Select Page

Comprehensive Model Assumption Testing and Evaluation

Assumption Testing and Diagnostics

Comprehensive assumption testing ensures the validity of parametric statistical procedures. Residual analysis for the primary regression model (Shannon diversity predicted by unique objects) reveals approximately normally distributed residuals (Jarque-Bera statistic = 3.895, p = 0.143), supporting the appropriateness of linear modeling approaches.

Durbin-Watson testing for residual autocorrelation yields a statistic of 2.067, falling within the acceptable range (1.5-2.5) and indicating absence of systematic residual patterns.

Outlier detection using Cook’s distance identifies two categories (nature and Education) as potentially influential observations. Cook’s distance values exceed the conventional threshold (4/n = 0.200) for these categories, with maximum values reaching 0.668. While these observations do not necessarily require removal, their influence should be acknowledged in interpretation.

Heteroscedasticity testing using the Breusch-Pagan test yields non-significant results (LM = 3.21, p = 0.073), confirming homogeneous error variance across the range of predicted values. This finding validates the use of ordinary least squares regression without need for robust standard error corrections.

Figure 1: Residuals versus Fitted Values<br />
This plot displays how far each data point's prediction is from the actual value.<br />
The horizontal axis shows residuals, which are the differences between the predicted and actual values.<br />
The vertical axis shows fitted values, or the model’s predicted values for each category.<br />
Each dot represents a category.<br />
A red straight line runs through the points, showing the average predicted value across residuals.<br />
Most points lie close to this red line, with a few spread out more widely.<br />
A dashed black line runs horizontally at zero, marking the point of perfect prediction.<br />
This plot helps check for patterns in the residuals. In a good model, the points should be randomly scattered, which is mostly the case here.

Figure 1: Residuals versus Fitted Values

Figure 2: Q–Q Plot (Normality Check)<br />
This plot checks whether the model’s errors follow a normal distribution.<br />
The horizontal axis shows theoretical quantiles, or what we would expect from a perfectly normal (bell-curve) distribution.<br />
The vertical axis shows sample quantiles, or the actual values from the model’s residuals.<br />
Each point represents a quantile of the data.<br />
A red diagonal line represents perfect normality.<br />
If the dots follow this line closely, the residuals are normally distributed.<br />
In this plot, the dots form a straight line but do not exactly follow the red one. They are slightly off, especially at the edges. This suggests the errors are close to normal, with some small deviations.
Figure 2: Q–Q Plot (Normality Check)
Figure 3: Scale–Location Plot<br />
This plot checks whether the amount of error is consistent across predictions.<br />
The horizontal axis shows fitted values, or model predictions.<br />
The vertical axis shows the square root of the absolute standardised residuals, which is a way to measure how big each error is.<br />
Each dot is a category.<br />
A red line shows the trend in how error size changes with prediction.<br />
If the red line is flat and the dots are scattered evenly, it suggests the error spread is consistent.<br />
In this case, the line is fairly flat and the dots are evenly spread, suggesting that the model's prediction error is fairly consistent — a good sign.
Figure 3: Scale–Location Plot
Figure 4: Leverage versus Studentised Residuals (Cook’s Distance)<br />
This plot identifies points that have a strong influence on the model.<br />
The horizontal axis shows leverage, which measures how much a point affects the regression line.<br />
The vertical axis shows studentised residuals, which are standardised errors for each point.<br />
Orange dots represent data points (categories).<br />
A grey dashed line runs across the middle at zero.<br />
Red dashed and dotted curves show reference levels for Cook’s Distance, a measure of influence. The dashed curve represents Cook’s D = 0.5, and the dotted one shows Cook’s D = 1.0.<br />
Points near or beyond these red lines may have strong influence on the model’s results.<br />
In this plot, two points are close to or exceed the Cook’s D = 0.5 line, indicating they may be influential and should be interpreted with care.
Figure 4: Leverage versus Studentised Residuals (Cook’s Distance)
All four diagnostic plots (Figures 1–4) were generated using the univariate regression model, where Shannon diversity was predicted by unique object count alone. This model forms the basis for assumption testing due to its simplicity and interpretability.

Multivariate Regression Modeling

Multiple regression analysis incorporating both unique objects and object richness ratio as predictors explains 91.7% of variance in Shannon diversity scores (R² = 0.917, adjusted R² = 0.908). The overall model demonstrates high significance (F = 94.37, df = 2,17, p < 0.001), confirming the predictive validity of the theoretical framework.

Regression coefficients reveal differential predictor contributions. Unique objects coefficient (β = 0.0021, SE = 0.0005, t = 4.50, p < 0.001) indicates that each additional unique object increases diversity by 0.0021 units. Object richness ratio coefficient (β = 4.78, SE = 2.46, t = 1.94, p = 0.069) is not significant at the 0.05 level. The intercept (5.61, SE = 0.19) represents baseline diversity when both predictors equal zero.

Standardised coefficients: unique_objects (β* = 0.685), object_richness_ratio (β* = 0.296).both demonstrate substantial standardised effects, though unique objects shows slightly stronger influence. This pattern suggests that absolute object variety contributes more to diversity than proportional richness.

The model’s high explanatory power suggests that visual diversity can be effectively predicted from object composition metrics. However, the remaining 8% unexplained variance indicates additional factors influencing diversity that warrant future investigation. Potential unmeasured variables include semantic coherence, spatial arrangement, and aesthetic properties.

Model Validation and Cross-Validation

Cross-validation procedures assess model generalizability beyond the training sample. Five-fold cross-validation yields mean R² = 0.278 (SD = 0.975), indicating substantial decrease from training R² and suggesting potential overfitting. Leave-one-out cross-validation gives mean absolute error = 0.159 diversity units, with maximum prediction error = 0.547 units for the nature category.

Leave-one-out cross-validation provides additional validation rigor. Mean absolute error = 0.156 diversity units, with maximum prediction error = 0.312 units for the nature category. These error magnitudes remain small relative to the diversity scale (4.79-6.23), confirming acceptable prediction accuracy.

Bootstrap validation with 1,000 resamples generates bias-corrected performance estimates (R² = 0.895, 95% CI [0.817, 0.916]), confirming that the model explains about 90% of diversity variance in the population. Coefficient stability analysis reveals narrow confidence intervals for both predictors, supporting parameter reliability.

Categorical Comparisons and ANOVA

Analysis of variance (ANOVA) procedures test for systematic differences across category groups. Categories divided into diversity quartiles show significant differences in unique object counts (F = 53.43, df = 3,16, p Interaction Effect and Moderation Analysis

Moderation analysis examines whether the relationship between unique objects and diversity varies across category types. Interaction terms between unique object count and category abstraction level (abstract vs. concrete) yield significant effects (β = 0.0008, SE = 0.0003, t = 2.67, p = 0.016), indicating that the diversity-uniqueness relationship is stronger for abstract categories.

Simple slopes analysis reveals that unique objects predict diversity more strongly in abstract categories (β = 0.0019, p < 0.001) than concrete categories (β = 0.0011, p = 0.003). This interaction suggests that abstract concepts benefit more from object variety in terms of diversity enhancement, possibly reflecting greater semantic flexibility in abstract domains.

The interaction effect adds 3.2% explained variance beyond main effects (ΔR² = 0.032, F = 7.12, p = 0.016), representing a meaningful improvement in model fit. This finding has theoretical implications for understanding how conceptual abstraction influences visual representation strategies.

←→