Introduction

Automatically gathering data about examinees’ behavior, often referred to as log file or process data, has become increasingly common in computerized large-scale assessments (LSAs). Such data include examinees’ response time to the test or its items, what they clicked on, and typing sequences, among others. This information can help improve item design, determine examinees’ engagement, and inform inferences about the construct (Oranje et al., 2017), primarily if the use of such data is based on an existing theory of the cognitive processes underlying people’s approach to a test (Brückner & Pellegrino, 2017). That way, log file data can provide evidence of the validity of the assessment (AERA, APA, & NCME, 2014). Moreover, they can help understand and reframe achievement differences in light of different test-taking strategies (Pohl et al., 2021).

The purpose of the current study is to provide a comprehensive review of common topics in existing research on log file or process data in LSAs. Given the value of educational and psychological theories in interpreting process data (Goldhammer & Zehner, 2017) and the relative lack of use of such theories in studies reporting such data in other contexts (Banihashem et al., 2018), we also focused on commonly used theories in the studies we reviewed. To meet these goals, we applied topic modeling (Blei, 2012), a natural language approach for automatically detecting topics in textual data.

This study would hopefully contribute to the existing literature in multiple ways: (a) reveal gaps in the literature (e.g., underexplored assessment domains or methods) by examining what has been studied so far, (b) highlight useful theories that could be relevant for future studies, (c) inform test developers of different ways in which log file data can be used to enhance assessment design, validation, and interpretation, and (d) introduce topic modeling as a method for literature review to the field of educational assessment and discuss its potential benefits and risks when used for that purpose.

Process data in LSAs

Log file data have become central in LSAs. They are routinely collected and published by many international and smaller-scale LSA programs, including the Programme for International Student Assessment (PISA; OECD, 2015), the Trends in Mathematics and Science Study (TIMSS; Fishbein et al., 2021), and the National Assessment of Educational Progress (NAEP; Bergner & von Davier, 2019). In their recent review of the literature, Reis Costa and Leoncio Netto (2022) discuss some seminal work in the field and show how they use different process variables to inform item-, person-, and system-level analyses.

There is no consensus on the definition of process data. Reis Costa and Leoncio Netto (2022), for example, define process data as “any type of information (e.g., response actions or timing) recorded on a computer platform into electronic files.” Namely, this view of process data relies on the technical aspects of how the data were collected. In contrast, Provasnik (2021) focuses on the content or purpose of the data and defines process data as “empirical data that reflect the process of working on a test question.” Such data are not necessarily recorded on a computer and may be gathered via cognitive interviews, eye tracking, brain imaging, etc. According to Provasnik (2021), then, Reis Costa and Leoncio Netto’s (2022) view of process data fits better to the term “log file data”, which highlights the source of the data rather than their purpose.

This work focuses on process data captured automatically in log files, log file data according to Provasnik (2021). However, while we agree that the distinction between log file and process data is crucial, it is relatively new and has not been consistently applied in the literature yet. It seems like much of the existing literature refers to log file data when using the term process data either with (e.g., Bergner & von Davier, 2019; Goldhammer et al., 2021a, 2021b) or without acknowledging their difference (e.g., Chen et al., 2019; Goldhammer et al., 2017a, 2017b).

In order to reflect that, we use the term “process data” to refer to any data automatically collected about test-takers’ response process. Such data typically include timing data (e.g., response time per item, the time between clicks) and the actions performed by the test-taker (keystrokes, mouse clicks, selection, navigation, etc.). They do not include automatically-collected data that goes beyond the response process like the device and setting in which the assessment took place (i.e., access-related data; Kroehne & Goldhammer, 2018), metadata, or the responses themselves. While such data are interesting and valuable, they are beyond the scope of our definition.

Theoretical models for process data in assessments

The underuse of theories is often discussed as an issue in learning analytics research (Khalil et al., 2023). Perhaps due to the wealth of data or the rapid methodological advancements of recent years, many studies using process data or other clickstream data seem to be purely empirical, focusing on methods and findings rather than on testing theoretical predictions. While data-driven studies are certainly important for describing phenomena and can be useful for informing theory development, particularly in nascent fields like learning analytics, theories are key for generating hypotheses, interpreting findings, and applying them in practice (Banihashem et al., 2019).

Theory seems to play a more important role in the field of educational assessment. For example, the most commonly acceptable definition of validity (AERA, APA, & NCME, 2014) refers to both empirical evidence and theory as important aspects of validity. Theoretical models describing how the target construct is structured, developed, and manifested are key in assessment construction (Rupp et al., 2010), and inform task design, validation, and interpretation (Borsboom et al., 2004).

Kane and Mislevy (2017) show how theories and process data can be used together for validation purposes. Cognitive theories that describe the mechanisms behind how performance in a certain task (e.g., responding to an item) is produced can be used to guide process data interpretation. Process data can then be used as evidence for said mechanisms, or as evidence that the mechanisms are not being used in practice (e.g., students are guessing). Motivational theories can also be used to explain why unexpected strategies are used, potentially contributing to preventing such issues in the future.

