Factor screening identifies which variables influence your outcome. Whether you are adjusting manufacturing parameters or evaluating clinical data, this process prevents the inefficiency of testing every possible combination. Lattice manages this through a structured three-stage flow: first, the LLM analyzes your request to identify relevant factors; second, our deterministic engine performs statistical calculations like ANOVA or non-parametric tests; finally, the LLM translates these technical outputs into plain English explanations. We prioritize data integrity by running a multi-level validation check before any calculation. If your data is imbalanced or non-normal, the system prompts you with specific options to resolve issues—such as data transformation or alternative tests—rather than silently altering your inputs. This ensures your results remain fully traceable and accurate to the original data.
When to choose this family
- You have a large set of candidate factors (10+) and want to identify the 3-6 most critical ones.
- You are in the early stages of a process optimization project and need to prioritize variables before running a complex experimental design.
- You have measured a response and need to determine which input factors have a statistically significant impact on it.
- You want to rank factors based on their influence to save time and resources on future testing.
How screening works
Screening uses statistical tests to compare how different settings of a factor affect your outcome. By calculating the variance between groups, it determines whether differences in your results are likely due to the factor itself or just random variation.
Tools like Pareto charts and main effect plots visualize these findings. A Pareto chart ranks factors by their contribution, while main effect plots show the average shift in your response as you move through a factor's levels, making it clear which parameters require your immediate attention.
Differences from other analytical methods
Unlike diagnostic modeling, which seeks to predict precise outcomes, this family focuses purely on sorting and reduction. Its primary output is not a predictive equation, but a decision-ready list of 'significant' vs 'not significant' variables.
While other families might look at the complex relationship between many variables simultaneously, this approach emphasizes individual factor impact. This makes it specifically suitable for the initial phase of optimization, where understanding the 'what' is more important than defining the 'how' of the entire system.
Common pitfalls to avoid
A frequent mistake is overlooking data quality before running tests. Analyzing sparse or highly imbalanced data can lead to misleading conclusions, where a factor appears significant simply due to a lack of samples in one category.
Another error is blindly trusting p-values without considering the context. Always check the effect size and the validation report provided by Lattice to ensure the findings are meaningful for your specific process and not just a result of high sample sensitivity.
Frequently asked questions
- What happens if my data fails the validation check?
- Lattice categorizes validation issues into three levels: warnings, confirmations, and blocks. If your data triggers a 'confirm' level (like non-normality), the system will stop and offer you specific choices, such as applying a log transformation, before continuing the analysis. A 'block' indicates the data is insufficient for a valid result, such as having too few samples for an ANOVA.
- Why doesn't the system automatically clean my data?
- We avoid automatic data modification because it obscures the true source of your results. If the engine performed Box-Cox transformations or deleted outliers behind the scenes, you would lose the ability to trace your results back to your raw observations. You are always in control of these decisions.
- Can I use these tools for more than just manufacturing processes?
- Yes. While the methodology is rooted in industrial process optimization, the underlying tools like ANOVA and Pareto ranking are applicable to any domain where you need to isolate influential variables, such as market research, clinical trials, or operational quality control.