# How to Choose Factor Levels for a Full Factorial Experiment in Biology

## Quick Answer

- Choose factor levels based on prior biological knowledge, pilot data, and the minimum effect size you need to detect, not arbitrary or convenient values.
- Set each factor to at least two levels, with spacing wide enough to produce measurable response differences while remaining biologically feasible and safe.
- Document your level-selection rationale before data collection, because post hoc adjustments compromise the statistical validity of a full factorial design.

## At a Glance

| Decision Point | Recommended Approach | Common Error | Consequence of Error |
| --- | --- | --- | --- |
| Number of levels per factor | Use 2 levels for screening, 3 or more for curvature detection | Using too many levels without replication | Unmanageable run counts and wasted resources |
| Level spacing | Base on prior literature and pilot data | Levels too close together | Response differences masked by noise |
| Level range | Span biologically meaningful extremes | Range too narrow or too wide | Missed effects or infeasible conditions |
| Replication | Include true replicates for error estimation | Confusing replicates with repeated measures | Inflated confidence in results |
| Blocking | Assign nuisance variables to blocks | Ignoring batch or day effects | Confounded treatment effects |
| Center points | Add for curvature detection in quantitative factors | Omitting center points | Missed nonlinear responses |

## Understanding Full Factorial Designs in Biology

A full factorial experiment tests all possible combinations of factor levels. In biology, these factors might be temperature, pH, nutrient concentration, light intensity, or treatment duration. The design allows you to estimate both the individual effect of each factor and the interactions between factors. For example, a study with three factors each at two levels requires eight experimental runs, and each run represents one unique combination of conditions.

The primary advantage of a full factorial design is its ability to reveal interactions. A factor that shows no effect on its own may become important when combined with another factor. This is common in biological systems where pathways respond to multiple inputs simultaneously. The cost of this completeness is the exponential growth in run count. Each additional factor doubles the number of runs when all factors have two levels, and triples it when all have three levels.

The choice of factor levels determines what the experiment can detect. If the levels are too close together, the response may not change enough to be distinguished from experimental noise. If the levels are too far apart, you may miss the biologically relevant range or create conditions that are not feasible or safe. The level selection process therefore requires a balance between statistical power, biological relevance, and practical constraints.

## Core Principles for Level Selection

### Use Prior Knowledge to Define the Factor Range

The starting point for any factor level selection is the existing literature and your own preliminary observations. Published studies in your specific biological system provide the most direct evidence for plausible factor ranges. If you are studying enzyme activity, for example, prior reports of the enzyme's optimal temperature and its stability limits should define your search window. The National Library of Medicine's research methods resources can help you locate authoritative biomedical and life-science references that establish these ranges [<a href="#ref-1">1</a>].

Your own pilot data also matter. A small pilot experiment with a few exploratory runs can reveal whether your proposed levels produce measurable differences. Pilot data help you avoid the common error of selecting levels that are too close together, where the response is flat across the range. They also help you avoid levels that are too extreme, where the biological system fails entirely.

When prior knowledge is limited, use a wider range with more levels. This approach is more expensive but reduces the risk of missing the biologically active region. You can narrow the range in a follow-up experiment once you know where the response changes.

### Define the Minimum Effect You Need to Detect

The spacing between factor levels must be large enough to produce a response change that you can detect with your measurement system. This minimum detectable effect depends on the variability of your response measurement and your sample size. If your measurement has high variability, you need larger level spacing to see a real difference above the noise.

A practical approach is to estimate the expected response difference between your proposed levels using prior data or published values. If the expected difference is smaller than your measurement noise, the experiment will not be able to distinguish the levels. You then have three options: increase the level spacing, reduce measurement noise, or increase replication.

### Balance Biological Feasibility with Statistical Needs

The statistically ideal level range may not be biologically feasible. For example, a temperature range that maximizes the statistical power might include temperatures that kill the organism or denature the assay. The range must remain within the limits where the biological system functions and where your measurement remains valid.

Feasibility also includes practical constraints such as the time required for each run, the cost of reagents, and the availability of equipment. A full factorial design with many factors and levels can quickly become impractical. The total number of runs is the product of the number of levels for each factor. A design with four factors at three levels requires 81 runs, which may be too many for a single experiment.