In spite of their importance in assessment, it is unclear whether theories are being used in practice in studies involving process data. As process data has not been consistently collected and studied until relatively recently, theories are likely underused in studies using process data, similar to the status of the learning analytics field as a whole. Therefore, one of the purposes of this study is to examine whether and which theories have been used in the field so far.

Topic modeling for literature reviews

Given the multitude of studies applying process data in the context of LSAs, we decided to apply a natural language processing (NLP) approach to classify the studies. Specifically, we used topic modeling (Blei, 2012), an unsupervised learning method used to detect latent commonalities, or topics, in a large group of documents (in our case, research papers). This approach seems to be particularly useful for a review that focuses on what the main topics studied in a field are, and it has been applied in contexts similar to ours. For instance, Chen et al. (2022) used topic modeling to identify themes in learning analytics research and found that these papers covered topics like web services for learning, flipped learning, and assessment.

Topic modeling is a name for a group of methods (see Vayansky & Kumar, 2020 for a review). We chose to use structural topic modeling (STM; Roberts et al., 2016), which allows for the inclusion of metadata. This approach was preferred because it enabled us to include the articles’ first author as an additional data source when identifying topics. Since authors tend to write about similar topics, we believe including the first author as an additional information source could lead to more stable topics. We now turn to an intuitive introduction to topic models. A more rigorous treatment is available elsewhere (e.g., Blei, 2012).

In any topic model, it is assumed that there is a known number of topics in the data, and each topic is related to words with different probabilities. So, in a dataset of restaurant menus, there can be a topic such as “breakfast” which is related to the occurrence of words like bagel or eggs, and a topic such as “drinks” which is indicated by words like tea or soda on the menu. Then, each document (in our case, menus) contains a number of words and the distribution of these terms emerges from a mixture of topics with different probabilities. So, a menu of a bagel shop is more likely to be generated from the topic “breakfast” than from the topic “drinks”, and it is more likely to include the words bagel or eggs. A menu of a café might be a mixture of both categories and may include eggs but also tea. In structural topic modeling, information about the restaurant’s location or head chef could also inform the identification of topics. For instance, a menu of a restaurant in New England might be more likely to be of the topic “seafood” than “barbecue.” In reality, the probabilities that words and documents are related to certain topics are unknown and have to be estimated from the observed documents.

In our case, the model identifies groups of articles that share similar content (i.e., words) and uses these common words to derive the topics and their relationships. Then, each paper is assigned a set of probabilities to be representative of these topics. Some papers may clearly belong to one single topic and not to the others (as was the case for our bagel shop menu), so they will be assigned with a high probability of being in that topic and a low probability of being in the others. Other papers may represent a mixture of several topics. Regardless, topic modeling is commonly used for grouping large collections of text documents by their content, making it particularly appropriate for literature reviews. In this study, we used topic modeling to identify common themes in the extant research on process data in LSAs.

The current study

There are no comprehensive literature reviews of the use of process data in LSAs. Existing reviews focus on the use of technology in LSAs (Zenisky & Sireci, 2002) or the use of process data in related fields (e.g., higher education; Viberg et al., 2018). In their recent work, Reis Costa and Leoncio Netto (2022) review selected papers related to process data in international LSAs. While valuable, this work does not introduce the full picture of how process data is used in LSAs research, as smaller-scale assessments are some of the most advanced users and researchers of such data (e.g., Bergner & von Davier, 2019). In addition, Reis Costa and Leoncio Netto (2022) use an ecological framework to analyze their target studies and do not examine the theories used in the studies themselves.

In the current review, we explore common topics in the literature on process data in LSAs as a whole. We also looked at whether and which theories have been used in existing studies. We believe that this is the first step in understanding what is already known, what remains to be explored, and how theories can inform future studies, thus contributing to other researchers interested in this topic. So, our research questions are:

RQ1: What are the main topics in studies involving process data in LSAs?

RQ2: To what extent are they informed by psychological or educational theories?

Methods

Our review follows Arksey and O’Malley’s (2005) framework for conducting rigorous scoping reviews. Scoping reviews aim to provide a snapshot of the field and describe the range of relevant studies without evaluating their quality. Since we were interested in the various topics being explored in the literature and did not aim to comment on their quality, and given the diverse foci of the relevant studies, we considered this approach to be appropriate. Arksey and O’Malley’s framework has five stages, which were repeated iteratively until a final sample of studies was identified. The five stages are presented below.

Identifying the research question

In this stage, the authors define their research questions and all of their variables. Our research questions are presented above. Our main variable, process data, is defined as any data automatically collected about test-takers’ response process. As discussed above, while we acknowledge the complexity and disagreement in the literature on what constitutes process or log file data, we found this definition to be simple and inclusive of the studies we were interested in.

