# How to Interpret Posterior Distributions in Biological Studies ## Key Takeaways - Interpret posterior distributions by reporting the full uncertainty interval (credible interval) alongside the median, as biological effect sizes (e.g., fold changes in gene expression, population growth rates) vary across plausible parameter values, not just the average. - For skewed posteriors, common with variance or rate parameters (e.g., variance in animal growth, infection transmission rates), the posterior median is a more robust central summary than the mean, which can be unduly influenced by extreme values. - Report the probability of direction (e.g., proportion of posterior supporting a positive treatment effect on animal weight gain) to convey the consistency of an effect, but recognize it does not quantify effect magnitude or practical significance. - Avoid misinterpreting credible intervals as frequentist confidence intervals; a credible interval is a direct probability statement about the parameter's plausible range given the model and data, crucial for assessing if a treatment effect plausibly exceeds a minimum clinically important difference. - Connect posterior distributions to decision thresholds (e.g., minimum effective dose for a vaccine, threshold for intervention in a conservation study) by calculating the probability of exceeding that threshold, transforming statistical uncertainty into actionable biological or management insights. --- ## Quick Answer - Interpret posterior distributions by reporting the full uncertainty interval, beyond the mean, because biological effect sizes vary across credible parameter values. - Use the posterior median and credible interval as primary summaries when distributions are skewed, which commonly occurs with variance and rate parameters. - Report probability of direction alongside intervals to communicate the proportion of the posterior supporting a directional effect, but recognize it does not measure effect magnitude. ## At a Glance | Posterior Summary | What It Answers | Biological Use Case | Common Misinterpretation | |---|---|---|---| | Posterior Mean | What is the average parameter value under the model? | Estimating expected growth rate across a population | Treating the mean as the only value when the distribution is wide | | Posterior Median | What is the central value that splits the distribution in half? | Reporting survival probabilities with skewed distributions | Ignoring the median when the mean and median diverge | | Credible Interval | What range contains the parameter with stated probability? | Estimating the range of plausible enzyme kinetic constants | Confusing it with a frequentist confidence interval | | Probability of Direction | What proportion of the posterior supports a positive or negative effect? | Assessing whether a treatment consistently increases yield | Using it alone without reporting effect size or interval width | ## The Core Problem: Why Means Alone Mislead Biologists A Bayesian analysis produces a full probability distribution for each parameter of interest. That distribution encodes everything the model and data say about the plausible values of that parameter. When researchers reduce that distribution to a single number, they discard the information that makes Bayesian methods useful for biological decision making. Consider a field experiment measuring the effect of a feed additive on weight gain in broiler chickens. The posterior distribution for the additive effect might have a mean of 50 grams. If the distribution is narrow, the evidence strongly supports a gain near 50 grams. If the distribution is wide, the same mean could be consistent with a gain anywhere from 10 to 90 grams. The biological decision changes completely depending on which situation applies. A producer deciding whether to adopt the additive needs to know the range of plausible outcomes, beyond the average. The same logic applies across biological research. Gene expression studies estimate fold changes. Ecological studies estimate population growth rates. Clinical studies estimate treatment effects. In every case, the posterior distribution contains the full evidence about the parameter. Reporting only the mean is like describing a field by its average soil pH without noting that some plots are acidic and others are alkaline. The practical consequence of ignoring uncertainty is overconfident decisions. A researcher who sees a posterior mean above zero may conclude the effect is real. But if the credible interval includes zero, the data do not rule out the possibility that the effect is absent. The distinction between statistical evidence and biological certainty is central to interpreting posterior distributions correctly. ## What a Posterior Distribution Actually Represents A posterior distribution is the updated belief about a parameter after combining prior information with observed data. The prior distribution encodes what was known before the experiment. The likelihood describes how the data relate to the parameter. Bayes theorem combines these two components to produce the posterior. The posterior distribution is a full probability distribution. It assigns probability density to every possible value of the parameter. The shape of that distribution carries information. A narrow distribution indicates that the data strongly constrain the parameter. A wide distribution indicates that substantial uncertainty remains. The posterior distribution is not a single estimate. It is a complete description of what the model and data imply about the parameter. Any single-number summary of that distribution is a reduction. The reduction is useful only if it preserves the information needed for the decision at hand. For biological parameters, the