Only a few screening tools exist for use with German orthography: The “LRS-Screening” (Endlich et al., 2019) used in the year preceding formal schooling employs 14 subtests to forecast word-spelling deficits by the end of first grade. Notably, it has a sensitivity of 0.74, a specificity of 0.68, and a positive predictive value (PPV) of 0.27. The “LRS-Screening” is administered individually in a traditional paper-and-pencil format and takes approximately 30 min plus additional scoring time. It lacks an engaging cover story, which potentially limits its appeal. Another screening tool commonly used prior to school entry is the “Bielefelder Screening” (Jansen et al., 1999). Although its manual reports promising predictive values, independent studies have failed to replicate these findings. The Phonological Awareness-Reading and Spelling Screening (PB-LRS; 80) is designed for group assessment during the final year of kindergarten. It boasts a sensitivity of 63%, a specificity of 87%, and a PPV of 36%, but it requires approximately 60 min to administer. Another established tool, the “A Tour of Hörhausen” (Martschinke et al., 2001), comprises phonological assessments individually at the beginning and midpoint of first grade. Prognostic validity studies based on a sample of 375 children revealed a specificity exceeding 80%, with sensitivity ranging from 38 to 48%. Assessment takes approximately 40 min. The only app-based screening is a tool called “Förderorientieres Schuleingangsscreening” (BMBWF, 2024), which is supported by the Austrian government and consists of tasks that concern phonological awareness and letter knowledge (Jöbstl et al., 2022). This forms part of a broader battery of seven modules that also address precursors of numeracy development, graphomotor skills and executive functions. This assessment is of acceptable validity and presented in a motivating format, but lacks important precursors for written language, such as vocabulary and grammar.

In summary, existing screening tools for predicting word spelling difficulties in German are time-consuming to administer or exhibit limited predictive efficacy, as detailed in comprehensive reviews (Marx and Lenhard, 2010). Language as a main factor is, in many cases, not one of the components screened.

The objective of the present study was to develop and validate an efficient school-entry screening tool for identifying spelling problems at the end of grade one in community school settings. In order to enable broad implementation, the tool had to be administrable by teachers and feasible in terms of interpretation and integration into everyday school life. Therefore, to facilitate assessment and interpretation of results, and following the general trend in psychological and educational assessment towards increasing integration of digital media in practice (Schaumburg, 2015), we developed an app-based screening approach, which in Austria can also be administered in the final year of kindergarten (Thompson, 2015). Unlike existing screening tools, this study considered a broad set of precursors assessed at school entry phonological information processing (initial phoneme detection, letter knowledge, and phonological working memory) and language (vocabulary and grammar) as predictors of spelling. Focusing on the validity of the selected set of predictors, this study evaluated diagnostic accuracy—broadly defined as the ability of the screening to discriminate between children with and without spelling problems. To investigate the generalizability of the screening, an examination was undertaken of the potential for variation in diagnostic accuracy according to sociodemographic variables. These include the child’s gender and first language, both of which are known to be associated with spelling as well as spelling precursors. The extant research demonstrates that children from low socioeconomic status (SES) backgrounds exhibit lower verbal academic performance than their peers from high-SES families (Sirin, 2005; White, 1982). Similarly, there is substantial evidence indicating that children from low-SES environments demonstrate weaker orthographic skills, particularly in spelling, in comparison to their higher-SES counterparts (Breit et al., 2016; Niemietz et al., 2023 for German-speaking countries). Large-scale international assessments such as PISA (OECD, 2023) indicate that students who have acquired the language of instruction as a second language, frequently due to a migration background, frequently demonstrate suboptimal performance in spelling. Specifically, research findings suggest that both second-language German learners (Lenhart et al., 2019) and students with a migration background (Henschel et al., 2023) exhibit comparatively weaker spelling abilities than their peers.