Definitions of LSAs can also be complex (see a discussion in Khorramdel et al., 2023). However, we again chose a rather simple definition of LSAs which allowed us to include a wide variety of assessments and therefore, studies: LSAs are any achievement or knowledge assessments administered to many test-takers, at least to an entire school district. Future studies might focus on more specific LSA definitions, e.g., those focusing on monitoring educational systems.

We posed no limitations on the age of the participants. We included both children (e.g., school achievement tests like TIMSS and PISA) and adults (e.g., licensure tests or the Program for the International Assessment of Adult Competencies [PIAAC]). Since we were not interested in the scale of the study but rather the type of assessment used, we included studies that used publicly available LSA items but were small-scale themselves (e.g., fewer than 100 participants).

Finally, we wanted to focus on educational assessments, so we chose to focus our search on the content domains of reading literacy, mathematics and numeracy, and science. These domains were selected as they are often the focus of LSA, though other content domains could also be examined in the future. We included studies not directly measuring these domains only if they were administered together with and were related to a test of one of these domains (e.g., a survey on the effort the student put into a mathematics assessment).

Identifying relevant studies

In this stage, the strategies used to find studies for review are defined. Our main method was to search electronic databases. We searched Scopus, ERIC, and Google Scholar from May to June 2022. We included several groups of keywords separated by the Boolean “AND”. The first involved the type of data we were interested in, and included the following terms separated by “OR”: “process data”, “response process”, “timing”, “clickstream”, “keystroke”, “log data”, “log files”, “activity logs”, “trace data”, “event data”, “response time”, “speededness”, “telemetry”, “sequence”, “learning analytics”, and “data mining”. The second group was related to assessment and included the terms “assessment”, “testing”, and “measurement” (separated by “OR”). Next, as we wanted to focus on specific assessment domains, we added these terms separated by “OR”: “reading”, “language”, “literacy”, “writing”, “numeracy”, “math”, “quantitative”, “science”, and “stem”. Finally, given a large number of irrelevant studies that were found, we added the context we were interested in, “education” OR “school”. In platforms that allowed for more focused searches, these terms were searched only in the title, keywords, and abstract of the papers to make sure only relevant results were identified. Otherwise, the full manuscript was searched. These steps resulted in 2718 articles from Scopus, 1094 from ERIC, and 1107 from Google Scholar, but after removing duplicates, 4677 papers remained.

After some steps related to study selection (see below), we examined the references of the remaining studies for titles that included the relevant keywords. We read those based on the study selection criteria and, if they were relevant, searched their references as well. This procedure was repeated until no new studies were added. This method added 158 papers to our list.

Study selection

Before the initial search, we developed several criteria to narrow down the results. First, we only included studies relevant to our RQ, namely, that relied on at least one type of process data in the context of an educational LSA. Note that while our initial content domains and search terms only included reading, math, and science, we identified many studies involving complex problem-solving (CPS) assessments, a topic related but not directly a part of any of these domains. We decided to include these studies given their ubiquity in the literature and CPS’ importance in the content areas we focused on (OECD, 2014).

We also limited our search to English-language papers that had some empirical components. That is, we included review papers or model development papers only if they included an application on empirical data. We did not limit ourselves in terms of the publication platform, except we did not include full books, dissertations, and test technical reports due to their volume; chapters in books and reports of studies done by test developers were included. Based on these criteria we read the papers’ titles and abstracts to determine their relevance. We excluded 4603 sources (see details in Table 1) and retained 74. Together with the 158 identified from these studies’ references, our final sample included 232 articles.

Table 1 Details on the excluded sources
Full size table

Charting the data

In this stage, the articles are coded for the variables of interest, usually by carefully reading them and recording variables relevant to the research questions. In our case, we first performed the topic modeling as a way of collating the articles (see “Collating, summarizing, and reporting the results” Section). We then manually coded their publication year and venue, first author, and the assessment they used. We also wrote a brief description of their methodology and some key findings. As we were interested in the articles’ theoretical basis, we also documented their use of theories. All of these were used to summarize the studies within each topic, as described in “Collating, summarizing, and reporting the results” Section.

Collating, summarizing, and reporting the results

As described in “Topic modeling for literature reviews” Section, we applied a structural topic model for an initial summarization of the data. We used the stm R package (Roberts et al., 2019) to extract the topics. Note that the topic modeling was conducted on the papers’ abstracts and not the full papers. This was done in order to focus on the most important terms within each paper rather than extraneous information that might result in meaningless topics (e.g., topics focusing on a specific assessment as opposed to the study’s content). We also excluded executive summaries which are usually longer than abstracts. So, we excluded 11 papers that did not have an abstract, and only 221 papers were included in the analysis.

We then performed several data-cleaning steps. We removed citations and section headings within structured abstracts (e.g., “background”, “conclusions”). This was done to minimize common phrases among papers that are not similar in their content. We also expanded contractions and spelled out all acronyms with the exception of test names (e.g., we kept TOEFL instead of Test of English as a Foreign Language, PIRLS for Progress in International Reading Literacy Study, etc.), as these are known by their abbreviated name, and expanding it could result in a wrong classification. Next, we used the automatic stm cleaning functions which perform standard NLP pre-processing such as setting all words to lowercase, removing punctuation and stopwords (function words like “the”), and stemming the words (recoding them to their most basic form, e.g., “responding” turned to “respons”).