posterior distribution often has a shape that is not symmetric. Variance parameters, rate parameters, and count parameters frequently produce skewed posteriors. In those cases, the mean and the median can differ substantially. The mean is pulled toward the tail of the distribution. The median is the value that splits the distribution into two equal halves. For a skewed posterior, the median is often a more representative central value than the mean. The posterior distribution also supports interval statements. A credible interval is a range of parameter values that contains a specified proportion of the posterior probability. A 95 percent credible interval contains 95 percent of the posterior probability. This is a direct probability statement about the parameter. It is not the same as a frequentist confidence interval, which has a different interpretation. ## Posterior Mean: The Average of the Distribution The posterior mean is the expected value of the parameter under the posterior distribution. It is the average of all parameter values weighted by their posterior probability. The mean is a natural summary when the posterior is symmetric and when the decision depends on the average outcome. In biological applications, the posterior mean is often used to estimate the expected effect size. For example, the mean of the posterior distribution for a treatment effect estimates the average effect across the population of interest. The mean is also used in cost-benefit calculations where the expected value matters. The mean has a limitation. It is sensitive to the tails of the distribution. If the posterior has a long tail toward large values, the mean is pulled in that direction. The mean may then be higher than the most probable value. The mean is also not a good summary when the posterior is multimodal, meaning it has two or more distinct peaks. In that case, no single central value describes the distribution well. For biological parameters that are constrained to be positive, such as growth rates or concentrations, the posterior is often skewed. The mean is pulled toward the upper tail. The median is a more robust central summary in that case. The mean is most useful when the posterior is approximately symmetric and when the decision depends on the average outcome. When the decision depends on the most probable value or on the range of plausible values, other summaries are more informative. ## Posterior Median: The Robust Central Value The posterior median is the value that divides the posterior distribution into two equal halves. Fifty percent of the posterior probability lies below the median, and fifty percent lies above it. The median is a robust summary because it is not sensitive to the tails of the distribution. For skewed posteriors, the median is often the preferred central summary. Consider a posterior for a variance parameter. Variance parameters are constrained to be positive and often have a long tail toward large values. The mean of such a distribution is pulled toward the tail, while the median is closer to the most probable values. The median is a better representation of the typical value. The median is also useful when the posterior is asymmetric for other reasons. For example, a posterior of a survival probability may be skewed when the data contain few events. The median provides a central value that is not distorted by the tail. The median is not always the best choice. When the posterior is symmetric, the mean and the median are approximately equal. In that case, the choice between them does not matter much. When the posterior is multimodal, the median may fall in a region of low probability between the peaks. In that case, the median is not a good summary. For biological reporting, the median is often paired with the credible interval. The median and the credible interval together describe the central value and the uncertainty. This pair is a common summary for parameters with skewed posteriors. ## Credible Intervals: The Range of Plausible Values A credible interval is a range of values that contains a specified proportion of the posterior probability. A 95 percent credible interval contains 95 percent of the posterior. The interval is a direct probability statement about the parameter. Given the model and the data, the parameter has a 95 percent probability of being within the interval. The credible interval is often confused with the frequentist confidence interval. The two have different meanings. A confidence interval is a statement about the procedure, not about the parameter. A 95 percent confidence interval means that the procedure produces intervals that contain the true value 95 percent of the time in repeated sampling. A credible interval is a statement about the parameter itself. The distinction matters for biological interpretation. The credible interval is computed from the posterior distribution. The most common type is the equal-tailed interval, which excludes the lower 2.5 percent and the upper 2.5 percent of the posterior. Another type is the highest posterior density interval, which is the narrowest interval that contains the specified probability. The highest posterior density interval is narrower than the equal-tailed interval when the posterior is skewed. The choice of interval type matters for skewed posteriors. The equal-tailed interval is easy to compute and interpret. The highest posterior density interval is more efficient in the sense that it is the shortest interval for the specified probability. For a symmetric posterior, the two intervals are the same. For a skewed posterior, they differ. The credible interval is used to assess whether a parameter is plausibly zero. If the credible interval excludes zero, the data support