4. Statistical Tools for Research
Learning Objectives
- Explain the roles of probability, hypothesis testing, and confidence intervals in biotechnology research
- Distinguish a null hypothesis from an alternative hypothesis and interpret a p-value correctly
- Identify which statistical tool fits common biotechnology tasks: sequence alignment, differential expression, GWAS
- Explain why correlation does not imply causation, using a biological example
- Evaluate a research claim by checking whether statistical significance and biological significance were both addressed
Quick Answer
Statistical tools are what let biotechnology researchers tell the difference between a genuine biological signal and random noise in their data. Concepts like probability, hypothesis testing, confidence intervals, and p-values form the foundation, while specialized software (DESeq2 for gene expression, PLINK for genome-wide association studies, BLAST for sequence comparison) applies these principles to specific types of biological data. Without statistics, a researcher looking at thousands of genes or millions of DNA bases would have no rigorous way to decide which differences are real and which are just the kind of variation you'd expect by chance — which is exactly why every credible biotechnology paper reports its statistical methods alongside its biological conclusions.
Why Statistics Is Unavoidable in Biotechnology
Modern biotechnology generates enormous datasets — a single RNA-seq experiment might measure the expression of 20,000 genes simultaneously. If you test all 20,000 genes for a "significant" difference between two conditions, some will look significant purely by chance, even if nothing biologically real is happening. Statistics gives researchers the tools to quantify that risk and control for it, rather than being fooled by patterns that don't actually exist.
Why It Matters
Without statistical rigor, a researcher could easily convince themselves that a random fluctuation is a breakthrough discovery. Statistics is the discipline that keeps enthusiasm honest.
Core Statistical Concepts
Probability Theory
Probability assigns a numerical likelihood to an event, forming the mathematical basis for everything that follows. If a coin is fair, the probability of heads is 0.5 — a simple example, but the same logic underlies far more complex calculations, like the probability of observing a given number of sequencing reads mapping to a gene by chance alone.
Hypothesis Testing
Hypothesis testing is a formal procedure for deciding whether observed data is consistent with a "no effect" scenario. It starts with a null hypothesis (H0) — the default assumption that there is no real difference or effect — and an alternative hypothesis (H1) — the claim being tested. For example, testing whether a drug affects average cell viability: H0 states the drug has no effect (mean viability is unchanged); H1 states it does. The test produces a p-value: the probability of seeing data this extreme (or more extreme) if H0 were actually true. A small p-value (conventionally < 0.05) is treated as evidence against H0.
Confidence Intervals
A confidence interval gives a plausible range for a population parameter, rather than a single point estimate, accounting for sampling variability. If a study measures a mean protein concentration of 180 units with a 95% confidence interval of (170, 190), it means: if the study were repeated many times, about 95% of such intervals would contain the true population mean.
Correlation vs. Causation
Two variables can move together (correlate) without one causing the other — a classic pitfall in observational biotechnology data. A gene's expression might correlate with disease severity simply because both are driven by a third, unmeasured factor (a confounder), not because the gene causes the disease.
Why It Matters
Reporting only "p < 0.05" without a confidence interval or effect size hides how precise or how large the finding actually is. Two studies can both report p < 0.05, but one might show a huge, clinically important effect while the other shows a tiny, practically meaningless one — the confidence interval and effect size are what reveal this difference.
Common Misunderstanding
A p-value is often misread as "the probability the null hypothesis is true." It is not — it's the probability of the observed data (or more extreme) given that the null hypothesis is true. This subtle distinction matters: p = 0.03 does not mean there's a 3% chance the drug has no effect; it means data this extreme would occur 3% of the time if the drug truly had no effect.
Common Statistical Tools in Biotechnology
- Sequence alignment: BLAST and Clustal Omega compare DNA/protein sequences for similarity, underpinning everything from species identification to variant calling.
- Differential expression analysis: DESeq2 and edgeR apply specialized statistical models suited to RNA-seq count data, correcting for the fact that raw counts don't follow a simple normal distribution.
- Genome-wide association studies (GWAS): Tools like PLINK test millions of genetic variants for statistical association with a trait or disease, requiring careful correction for testing so many hypotheses at once.
- Phylogenetic analysis: RAxML and MrBayes use statistical models of sequence evolution to infer evolutionary trees from genetic data.