Following these steps, we had to choose the optimal number of topics for the analysis. We did so by estimating models with five to ten topics. These numbers were selected because they would, assuming a uniform distribution, lead to an expected number of 20–40 papers per topic, which we considered reasonable for succinctly summarizing the literature. These five to ten topic models were then evaluated using their semantic coherence and FREX (frequency-exclusivity), two commonly used measures of topic models’ quality (e.g., Pandur et al., 2020). Semantic coherence represents how often the most common words of a topic occur with other common words, or how consistently these words cooccur. In a good-quality model, we expect to see high levels of coherence. For example, if a topic’s most common words are “log” and “file”, we would like them to appear together in most of the papers assigned to this topic.

A topic’s exclusivity is the extent to which the most common words in a topic are distinct from the words in other topics. This is a measure of how unique the topics are, and we are interested in exclusive, easily distinguishable topics. For example, if one topic had the words “response” and “time”, we would like these words to be rare in other topics. Exclusivity alone, however, is inflated by rare words; if a word appears only in two documents in the whole dataset and the documents belong to the same topic, that topic will have an unjustifiably high exclusivity value. So, we used FREX which takes into account the words’ frequency in the data overall when determining how different topics are. However, it is important to note that some overlap between topics should be expected even when FREX values are high, especially in this case where the texts share many similar words. A more formal treatment of semantic coherence and FREX is available in Mimno et al. (2011) and Bischof and Airoldi (2012), respectively.

Note that while we want models with high topic coherence as well as high topic FREX, there is usually a tradeoff between the two. Models with few topics tend to have high coherence values but have low FREX values because, while it is likely that the common words in each topic appear in many documents within that topic, it is also likely that they are not unique to that topic. There are also no acceptable thresholds for these measures as they heavily depend on the specific dataset. So, to select the optimal model, we plotted the coherence and FREX values for each of the five to ten topics models we estimated and chose a model that had relatively high values on both measures. We also preferred a more parsimonious solution, namely, with fewer topics.

Once we selected the number of topics we estimated our final model. This resulted in a vector of probabilities for each paper representing how likely it is to belong to each of the topics. For each paper, we chose the topic with the highest probability and assigned the paper to that topic. Then, we interpreted the models based on their most common words and some representative examples. The resulting topics and the content of the papers assigned to each topic are reported in the Results section. After each paper was assigned to a topic, we read all the abstracts in each topic to identify prominent subtopics. This way, the topic model assisted us in the initial identification of themes in the reviewed studies but we were also able to delve into more detailed types of studies. We also searched specifically for theories and findings in the respective sections of the papers. Table 2 presents this procedure.

Table 2 Summary of the data summarization stage
Full size table

Results

Figure 1 presents the mean semantic coherence and FREX of the models with five to ten topics. Based on these results and an examination of the topics themselves, we chose a model with six topics. The topics’ most common words and representative examples led us to label the topics as response time models, response time-general, aberrant test-taking behavior, action sequences, complex problem-solving, and digital writing.

Fig. 1
figure 1

The topic models’ coherence and FREX

Full size image

Table 3 presents the topics’ most common words, first authors, publication venues, and assessments used. The most prolific first authors in our sample overall were S. L. Wise (with 19 papers), W. J. van der Linden (8), and F. Goldhammer (7). The most common publication venues were the British Journal of Mathematical and Statistical Psychology (14), Applied Measurement in Education (13), and Frontiers in Psychology (12). The most common assessments used were PISA (40), PIAAC (26), and Measures of Academic Progress (MAP) Growth (20), though 32 studies did not name the specific assessment used. Table 4 presents descriptive statistics of the probabilities of the papers assigned to each topic. Appendix A presents the papers’ probabilities of being in each topic and the topic they were assigned to. Below we present these topics and the findings of studies assigned to these topics.

Table 3 The six topics and common unique words, first authors, publication venues, and assessments
Full size table
Table 4 Topic probabilities’ descriptive statistics
Full size table

Response time models

This topic includes mostly methodological papers where a new model for response time analysis is developed or evaluated. They were published between 1997 (Jansen, 1997) and 2021 (DeCarlo, 2021; Rios & Soland, 2021). These papers are rarely based on a theoretical foundation but many of them expand on existing measurement models (see below).

Many of the articles in this group introduce new models combining item responses and response times. They are usually a type of item response theory (IRT) model (e.g., Jeon & De Boeck, 2019; Pokropek, 2016; Wang & Hanson, 2005) or mixture models that identify different populations within the data (e.g., Liu et al., 2020; Ulitzsch et al., 2020c; Wang & Xu, 2015). Of specific note given its impact on the field is van der Linden’s (2007) hierarchical speed-accuracy framework. This framework suggests combining any IRT model that suits the data, a model describing the response time distribution (e.g., a lognormal distribution; van der Linden, 2006), and a higher-level model that accounts for any dependencies among the parameters of these models at the population level (e.g., high-ability test-takers also tend to work faster).