a nonzero effect. If the interval includes zero, the data do not exclude the possibility of no effect. This is a common use of the interval in biological studies. The width of the credible interval is a measure of uncertainty. A narrow interval indicates that the data strongly constrain the parameter. A wide interval indicates that the data are not very informative. The width depends on the sample size, the variability of the data, and the prior. ## Probability of Direction: The Proportion of the Posterior on One Side The probability of direction is the proportion of the posterior distribution that lies on a specified side of a threshold. The most common threshold is zero. The probability of direction is then the proportion of the posterior that is positive or the proportion that is negative. The probability of direction is a measure of the consistency of the effect. A probability of direction of 0.95 means that 95 percent of the posterior supports a positive effect. The remaining 5 percent supports a negative effect. The probability of direction is a measure of the evidence for the sign of the effect. The probability of direction is not a measure of the magnitude of the effect. A probability of direction of 0.99 can be associated with a tiny effect or a large effect. The probability of direction does not distinguish between the two. The probability of direction is also not a measure of the practical importance of the effect. The probability of direction is related to the credible interval. If the credible interval excludes zero, the probability of direction is high. If the credible interval includes zero, the probability of direction is lower. The probability of direction is a continuous measure, while the credible interval is a binary statement about whether zero is included. The probability of direction is useful for reporting the consistency of an effect. It is often reported alongside the credible interval and the median. The combination of the median, the credible interval, and the probability of direction provides a fuller description of the posterior than any single summary. The probability of direction has a limitation. It does not account for the magnitude of the effect. A probability of direction of 0.95 for a tiny effect is not the same as a probability of direction of 0.95 for a large effect. The probability of direction should be reported with the effect size and the credible interval. ## Reporting Posterior Summaries in Biological Papers The reporting of posterior summaries in biological papers should follow the same principles as the reporting of any statistical analysis. The methods should describe the model, the prior, and the computational approach. The results should report the posterior summaries with enough detail for the reader to interpret the evidence. The minimum reporting standard is the median and the credible interval for each parameter of interest. The median is the central summary and the credible interval is the uncertainty summary. The probability of direction can be reported as an additional summary of the consistency of the effect. The reporting should also include the prior distribution. The prior is a component of the model and the posterior depends on it. The prior should be described in the methods. The sensitivity of the results to the prior should be assessed when the prior is informative. The reporting should follow the relevant reporting guidelines. The [EQUATOR Network](https://www.equator-network.org/) provides a list of reporting guidelines for different study types. The use of a reporting guideline improves the transparency of the methods and the results. The guideline should be selected based on the study design. The reporting should also follow the publication ethics standards. The [Committee on Publication Ethics](https://publicationethics.org/core-practices) core practices cover authorship, peer review, data, conflicts, and misconduct. The reporting of the statistical analysis is part of the responsible conduct of research. The data and the code should be shared when possible. The [NIH Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy) describes the expectations for data sharing. The sharing of data and code allows other researchers to reproduce the analysis. The reproducibility of the analysis is part of the transparency of the research. ## Practical Workflow for Interpreting a Posterior Distribution The interpretation of a posterior distribution should follow a structured workflow. The workflow ensures that the interpretation is complete and that the limitations are considered. The first step is to examine the full posterior distribution. The distribution should be plotted. The plot shows the shape of the distribution, including any skewness or multimodality. The plot is the most informative summary of the posterior. The second step is to compute the posterior median and the credible interval. The median is the central summary and the credible interval is the uncertainty. The median and the credible interval are the standard summaries for reporting. The third step is to compute the probability of direction. The probability of direction is the proportion of the posterior on the relevant side of the threshold. The probability of direction is a measure of the consistency of the effect. The fourth step is to interpret the summaries in the context of the biological question. The median is the estimate of the effect. The credible interval is the range of plausible values. The probability of direction is the consistency of the effect. The interpretation should consider the magnitude of the effect and the uncertainty. The fifth step is to assess the sensitivity of the results to the prior. The prior is part of the model. The