Deviation, Epistatic

For each example, statistical models describe the pattern of obtaining new or unique items with incremental increases in sample size. Individual lists were first analyzed with Flame [31 ,32 ] to provide the list of unique items for each example and the Smith [14 ] and Sutrop [15 ] item salience scores. Duplicate items due to spelling, case errors, spacing, or variations were combined.
To help develop an interviewing stopping rule, a simple model was used to predict the unique number of items contributed by each additional respondent. Generalized linear models (GLM, log-linear models for count data) were used to predict the unique number of items added by each respondent (incrementing sample size), because number of unique items added by each respondent (count data) is approximately Poisson distributed. For each example, models were fit with ordinary least squares linear regression, Poisson, and negative binomial probability distributions. Respondents were assumed to be in random order, in the order in which they occurred in each dataset, although in some cases they were in the order they were interviewed. Goodness-of-fit was compared across the three models with minimized deviants (the Akaike Information Criterion, AIC) to find the best-fitting model [33 ]. Using the best-fitting model for each example, the point of saturation was estimated as the point where the expected number of new items was one or less. Sample size and domain size were estimated at the point of saturation, and total domain size was estimated for an infinite sample size from the model for each example as the limit of a geometric series (assuming a negative slope).
Because the GLM models above used only incremental sample size to predict the total number of unique items (domain size) and ignored variation in the number of items provided by each person and variation in item salience, an additional analysis was used to estimate domain size while accounting for subject and item heterogeneity. For that analysis, domain size was estimated with a capture-recapture estimation technique used for estimating the size of hidden populations. Domain size was estimated from the total number of items on individual lists and the number of matching items between pairs of lists with a log-linear analysis. For example, population size can be estimated from the responses of two people as the product of their number of responses divided by the number of matching items (assumed to be due to chance). If Person#1 named 15 illness terms and Person#2 named 31 terms and they matched on five illnesses, there would be 41 unique illness terms and the estimated total number of illness terms based on these two people would be (15 x 31) /5 = 93.
A log-linear solution generalizes this logic from a 2 x 2 table to a 2^K table [34 ]. the capture–recapture solution estimates total population size for hidden populations using the pattern of recapture (matching) between pairs of samples (respondents) to estimate the population size. An implementation in R with GLM uses a log-linear form to estimate population size based on recapture rates (Rcapture [35 ,36 ]). In this application, it is assumed that the population does not change between interviews (closed population) and models are fit with: (1) no variation across people or items (M₀); (2) variation only across respondents (M_t); (3) variation only across items (M_h); and (4) variation due to an interaction between people and items (M_ht). For each model, estimates were fit with binomial, Chao’s lower bound estimate, Poisson, Darroch log normal, and gamma distributions [35 ]. Variation among items (heterogeneity) is a test for a difference in the probabilities of item occurrence and, in this case, is equivalent to a test for a difference in item salience among the items. Due to the large number of combinations needed to estimate these models, Rcapture software estimates are provided for all four models only up to a sample of size 10. For larger sample sizes (all examples in this study had sample sizes of 20 or larger), only model 1 with no effects for people or items (the binomial model) and model 3 with item effects (item salience differences) were tested. Therefore, models were fit at size 10, to test all four models and then at the total available sample size.

We performed two experiments. Experiment 1 was designed to investigate whether there is any TB difference in human-human interaction and human-computer interaction under the simultaneous variation of both a cover story (i.e., belief in human-human interaction) and a stimulus. Experiment 2 was set up to disentangle the differential contributions of a cover story and stimulus. By including a confederate, participants were led to believe they were interacting with another human being. Involving a confederate has been shown to convince participants that they are really interacting with another person and thereby simulate a realistic and ecologically valid interactive social situation (Pfeiffer et al., 2014 (link); Schilbach et al., 2010 ). To this end, participants were introduced to another person of the same gender and similar age as their partner for the study prior to participating in the experiment. In fact, the partner was a confederate of the experimenter and not active during the experiment. Instead, the entire experimental procedure was computer controlled.
After arriving at the test site, participants spent several minutes with their confederates for general information and informed consent, prior to being separated by a mock coin toss made out between participant and confederate. For the toss, confederates were instructed to always let the participants choose. The coin toss was rigged in favor of the participant who always won. Subsequent instructions heavily emphasized the interactive nature of the experiment by employing repeated mentions of the interaction partner and the repeated use of the words “interactive,” “together,” “cooperation.” Participants were instructed that they would act as the active part in an interactive experiment and that they would give orders to their partner via their computer by pressing either the left or the right arrow key. Thereby, the confederate would always act as the reactive partner. The confederate allegedly would be seated in front of an eye-tracker measuring their eye movements and depicting them in real time on the participants’ screen. Participants were told that the partners would be instructed to react to their orders by responding as quickly as possible by looking either to the right or the left, corresponding to the pressed arrow key, and that it was the participants’ task to estimate their partner’s reaction time.