- Structure and function prediction: Tools like AlphaFold rely on statistical/machine-learning models trained on known protein structures to predict new ones.
Real-World Example
A genome-wide association study (GWAS) testing for genetic variants linked to heart disease might test a million SNPs simultaneously. Because testing that many hypotheses inflates the chance of false positives, researchers apply a stricter significance threshold (multiple-testing correction, such as Bonferroni correction or false discovery rate control) before declaring any SNP a real hit — then annotate the significant variants and run pathway enrichment analysis to interpret what genes and biological processes are implicated.
Key Terms
| Term | Definition | Related Concept |
|---|---|---|
| Null Hypothesis (H0) | The default assumption of no real effect or difference | Alternative Hypothesis |
| Alternative Hypothesis (H1) | The claim being tested, stating a real effect or difference exists | Null Hypothesis |
| P-value | The probability of observing data this extreme (or more) if the null hypothesis were true | Significance Level |
| Confidence Interval | A range of plausible values for a population parameter, given sampling variability | P-value |
| Multiple-Testing Correction | A statistical adjustment applied when many hypotheses are tested at once, to control false positives | GWAS, Bonferroni Correction |
| Correlation | A statistical association between two variables that move together | Causation, Confounding Variable |
| Confounding Variable | An unmeasured factor that influences both variables in an observed correlation, creating a misleading association | Correlation |
| Effect Size | A measure of the magnitude of a difference or relationship, independent of sample size | P-value, Fold-Change |
Common Mistakes
Misconception: A p-value tells you the probability that the null hypothesis is true. Why it's wrong: A p-value is calculated assuming the null hypothesis is true — it describes how surprising the observed data would be under that assumption, not the probability that the assumption itself is correct. Correct understanding: A small p-value means the observed data would be unlikely if there were truly no effect, providing evidence against H0 — but it is not a direct probability statement about H0 itself.
Misconception: If two variables are strongly correlated in biological data, one must be causing the other. Why it's wrong: A correlation can arise from a third, unmeasured confounding variable that independently affects both, or simply from coincidence in a large enough dataset. Correct understanding: Correlation only justifies further investigation (e.g., a controlled experiment); establishing causation requires an experimental design with proper controls, not just an observed statistical association.
Misconception: Testing thousands of genes for significance and using the standard p < 0.05 threshold for each one is a valid way to find real differentially expressed genes. Why it's wrong: With 20,000 genes tested at p < 0.05 each, roughly 1,000 genes would appear "significant" purely by chance even if nothing biological were happening, because the 5% false-positive rate applies per test. Correct understanding: Multiple-testing correction (e.g., false discovery rate control via the Benjamini-Hochberg procedure) must be applied when testing many hypotheses simultaneously, to keep the overall false-positive rate under control.
Comparison and Connections
| Concept | What It Tells You | What It Doesn't Tell You |
|---|---|---|
| P-value | How surprising the data would be if H0 were true | Whether the effect is large or biologically meaningful |
| Confidence Interval | A plausible range for the true effect size | Whether the effect is statistically significant on its own (though overlap with zero suggests non-significance) |
| Effect Size (e.g., fold-change) | The magnitude of the observed difference | Whether that magnitude could have arisen by chance |
| Correlation | Whether two variables move together | Whether one variable causes the other |
Practice Questions
Recall
-
Define the null hypothesis and the alternative hypothesis. Look for: null hypothesis (H0) assumes no real effect/difference exists; alternative hypothesis (H1) is the claim being tested that a real effect/difference does exist.
-
What does a 95% confidence interval of (170, 190) for a mean protein concentration actually mean? Look for: if the study were repeated many times, about 95% of the resulting confidence intervals would contain the true population mean — it is not "95% probability the true value is in this specific interval."
Understanding
-
Explain why testing 20,000 genes at a standard p < 0.05 threshold without correction is statistically problematic. Look for: each test has a 5% chance of a false positive under H0, so testing 20,000 genes would yield roughly 1,000 false positives by chance alone; multiple-testing correction is needed to control this.
-
Why can't an observed correlation between a gene's expression and a disease's severity, by itself, prove the gene causes the disease? Look for: an unmeasured confounding variable could independently affect both the gene's expression and disease severity, creating an association without direct causation; only a controlled experiment can more confidently establish causation.