In summary, the following research questions guided this study:

This prospective study followed children from the beginning to the end of first grade and is part of a longitudinal and ongoing project called SCHNAPP (Detection of risks factors for reading and writing development, see Schöfl et al., 2022). The sample was recruited from various community-based schools in the district of the Upper Austrian capital. The principals of 14 schools were telephoned and then visited in person, and invited to participate in the study project. All of them agreed to participate and informed their teachers of first grade classes (n = 32). Parents of 557children (96.7%) provided permission for their child to participate in the study.

Precursors of spelling—phonological information processing and language—were assessed in the first weeks of first grade, before the formal teaching of reading and spelling had begun. In detail, individualized screening of precursors started in autumn 2022 within the first 2 weeks of the school year. Within 3 weeks, 85% of the sample had been assessed. In the two subsequent weeks, children who had been ill or unable to attend within the first testing period were surveyed. Spelling was assessed by a test administered in a classroom setting at the end of first grade. Twenty children were excluded from the analysis because they were listed as (preschoolers¹) in their first year of learning and did not advance to the first grade at the end of the school year. Eight children were diagnosed during the school year as having special educational needs (SEN). The spelling test was not administered to these children as they had not yet acquired the requisite knowledge of the alphabet. A total of 7 children who were part of the target population for the screening could not be tested (4 children changed school and 3 children were ill, at the first and second test date). Thus, our study used data from n = 522 children (=557 (with parental consent) – 20 (preschoolers) – 8 (SEN students) – 7 (missing tests)).² For these children all variables used in the current study were available. Figure 1 shows the recruitment pathways and timeline.

The distribution of the children in the final study sample was heterogeneous. The proportion of girls in the sample was 46.7%, which equals the proportion of girls among Upper Austrian first graders in 2022 [χ²(1) = 0.00, p > 0.05; Statistik Austria, 2024]. 73.2% of the children assessed spoke German exclusively as their first language, which also corresponds to the population value (i.e., Upper Austrian first graders) of 73.1% [χ²(1) = 0.01, p > 0.05; Statistik Austria, 2024]. The sample had parents from all educational backgrounds. In about four out of 10 families (42.7%) at least one parent held a university degree. In 21.3% of the families, the highest educational degree was a university entrance qualification. About a third (30.1%) reported a vocational education or training, and in 5.9% of the families neither parent had any qualification beyond compulsory schooling. Note that, comparison of the proportion of parents with the highest educational level in our sample with that of the Upper Austrian parent population of 4th graders in the school year 2017/18³ shows that families in which at least one parent has a university degree are overrepresented (42.7% vs. 26.5%) and parents with vocational education or training are underrepresented [30.1% vs. 45.8%; χ²(3) = 77.50, p < 0.001]. Thus, our study sample was representative in terms of gender and first language, but not in terms of parental education, which is probably due to the sample having been recruited from many schools in urban areas, where more parents with higher education live (see also Elliott, 2018, p. 245; Yulianti et al., 2023).

Before summer vacation, teachers were informed by the research team about the procedure and received an information letter for the parents or legal guardians, including consent forms and questions about parents’ educational background, children’s first language, and language use. Teachers entered the IDs on a digital platform for use on a tablet. The investigators, all of whom were student teachers, were enrolled in a student seminar where they learned about the tasks and the testing procedure and for which they received academic credits (towards their study program). Phonological information processing and language skills at the beginning of grade 1 were assessed one on one, with child and instructor seated opposite each other across a table. After a brief welcome, the child was handed a tablet running an app that introduced SCHWUPP the friendly dragon. App navigation was designed such that the child could use it independently. If needed, investigators could help, for instance, to move SCHWUPP across the screen. All instructions essential for the child were recorded as audio files, opened automatically, and repeated if necessary. A yellow background in the app signaled to the test administrator that the child was making test selections independently (e.g., in the phonological tasks). A gray background meant that the test administrator should take the tablet to read and then give instructions. The assessment, including all subtests, took an average of 15.4 min (SD = 4.3) per child. Pauses were possible but rarely needed. The spelling test is also tablet-based (see below). It was administered by the research team and student teachers in a classroom setting.

