 -0.0075∗ -0.0016 -0.0017 -0.0005 (0.0039) (0.0040) (0.0032) (0.0025) (0.0025) hh_income 0.0037 0.0016 0.0005 0.0007 0.0002 (0.0028) (0.0029) (0.0023) (0.0018) (0.0018) int_trauma_exp -0.0004 -0.0033 -0.0009 -0.0019 -0.0012 (0.0020) (0.0021) (0.0017) (0.0013) (0.0013) Fit statistics Observations 667 667 667 667 667 R2 0.11546 0.07145 0.03663 0.01581 0.01706 Adjusted R2 0.09923 0.05441 0.01895 -0.00225 -0.00097 F-test 7.1142 4.1934 2.0723 0.87541 0.94597 F-test, p-value 2.85 × 10−12 2.22 × 10−6 0.01688 0.57210 0.50002 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 56 Table 22: Bias tests for Capacity: Budget High Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.0092 -0.0277 -0.0472∗∗∗ 0.0070 -0.0051 (0.0109) (0.0169) (0.0164) (0.0105) (0.0109) data_roundR3 0.0140 -0.0153 0.0346∗∗ 0.0046 0.0027 (0.0105) (0.0162) (0.0158) (0.0101) (0.0105) refugee -0.0067 0.0035 -0.0076 0.0021 -0.0036 (0.0069) (0.0106) (0.0103) (0.0066) (0.0068) num_child 0.0008 -0.0013 -0.0043∗ -0.0003 0.0003 (0.0017) (0.0027) (0.0026) (0.0017) (0.0017) hh_head_sex -0.0030 -0.0003 -0.0166∗ 0.0110∗ -0.0026 (0.0062) (0.0096) (0.0093) (0.0060) (0.0062) hh_head_age 0.0001 -0.0001 0.0005 1.9 × 10−5 −6.18 × 10−5 (0.0003) (0.0004) (0.0004) (0.0003) (0.0003) parent_eduyears 0.0013∗ 0.0005 0.0007 0.0002 0.0006 (0.0007) (0.0011) (0.0010) (0.0007) (0.0007) parent_reledu 0.0018 -0.0113 -0.0253 0.0002 0.0042 (0.0118) (0.0182) (0.0177) (0.0113) (0.0117) eld_sex 0.0029 0.0068 -0.0158∗∗ -0.0107∗∗ -0.0018 (0.0048) (0.0074) (0.0072) (0.0046) (0.0048) eld_age -0.0007∗ −1.18 × 10−5 -0.0007 -0.0004 −5.27 × 10−5 (0.0004) (0.0006) (0.0006) (0.0004) (0.0004) hh_asset_index -0.0021 -0.0007 0.0004 0.0033∗ 0.0018 (0.0020) (0.0030) (0.0030) (0.0019) (0.0020) hh_income 0.0036∗∗ 0.0062∗∗∗ 0.0011 0.0031∗∗ 0.0011 (0.0014) (0.0022) (0.0021) (0.0014) (0.0014) int_trauma_exp 0.0018∗ 0.0014 -0.0004 0.0004 0.0011 (0.0010) (0.0016) (0.0015) (0.0010) (0.0010) Fit statistics Observations 667 667 667 667 667 R2 0.03223 0.02420 0.04615 0.03678 0.01432 Adjusted R2 0.01447 0.00629 0.02865 0.01910 -0.00377 F-test 1.8149 1.3515 2.6371 2.0807 0.79180 F-test, p-value 0.04242 0.18482 0.00188 0.01636 0.65928 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 57 Table 23: Bias tests for Capacity: Budget Low Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) 0.0064 -0.1465∗∗∗ -0.0080 0.0140 0.0073 (0.0145) (0.0227) (0.0208) (0.0177) (0.0170) data_roundR3 -0.0225 0.0435∗∗ -0.0352∗ -0.0153 -0.0080 (0.0139) (0.0218) (0.0200) (0.0170) (0.0163) refugee -0.0102 -0.0141 -0.0314∗∗ -0.0223∗∗ -0.0230∗∗ (0.0091) (0.0142) (0.0130) (0.0111) (0.0106) num_child 0.0025 0.0023 -0.0014 0.0050∗ 0.0059∗∗ (0.0023) (0.0036) (0.0033) (0.0028) (0.0027) hh_head_sex -0.0144∗ -0.0233∗ -0.0098 -0.0055 -0.0008 (0.0083) (0.0129) (0.0118) (0.0100) (0.0096) hh_head_age −4.17 × 10−5 0.0004 -0.0009∗ -0.0005 -0.0003 (0.0004) (0.0006) (0.0005) (0.0004) (0.0004) parent_eduyears -0.0009 0.0010 0.0007 -0.0011 -0.0006 (0.0009) (0.0014) (0.0013) (0.0011) (0.0011) parent_reledu 6.56 × 10−5 -0.0070 -0.0189 -0.0134 -0.0064 (0.0156) (0.0244) (0.0224) (0.0190) (0.0183) eld_sex 0.0079 0.0261∗∗∗ 0.0017 0.0120 0.0083 (0.0064) (0.0099) (0.0091) (0.0078) (0.0074) eld_age 0.0003 -0.0011 0.0001 0.0002 −3.12 × 10−5 (0.0005) (0.0008) (0.0007) (0.0006) (0.0006) hh_asset_index 0.0006 -0.0012 -0.0017 -0.0016 -0.0035 (0.0026) (0.0041) (0.0037) (0.0032) (0.0031) hh_income 0.0006 -0.0003 -0.0002 0.0008 5.47 × 10−5 (0.0019) (0.0029) (0.0027) (0.0023) (0.0022) int_trauma_exp 0.0015 0.0028 0.0016 0.0015 0.0007 (0.0014) (0.0021) (0.0020) (0.0017) (0.0016) Fit statistics Observations 667 667 667 667 667 R2 0.02488 0.03241 0.04124 0.02117 0.01903 Adjusted R2 0.00699 0.01465 0.02365 0.00321 0.00103 F-test 1.3905 1.8253 2.3443 1.1790 1.0575 F-test, p-value 0.16518 0.04092 0.00601 0.29415 0.39394 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 58 Table 24: Bias tests for Capacity: Awareness Information High Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.2264∗∗∗ -0.0253 -0.2251∗∗∗ -0.0274 -0.0159 (0.0324) (0.0232) (0.0318) (0.0185) (0.0182) data_roundR3 0.0049 -0.0259 0.0386 0.0093 -0.0104 (0.0311) (0.0223) (0.0305) (0.0178) (0.0175) refugee -0.0197 0.0072 -0.0064 0.0175 0.0133 (0.0203) (0.0146) (0.0199) (0.0116) (0.0114) num_child 0.0017 -0.0035 0.0004 -0.0011 -0.0010 (0.0051) (0.0037) (0.0050) (0.0029) (0.0029) hh_head_sex 0.0345∗ 0.0060 0.0372∗∗ -0.0074 -0.0044 (0.0184) (0.0132) (0.0181) (0.0105) (0.0103) hh_head_age 0.0006 0.0004 0.0011 −1.72 × 10−5 -0.0002 (0.0008) (0.0006) (0.0008) (0.0005) (0.0004) parent_eduyears -0.0036∗ -0.0004 -0.0032 0.0022∗ 0.0029∗∗ (0.0020) (0.0015) (0.0020) (0.0012) (0.0011) parent_reledu -0.0068 -0.0314 0.0089 0.0089 -0.0032 (0.0349) (0.0250) (0.0342) (0.0199) (0.0196) eld_sex 0.0319∗∗ 0.0013 0.0006 0.0128 0.0147∗ (0.0142) (0.0102) (0.0139) (0.0081) (0.0080) eld_age 0.0001 -0.0005 -0.0004 -0.0004 3.4 × 10−6 (0.0011) (0.0008) (0.0011) (0.0007) (0.0006) hh_asset_index -0.0014 0.0010 0.0018 0.0048 0.0041 (0.0058) (0.0042) (0.0057) (0.0033) (0.0033) hh_income 0.0019 0.0011 0.0037 -0.0003 0.0004 (0.0042) (0.0030) (0.0041) (0.0024) (0.0024) int_trauma_exp 0.0047 0.0021 0.0032 0.0004 0.0011 (0.0031) (0.0022) (0.0030) (0.0017) (0.0017) Fit statistics Observations 667 667 667 667 667 R2 0.02910 0.02823 0.02511 0.01993 0.02743 Adjusted R2 0.01128 0.01040 0.00723 0.00194 0.00959 F-test 1.6334 1.5831 1.4040 1.1081 1.5372 F-test, p-value 0.07802 0.09173 0.15880 0.35014 0.10605 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 59 Table 25: Bias tests for Capacity: Awareness Information Low Dependent Variables: ChatGPT Llama213Bchat Llama213B iQualChatGPTaug iQual Model: (1) (2) (3) (4) (5) Variables (Intercept) -0.3240∗∗∗ -0.6286∗∗∗ -0.4136∗∗∗ -0.0033 -0.0087 (0.0300) (0.0318) (0.0340) (0.0076) (0.0088) data_roundR3 