results should be robust to reasonable changes in the prior. The sensitivity analysis is part of the assessment of the robustness of the results. The sixth step is to report the results. The report should include the posterior summaries, the prior, and the sensitivity analysis. The report should follow the relevant reporting guideline. ## Records and Measurements for Posterior Interpretation The interpretation of a posterior distribution should be documented in the analysis records. The records should include the model specification, the prior, the data, and the posterior summaries. The records allow the analysis to be reproduced and the interpretation to be checked. The model specification should include the likelihood and the prior. The likelihood describes the relationship between the data and the parameters. The prior describes the knowledge before the data. The model specification should be recorded in the methods. The data should be recorded in a format that allows the analysis to be reproduced. The data should include the variables, the units, and the missing values. The data should be stored in a repository when possible. The posterior summaries should be recorded for each parameter of interest. The summaries should include the median, the credible interval, and the probability of direction. The summaries should be recorded with the units and the interpretation. The records should also include the software and the version of the software. The software and the version affect the results. The records should include the random seed if the analysis uses simulation. The records should be stored in a way that allows the analysis to be reproduced. The records should be shared when possible. The sharing of the records allows the analysis to be checked by other researchers. ## Common Failure Patterns in Interpreting Posterior Distributions The interpretation of posterior distributions has several common failure patterns. The patterns are mistakes that lead to incorrect conclusions. The patterns should be avoided. The first failure pattern is reporting only the posterior mean. The mean is a summary of the distribution, but it does not describe the uncertainty. The mean alone does not allow the reader to assess the range of plausible values. The mean should be reported with the credible interval. The second failure pattern is treating the credible interval as a confidence interval. The two intervals have different meanings. The credible interval is a probability statement about the parameter. The confidence interval is a statement about the procedure. The confusion leads to incorrect interpretations. The third failure pattern is using the probability of direction as a measure of the magnitude of the effect. The probability of direction is a measure of the consistency of the effect. The probability of direction does not describe the magnitude of the effect. The probability of direction should be reported with the median and the credible interval. The fourth failure pattern is ignoring the prior. The prior is part of the model. The posterior depends on the prior. The prior should be reported and the sensitivity to the prior should be assessed. The fifth failure pattern is overinterpreting the results. The posterior is a description of the parameter under the model and the data. The posterior is not a description of the true value of the parameter. The posterior is conditional on the model and the prior. The sixth failure pattern is not reporting the model. The model is the basis of the posterior. The model should be reported in the methods. The model should be described in enough detail for the reader to understand the analysis. ## Limitations of Posterior Summaries The posterior summaries have limitations that should be considered. The limitations are part of the interpretation of the results. The posterior mean is sensitive to the tails of the distribution. The mean is pulled toward the extreme values. The mean is not a robust summary for skewed distributions. The median is a more robust choice for skewed distributions. The credible interval is a statement about the posterior. The credible interval is not a statement about the true value of the parameter. The credible interval is conditional on the model and the data. The credible interval is not a guarantee that the parameter is in the interval. The probability of direction is a measure of the consistency of the effect. The probability of direction does not describe the magnitude of the effect. The probability of direction is not a measure of the practical importance of the effect. The posterior summaries are conditional on the model. The model is a simplification of the biological process. The model may not capture all the relevant aspects of the process. The posterior summaries are conditional on the model. The posterior summaries are conditional on the data. The data are a sample from the population. The data may not be representative of the population. The posterior summaries are conditional on the data. The posterior summaries are conditional on the prior. The prior is the knowledge before the data. The prior may not be accurate. The posterior summaries are conditional on the prior. ## Safety and Regulatory Context The interpretation of posterior distributions has a safety and regulatory context in biological studies. The context applies to studies that inform decisions about health, safety, and the environment. The context is part of the responsible conduct of the research. The [National Institutes of Health](https://grants.nih.gov/) provides the policy for the conduct of research. The policy covers the application, the review, and the award management. The policy also covers the conduct of the research. The interpretation of