First, we performed bivariate analyses. We report point-biserial correlations r_pb between the predictors and spelling as well as area under the curve (AUC) values based on receiver operator characteristic (ROC) curve analyses to evaluate the diagnostic accuracy of each predictor. AUCs are global measures of diagnostic accuracy. Values ≥0.9 are regarded as excellent, AUCs ≥0.8 and <0.9 as good, AUCs ≥0.7 and <0.8 as fair, and tests with AUCs <0.7 as poor (Swets, 1988). The pROC package (Robin et al., 2011) in R was used for ROC curve analysis.

Second, we applied multivariate method to simultaneously evaluate the predictive power of spelling precursors. We performed a logistic regression with adaptive variable selection to identify important predictors for inclusion in the screening. Specifically, we used the least absolute shrinkage and selection operator (LASSO; McNeish, 2015; Friedman et al., 2010; for an application of LASSO for the selection of screening variables, see Schöfl et al., 2022; Silverman et al., 2021) implemented in the glmnet R Package (Friedman et al., 2021). LASSO is a statistical learning procedure that serves two purposes: selection of important predictors and attenuation of overfitting. Notably, standard variable selection procedures such as backward and forward selection tend to increase overfitting, which limits the generalizability of the regression results and thus also of a screening algorithm to be developed because sample regression estimates are a blend of signal and noise; sample estimates are therefore greater than population estimates. To counter inflation of the sample estimates and improve generalizability of the results, LASSO applies a penalty term (λ) to the likelihood function. As in backward or forward selection procedures, unimportant predictors are excluded from the regression model. Note that LASSO does not allow any conclusions to be drawn about the statistical significance of a predictor. Although methods for calculating p-values for linear LASSO models have been developed, no such methods are available for logistic models. The predictors selected must be considered important irrespective of whether their regression estimates are statistically significant (McNeish, 2015; Silverman et al., 2021). We z-scored the predictors to make their effects comparable. Thus, the estimates are standardized. As noted, LASSO uses a penalty term (λ) to counter overfitting and zero-out predictors. Following the suggestions in the literature (James et al., 2013), we used 10-fold cross-validation to select the penalty term. Specifically, we chose the λ-value that resulted in the largest AUC (λ_{AUC_max}) to ensure the highest diagnostic accuracy of the prediction model. However, since the maximal AUC may in some cases insufficiently attenuate overfitting, we additionally report the results for a second λ value. We applied the one standard error rule (λ_{AUC_1SE}; selecting the highest value of λ at which the AUC is within one standard error of the greatest AUC; see, e.g., James et al., 2013, p. 214), which selects a model that has a comparable AUC to the model with the highest AUC, but which is more parsimonious (i.e., has a smaller number of predictors). Tenfold cross-validation randomly splits the data into 10 groups (data sets), where one set is used for validation and the remaining nine sets are used for training in the statistical learning procedure. Given this randomness, the results vary with each re-estimation of the model (James et al., 2013). To make our results reproducible and evaluate their robustness, we set five different random seeds and averaged the resulting estimates. For comparison we also report the results of a classical logistic regression using all predictors simultaneously. Finally, we used the estimates of the LASSO models to calculate the probability of spelling deficits at the end of grade 1 (i.e., spelling scores one SD below the mean). These scores, ranging from 0 to 1, were used as screening measures in subsequent analyses.

Third, by applying a bootstrapped test for the comparison of AUC values of paired ROC curves, we compared the diagnostic accuracies of the screening scores λ_{AUC_max} and λ_{AUC_1SE} (Robin et al., 2011).

Fourth, we evaluated whether the ROC curves differed (i) between girls and boys and (ii) between children with and without German as their first language. Differences in ROC curves and AUC values indicate variations in diagnostic accuracy and therefore limit the generalizability of screening (Youngstrom, 2014). We used a bootstrapped test for unpaired ROC curves to compare AUC values between two groups. Further, we applied the Venkatraman (2000) permutation test implemented in the pROC R-package (Robin et al., 2011), which compares actual ROC curves. Note that two ROC curves that do not differ in screening scores have the same sensitivity and specificity across groups, and thus no group-specific cut-off values are required.