Many of the rest of the articles extend or develop tools to be used within van der Linden’s (2007) framework. Some relax van der Linden’s model’s assumptions, specifically that response time and accuracy are conditionally independent, allowing for correlations between response time and accuracy even after accounting for the overall ability and speed traits (Bolsinova et al., 2017a, 2017b; Meng et al., 2015). Others extend it to a multidimensional case (Lu et al., 2020; Man et al., 2019; Zhan et al., 2018). And yet others propose other additions such as fit statistics (Fox & Marianti, 2017), hypothesis testing (van der Linden & Glas, 2010), or invariance testing procedures (Glas & van der Linden, 2010) that can be applied to that model.

Response time-general

This group included papers that used response time not necessarily as a part of a methodological paper, though other than the use of response time, we were not able to find another common theme within this group. These papers were published between 1994 (Bergstrom et al., 1994; Parshall et al., 1994) and 2022 (Araneda et al., 2022). Most of these studies did not rely on theoretical grounds, but some mention the dual-processing model (Evans & Stanovich, 2013). This model divides higher-order thinking into processes that are performed rapidly and intuitively or slowly and methodically.

A common subtopic links response times with item or person characteristics. Some item characteristics that were found to be associated with longer response times are item difficulty, discrimination, length, late position in the test, and cognitive complexity (Bergstrom et al., 1994; Halkitis et al., 1996; Zenisky & Baldwin, 2006). Findings are less consistent with respect to person characteristics (Parshall et al., 1994; Zenisky & Baldwin, 2006). Finally, studies had different conclusions regarding the link between ability or response accuracy and response times, where some found that longer response times are associated with lower or higher ability whereas others found no association (Bergstrom et al., 1994; Bridgeman & Cline, 2000). This might be attributed to the behaviors described in the dual processing theory, where the association between ability and speed depends on whether the task requires slow vs. fast processing (also known as the time-on-task effect; Goldhammer et al., 2014).

Another set of studies involved item design. These studies looked at innovative item designs or layouts and how they differ from traditional multiple-choice items, mainly in terms of response time. For example, they tested the effects of innovative designs on response accuracy and speed with different results depending on the item type (Ponce et al., 2021; Saß et al., 2012). Other studies compared the information and efficiency (information per time unit) of these items, finding that they are at least as informative as multiple-choice items (Jodoin, 2003; Ponce et al., 2020; Wan & Henly, 2012).

Finally, a set of studies focused on measurement issues related to response time. These studies, for example, determined the reliability of a speed factor (Davison et al., 2012) and its invariance (Shin et al., 2020). They also used different methods to detect aberrant behaviors in a test (speededness or item preknowledge) and their potential impact on test design (Margolis et al., 2020; Qian et al., 2016).

Aberrant test-taking behaviors

The topic of Aberrant test-taking behaviors includes papers identifying and/or using information about participants’ actions that reflect disengagement or guessing. Unlike a subgroup of the previous topic, these studies did not necessarily focus on response time to identify aberrant test-taking behaviors (though many of them did). The studies assigned to this group were published between 1997 (Schnipke & Pashley, 1997; Schnipke & Scrams, 1997) and 2021 (e.g., Gorgun & Bulut, 2021).

About half of the studies do not mention theories. Among the rest, some prominent theories are the expectancy-value theory (Wigfield & Eccles, 2000) and the demands-capacity model (Wise & Smith, 2011). Broadly, the expectancy-value theory suggests that whether a person engages in a behavior (i.e., working hard on a test) depends on the behavior’s subjective value to that person and their expectations that they can succeed in performing it. The demands-capacity model posits that the effort an examinee exerts when responding to an item depends on what the item demands or the resources it requires and on the examinee’s capacity or willingness to work on the item. It is easy to see how both theoretical frameworks are relevant when studying examinees’ test-taking efforts. Schnipke and Scrams’ (1997) conceptualization of test-taking behavior as either solution behavior or rapid guessing behavior is also often mentioned in this category.

A common subtopic in this group involves the comparison of different response time thresholds for identifying non-effort in a given item or examinee; usually, non-effort is defined as spending less than a certain amount of time on an item or a group of items, and there are many methods in determining that amount of time. The most common item thresholds were absolute (e.g., below two seconds; Soland et al., 2019a, 2019b) or normative (e.g., 10% of the average time all test-takers spend on the item; Wise & Ma, 2012). There were other, more complex ones, for example, a visual inspection of the empirical response time distribution to identify item response times that were more common than expected when assuming solution behaviors (Lee & Jia, 2014), an examination of when response accuracy is below chance (Wise, 2019), fitting mixture models that separate effortful from non-effortful responses (Soland et al., 2021), and combining response times with the number of actions a test-taker took (Sahin & Colvin, 2020). Each of these methods has its advantages and disadvantages, and different studies disagree on which one is the best and in what context.

