The Effect of Mobile Devices on the Development and Health of Young Children

Quality assurance Researchers undertaking the measurements are highly experienced in the procedures and analysis. The investigators are utilising only validated tools that are well-established in the literature. The exposure and use of interactive electronic devices (e.g., tablets and smartphones) will be measured by an app that will access information directly from the participant's devices (Android or iOS) using mobile sensing software to capture screen time and app usage (Lind et al. 2023). The app has been successfully used to measure smartphone use in children (Bagot et al., 2022; Wade et al., 2021).

The measurement of emerging abilities, our primary outcome measure, has been validated and widely used in the literature (Howard et al. 2017). The investigators will use an administration fidelity checklist to assess practice across researchers before the start of data collection. A range of secondary outcomes will be recorded (i.e. BMI (calibrated scales and repeated measures), movement behaviour (accelerometry Cliff et al. 2024), motor development (NIH toolbox, Reuben et al., 2013), parent and child interaction (StimQ2, Cates et al. 2023). These measurements have been used in the SUNRISE study, which is collecting data on children at this age in 63 countries around the world (https://sunrise-study.com/) and published in a protocol (Okely et al., 2021).

Training videos and support have been provided by the SUNRISE team and have been used to inform our study protocol. Since child development is influenced by a number of covariates, the investigators have selected the most commonly reported in the literature and suggested by our PPI group (i.e. sex, age, ethnicity, maternal education, parenting style (PSDQ- SV. Robinson et al. 2001), attendance to childcare, parental addiction to smartphones (SAS-SV, Kwon et al., 2013) and screen viewing policy.

Data processing Data capture forms have been created for applicable measures to standardise data input and aid with data entry. In terms of data processing and analysis. The investigators will conduct preliminary data cleaning, exploring whether values of continuous variables are within range, plausibility of means and standard deviations, and validity of coded categories. The investigators will assess data distributions and identify any univariate outliers from graphical methods and from cases with very large, standardised scores disconnected from other scores and multivariate outliers by graphical methods and inspection of leverage/Mahalanobis distances, discrepancy and influence statistics. Any possible errors will be investigated on an individual basis. The investigators will also investigate the extent, pattern and nature of missing data. If the proportion of missing data is small (below 5%), they will consider complete case analyses. If the amount or pattern of missing data precludes complete case analysis, the investigators will consider data imputation. Multiple imputations will be used due to their robustness to the type of data missingness. If imputation is conducted, sensitivity studies will be conducted by comparing results derived from data with and without imputation.

Data analysis Descriptive analysis The sample will be summarised descriptively. The investigators will report the number of children in each educational unit and primary and secondary outcomes by time point (baseline and follow-up). For continuous outcomes, summary information will be presented as means (standard deviations (SD); ranges). For categorical outcomes, summary information will be presented as frequencies (percentages). The need for variable transformations to stabilise variance or achieve Normality will be assessed.

Inferential analysis The following test variables will be considered for the analysis: 1) IED duration; 2a) IED mode educational; 2b) IED mode age-appropriate.

1. Multilevel regression modelling of primary outcome The investigators conceptualised a 2-level random intercepts multilevel model, with children clustered within educational units. This model will consider EA at the follow-up to be the outcome measure designed to answer the primary research questions: How is IED duration (hours per day) and mode (educational vs. non-educational; age-appropriate vs. non-age-appropriate) at baseline, controlling for child-level covariates (including EA baseline scores) and educational unit-level covariates, associated with EA at follow-up in a multilevel model in which children are clustered within educational units? The variance partition coefficient (VPC; the proportion of residual variance associated with each level of the model) will be assessed via a null model before proceeding to a covariate model. The investigators will consider merging the levels in the model if VPC statistics reveal negligible clustering effects (negligible residual variance at the level of the educational unit). A non-linear multiple regression modelling will be conducted if there is evidence for a non-linear relationship between the level of IED usage at baseline and EA at follow-up, allowing for a single maximum (corresponding to optimum levels of baseline IED usage) or plateau; otherwise, linear regression models will be conducted. If data indicates an optimum IED level of usage associated with a specific maximum value of EA score at follow-up, alternative non-linear functional forms will be considered with maxima or plateauing features, including polynomial (e.g. quadratic) and logarithmic forms, and piecewise functions. The investigators will compare the fit of multiple distributions in the vicinity of any turning point and select the best-fitting model in this region to maximise the accuracy with which the maximum value of EA may be obtained. Confidence intervals will be fitted around the function to derive a range of values for the maximum value.