## Number of Levels per Factor

### Two Levels for Screening Experiments

Two levels per factor is the most common choice for screening experiments. This design requires the fewest runs and provides the basic information about whether each factor has an effect and whether factors interact. The two levels should represent the low and high ends of the biologically meaningful range.

The limitation of a two-level design is that it cannot detect curvature in the response. If the response increases to a peak and then decreases, a two-level design will only show the difference between the two endpoints. You will not know whether the response is linear or curved between those points. This limitation is acceptable when you are screening many factors and plan to follow up with a more detailed experiment on the important ones.

### Three or More Levels for Curvature Detection

Three levels per factor allow you to detect curvature. The three levels typically include a low, a center, and a high value. If the response at the center point differs from the midpoint of the low and high responses, the response is curved. This information is valuable when you need to identify an optimal factor level instead of just determine whether the factor has an effect.

The cost of three levels is a substantial increase in the number of runs. A design with three factors at three levels requires 27 runs, compared to 8 runs for the same factors at two levels. The additional runs are justified when the factor is known to have a nonlinear effect or when you need to identify an optimum.

### Center Points as a Compromise

Center points provide a way to detect curvature without running a full three-level design. In a two-level factorial design, you add one or more runs at the midpoint of all factors. The center point runs are repeated to estimate the experimental error. If the average response at the center points differs significantly from the average of the factorial runs, the response has curvature.

Center points are a cost-effective addition to a two-level design. They require only a few extra runs and provide information about whether a more detailed experiment is needed. They also provide a direct estimate of experimental error, which is useful for statistical tests.

## Spacing Between Levels

### Use the Expected Response Curve to Set Spacing

The spacing between levels should be based on the expected shape of the response curve. If the response is expected to change gradually across the range, wider spacing is needed to produce a detectable difference. If the response is expected to change sharply, narrower spacing can be used.

The most reliable way to estimate the expected response curve is from pilot data. A pilot experiment with several levels across the range can show where the response changes and where it is flat. This information allows you to place your factorial levels in the region where the response is changing, which maximizes the power of the experiment.

### Avoid Levels That Are Too Close Together

Levels that are too close together produce responses that are indistinguishable from experimental noise. This is a common error in biology, where researchers choose levels based on convenience or convention instead of on the expected response. The result is a null finding that reflects the design instead of the biology.

The solution is to use prior data to estimate the expected response difference between your proposed levels. If the expected difference is smaller than the measurement noise, the levels are too close. Increase the spacing until the expected difference is at least several times the noise.

### Avoid Levels That Are Too Far Apart

Levels that are too far apart can also be problematic. If the high level is outside the biologically functional range, the response may be zero or the organism may die. This creates a response that is not informative about the factor's effect within the functional range.

The high level should be at the upper end of the biologically meaningful range, not beyond it. The low level should be at the lower end of the meaningful range. The range should be wide enough to produce a detectable response difference but narrow enough to keep the system functional.

## Replication and Its Role in Level Selection

### True Replicates versus Repeated Measurements

Replication is the number of independent experimental units at each factor combination. A true replicate is a separate experimental unit that is independently prepared and measured. A repeated measurement is the same experimental unit measured multiple times. Only true replicates provide an estimate of the experimental error that is used in statistical tests.

The distinction matters for level selection because the error estimate determines the minimum detectable difference. If you use repeated measurements as if they were replicates, you will underestimate the error and overestimate the precision of your results. This can lead you to choose level spacing that is too narrow for the true variability of the system.

### How Many Replicates Do You Need

The number of replicates depends on the variability of the measurement and the size of the effect you need to detect. More replicates reduce the error of the mean and allow you to detect smaller differences. The tradeoff is the cost of each additional replicate.

A common approach is to start with three replicates per combination and increase the number if the variability is high. The pilot data can provide an estimate of the variability, which allows you to calculate the number of replicates needed to detect your minimum effect. This calculation is a standard part of experimental design and is supported by the research methods literature available through the National Library of Medicine [<a href="#ref-1">1</a>].