Methods for identifying disengaged test-takers are less involved and are usually based on the percent of items where disengagement was identified based on one or more of the aforementioned thresholds (e.g., a person who was determined to be non-effortful in over X% of the items is considered non-effortful, where X is often 10 or 15; Wise, 2014). A different approach which is more suitable for survey items was proposed by Soland et al. (2019a, 2019b), where multiple factors like responses to reverse-coded items and distribution of the selected response options were examined to determine participants’ effort.

Another subtopic involved work on using statistical methods for identifying disengagement. For instance, Ulitzsch et al. (2020a) developed a model for identifying disengaged participants based on rapid guesses and non-responses, and Wise and Kuhfeld (2021a) developed a method for identifying semi-effortful responses. Related, some works compared different methods or models for identifying disengagement in terms of their convergent and discriminant validity (Rios et al., 2014) or fit to data (Wise & Kingsbury, 2016). Other studies explored the statistical effects of non-effort on ability estimation at the individual- (Rios et al., 2017) or school level (Wise et al., 2020) or in comparing group scores (Rios, 2021) or students’ growth (Wise, 2015).

Other papers were more applied rather than model- or method-focused. For instance, several studies compared the levels of disengagement or rapid guessing across groups, including by country, language, school track, gender, and race (Goldhammer et al., 2016; Kroehne et al., 2020; Lindner et al., 2019; Los et al., 2020; Soland, 2018a). Other studies tested the link between disengagement and performance (e.g., Pools & Monseur, 2021) or other variables of interest, including student characteristics (e.g., demographics, school achievement, and school engagement; Jensen et al., 2018; Soland & Kuhfeld, 2019; Soland et al., 2019a, 2019b) and item properties (e.g., item difficulty, length, position, or content; Kroehne et al., 2020; Michaelides et al., 2020; Setzer et al., 2013). Note that these studies are different from those mentioned above as they do not test the effects of disengagement on ability estimates but rather use the disengagement measure itself and compare it across groups or use it for prediction purposes.

Action sequences

This topic includes studies that used various methods to examine test-takers’ action sequences, namely, the number and order of actions (e.g., clicks) within the assessment platform. For example, Lee and Haberman (2016) used these data to see if test-takers respond to items in the intended order. The studies were published between 2010 (Bennett et al., 2010) and 2021 (e.g., He et al., 2021). They rarely rely on a theory except for a few studies using models of the constructs they target.

The largest subtopic in this group involved score prediction based on process data. Some studies used distance-based methods, namely, they identified how different examinees’ behavior was from the most efficient solution and linked it to performance (Hao et al., 2015; He et al., 2019a, 2019b). Others identified specific actions or series of actions that were associated with success in the item goal-directed behavior (He & von Davier, 2015), number of attempts (Stadler et al., 2019), systematic control of variables (known as the vary-one-thing-at-a-time or VOTAT strategy; Tschirgi, 1980; e.g., Han et al., 2019), etc.

Group comparison and test-takers clustering were also common types of studies. Some studies compared the actions taken by groups based on demographics (Bennett et al., 2010; Liao et al., 2019) or by their performance (Zehner et al., 2020). Others identified groups within the data. For instance, He et al., (2019a, 2019b) detected three groups based on actions in a CPS item and labeled them as low-, medium-, and high-effort, and Xu et al. (2018) found different incorrect strategies when clustering participants’ action sequences.

Finally, there were several methodological studies included in this group. Some studies explored the use of dimension reduction techniques (Tang et al., 2020, 2021) before using the reduced data for other purposes such as prediction. Other studies demonstrated models (e.g., Cui et al., 2020 and the adaptive neuro-fuzzy inference system, Kroehne & Goldhammer, 2018 and finite state machine, or Shu et al., 2017 and Markov-IRT). And others compared the performance of several models (Qiao & Jiao, 2018; Ramalingam & Adams, 2018).

Complex problem-solving

The papers assigned to this topic are mostly about test-takers’ interactions with items designed to measure CPS skills. Like the previous group, many of them used action sequences, but here the focus was on CPS as a construct and not the type of data, distinguishing this topic from the previous one. They were published between 2013 (Goldhammer et al., 2013; Greiff et al., 2013) and 2022 (Ulitzsch et al., 2022). Most of these studies did not rely on an existing psychological or educational theory and were more exploratory, but the dual-processing theory was mentioned more than once (e.g., Scherer et al., 2015). In addition, some papers sometimes mention frameworks of the target construct that were used to design the items or identify key response processes (e.g., the OECD framework for CPS as consisting of knowledge acquisition and application; Eichmann et al., 2019).

Most of the studies in this category used actions taken by examinees and linked them to their performance. Some of the studies focused on relatively simple predictors like the number of actions performed and response time (Scherer et al., 2015; Vörös & Rouet, 2016). Others looked at more complex patterns based on action sequences that had real-world interpretations. For example, Greiff et al. (2015) used the log data in a PISA CPS item to identify the VOTAT strategy. This strategy was then found to be related to individual- and country-level performance on that item. VOTAT was very commonly used in the reviewed studies and consistently found to predict better performance (e.g., Wüstenberg et al., 2014). Other problem-solving behaviors were detected, as well (e.g., preparation vs. execution; Eichmann et al., 2020a, 2020b).