the posterior is part of the conduct of the research. The [Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy) describes the expectations for data sharing. The policy applies to the data that support the research. The data should be shared to allow the analysis to be reproduced. The sharing of the data is part of the transparency of the research. The [Committee on Publication Ethics](https://publicationethics.org/core-practices) describes the core practices for the publication of the research. The practices cover the authorship, the peer review, the data, the conflicts, and the misconduct. The interpretation of the posterior is part of the reporting of the research. The reporting should be transparent and accurate. The [ORCID](https://info.orcid.org/researchers) provides the researcher identity. The identity is used to link the researcher to the research. The identity is used to the research record. The identity is part of the transparency of the research. The [Research Methods Resources](https://www.ncbi.nlm.nih.gov/books) provides the reference for the research methods. The reference covers the methods for the research. The methods include the statistical analysis. The interpretation of the posterior is part of the statistical analysis. ## Professional Escalation Criteria The interpretation of a posterior distribution may require the escalation to a professional in some situations. The escalation is appropriate when the interpretation is beyond the expertise of the researcher. The escalation is appropriate when the results are consequential. The escalation is appropriate when the posterior is not well-behaved. The posterior may be multimodal or have a heavy tail. The interpretation of the posterior may require the expertise of a statistician. The escalation is appropriate when the posterior is not well-behaved. The escalation is appropriate when the prior is not well-justified. The prior is part of the model. The prior should be justified. The escalation is appropriate when the prior is not well-justified. The escalation is appropriate when the results are consequential. The results may inform a decision about the health, the safety, or the environment. The escalation is appropriate when the results are consequential. The escalation is appropriate when the results are used to make a decision. The escalation is appropriate when the results are not reproducible. The results may not be reproducible because of the data or the code. The escalation is appropriate when the results are not reproducible. The escalation is appropriate when the results are not reproducible. The escalation is appropriate when the results are not consistent with the biological knowledge. The results may not be consistent with the biological knowledge. The escalation is appropriate when the results are not consistent with the biological knowledge. The escalation is appropriate when the results are not consistent with the biological knowledge. ## Decision Thresholds: Converting Posterior Uncertainty into Management Actions A posterior distribution becomes practically useful only when it is connected to a decision threshold. In biological research, decisions rarely hinge on whether a parameter is exactly zero. They hinge on whether an effect is large enough to justify a change in practice, a further investment, or a regulatory action. A posterior mean of 50 grams for a feed additive effect is not actionable by itself. The actionable question is whether the plausible range of effects, as described by the full posterior, supports a decision to adopt the additive, to run another trial, or to abandon the line of inquiry. The decision threshold approach requires the researcher to define, before examining the posterior, the effect size that would change a decision. This threshold is a biological or economic judgment, not a statistical one. For a growth trial, the threshold might be the minimum weight gain that pays for the cost of the additive. For a conservation study, the threshold might be the minimum population growth rate that avoids a management intervention. For a clinical study, the threshold might be the minimum improvement in a biomarker that justifies a treatment recommendation. The threshold converts the posterior distribution from a description of uncertainty into a tool for action. ### Defining the Decision Threshold Before Analysis The decision threshold must be defined before the posterior is examined. Defining the threshold after seeing the results invites the researcher to choose a threshold that makes the results look favorable. This is a form of post hoc decision making that undermines the credibility of the analysis. The threshold should be recorded in the analysis plan, along with the rationale for the chosen value. The threshold should be expressed in the same units as the parameter of interest. If the parameter is a weight gain in grams, the threshold is a number of grams. If the parameter is a survival probability, the threshold is a probability. If the parameter is a gene expression fold change, the threshold is a fold change. The threshold should be a single value or a range of values that separates the decision space into distinct actions. The threshold should be justified by the biological or economic context. The justification should be recorded. For example, a threshold for a feed additive might be based on the cost of the additive and the value of the weight gain. A threshold for a conservation intervention might be based on the population viability analysis. A threshold for a clinical biomarker might be based on the minimal important difference established in the literature. The justification makes the threshold transparent and defensible. The threshold should be distinguished