Fifth, we used the R-OptimalCutpoints package (López-Ratón et al., 2014) to evaluate several cut-offs that can be used in practice. We report the following diagnostic accuracy statistics: sensitivity (Se; rate of children with spelling deficits identified correctly by the screening process), specificity (Sp; rate of children without spelling deficits identified correctly by the screening), positive predictive value (PPV; rate of screen-positives that show spelling deficits), negative predictive values (NPV; rate of screen-negatives that do not show spelling deficits), and diagnostic likelihood ratios for positive and negative screening results (DLR+ and DLR−, respectively). Unlike predictive values, DLR+ and DLR− are not associated with the prevalence of the problem (here: spelling deficits) under investigation (Pepe, 2004). DLR+ indicates the multiplicative change in the pre-screening odds of scoring one SD below the mean in the spelling test given a positive screening result (i.e., post-screening odds = DLR+ × pre-screening odds). DLR− is the change in the pre-screening odds of scoring one SD below the mean given a negative screening result (post-screening odds = DLR− × pre-screening odds). In the medical literature, DLR+ values ≥10 and DLR− ≤ 0.1 indicate great changes in pre-screening odds, DLR+ ≤ 10 and > 5, and DLR− > 0.1 and ≤ 0.2 indicate moderate changes, DLR+ ≤ 5 and > 2, and DLR− > 0.2 and ≤ 0.5 indicate small changes. DLR+ < 2 and DLR− > 0.5 are rarely important (Jaeschke et al., 1994).

Early letter knowledge is a robust surface variable in spelling (Cordewener et al., 2012), comparable to rapid naming in reading (Lervåg and Hulme, 2009). The strong relationship between letter knowledge and spelling skills might reflect an already existing access to written language itself. Children who know many letters might be able to show—at least implicitly—some elementary understanding of the basic phonological principle that is characteristic of alphabetical writing systems (Treiman, 2013).

Correlations between early letter knowledge and reading acquisition were described about 20 years ago (Foulin, 2005; Leppänen et al., 2008; Muter et al., 2004). Letter knowledge reflects seemingly implicit knowledge about the structure of written language. Simple training that focuses only on individual letters, however, is not effective for reading acquisition (in German: Scheerer-Neumann, 2006). Intervention studies instead support the linking hypothesis that it is the combination of letter knowledge, phonological exercises and subsequently the writing of syllables and words (Moraske et al., 2018). Isolated and exhaustive practice of letter knowledge at the preschool level is not a useful approach. Although the predictive power of phonological awareness is weaker than expected, the importance of PA for school exercises remains significant. In particular, the training of letters in combination with phonological awareness exercises should be emphasized (Moraske et al., 2018). To support children in this skill, Schnitzler (2008) suggested a variety of exercises, for instance, identifying words with the same initial phoneme or changing the initial phonemes in words. According to Mayer (2016), phonological awareness of smaller units, such as phonemes, develops during written language acquisition. In pratice, we suggest working with the initial phoneme at the beginning of written language acquisition rather than with various phoneme positions within the word which are more difficult to detect (Schnitzler, 2008). The combination of phonological awareness (i.e., detecting phonemes) and the basic alphabetic principle (i.e., letter-sound correspondence) is crucial for written language acquisition (Moraske et al., 2018). As the children are gradually introduced to letters and their names and sounds, initial phoneme detection is therefore to be systematically combined with the initial letter. This could be supported visually by the house-and-garage model mentioned above (Röber, 2013).

Regarding written language acquisition in grade 1, we infer that children with deficits in letter knowledge might benefit from integrated letter training that allows systematic insights not only into formal aspects, but also into aspects of the positioning of a letter within the syllable and the written canonical trochaic form of the German writing system. The letter <e> and its pronunciation depending on whether it is positioned in the strong or weak syllable might be a good example (Bredel, 2009; Fuhrhop and Peters, 2013). In practice, this could involve using use the house-and-garage model [originally designed for German trochaic words by Röber (2013)] to visualize for children the position of the latter within the syllable.