Other studies used process data to detect latent groups or to compare the behaviors of existing groups based on performance or demographics. For instance, based on their action sequences in a CPS item, Teig et al. (2020) distinguished between those who consistently used VOTAT, those starting to use VOTAT, and disengaged test-takers. Eichmann et al. (2020a, 2020b) compared different groups’ performance and strategies taken in a PISA CPS item. They found that boys interacted more with the item and had more exploration behaviors, which explained why boys were more successful than girls on that item.

Digital writing

In this category, the studies involved writing assessments and examinees’ interactions with them. These studies were published between 2008 (Crawford et al., 2008) and 2022 (Bennett et al., 2022; Gong et al., 2022; Talebinamvar & Zarrabi, 2022). Most studies relied on some theoretical model of writing. One example is Hayes’ (2012) work which describes the writing stages as planning, translating (the plan into language), transcribing (into written language), and evaluating (e.g., Guo et al., 2020). Another example is Kellogg’s (1996, 2001) works where writing consists of formulation, execution, and monitoring (e.g., Révész et al., 2019). As many studies use data from a specific assessment, the CBAL, they also often describe the model used to design this assessment (Deane et al., 2015; e.g., Zhang et al., 2016).

Studies about digital writing covered several subtopics, the largest one involved comparing the writing patterns of existing groups. For example, Bennett et al. (2021) found gender differences in the response process (typing speed, productivity, and specific types of editing behaviors) even after controlling for performance. Deane, Roth, et al. (2018) took a different approach and manually classified test-takers’ behaviors into stages (e.g., drafting, copy-editing) and found that they were associated with different typing behaviors. For instance, during drafting, test-takers wrote fluently but paused between sentences (reflecting thinking about the next sentence), but during copy-editing, typing was slow and edits were focused on individual words.

Other prominent subtopics were the association of process features with performance and identifying behavioral patterns based on those features. For instance, Bennett et al. (2020) found that high-performers typed quicker, started more words, and spent more time on the task, but did not differ from low-performers in terms of editing behaviors. In an example of finding behavior patterns, Talebinamvar and Zarrabi (2022) used typing pauses and typing fluency (among other measures) and identified five clusters of test-takers: strategic planners, rapid writers, emerging planners, average writers, and low performers. These clusters were associated with performance – strategic planners had the highest performance, and low performers had the lowest.

Discussion

This review explored common topics within the literature on process data in studies about LSAs, applying a topic modeling approach for the initial theme extraction process. We identified six overarching topics, response time models, response time-general, aberrant test-taking behavior, action sequences, CPS, and digital writing. We also discussed the use of theory in the reviewed topics as well as common subtopics within them.

Several observations can be made based on our findings. First, it seems that some of our topics focused on response time and some focused on other data sources. The ubiquitous use of response time is unsurprising since, unlike typing, for example, total item response time is relevant in any type of item response process, and is relatively easy to record. This is also evident by the dates of the studies, as the topics related to response time include much older studies in comparison with the other topics (e.g., the response time–general group has studies from 1994 while the oldest study in the digital writing group is from 2008). Notably, studies rarely use other types of response time such as the time before the first action, the time between revisions, etc. That information is more commonly used in papers on digital writing and might be useful in analyzing other types of items as well, including relatively simple multiple-choice items. So, we suggest that this information should be collected in any item and made available to researchers.

Second, many of the reviewed studies were methodology-focused, namely, they proposed, evaluated, or improved statistical models for estimating examinees’ abilities. Again, perhaps unsurprisingly given their availability, most of these models augment existing measurement models with item response time data rather than other types of process data. Future studies could explore the use of other timing or action data to enhance ability estimates.

Third, we presented here topics that were common in the literature, but what is less studied could also be of interest to researchers. While we identified specific topics related to writing and CPS, reading and mathematics did not have their specific group, but were rather included in other topics (e.g., Salles et al.,’s, 2020 study on a mathematics assessment was included in the Action sequences topic). This suggests that there are relatively few studies on these constructs or that they do not use a unique terminology distinguishing them from other studies. Assuming this is not an artifact of our search strategy, we can hypothesize a few reasons for this finding.

With respect to reading, it is more difficult to capture the process automatically without eye-tracking devices or other types of specialized hardware compared to typing or reacting within a given time. Clicks, mouse movements, and scrolling are more indirect measures, and while they can still be useful in studies on digital reading, they may be more obvious choices for capturing the writing process or online navigation. Nevertheless, even simple measures such as response time can serve to assess reading fluency as described in theories of reading such as the Information Processing Theory (LaBerge & Samuels, 1974), and should be used more often in reading LSAs.

