1.1 Traditional self-reported measures of reading time

Traditionally, reading time has been assessed using retrospective self-report questionnaires [20]. These measures show small to moderate associations with reading outcomes [12] and capture both quantitative aspects (e.g., time and frequency) and qualitative ones (e.g., text type, parent-child interaction), which differentially impact reading motivation and skill development [22–25]. However, retrospective questionnaires require complex cognitive judgments and are therefore vulnerable to multiple source of bias, including subjective interpretations of frequency and duration, memory and reference biases, and social desirability effects, often leading to overestimation of actual reading behavior [26–33].

As retrospective self-reported measures appear too coarse-grained, some researchers have turned to more fine-grained approaches, such as reading diaries. These involve daily reporting of reading activities, capturing not only the time spent reading but also qualitative aspects, such as the type of material read [34–36]. Such diaries have been found to be reliable predictors of reading outcomes and motivation [35,37–39]. Although they reduce some limitations of retrospective questionnaires, their data quality depends on compliance and timing of completion, as delayed entries may reintroduce retrospective bias and increase missing data, particularly over longer collection periods [40–42]. In addition, traditional paper-and-pencil formats impose a substantial burden on participants, which may negatively affect adherence and data quality [43].

1.2 Ambulatory assessment of reading time

To address these limitations, some researchers have proposed using ambulatory assessment [44,45], defined as the use of digital tools to track behaviors in real time and in natural settings. Thanks to technological advances, ambulatory assessments can now be easily carried out using mobile apps on participants’ phones, as already demonstrated in the field of health [46,47]. More specifically, these methods have recently been successfully applied to assess screen time, demonstrating stronger psychometric properties than self-reported measures (for a review, see [48]).

However, to date, only one study has used ambulatory assessment to measure reading time, despite several advantages highlighted by Locher and collaborators [49]. First, smartphone ownership is common, even in disadvantaged families, enabling continuous data collection through a dedicated mobile app. This provides a convenient way to complete a reading diary at any time. Moreover, mobile apps offer high flexibility by allowing questions to be tailored, either by removing unnecessary items or by adding new ones based on participants’ responses or profiles. Another advantage is the ability to detect when entries are completed retrospectively, allowing for a more precise evaluation of data quality. Finally, mobile apps may be easier to use than paper-based methods, promoting broader participation and increasing the generalizability of results. Locher and collaborators [49] demonstrated that using a mobile app to track the reading behavior of university students is a reliable approach that reduces the burden of diary completion. Importantly, they found that app-based data and retrospective questionnaires were closely related, although reading time was overestimated in the questionnaire. This suggests that mobile apps may help reduce social desirability bias, which contributes to such overestimations.

1.3 The present study

The present study aimed to extend this approach in several ways. First, we developed a mobile application that allowed participants to time the duration of reading activities using a stopwatch function, rather than reporting it retrospectively. We believed that this method would encourage greater attention to accurately reporting reading activity, while reducing retrospective bias and providing a more objective measure of reading time. To distinguish this approach from traditional self-reports, and to draw a parallel with studies using logged screen time in digital media research, we refer to this method as a logged measure of reading time. Second, the study focused on primary school children. Since children do not necessarily own smartphones, we targeted reading time at home, where parents could use the mobile app to record their child’s reading activities. Finally, we assessed children’s reading fluency to examine its relationship with recorded reading times. A total of 109 children from Grade 1 to Grade 5 and their parents first completed a paper-and-pencil questionnaire about reading behavior, then used a mobile app for 14 days to log the duration of reading activities in real time. Children’s reading fluency was assessed at school. The study aimed, first, to evaluate the degree of convergence between self-reported and app-recorded reading time, and second, to assess the extent to which each measure contributes to the prediction of reading fluency.

2.1. Participants

A total of 109 children (49 female, 60 male) from a public elementary school (Grade 1: n = 20; Grade 2: n = 29; Grade 3: n = 29; Grade 4: n = 13; Grade 5: n = 18) were tested in the middle of their school year (from 01/01/2024 to 30/04/2024). All participants were native French speakers with normal or corrected-to-normal vision. Caregivers provided written informed consent prior to the experiment. In addition, children provided verbal assent before participation. The experimenter explained the study in age-appropriate language, emphasized that participation was voluntary, and informed children that they could withdraw at any time without consequences. This procedure, including the use of verbal child assent, was reviewed and approved by the French Ethics Committee Review Board (2023-01-05-04). The experiment was conducted in accordance with relevant guidelines and regulations, as well as the Declaration of Helsinki.