   The investigators will conduct both non-fully adjusted and adjusted models, with adjusted models adjusted for all covariates at each level of the model. Non-adjusted models will include (i) the single determinant of the level of IED usage (duration) at baseline, (ii) the determinants of IED predominant mode at baseline (as defined above), and (iii) determinants of IED duration and mode. These determinants will all be added in at the child level of the model. First-order interactions will be included within unadjusted models to capture any differential effects in assessing levels of IED usage with differing predominant modes of use. Any interaction revealed to be of substantive importance will be retained in a re-cast model alongside all main effects. Adjusted models will be based on the included variables of both IED duration and IED predominant mode, any interactions of substantive importance and all controlling covariates at the appropriate level of the model. The investigators will not use automated modelling strategies for variable selection and will retain all main effects in the adjusted model. However, collinearity will be assessed in adjusted models and consider deletion of controlling covariates if excessive collinearity is apparent (variance inflation factor >=5 for any variable).

2. Multilevel regression modelling of secondary outcomes The investigators will conduct multiple linear regression modelling on all numerical secondary outcomes and multiple logistic regression modelling for the binary secondary outcome measure, using the same model structure as for the main analysis of the primary outcome. If any evidence is revealed for non-linearity between the level of IED duration and the secondary outcomes, The investigators will consider non-linear modelling for the analysis of the primary outcome; else, they will consider linear modelling. In both cases, the same set of covariates and interactions defined for the primary outcome will be adjusted. No adjustment for multiple analyses will be made; hover, all analyses will be planned a priori and reported in full.

3. Multilevel regression modelling: subsidiary analysis To explore the socio-ecological correlates of IED duration at the individual (gender, age and ethnicity), interpersonal (maternal education, parenting style and smartphone addiction), and organisational (childcare policy and attendance) levels, will be conducted as a subsidiary analysis, considering the above set of variables as determinants of IED using the variables that are reported as controlling variables.

   Random intercepts and multiple linear regression modelling on the outcome of IED duration will be conducted at baseline. A 2-level hierarchical structure will be used, with variables designated as Individual or Interpersonal attached at the lower (child) level and variables designated as Organisational attached at the upper (educational unit) level.

   Any relationships revealed during this process will be used in future modelling to generate hypotheses within a wider structural equation modelling framework.

4. Sensitivity analysis and data reporting For the primary outcome, the plan is to conduct sensitivity analyses to assess the sensitivity of the model to certain assumptions, as mentioned above. The investigators will compare parameter estimates of tested variables in unadjusted and adjusted models. For the multilevel modelling of the relationship between IED duration and EA, both random slopes and random intercepts models will be used to assess variation in slopes between higher-level units. If data imputation is viable, models will be run with and without imputed data (see 'Data cleaning and assessment of missing data').

For linear and non-linear regression models of continuous numerical outcomes, in the main and subsidiary analyses, the investigators will report all unstandardised parameter estimates with associated 95% confidence intervals and p-values. For logistic regression models, all odds ratios with associated 95% confidence intervals and p-values will be reported. If evidence is revealed for a non-linear trend, the investigators will report the functional form of the best-fitting curve and identify the location of any maximum or commencement of plateauing effects. All regression modelling assumptions will be checked, including homogeneity of variance and normality of outcome variables for each value of an independent variable, using residual analysis. Statistical analyses will be conducted using MLwiN version 3.06 (Charlton et al., 2022) and Stata version 17 (Stata Corp 2021).