## Blocking and Nuisance Factors

### Identifying Nuisance Factors

Nuisance factors are variables that affect the response but are not the focus of the experiment. Examples include the day of the experiment, the batch of reagents, the position of the sample in the incubator, or the person performing the measurement. These factors can add variability to the experiment and obscure the effects of the factors you are studying.

The first step is to identify the nuisance factors that are likely to affect your system. This requires knowledge of your experimental system and the sources of variability that have been observed in prior work. If you are unsure, you can run a pilot experiment with the same protocol and measure the variability across days or batches.

### Assigning Nuisance Factors to Blocks

Blocking is a technique that assigns nuisance factors to groups of runs that are as similar as possible. For example, if you know that the day of the experiment affects the response, you can run all the combinations for one day in a block and the remaining combinations on another day. The block effect is then accounted for in the statistical analysis.

Blocking is particularly important in full factorial designs because the number of runs can be large. If the runs are spread over many days or batches, the nuisance variability can be substantial. Blocking reduces this variability and increases the power of the experiment.

### The Cost of Ignoring Blocking

If you ignore a nuisance factor that is present, the variability from that factor becomes part of the experimental error. This increases the error and makes it harder to detect the effects of the factors you are studying. The result is a less powerful experiment that may fail to detect real effects.

The solution is to identify the nuisance factors before the experiment and include them in the design. This requires planning and may require additional runs to accommodate the blocks. The additional runs are usually worth the cost because they improve the quality of the results.

## Practical Workflow for Choosing Factor Levels

### Step 1: Review the Literature and Prior Data

Begin by collecting all available information about the factors you plan to study. This includes published studies, your own prior experiments, and any preliminary data. The goal is to establish the biologically meaningful range for each factor and the expected response shape.

The National Library of Medicine research methods resources can help you locate authoritative references for your specific biological system [<a href="#ref-1">1</a>]. The EQUATOR Network provides reporting guidelines that can help you understand what information is expected in a well-designed study [<a href="#ref-2">2</a>].

### Step 2: Define the Minimum Effect You Need to Detect

Decide on the smallest effect that is biologically or practically important. This is the effect that you would not want to miss. The minimum effect size should be based on the purpose of the experiment, not on what is statistically convenient.

### Step 3: Estimate the Experimental Error

Use pilot data or prior studies to estimate the variability of your measurement. This estimate is needed to calculate the number of replicates and the level spacing required to detect the minimum effect.

### Step 4: Select the Number of Levels

Choose the number of levels for each factor based on whether you need to detect curvature. Use two levels for screening and three levels for curvature detection. Consider center points as a compromise.

### Step 5: Set the Level Values

Set the level values based on the expected response range and the minimum detectable difference. The levels should be within the biologically functional range and spaced widely enough to produce a detectable response difference.

### Step 6: Plan for Replication and Blocking

Determine the number of replicates needed for each combination and identify the nuisance factors that require blocking. Include these in the design before the experiment begins.

### Step 7: Document Your Rationale

Write down the rationale for each level selection before the experiment. This documentation is important for reproducibility and for the reporting of the study. The EQUATOR Network provides reporting guidelines that specify what information should be included in the methods section of a paper [<a href="#ref-2">2</a>].

## Records and Measurements

### What to Record for Level Selection

The documentation of your level selection should include the following information for each factor:

- The biological rationale for the range
- The prior data or literature that supports the range
- The minimum effect size you need to detect
- The estimated experimental error
- The number of levels and the specific values
- The number of replicates per combination
- The blocking scheme

This information should be recorded before the experiment begins. It provides the basis for the statistical analysis and for the interpretation of the results.

### What to Record During the Experiment

During the experiment, record the actual conditions for each run, including the factor levels, the date and time, the equipment used, and the operator. This information is needed to identify any problems that may have occurred during the experiment.

The NIH Data Management and Sharing Policy describes the expectations for data management and sharing in NIH-funded research [<a href="#ref-3">3</a>]. While this policy applies to NIH-funded research, the principles of good data management apply to all research. The data from your experiment should be organized, documented, and stored in a way that allows for verification and reuse.