2.2. Measures of reading time

2.3. Measure of reading fluency

To assess reading fluency, we used the Alouette Test [54], a standardized reading assessment widely used in France for children aged 5–14. In this task, participants are asked to read aloud a 265-word text within a time limit of three minutes (180 seconds), aiming for both speed and accuracy. The text is designed to minimize reliance on semantic cues by presenting syntactically correct but nonsensical sentences (e.g., “Le printemps a mis ses nids,” [Spring has put on its nests]), thereby targeting decoding skills. It contains rare vocabulary, irregular spellings, and orthographic challenges such as silent letters and phonologically confusable items. Three main variables were recorded: the number of correctly read words (M), total reading time in seconds (TL), and the number of reading errors (E). These variables were used to compute the reading fluency index (CTL = [(M-E))/TL * 180]).

3.1. Preprocessing of reading time data

3.2. Comparisons between self-reported and logged measures of reading time

A comparison of reading time estimates revealed substantially higher weekly durations for the self-reported measure (M = 6.26 hours, SD = 2.78) than for the logged measure (M = 2.11 hours, SD = 1.80). This difference was highly significant (t = 12.97, p < .001, d = 1.64). Reading time did not significantly vary across grade levels (see Table 1; all pairwise comparisons with p > .05).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Mean weekly reading time (in hours) by grade level, measured using a mobile application (logged measure) and a paper-pencil questionnaire (self-reported measure). For each grade, the table presents the mean and standard deviation (in parentheses), as well as the observed range [minimum; maximum].

https://doi.org/10.1371/journal.pone.0344853.t001

Despite large differences in absolute levels, the two measures were moderately correlated (r = .45, p < .001), suggesting that the two measures capture partly overlapping constructs. To further evaluate the consistency of individual rankings across the two measurement methods, we computed an intraclass correlation coefficient (ICC) using a mixed-effects model implemented via the R package psych [56]. The ICC indicated a significant but relatively low level of agreement between the two measures (F (106, 106) = 2.42, p < .01).

To further assess the agreement between the two measurement methods, we conducted a Bland–Altman analysis [57]. The 95% limits of agreement (grey dotted lines) ranged from approximately 0 to −8 hours, indicating wide individual differences and poor agreement, consistent with the ICC results. Fig 1 shows that self-reported reading time was overestimated by approximately 4 hours per week (grey dashed line). A systematic bias was observed for most participants (grey points), with the self-reported measure consistently reporting higher values. This bias appeared proportional, as the variability in differences increased with higher mean reading durations. A linear regression confirmed this trend, showing a significant proportional bias (grey solid line).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Bland-Altman plot illustrating the agreement between logged and self-reported reading time (in hours).

The x-axis represents the mean of the two measures for each participant, and the y-axis shows the difference between logged and self-reported reading times. The solid black line represents the linear regression of the differences on the means, indicating a systematic bias. The dashed line indicates the average difference (bias), and the dotted lines indicate the limits of agreement.

https://doi.org/10.1371/journal.pone.0344853.g001

3.3. Relation between reading time and reading fluency

To examine the relationship between reading time and reading fluency, we first conducted Pearson correlation analyses. Weekly logged reading time recorded via the mobile application showed a moderate and significant correlation with fluency scores (r = .39, p < .001). Self-reported reading time from the paper-based questionnaire was also positively correlated with fluency, although to a lesser extent (r = .25, p < .05).

We then ran a series of linear regression models to compare the predictive power of the two measurement methods, controlling for age, gender, and mother’s professional occupation. In the self-reported model, weekly reading time significantly predicted fluency performance (b = 0.09, SE = 0.04, β = 0.23, partial R² = .06, p < .05). Mother’s occupation also emerged as a significant predictor (b = −0.21, SE = 0.08, β = −0.23, partial R² = .06, p < .05). This model accounted for 9.2% of the variance in fluency scores (adjusted R² = .092).

In contrast, the logged model showed that objectively measured weekly reading time was a stronger and more reliable predictor of fluency (b = 0.23, SE = 0.06, β = 0.37, partial R² = .13, p < .001), while mother’s occupation was not significant (b = –0.15, SE = 0.08, β = −0.17, partial R² = .03, p = .070). The model explained 16.3% of the variance in reading fluency (adjusted R² = .163).

Finally, in a combined model including both reading time measures, only the logged reading time remained a significant predictor of fluency (b = 0.21, SE = 0.07, β = 0.33, partial R² = .08, p < .01), whereas the self-reported estimate no longer reached significance (b = 0.03, SE = 0.04, β = 0.08, partial R² = .01, p = .46). Mother’s occupation again failed to reach significance (b = –0.16, SE = 0.08, β = −0.17, partial R² = .03, p = .066). The model explained 15.9% of the variance in reading fluency (adjusted R² = .159).