from the statistical null value of zero. The null value is the value that would indicate no effect. The decision threshold is the value that would indicate a practically important effect. These two values are often different. A posterior that excludes zero may still have most of its mass below the decision threshold. In that case, the data support the existence of an effect but not an effect large enough to change a decision. The distinction between statistical evidence and practical importance is central to the decision framework. ### The Posterior Probability of Exceeding the Threshold Once the decision threshold is defined, the posterior distribution can be used to compute the probability that the parameter exceeds the threshold. This probability is the proportion of the posterior distribution that lies above the threshold. It is a direct statement about the evidence for a practically important effect. The probability of exceeding the threshold is computed from the posterior samples. If the posterior is represented by a set of samples, the probability is the proportion of samples that are greater than the threshold. If the posterior is represented by a parametric distribution, the probability is the area under the distribution above the threshold. The computation is straightforward and should be reported with the same precision as the other posterior summaries. The probability of exceeding the threshold is different from the probability of direction. The probability of direction is the proportion of the posterior on one side of zero. The probability of exceeding the threshold is the proportion of the posterior on one side of the decision threshold. The two probabilities are equal only when the threshold is zero. When the threshold is positive, the probability of exceeding the threshold is smaller than the probability of direction. When the threshold is negative, the probability of exceeding the threshold is larger. The probability of exceeding the threshold is also different from the credible interval. The credible interval describes the range of plausible values. The probability of exceeding the threshold describes the evidence for a specific action. A wide credible interval can be associated with a high or low probability of exceeding the threshold, depending on where the interval sits relative to the threshold. ### The Decision Rule The decision rule connects the probability of exceeding the threshold to a specific action. The rule should be defined before the analysis. The rule should specify the probability level that triggers each action. The rule should also specify the action to be taken when the probability falls in the intermediate range. A common decision rule uses two probability levels. If the probability of exceeding the threshold is above a high level, such as 0.95, the evidence supports adopting the effect. If the probability of exceeding the threshold is below a low level, such as 0.05, the evidence supports rejecting the effect. If the probability is between the two levels, the evidence is insufficient and the decision is to collect more data or to defer the decision. The choice of the probability levels depends on the consequences of the decision. If the cost of a false positive is high, the high level should be set higher. If the cost of a false negative is high, the low level should be set lower. The levels should be justified and recorded in the analysis plan. The decision rule should be applied to the posterior distribution, not to the point estimate. A posterior mean that is above the threshold does not guarantee that the probability of exceeding the threshold is high. The posterior mean is the average of the distribution. The probability of exceeding the threshold depends on the shape of the distribution and the location of the threshold relative to the distribution. ### Worked Example with a Growth Trial Consider a field trial of a feed additive in broiler chickens. The parameter of interest is the average weight gain attributable to the additive. The researcher defines the decision threshold as 40 grams. A gain of less than 40 grams does not justify the cost of the additive. A gain of 40 grams or more justifies the cost. The posterior distribution for the weight gain has a mean of 50 grams and a 95 percent credible interval from 20 to 80 grams. The probability of direction is 0.99, meaning that 99 percent of the posterior supports a positive gain. The probability of exceeding the threshold of 40 grams is 0.65, meaning that 65 percent of the posterior supports a gain of at least 40 grams. The interpretation is different from the interpretation based on the mean alone. The mean is exactly at the threshold. The credible interval includes values below the threshold. The probability of exceeding the threshold is 0.65, which is below the high level of 0.95. The decision rule would not support adopting the additive based on this evidence. The evidence is insufficient to justify the cost. Now consider a second scenario. The posterior distribution has a mean of 40 grams and a 95 percent credible interval from 30 to 50 grams. The probability of direction is 0.99. The probability of exceeding the threshold of 40 grams is 0.50. The decision rule would still not support adopting the additive, because the probability of exceeding the threshold is only 0.50. Now consider a third scenario. The posterior distribution has a mean of 45 grams and a 95 percent credible interval from 35 to 55 grams. The probability of exceeding the threshold of 40 grams is 0.85. The decision rule with a high level of 0.95 would still not support adopting the additive. The decision rule with a high level of 0.80 would support adopting the additive. The choice of the probability level changes