## Common Failure Patterns

### Levels Too Close Together

The most common failure pattern is selecting levels that are too close together. This results in a response that does not change enough to be detected. The experiment shows no effect, but the conclusion is not valid because the design was not capable of detecting the effect.

The solution is to use prior data to estimate the expected response difference and to ensure that the difference is larger than the measurement noise. If the expected difference is not known, use a wider spacing or a pilot experiment.

### Levels Too Far Apart

The opposite failure is selecting levels that are too far apart. This can result in the high level being outside the biologically functional range. The response at the high level may be zero or the organism may die, which does not provide useful information about the effect of the factor.

The solution is to base the range on the biologically meaningful range and to verify that the high level is within the functional range. A pilot experiment can help you identify the functional range.

### Ignoring Nuisance Factors

A third failure is ignoring nuisance factors that affect the response. This increases the experimental error and reduces the power of the experiment. The result is a failure to detect real effects.

The solution is to identify the nuisance factors before the experiment and include them in the design through blocking. This requires knowledge of the experimental system and the sources of variability.

### Using Repeated Measurements as Replicates

A fourth failure is using repeated measurements as replicates. This underestimates the error and overestimates the precision of the experiment. The result is a conclusion that is not supported by the actual variability of the experiment.

The solution is to ensure that the replicates are independent experimental units. Each replicate should be a separate run of the experiment, not a repeated measurement of the same run.

## Limitations of Full Factorial Designs

### The Number of Runs Grows Quickly

The most significant limitation of a full factorial design is the number of runs. The number of runs is the product of the number of levels for each factor. A design with five factors at two levels requires 32 runs, and a design with five factors at three levels requires 243 runs. This can be impractical for many biological experiments.

The solution is to use a fractional factorial design, which tests only a subset of the combinations. Fractional designs are less efficient at estimating interactions but require fewer runs. The choice between a full and fractional design depends on the number of factors and the importance of the interactions.

### The Design Assumes a Specific Response Model

A full factorial design assumes that the response can be described by a model that includes the main effects and interactions. If the response has a more complex shape, such as a threshold or a peak, the design may not capture it. The use of center points can detect curvature, but a more complex response may require a response surface design.

### The Results Are Specific to the Levels Chosen

The results of a full factorial experiment are specific to the levels chosen. If the levels are changed, the results may change. This is why the level selection must be based on the biologically meaningful range and the purpose of the experiment.

## Safety and Regulatory Context

### Biosafety Considerations

The choice of factor levels can affect the safety of the experiment. For example, a factor that increases the growth of a pathogen may also increase the risk of exposure. The levels should be chosen to minimize the risk of exposure while still providing useful information.

The biosafety level of the experiment should be determined before the experiment begins. The level depends on the organism and the procedures used. The factor levels should be consistent with the biosafety level of the laboratory.

### Ethical and Regulatory Requirements

The design of the experiment should comply with the ethical and regulatory requirements that apply to the research. This includes the requirements for the use of animals, human subjects, or hazardous materials. The NIH Grants and Funding pages provide information about the requirements for NIH-funded research [<a href="#ref-4">4</a>].

The Committee on Publication Ethics core practices describe the ethical standards for the conduct and reporting of research [<a href="#ref-5">5</a>]. These practices include the responsible conduct of research, the reporting of data, and the handling of misconduct. The design of the experiment should be consistent with these practices.

## Professional Escalation Criteria

### When to Seek Help

If you are uncertain about the level selection for your experiment, you should seek help from a statistician or a colleague with experience in experimental design. This is particularly important when the experiment is large or when the consequences of a poor design are significant.

### When to Redesign the Experiment

If the pilot data show that the expected response difference is smaller than the measurement noise, you need to redesign the experiment. The redesign may involve increasing the level spacing, reducing the measurement noise, or increasing the number of replicates.

### When to Stop the Experiment

If the experiment is producing results that are not interpretable because of the design, you should stop the experiment and redesign it. Continuing an experiment with a poor design wastes resources and produces results that cannot be interpreted.