0.0630∗∗ 0.1408∗∗∗ 0.0198 -0.0005 0.0025 (0.0287) (0.0305) (0.0326) (0.0073) (0.0085) refugee 0.0177 0.0436∗∗ 0.0255 0.0101∗∗ 0.0071 (0.0188) (0.0199) (0.0213) (0.0048) (0.0055) num_child 0.0028 0.0036 -0.0016 0.0003 -0.0009 (0.0048) (0.0050) (0.0054) (0.0012) (0.0014) hh_head_sex -0.0284∗ -0.0319∗ -0.0647∗∗∗ -0.0033 -0.0071 (0.0170) (0.0181) (0.0193) (0.0043) (0.0050) hh_head_age -0.0005 -0.0004 -0.0007 -0.0001 0.0001 (0.0007) (0.0008) (0.0008) (0.0002) (0.0002) parent_eduyears 0.0052∗∗∗ 0.0063∗∗∗ 0.0038∗ -0.0007 -0.0013∗∗ (0.0019) (0.0020) (0.0021) (0.0005) (0.0006) parent_reledu 0.0506 0.0068 -0.0011 -0.0082 -0.0116 (0.0322) (0.0342) (0.0365) (0.0082) (0.0095) eld_sex -0.0089 0.0080 -0.0116 0.0012 0.0006 (0.0131) (0.0139) (0.0149) (0.0033) (0.0039) eld_age -0.0002 -0.0004 -0.0004 0.0001 −3.79 × 10−6 (0.0011) (0.0011) (0.0012) (0.0003) (0.0003) hh_asset_index 0.0053 0.0078 0.0094 0.0013 0.0005 (0.0054) (0.0057) (0.0061) (0.0014) (0.0016) hh_income 0.0009 0.0024 0.0025 0.0009 0.0012 (0.0039) (0.0041) (0.0044) (0.0010) (0.0011) int_trauma_exp -0.0004 -0.0005 −5.23 × 10−5 0.0004 -0.0002 (0.0028) (0.0030) (0.0032) (0.0007) (0.0008) Fit statistics Observations 667 667 667 667 667 R2 0.05864 0.15437 0.04177 0.01835 0.01859 Adjusted R2 0.04137 0.13885 0.02419 0.00034 0.00058 F-test 3.3950 9.9489 2.3757 1.0187 1.0323 F-test, p-value 7.7 × 10−5 4.8 × 10−18 0.00532 0.42946 0.41682 IID standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 60 Figure 7: Refugee and Male eldest child coefficients for all annotations Aspiration: Religious Aspiration: Secular Ambition: No Ambition Ambition: Job Secular Refugee status Male eldest child 00 01 02 03 50 25 00 25 50 4 2 00 02 50 25 00 25 .0 .0 0. 0. 0. 0. .0 .0 0 0 0 0. 0. .0 .0 0 0 −0 −0 0. 0. 0. 0. 0. −0 −0 −0 −0 Ambition: Vocational Training Ambition: Entrepreneur Ambition: Education High Ambition: Education Neutral Refugee status Male eldest child 1 00 01 02 2 1 00 01 02 0 5 0 5 0 2 00 02 04 06 .0 .0 .0 5 2 00 02 05 .0 0. 0. 0. 0. 0. 0. .0 .0 0. 0. 0. 0. −0 −0 −0 −0 0. 0. 0. −0 −0 Ambition: Education Low Ambition: Education Religious Ambition: Marriage Ambition: Migration Model Expert annotations Refugee status iQual iQual + ChatGPT aug ChatGPT Male eldest child Llama−2 13B Llama−2 13B chat 01 00 01 02 00 01 02 04 02 00 02 0 5 0 5 0 00 00 01 01 02 . 0. 0. 0. 0. 0. 0. . . 0. 0. −0 −0 −0 0. 0. 0. 0. 0. Capacity: Reliance On God Capacity: Ability High Capacity: Ability Low Capacity: Budget High Refugee status Male eldest child 4 2 00 02 5 0 5 0 1 00 01 2 1 00 01 .0 .0 02 00 02 05 .0 .0 .0 0. 0. 0. 0. 0. 0. −0 −0 −0 −0 −0 . 0. 0. 0. −0 Capacity: Budget Low Capacity: Awareness Information High Capacity: Awareness Information Low Refugee status Male eldest child 2 1 0 1 25 0 5 0 4 3 2 1 0 1 .0 .0 0 0 00 02 05 .0 .0 .0 .0 0 0 0. 0. .0 0. 0. −0 −0 −0 −0 −0 −0 0. 0. 0. −0 Coefficient Note : 61