4.1. Data quality of the log-based measure

The log-based measure showed satisfactory data quality, with acceptable internal consistency and reporting stability over the two-week period. However, data quality appeared lower than that reported by Locher and collaborators [49] with university students. Several explanations can be considered.

First, logging relies heavily on parental compliance. Parents must consistently initiate and terminate the stopwatch for each reading episode, which may result in omissions or inaccuracies, particularly during busy routines or over extended monitoring periods. Second, independent reading episodes occurring outside parental supervision may go unrecorded. This limitation is likely to increase with children’s age and autonomy, potentially leading to a systematic underestimation of reading time in older children. Third, the use of a stopwatch, as opposed to manual time entry, may introduce additional noise—for example when the timer is started unintentionally or not stopped immediately after a reading episode. Despite these limitations, the logged reading times remained very similar to those reported by Locher and collaborators [49] with university students. Importantly, these limitations indicate that app-based logged measure should not be interpreted as exhaustive records of all reading activity. Rather, they should be considered conservative, behaviorally anchored estimates of supervised reading time.

4.2. Data quality of self-reported measure

Contrary to the log-based measure, self-reported measure showed clear signs of overestimation. This overestimation may be due to social desirability bias, which could have been stronger in this study because total reading time included self-reports for three types of reading activities (shared reading, independent reading, and reading for homework), potentially amplifying the effect. In this regard, research on reading time rarely includes controls for social desirability bias [58], even though several methods have been developed to address it [59]. In contrast, the logged measure seems to be less sensitive to social desirability bias, as it is more difficult or impractical to inflate reading time when using a stopwatch.

Another, non-exclusive explanation for the overestimation in self-reports is that parents may have difficulty clearly distinguishing between different types of reading activities. As a result, a single reading event may be reported under multiple categories (e.g., both shared reading and reading for homework), leading to double counting and an inflated total reading time. The rationale for separating reading activities was to capture the diversity of reading practices across the primary school years. For example, Grade 1 children are more likely to engage in shared reading and less in independent reading, while the opposite tends to be true in Grade 5 [51,52]. Using a single global question that explicitly includes all types of reading activities might have reduced overestimation. In the same vein, it is possible that parents overestimate reading time because children appear to be reading (e.g., holding a book) without sustained engagement in decoding or comprehension. This limitation should be more pronounced for self-reports than for log-based measures. Stopwatch-based logging requires an explicit initiation of reading episodes, thereby reducing reliance on inference or routine-based assumptions. However, the present design did not allow for direct verification of children’s cognitive engagement during reading episodes. Neither the questionnaire nor the app-based logging can distinguish between active reading and more superficial engagement.

4.3. What do these two measures really capture?

At first glance, the moderate correlation between the two measures suggests that logged and self-reported data capture only partially overlapping constructs, in line with findings from digital media use research, where similar discrepancies between logged and self-reported measures have been observed [48]. This is further supported by the low level of agreement between them, indicating that the measures are not interchangeable for consistently ranking participants. The Bland–Altman analysis revealed a systematic bias, with the self-reported measure overestimating reading time for nearly all participants by an average of four hours. This finding supports the idea that self-reports are sensitive to social desirability bias, as previously discussed. However, this explanation is not entirely sufficient, as the bias was also proportional to the total reading time reported by participants. One possible explanation is that frequent readers tend to overreport their reading time, as it may be more difficult to estimate the duration of habitual activities embedded in daily routines [26]. It is also possible that less skilled readers, due to the cognitive effort required and reduced enjoyment, perceive time as passing more slowly, which may lead to overestimation in self-reported reading time [60].

The proportional nature of this systematic bias remains an open question, as other research has identified different patterns. In the field of digital media use, a regression-to-the-mean bias—where heavy users tend to underreport their activity, while light users tend to overreport it—has been described in several studies [61–63]. As acknowledged by Scharkow [63], several factors may influence the direction and nature of systematic bias, including gender, cognitive abilities, social desirability, behavior frequency, the specificity of the construct measured, and methodological limitations of the instruments. Systematic bias can distort not only correlations with other variables but also estimates of means, variances, and proportions. Future studies using logged data to measure reading time should aim to disentangle the contribution of these factors to better understand the biases affecting self-reported data, which appear less suitable for accurately capturing actual reading time.