## A Decision Log System for Factor Level Selection

The existing workflow describes what to record, but researchers often struggle with how to structure the decision process so that level choices remain defensible after the experiment is complete. A decision log is a practical tool that forces you to write down the reasoning for each level choice before data collection begins. This log serves as a pre-registration of your design decisions and provides a clear record that can be reviewed by collaborators, statisticians, or journal reviewers. The log also helps you identify weak assumptions early, when they can still be corrected.

### The Structure of a Factor Level Decision Log

Create one table for each factor in the experiment. Each table should contain the following rows: the proposed low level, the proposed high level, the proposed center level if applicable, the source of the range, the expected response difference, the estimated measurement noise, the minimum detectable effect, and the feasibility check. The source of the range should cite the specific publication, pilot experiment, or prior dataset that supports the chosen values. The expected response difference is your quantitative prediction of how much the response will change between the low and high levels. The estimated measurement noise comes from pilot data or published values for the same assay. The minimum detectable effect is the smallest change you need to detect to answer your biological question.

The log should be completed before any experimental runs are performed. If you cannot fill in a row because you lack the information, that is a signal that you need to run a pilot experiment or consult additional literature before committing to the full factorial design. The National Library of Medicine research methods resources can help you locate the authoritative references needed to fill these gaps [<a href="#ref-1">1</a>].

### A Worked Example from Molecular Biology

Consider a researcher studying the effect of a chemical inducer on gene expression in a cell line. The factor is the inducer concentration, and the response is the fluorescence signal from a reporter gene. The researcher has prior data showing that the response increases with concentration and then plateaus. The pilot data show that the measurement noise, expressed as the standard deviation of replicate measurements, is 5 percent of the mean signal.

The researcher wants to detect a 20 percent change in the response, which is the minimum effect that would be biologically meaningful for the downstream application. The decision log would record the following: the low level is set at a concentration that produces a response near the baseline, the high level is set at a concentration that produces a response near the plateau, and the expected response difference between these levels is 80 percent of the maximum signal. The expected difference of 80 percent is well above the minimum detectable effect of 20 percent and is more than ten times the measurement noise of 5 percent. The feasibility check confirms that both concentrations are within the solubility limits of the inducer and do not affect cell viability.

This log entry shows that the level spacing is justified by the expected response curve, the measurement noise, and the minimum effect of interest. If the expected difference had been only 10 percent, the log would reveal that the levels are too close together to detect the minimum effect of 20 percent. The researcher would then need to widen the range or reduce the measurement noise.

### A Worked Example from Ecology

In an ecological study, the factor is the density of a competitor species and the response is the growth rate of a focal species. The researcher has field observations suggesting that the growth rate declines as competitor density increases. The pilot data show that the measurement noise is high, with a standard deviation of 15 percent of the mean growth rate.

The minimum effect of interest is a 30 percent reduction in growth rate, because smaller reductions are not considered ecologically significant. The decision log records the low density as the natural background level and the high density as the maximum density observed in the field. The expected response difference between these levels is 40 percent based on the field observations. The expected difference of 40 percent is above the minimum detectable effect of 30 percent but is less than three times the measurement noise of 15 percent.

This log reveals a potential problem. The expected difference is only 2.7 times the noise, which may not provide sufficient statistical power with a reasonable number of replicates. The researcher can respond by increasing the number of replicates, reducing the measurement noise through more careful measurement protocols, or widening the density range if the field observations support it. The decision log makes this tradeoff explicit before the experiment begins.

### How to Use the Log to Detect Design Failures

The decision log is beyond a documentation tool. It is a diagnostic tool that can identify design failures before you spend resources on the full experiment. The key diagnostic is the ratio of the expected response difference to the measurement noise. If this ratio is below 2, the design is likely to fail because the response difference will be indistinguishable from noise. If the ratio is between 2 and 4, the design may work but will require a larger number of replicates. If the ratio is above 4, the design is likely to detect the effect with a modest number of replicates.

The second diagnostic is the comparison between the expected response difference and the minimum detectable effect. If the expected difference is smaller than the minimum effect you need to detect, the design cannot answer your biological question. You must either widen the levels or reconsider the minimum effect size. The minimum effect size should be based on the purpose of the experiment, not on what is convenient to detect.