The “MICE” package in R 4.1.3 was used to perform multivariate imputation of missing values [24 (link)]. Participants in both the case and control groups were statistically analysed. For each group, factors that may impact the risk of CV events were discussed independently. Continuous variables of normal distribution were described by the mean ± standard deviation ( $\overline{X}$  ± S), and continuous variables that did not meet normal distribution were described by the median and interquartile range (M (IQR)). The statistical description of categorical variables was performed using percentages. Baseline characteristics between the case and control groups were compared by t-test for continuous variables of normal distribution, rank-sum test for continuous variables of non-normal distribution, and chi-square test for categorical variables. The Cox proportional hazard model was used for multivariate analysis, and collinearity diagnosis was performed to screen the independent risk factors affecting the outcome events. Set the terms for the interaction between the variables to be tested and time (time-dependent covariates). The violation was determined by the statistical significance of the Proportional Hazards Assumption (PH Assumption). The Akaike information criterion (AIC criteria) was used to calculate Cox regression for the highest amount of variation with the fewest possible independent variables. The “stepAIC” function of package “MASS” in R was used to screen variables, the direction of stepwise regression was “both”. The HR values and their 95% confidence intervals (95% CIs) were used to assess whether the indicators were associated with the outcome events as well as the strength and direction of the association.
Cox proportional hazard model was used to fit the survival data, and a prediction model including environmental and genetic data was constructed. The duration of survival was measured in months. The prediction model was evaluated from two aspects of discrimination and calibration. The Area Under Curve (AUC) was used to measure the discrimination of the prediction model. Considering the characteristics of survival data, a time-dependent ROC curve was used to assist in the diagnosis of discrimination. The receiver operating characteristic curve (ROC curve) was plotted with the true positive rate (sensitivity) as the ordinate and the false positive rate (1-specificity) as the abscissa. The cut-off value and its corresponding sensitivity and specificity were marked on the ROC curve, and the Youden's index (sensitivity + specificity−1) was calculated [25 (link)]. The AUC of the model above 0.7 was considered good discrimination. Brier score and calibration plot were used to measure the calibration of the prediction model. Brier score of the model between 0 and 0.25 was considered good calibration. The abscissa of the calibration plot was the predicted probability and the ordinate was the actual probability. Ideally, the calibration curve was a diagonal line (predicted probability equals actual probability). When the curve was below the diagonal line, the predicted probability was higher than the actual probability; conversely, it was less than the actual probability. The tenfold cross-validation method was used to internally validate the model. Interaction effects between independent variables were also considered. To explore whether the effect of BP on CV events was contingent on variations in CNVs, the interaction terms of two BP variation-related CNVs loci with baseline SBP and DBP were developed respectively.
Then we conducted sensitivity analysis to selection of alternative parameters. This prediction model was based on stepwise regression and AIC criteria, and was of professional significance. To quantify the improvement offered by CNV loci, we used net reclassification improvement (NRI) and integrated discrimination improvement (IDI) to examine the model performance. The risk stratification threshold of category-based NRI was 20% [26 (link), 27 (link)]. If NRI > 0, the model performance has improved after adding CNV loci; if NRI < 0, the model performance has decreased; if NRI = 0, the model was considered not changed. IDI was judged in the same way as NRI.
The nomogram drawing with R was used to visualise the model. Using multivariate regression analysis, several predictors were combined, and calibrated line segments were drawn in a certain proportion on the same plane to indicate the link between variables in the prediction model. In the Cox model, each influencing factor was given a score depending on its contribution to the outcome variable. After then, the total score was obtained by adding all the scores together. Finally, the function conversion relationship between the overall score and the likelihood of the result event was utilised to calculate the predicted probability of each particular outcome event.