For German spelling, oral grammar competence is an important precursor (von Goldammer et al., 2011), which is also confirmed by in the Schnapp spelling test. This might be explained by the strong grammatical (syllabic, morphological and syntactical) principles in the German writing system that act in concert. Theoretical approaches to grammar teaching and recent writing system research in German assume a strong relationship between grammar and literacy skills in general (Bredel, 2015b for an overview of the relationship between grammar and written language). Furthermore, the ability to recognize certain patterns in written language units is important and develops during written language acquisition. Dealing with written language has a number of prerequisite metalinguistic skills. Learning an alphabetical writing system then requires the abilities to identify sounds and to recognize how these sounds are transferred into written language. The ability to recognize certain patterns in written language units is also important (Haueis, 2015). Sufficient knowledge of grammar helps to detect syllable and morphological patterns in German word spelling and later, when dealing with sentences, to detect syntactic patters, such as space between words, German capitalization, and punctuation, when dealing with sentences (Bredel, 2015c). Morphological awareness is often cited as important for German word spelling (Görgen et al., 2021). Morphemes are units that are not only relevant for stem consistency, but also for identifying inflections. The latter is often placed in the weak schwa syllable. Stable access to stem consistency and how inflection is managed might improve spelling skills in general.

Grammar as a key precursor for German spelling adds weight to the argument that grammatical structures are very important in the German writing system. For children who show deficits when screened, knowledge about the canonical trochaic foot and related patterns allows them to gain insights into morphological principles far more easily (Röber, 2013; Bredel et al., 2017). Furthermore, working with syntactical structures, such as German V2-sentences, helps children to understand how sentences are constructed. To support language development in a practical way, children can be introduced to the topological field model (Granzow-Emden, 2015). For first graders, pictures or word cards could be used to form a spoken sentence instead of writing down sentences in the model and reading them aloud. Visualizing sentence structure promotes metalinguistic skills in children (Topalovic and Michalak, 2015).

The screening showed acceptable diagnostic accuracy [AUC = 0.757, DeLong 95% CI (0.698; 0.816)]. However, to achieve the desired sensitivity (>0.80), we must accept a relatively high rate of false positives (specificity = 0.57), which results in about one out of two pupils being screened positive. With a focus on intervention, this large group of screen-positives might thus challenge or even overtax teachers and school resources in providing support for these children. The problem of a high false-positive rate is not unique to our screening test (Cogo-Moreira et al., 2023; Duff et al., 2015; Johnson et al., 2009). Future research should therefore aim to improve specificity, for instance, by adding further subtests to the screening tool (Johnson et al., 2009). For example, reading socialization via parents or friends, motivational aspects, and further predictive child-related variables could be considered (only language-related variables were included in our approach). Specific attention processes or visual approaches were not considered, as the requisite data were not collected in this study. Regarding phonological awareness, only one subtest was included in the Schnapp screening battery. A wider approach to phonological awareness including both easier subtests, such as rhyming and syllable counting, and more difficult ones (e.g., manipulating phonemes) following the hierarchical structure of phonological tasks by Schnitzler (2008) should be analyzed.

Some researchers (Kim et al., 2013; Stainthorp et al., 2013) found orthographic awareness to be a significant precursor of early spelling skills. Actually, we do not include any task to tap orthographic awareness. Subsequent studies should add this variable for screening purposes.

Finally, continuous surveillance of spelling development is required because there may be various pathways to spelling difficulties. Many children with early difficulties do not go on to develop spelling problems, in the same way that many children who initially screen negative may later demonstrate spelling difficulties later, analogously to early screening for reading problems (Duff et al., 2015).

In summary our didactical implication to support written language acquisition for children in primary school is related to didactical implementations (Röber, 2013; Röber et al., 2019; Bredel, 2016; Bredel et al., 2017 based on the graphematical theory Primus, 2010; Evertz, 2016; Maas, 2022; Eisenberg, 2020; Fuhrhop and Peters, 2023)—both mentioned above. Children gain a comprehensive insight into the patterns of written language such as different sounds depending on letter positioning within the graphematic foot, how strong and weak syllables are presented in written language and how morphological stems are derivated from the graphematic foot, which is the initial point for constructing sentences.