That there was not a separate topic for mathematics was surprising because some mathematics subjects seem to lend themselves well to studies involving process data (e.g., geometry; von Davier et al., 2023), and can be easily related to elements of relevant theories such as the Mathematical Problem-Solving model (e.g., self-monitoring, specific problem-solving strategies; Schoenfeld, 2013). Other researchers may want to explore this subject further.

Finally, we found very few studies on the use of process data in surveys. Admittedly, these were not at the center of our exploration, so it is possible we did not include relevant studies in the current review. Nevertheless, we wanted to emphasize that process data have the potential to assess (dis)engagement in surveys in a way that conforms with existing motivational theories and that can inform estimates of the targeted constructs (e.g., Soland et al., 2019a, 2019b).

Related to theories, the field of learning analytics is seeing a transition from data-driven to theory-driven research, as theories “contribute to understanding and explaining phenomena or make visible causal relationships between variables, as well as allowing us to make rational inferences about future events” (Khalil et al., 2023). We were interested in the extent to which similar studies on LSAs rely on educational or psychological theories. While many of the reviewed studies were data-driven or methodology-focused, lacking a theoretical basis, several other studies did use relevant theories, either motivational (expectancy-value theory) or cognitive (dual processing theory, demands-capacity model). Some studies that focused on validating a specific assessment also introduced theories on how its target construct should be measured. Given how educational assessments start with defining what one wants to measure, and process data is often used to gather evidence related to this question (Knight & Shum, 2017), it makes sense that the use of theories is relatively common in the field of educational measurement, but there is still room for improvement.

This study as a whole demonstrates the strengths of using topic modeling for literature reviews. Having themes automatically generated from abstracts saved time on multiple readings and analyses of those abstracts. Those interested in an overall impression of the research topics in a certain field, as was our purpose in this review, may find this approach useful especially if the number of studies to analyze is substantial. We also searched for subtopics within the automatically-generated topics to get a more fine-grained understanding of what researchers in the field are interested in, but dividing papers into subtopics was easier when knowing they all share an overarching theme.

However, this approach is not without its limitations. A notable one is that some of the studies were not categorized by the topic model as one would expect. We will discuss an illustrative example that demonstrates the limitations of topic modeling for literature reviews. Our model placed Mayerl’s (2005) study on respondents’ baseline speed into the aberrant test-taking behavior group, and not the response time-general group as one would expect given that no aberrant behaviors were at the center of the study (many studies had the opposite issue, as discussed in  “Aberrant test-taking behaviors” Section). Looking at the probabilities associated with this paper, it indeed looks like this paper had rather similar probabilities of being generated from many groups: it had a probability of 0.26 to be generated from aberrant test-taking behavior, 0.22 to be generated from response time-general, but also a probability of 0.24 and 0.20 to be generated from response time models and CPS, respectively. Evidently, this paper’s categorization was not clear-cut. Examining the paper’s abstract suggests that the topic of aberrant test-taking behavior may have been preferred because this abstract included words like “speed” which is associated with speededness, and “bias” which is a possible result of aberrant behavior during testing.

This demonstrates how the use of specific words could drastically impact where documents are placed, a serious limitation if one wishes to assign papers into groups and study trends within these groups (as was reported for example by Chen et al., 2022). Relatively minor linguistic changes or small differences in probabilities could lead to drastic differences in where a paper is assigned. This may lead to a substantial impact when trying to compare the topics’ sizes or changes in these topics’ popularity in different years or platforms. So, although we provided information on the topics’ sizes for the interested reader, we suggest that topic modeling should be used mostly for identifying themes within a body of literature rather than for comparing how many studies were related to each topic.

Beyond the built-in limitations of topic modeling for literature reviews, our specific application involved a relatively small dataset. While hundreds of documents can be used in a topic model, much larger datasets are usually required for a stable model (Zhao et al., 2015). However, since we used topic modeling as an initial way of extracting repeating topics in the field and were less interested in categorizing our documents, we believe topic modeling is appropriate and useful, as demonstrated by our rather interpretable results.

In addition to the limitation related to topic modeling, this study had other limitations. For example, since we wanted to be as comprehensive as possible, we included any relevant empirical study we could find. This resulted in several issues. First, we did not limit ourselves to peer-reviewed publications nor did we evaluate the quality of the studies. While it is our impression that the studies were of very high quality, we did not formally examine this assertion. Second, we included book chapters and technical reports that sometimes include data and analyses that were also presented in other venues like journal articles. However, since our focus was not on how many studies fell under each topic or subtopic, we feel that this had a minimal effect on our findings.

Conclusion

In this work, we used topic modeling to identify common themes in the literature using process data in LSAs. We identified six groups: response time models, response time-general, aberrant test-taking behavior, action sequences, CPS, and digital writing. Based on these groups, we suggested several important new directions for future research, including research using different types of process data and research in reading and math. We hope that our work will inspire other researchers to pursue these and other directions not suggested here. Doing so can improve our understanding of test-takers’ response processes, thus enhancing the validity of LSAs as well as what is known about how people solve problems more broadly.