The third diagnostic is the feasibility check. The log should record whether each level is within the biologically functional range and whether the experimental conditions are safe and practical. If the high level is outside the functional range, the response may be zero or the organism may die. This creates a response that is not informative about the factor's effect within the functional range.

### A Scoring System for Level Selection

To make the decision log more systematic, you can assign a score to each factor based on the quality of the evidence supporting the level selection. The score has three components. The first component is the source of the range, with a score of 3 for direct prior data from the same system, 2 for published studies in a similar system, and 1 for expert opinion or no direct evidence. The second component is the expected response difference relative to the minimum detectable effect, with a score of 3 if the expected difference is at least twice the minimum effect, 2 if it is between 1 and 2 times the minimum effect, and 1 if it is below the minimum effect. The third component is the expected difference relative to the measurement noise, with a score of 3 if the ratio is above 4, 2 if the ratio is between 2 and 4, and 1 if the ratio is below 2.

A total score of 7 to 9 indicates that the level selection is well supported and the experiment is likely to succeed. A total score of 4 to 6 indicates that the level selection has some support but may require more replicates or a wider range. A total score of 3 or below indicates that the level selection is not adequately supported and the experiment should be redesigned before data collection.

This scoring system is not a substitute for statistical power calculations. It is a quick screening tool that helps you decide whether the level selection is worth the effort of a formal power analysis. The formal power analysis requires an estimate of the variance and the effect size, which are the same quantities recorded in the decision log.

### When to Escalate to a Statistician

The decision log also provides a clear trigger for professional escalation. If the score for any factor is 3 or below, or if the expected response difference is below the minimum detectable effect, you should consult a statistician before proceeding. The statistician can help you determine whether a different design, such as a fractional factorial or a response surface design, would be more appropriate.

The decision log should be shared with the statistician as part of the consultation. The log provides the statistician with the information needed to assess the design and recommend changes. The statistician can also help you calculate the number of replicates needed to detect the minimum effect given the estimated noise.

### The Decision Log as a Reporting Tool

The decision log also serves as the basis for the methods section of your paper. The EQUATOR Network provides reporting guidelines that specify what information should be included in the methods section of a study [<a href="#ref-2">2</a>]. The decision log contains the information needed to report the level selection in a transparent way. You can include the log as a supplementary table or summarize the key decisions in the methods section.

The Committee on Publication Ethics core practices describe the ethical standards for the conduct and reporting of research [<a href="#ref-5">5</a>]. These practices include the responsible conduct of research and the reporting of data. The decision log supports these practices by providing a clear record of the design decisions and the rationale behind them.

### The Decision Log and Data Management

The decision log should be stored with the experimental data and made available to other researchers. The NIH Data Management and Sharing Policy describes the expectations for data management and sharing in NIH-funded research [<a href="#ref-3">3</a>]. While this policy applies to NIH-funded research, the principles of good data management apply to all research. The decision log is part of the metadata that describes how the data were generated.

The decision log should be stored in a format that is accessible and readable. A spreadsheet or a text file is sufficient. The log should include the date it was created and the name of the person who created it. This information is important for verifying the timeline of the design decisions.

### The Decision Log and Researcher Identity

The decision log also supports the researcher's professional record. The ORCID for Researchers pages describe how researchers can maintain a record of their work and contributions [<a href="#ref-6">6</a>]. The decision log is part of the research record that demonstrates the rigor of the experimental design. It can be shared with collaborators, reviewers, or funders as evidence of the quality of the research.

The decision log is not a publication in itself, but it is a component of the research record that supports the publication. The log provides the evidence that the level selection was based on prior knowledge and a systematic assessment of the expected response, the measurement noise, and the minimum detectable effect.

### A Template for the Decision Log

The following template can be adapted for any factor in a full factorial experiment. The template is a table with one row per factor. The columns are the factor name, the low level, the high level, the center level, the source of the range, the expected response difference, the estimated noise, the minimum detectable effect, the feasibility check, and the score.

| Factor | Low Level | High Level | Center Level | Source of Range | Expected Difference | Estimated Noise | Minimum Effect | Feasibility | Score |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Inducer concentration | 0.1 uM | 10 uM | 1 uM | Pilot study | 80 percent | 5 percent | 20 percent | Cell viability confirmed | 9 |
| Temperature | 20 C | 30 C | 25 C | Published study | 40 percent | 15 percent | 30 percent | Organism survives | 6 |

The template is a starting point. You can add rows for additional factors and columns for additional information. The key is to complete the log before the experiment begins and to use it to identify design problems before you spend the resources on the full experiment.