the decision. The example shows that the posterior mean alone does not determine the decision. The decision depends on the threshold, the probability of exceeding the threshold, and the probability levels in the decision rule. The decision framework makes the connection between the posterior and the action explicit. ### Recording the Decision Framework The decision framework should be recorded in the analysis records. The records should include the threshold, the justification for the threshold, the probability levels, and the decision rule. The records should also include the posterior distribution and the computed probability of exceeding the threshold. The records should be stored with the data and the code. The records allow the decision to be reproduced and the reasoning to be checked. The records should be shared when possible. The sharing of the records is part of the transparency of the research. The [NIH Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy) describes the expectations for data sharing. The policy applies to the data that support the research. The decision records are part of the data that support the research. The records should be shared to allow the analysis to be reproduced. The decision framework should also be reported in the paper. The methods should describe the threshold, the probability levels, and the decision rule. The results should report the probability of exceeding the threshold for each parameter of interest. The discussion should interpret the decision in the context of the biological question. ### Common Failure Patterns in Decision Thresholds The decision framework has several common failure patterns. The patterns are mistakes that lead to incorrect decisions. The patterns should be avoided. The first failure pattern is defining the threshold after the analysis. The threshold is chosen to make the results look favorable. The threshold should be defined before the analysis. The threshold should be recorded in the analysis plan. The second failure pattern is using the probability of direction as the decision probability. The probability of direction is the proportion of the posterior on one side of zero. The decision probability is the proportion of the posterior on one side of the threshold. The two probabilities are different when the threshold is not zero. The decision probability should be computed and reported. The third failure pattern is using the credible interval as the decision rule. The credible interval describes the range of plausible values. The decision rule uses the probability of exceeding the threshold. The credible interval does not directly give the probability of exceeding the threshold. The probability should be computed from the posterior. The fourth failure pattern is ignoring the probability levels. The decision rule requires the probability levels to be defined. The levels should be defined before the analysis. The levels should be recorded in the analysis plan. The fifth failure pattern is not recording the decision. The decision should be recorded in the analysis records. The records should include the threshold, the probability levels, and the decision rule. The records should be shared when possible. ### Professional Escalation Criteria for Decision Thresholds The decision framework may require the escalation to a professional in some situations. The escalation is appropriate when the threshold is not well-justified. The threshold should be justified by the biological or economic context. The escalation is appropriate when the threshold is not well-justified. The escalation is appropriate when the probability levels are not well-justified. The probability levels should be justified by the costs of the false positive and the false negative. The escalation is appropriate when the probability levels are not well-justified. The escalation is appropriate when the decision is consequential. The decision may inform a decision about the health, the safety, or the environment. The escalation is appropriate when the decision is consequential. The escalation is appropriate when the decision is used to make a decision. The escalation is appropriate when the posterior is not well-behaved. The posterior may be multimodal or have a heavy tail. The interpretation of the posterior may require the expertise of a statistician. The escalation is appropriate when the posterior is not well-behaved. The escalation is appropriate when the results are not reproducible. The results may not be reproducible because of the data or the code. The escalation is appropriate when the results are not reproducible. The escalation is appropriate when the results are not reproducible. The escalation is appropriate when the results are not consistent with the biological knowledge. The results may not be consistent with the biological knowledge. The escalation is appropriate when the results are not consistent with the biological knowledge. The escalation is appropriate when the results are not consistent with the biological knowledge. ## Frequently Asked Questions ### What is the difference between a posterior mean and a posterior median? The posterior mean is the expected value of the parameter under the posterior distribution. The posterior median is the value that splits the posterior into two equal halves. The mean is sensitive to the tails of the distribution. The median is not sensitive to the tails. For a skewed posterior, the median is often the more representative central value. ### How do I interpret a 95 percent credible interval? A 95 percent credible interval contains 95 percent of the posterior probability. Given the model and the data, the parameter has a 95 percent probability of being in the interval. The credible interval