4.4. How are these two measures associated with reading fluency?

Both measures were significantly correlated with reading fluency, with the logged measure showing the strongest association. Regarding this difference, it is possible that mean imputation may have slightly attenuated correlations involving the self-reported measure. However, given the very small proportion of missing data and the use of grade-level imputation, this effect is likely negligible and unlikely to explain the observed superiority of the logged measure. Importantly, available evidence indicates that the present sample size was adequate to investigate associations between different measures of reading time and reading fluency. Based on prior literature, Torppa et al. (2020) reported correlations between leisure reading and reading fluency ranging from r = .28 to  .41 (M ≈ .35) in children from Grade 1 to Grade 4 [16]. These findings converge with the meta-analysis by Mol and Bus (2011) [12], which showed that associations with reading fluency are generally stronger than those observed for more constrained basic decoding skills. Adopting a conservative expected correlation of r = .28 and a conventional target power of 80% [64], an a priori power analysis using the pwr.r.test function from the R package pwr indicated that approximately 97 participants would be required (two-sided α = .05). With the present sample size (N = 109), the corresponding post hoc power based on the observed correlation (r = .25) is approximately 75%, and the sample size required to detect this effect with 80% power would be approximately 123 participants.

This predictive advantage was further confirmed in the regression analyses: when both measures were included in the model, only the logged measure remained a significant predictor. This suggests that the self-reported measure no longer provided meaningful information, likely due to shared variance with the logged measure. This finding is consistent with the literature, which shows that daily measures of reading time are often considered the gold standard for assessing print exposure. They limit retrospective bias and offer superior content, convergent, and criterion validity [35,39]. Their predictive power is comparable to that of print exposure checklists, which have been shown to outperform self-reported reading time [12].

Moreover, the fact that mother’s occupation was a significant predictor only in the self-reported model, but not in the logged model, supports the idea that more precise behavioral tracking can reduce reliance on proxy variables. Although this issue has not, to our knowledge, been directly examined in the field of literacy, it has been addressed in other domains. For example, in the insurance industry, behavioral tracking—such as installing devices in cars to record real-time driving behaviors (e.g., duration, distance, speed)—allows for a more accurate assessment of risk than relying on proxies like income or occupation [65]. It is well established that proxy variables can introduce measurement bias and may fail to capture the underlying construct. Taken together, these findings help explain the added value of the logged measure over the traditional self-reported measure.

4.5 Implications and challenges for future research

Despite these promising results, several avenues for improving the reliability of app-based logged measures can be identified. First, data quality could be improved. Allowing retrospective entries may reduce missing data. However, it may also encourage participants to abandon real-time logging. This would reintroduce recall bias and create heterogeneous data types [58,59]. A better solution would be to include a post-study compliance questionnaire. This would allow researchers to assess how consistently families used the stopwatch function. An annotation feature could also be added. Users could then report recording errors, such as forgetting to stop the timer.

Second, as discussed earlier, parental involvement may have influenced data completeness. In the present study, parents were asked to record their child’s reading activities. It is possible that the degree of parental supervision varied depending on the child’s age and level of autonomy. It may also have differed according to socioeconomic background. Such variability could have affected the proportion of missing data and the overall consistency of logging. Future studies should therefore include specific measures assessing children’s autonomy in both reading practices and recording behaviors. An open question remains whether child self-recording would provide more complete or more reliable data. The answer likely depends on developmental factors and the child’s level of autonomy.

Third, engagement and scalability must be considered. In the present study, research assistants supported families. This support improved compliance but limits scalability. Families more familiar with digital tools may also have been more likely to participate. This raises potential selection bias. These issues could be mitigated by improving the interface and simplifying the user experience. Gamification features may also enhance engagement. However, researchers should monitor possible reactivity effects induced by such features.

Improved engagement could make longer data collection periods more feasible. In the present study, we selected a 14-day period based on prior diary research [40,49]. Because app-based logging reduces participant burden, extended recording periods may be achievable without compromising data quality. This would make the method suitable for longitudinal designs and allow researchers to examine learning dynamics in greater detail.

App-based logging also opens broader perspectives for cross-cultural research. Applications can be developed in multiple languages. This may facilitate participation from families who do not speak the language of schooling [19]. Such an approach is particularly relevant because literacy practices in the home language can transfer to school-based literacy skills [66,67].

Finally, this approach is not limited to home literacy research. App-based logging could be implemented in school contexts. For example, it could complement existing measures of Academic Learning Time [68]. Current methods rely on teacher self-reports or direct classroom observations. The former are efficient but prone to bias. The latter are resource intensive. Real-time logging by teachers may offer a useful compromise.

4.6. Conclusion

While our findings may appear to argue for abandoning questionnaires, we believe they instead open new avenues for improving these measures. Rather than opposing the two approaches, their combined use in future studies could help identify the various factors that contribute to systematic biases, as widely documented in the literature. This, in turn, would support the development of statistical techniques to adjust for these influences [48,69]. Strengthening the predictive validity of questionnaires through such corrections would preserve their value as practical and scalable tools. Ultimately, integrating real-time and self-reported measures holds strong potential for advancing research on reading practices—both in ecological and methodological terms.