### The Decision Log and the Full Factorial Design

The decision log is particularly important for full factorial designs because the number of runs grows quickly with the number of factors and levels. A design with four factors at two levels requires 16 runs, and a design with four factors at three levels requires 81 runs. The cost of a poor level selection is multiplied by the number of runs. The decision log helps you avoid the cost of a failed experiment by identifying the design problems before the experiment begins.

The decision log also helps you compare the level selections across factors. In a full factorial design, the level selection for one factor affects the interpretation of the other factors. The decision log provides a consistent framework for evaluating the level selection for each factor and for identifying the factors that need more attention.

### The Decision Log and the Pilot Experiment

The decision log is most useful when it is based on pilot data. The pilot experiment provides the estimates of the measurement noise and the expected response difference that are needed to complete the log. The pilot experiment should be designed to provide these estimates with a small number of runs. A pilot experiment with three to five levels across the range of each factor is usually sufficient.

The pilot experiment should be run under the same conditions as the full experiment. The measurement protocol should be the same, and the equipment should be the same. The pilot data should be analyzed to estimate the noise and the response shape. The results of the pilot analysis are then recorded in the decision log.

### The Decision Log and the Minimum Effect

The minimum detectable effect is the smallest effect that is biologically or practically important. The decision log requires you to state this effect explicitly. This is a critical step because it forces you to think about the purpose of the experiment. The minimum effect should not be based on what is statistically convenient. It should be based on the biological question that the experiment is designed to answer.

If you cannot define a minimum effect, the experiment may not be worth doing. The decision log makes this explicit. If the minimum effect is not defined, the score for the expected difference relative to the minimum effect cannot be calculated. This is a signal that the experiment needs more thought before it is designed.

### The Decision Log and the Measurement Noise

The measurement noise is the variability of the response measurement under the same conditions. The decision log requires you to estimate this noise from pilot data or prior studies. The noise estimate is used to calculate the ratio of the expected difference to the noise. This ratio is a key indicator of the power of the experiment.

The noise estimate should be based on true replicates, not repeated measurements. A repeated measurement of the same sample underestimates the noise because it does not include the variability between samples. The decision log should record the basis of the noise estimate so that the estimate can be evaluated.

### The Decision Log and the Feasibility Check

The feasibility check is the final component of the decision log. The check confirms that each level is within the biologically functional range and that the conditions are safe and practical. The feasibility check should be based on the prior knowledge of the system and the practical constraints of the laboratory.

The feasibility check is important because the statistically ideal level range may not be biologically feasible. For example, a temperature range that maximizes the statistical power may include temperatures that kill the organism. The feasibility check prevents the selection of levels that are not practical.

### The Decision Log and the Professional Escalation

The decision log provides a clear basis for professional escalation. If the score for any factor is 3 or below, or if the expected difference is below the minimum effect, you should consult a statistician. The statistician can help you determine whether the design can be improved or whether a different design is needed.

The decision log should be shared with the statistician as part of the consultation. The statistician can review the log and provide recommendations. The statistician can also help you calculate the number of replicates needed to achieve the minimum effect with the estimated noise.

### The Decision Log and the Reporting

The decision log is a record of the design decisions that should be reported in the study. The EQUATOR guidelines provide reporting guidelines for the methods section of a study [<a href="#ref-2">2</a>]. The decision log provides the information needed to report the level selection in a transparent way.

The decision log can be included as a supplementary table in the publication. The log provides the evidence for the level selection and the rationale behind the decisions. The log also provides the information needed for the replication of the experiment by other researchers.

### The Decision Log and the Data Management

The decision log should be stored with the experimental data. The NIH Data Management and Sharing Policy describes the expectations for data management and sharing in NIH-funded research [<a href="#ref-3">3</a>]. The decision log is part of the data that describes how the data were generated.