is a direct statement about the parameter. The credible interval is not the same as a frequentist confidence interval. ### What is the probability of direction in a Bayesian analysis? The probability of direction is the proportion of the posterior distribution that lies on a specified side of a threshold. The most common threshold is zero. The probability of direction is a measure of the consistency of the effect. The probability of direction does not describe the magnitude of the effect. ### Should I report the posterior mean or the posterior median? The choice depends on the shape of the posterior. The median is more robust for skewed distributions. The mean is appropriate for approximately symmetric distributions. The median and the credible interval are a common summary for skewed posteriors. The mean and the credible interval are a common summary for symmetric posteriors. ### How do I choose the prior for a Bayesian analysis? The prior should reflect the knowledge before the data. The prior should be justified in the methods. The sensitivity of the results to the prior should be assessed. The prior is part of the model. The posterior depends on the prior. ### What is the difference between a credible interval and a confidence interval? A credible interval is a statement about the parameter. A 95 percent credible interval contains 95 percent of the posterior probability. A confidence interval is a statement about the procedure. A 95 percent confidence interval contains the parameter in 95 percent of the repeated samples. The two intervals have different meanings. ### How do I report the posterior summaries in a paper? The report should include the median and the credible interval for each parameter of interest. The report should include the probability of direction when relevant. The report should include the prior and the model. The report should follow the relevant reporting guideline. ### What should I do if the posterior is not well-behaved? The posterior may be multimodal or have a heavy tail. The interpretation of the posterior may require the expertise of a statistician. The escalation is appropriate when the posterior is not well-behaved. The escalation is appropriate when the results are consequential. ## Using the Evidence | Source | Best use in this topic | Important limitation | |---|---|---| | [Research Methods Resources](https://www.ncbi.nlm.nih.gov/books) | official guidance | Check the linked page for current local requirements | | [EQUATOR Network](https://www.equator-network.org/) | official guidance | Check the linked page for current local requirements | | [Core Practices](https://publicationethics.org/core-practices) | official guidance | Check the linked page for current local requirements | ## Related Bioinformatics Guides - [Spatial Transcriptomics Data Analysis: A Practical Workflow from Raw Data to Biological Insights](/knowledge/bioinformatics/spatial-transcriptomics-data-analysis-a-practical-workflow-from-raw-data-to-biological-insights) - [Gene Set Enrichment Analysis in R: A Practical Tutorial for Interpreting Omics Data](/knowledge/bioinformatics/gene-set-enrichment-analysis-in-r-a-practical-tutorial-for-interpreting-omics-data) - [Metabolomics Data Analysis in R: A Practical Workflow](/knowledge/bioinformatics/metabolomics-data-analysis-in-r-a-practical-workflow) - [Microbiome Data Analysis in R: A Practical Guide for Compositional Data](/knowledge/bioinformatics/microbiome-data-analysis-in-r-a-practical-guide-for-compositional-data) - [How to Interpret Gene Set Enrichment Analysis Results](/knowledge/bioinformatics/how-to-interpret-gene-set-enrichment-analysis-results) ## Related Clinical & Scientific Guides * [A Practical Guide to Detecting Antimicrobial Resistance Genes in Shotgun Metagenomic Data](/knowledge/bioinformatics/a-practical-guide-to-detecting-antimicrobial-resistance-genes-in-shotgun-metagenomic-data) * [Computational Immunology: Modeling the Immune System](/knowledge/bioinformatics/computational-immunology-modeling-the-immune-system) * [How to Set Hard Filters for Germline Variant Calling: A Practical Guide to GATK Best Practices](/knowledge/bioinformatics/how-to-set-hard-filters-for-germline-variant-calling-a-practical-guide-to-gatk-best-practices) ## References and Further Reading - [Research Methods Resources](https://www.ncbi.nlm.nih.gov/books). National Library of Medicine. - [EQUATOR Network](https://www.equator-network.org/). EQUATOR Network. - [Core Practices](https://publicationethics.org/core-practices). Committee on Publication Ethics. - [NIH Grants and Funding](https://grants.nih.gov/). National Institutes of Health. - [ORCID for Researchers](https://info.orcid.org/researchers). ORCID. - [Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy). National Institutes of Health. - [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information. - [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute. - [Global age-sex-specific all-cause mortality and life expectancy estimates for 204 countries and territories and 660 subnational locations, 1950-2023: a demographic analysis for the Global Burden of Disease Study 2023.](https://pubmed.ncbi.nlm.nih.gov/41092927). Lancet (London, England), 2025. - [Global burden of enteric infectious diseases, diarrhoeal diseases, and corresponding aetiologies, 1990-2023: a systematic analysis for the Global Burden of Disease Study 2023.](https://pubmed.ncbi.nlm.nih.gov/42229499). The Lancet. Infectious diseases, 2026. - [Measuring strain in the exoskeleton of spiders-virtues and caveats.](https://pubmed.ncbi.nlm.nih.gov/33459819). Journal of comparative physiology. A, Neuroethology, sensory, neural, and behavioral physiology, 2021. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.