The decision log should be stored in a format that is easy to read and share. The log should be stored in a text file or a spreadsheet. The log should include the date it was created and the version number. The log should be updated if the design is changed.

### The Decision Log and the Researcher Record

The decision log is part of the researcher record. The ORCID for researchers provides a way to maintain a record of the work and contributions [<a href="#ref-6">6</a>]. The decision log is a component of the research record that demonstrates the rigor of the research. The log can be shared with collaborators, reviewers, or funders as evidence of the quality of the research.

The decision log is not a publication in itself, but it is a component of the research record that supports the publication. The log provides the evidence for the level selection and the rationale behind the decisions. The log is a practical tool that helps you make defensible decisions about the factor levels in a full factorial experiment.

## Frequently Asked Questions

### How many levels should I use for each factor in a full factorial experiment?

Use two levels per factor for screening experiments where you need to identify which factors have an effect. Use three levels when you need to detect curvature in the response. Center points can provide curvature detection with fewer runs than a full three-level design.

### What is the difference between a replicate and a repeated measurement?

A replicate is an independent experimental unit that is run separately. A repeated measurement is the same experimental unit measured multiple times. Only true replicates provide an estimate of the experimental error that is needed for statistical tests.

### How do I choose the spacing between factor levels?

Base the spacing on the expected response difference and the measurement noise. The expected difference between the levels should be larger than the measurement noise. Use pilot data or prior literature to estimate the expected difference.

### What is a center point and why should I add it?

A center point is a run at the midpoint of all factors. It is used to detect curvature in the response. If the response at the center point differs from the midpoint of the two endpoints, the response is curved.

### What is blocking and when should I use it?

Blocking is a technique for assigning nuisance factors to groups of runs that are as similar as possible. Use blocking when you know that a nuisance factor, such as the day or the batch, affects the response. Blocking reduces the error and increases the power of the experiment.

### How many replicates do I need?

The number of replicates depends on the variability of the measurement and the minimum effect you need to detect. Use pilot data to estimate the variability and calculate the number of replicates needed to detect the minimum effect.

### What should I do if my pilot data show no response difference?

If the pilot data show no response difference across the proposed levels, the levels are too close together. Increase the spacing between the levels or use a wider range. If the response is still flat, the factor may not have an effect in the range you are studying.

### How do I document my level selection?

Record the rationale for each level selection before the experiment. Include the prior data or literature that supports the range, the minimum effect size, the estimated error, and the number of replicates. This documentation is important for the reproducibility of the experiment.

## Related Diagnostic Guides

- [Sanger Sequencing Steps: A Practical Guide for Accurate Results](/knowledge/diagnostics/molecular/sanger-sequencing-steps-a-practical-guide-for-accurate-results)
- [Sequencing Primer Design: Key Considerations for Reliable Results](/knowledge/diagnostics/molecular/sequencing-primer-design-key-considerations-for-reliable-results)
- [Biosafety Levels 1-4: A Comparative Framework for Diagnostic Laboratory Design](/knowledge/diagnostics/emerging-tech/biosafety-levels-1-4-a-comparative-framework-for-diagnostic-laboratory-design)

## References and Further Reading

<a id="ref-1"></a>[<a href="#ref-1">1</a>] [Research Methods Resources](https://www.ncbi.nlm.nih.gov/books). National Library of Medicine.

<a id="ref-2"></a>[<a href="#ref-2">2</a>] [EQUATOR Network](https://www.equator-network.org/). EQUATOR Network.

<a id="ref-3"></a>[<a href="#ref-3">3</a>] [Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy). National Institutes of Health.

<a id="ref-4"></a>[<a href="#ref-4">4</a>] [NIH Grants and Funding](https://grants.nih.gov/). National Institutes of Health.

<a id="ref-5"></a>[<a href="#ref-5">5</a>] [Core Practices](https://publicationethics.org/core-practices). Committee on Publication Ethics.

<a id="ref-6"></a>[<a href="#ref-6">6</a>] [ORCID for Researchers](https://info.orcid.org/researchers). ORCID.

> This article is educational and does not replace validated laboratory procedures, institutional biosafety review, manufacturer instructions, or professional interpretation.