Application
-
A researcher finds p = 0.001 for a gene's expression difference between healthy and diseased tissue, but the actual expression change is only 1.05-fold. How should they interpret this? Look for: the result is statistically significant but the effect size is tiny and likely not biologically meaningful; large sample sizes can make trivial differences statistically significant, so the fold-change should be considered alongside the p-value.
-
A GWAS tests 500,000 SNPs for association with a trait. Explain what correction the researcher should apply and why. Look for: multiple-testing correction (e.g., Bonferroni correction, adjusting the significance threshold to roughly 0.05/500,000, or false discovery rate control), because testing that many hypotheses would otherwise produce a large number of false-positive associations by chance.
Analysis
-
Two RNA-seq studies both report p < 0.05 for the same gene, but Study A has a 95% CI of (1.1, 1.3)-fold change and Study B has a 95% CI of (1.1, 8.5)-fold change. Which result would you trust more, and why? Look for: Study A's result is more precise (narrower confidence interval), suggesting a more reliable estimate of the true effect; Study B's wide interval suggests high uncertainty, possibly due to a smaller sample size or more variable data, so its point estimate is less trustworthy even though it's still "significant."
-
A news article claims "eating food X causes cancer" based on a study that found people who ate more of food X had higher cancer rates, in an observational survey. Critique this claim using the concepts from this page. Look for: this is a correlational, observational finding, not causal; the study cannot rule out confounding variables (e.g., people who eat more of food X might also smoke more, exercise less, or differ in age); establishing causation would require a controlled experimental design, which is often unethical or impossible for something like diet and cancer, so causal language is unjustified from this evidence alone.
FAQ
Q: Why is 0.05 the "standard" significance threshold, and is it always the right choice? It's a historical convention, not a universal law of statistics. In fields testing many hypotheses at once (like GWAS), a much stricter threshold is used to control false positives. Some researchers argue 0.05 is too lenient for high-stakes claims and recommend reporting exact p-values and effect sizes rather than a strict yes/no cutoff.
Q: What's the difference between statistical significance and biological/clinical significance? Statistical significance means the observed difference is unlikely to be due to chance alone, given the sample size. Biological or clinical significance means the size of that difference actually matters in a real-world sense. A study can have one without the other — especially with very large sample sizes, where even trivial differences become statistically significant.
Q: Why do bioinformatics tools like DESeq2 use specialized statistical models instead of a standard t-test? RNA-seq count data doesn't follow a normal distribution — it's discrete, often has many low counts, and its variance changes with its mean (a property called overdispersion). DESeq2 and edgeR use models (like the negative binomial distribution) specifically suited to this kind of count data, giving more accurate p-values than a standard t-test would.
Q: If a result isn't statistically significant, does that mean there's definitely no effect? No. "Failing to reject the null hypothesis" simply means the study didn't find sufficient evidence of an effect — this could be because there truly is no effect, or because the sample size was too small to detect a real but modest effect (low statistical power).
Q: How do I know which statistical test to use for my data? It depends on your data type and question: comparing two group means typically calls for a t-test; comparing more than two groups calls for ANOVA; testing association between two categorical variables calls for a chi-square test; and specialized biological data (like RNA-seq counts) often requires purpose-built tools rather than classical tests.
Quick Revision
- Probability theory underlies all statistical inference — it assigns numerical likelihoods to outcomes.
- Hypothesis testing compares a null hypothesis (no effect) against an alternative (an effect exists), producing a p-value.
- A p-value is the probability of the observed data (or more extreme) given H0 is true — not the probability that H0 itself is true.
- A confidence interval gives a plausible range for the true parameter, and its width reflects the precision of the estimate.
- Correlation does not imply causation — confounding variables can create an association without direct cause and effect.
- Testing many hypotheses at once (e.g., 20,000 genes, a GWAS with millions of SNPs) requires multiple-testing correction to control false positives.
- Statistical significance (p-value) and biological/clinical significance (effect size) are distinct — both should be reported and interpreted together.
- Specialized tools (DESeq2, edgeR, PLINK) apply statistical models tailored to specific biological data types, rather than generic tests.
- Failing to reject the null hypothesis does not prove there is no effect — it may just reflect insufficient statistical power.
- Choosing the right statistical test depends on the data type and the specific question being asked.
Related Topics
Prerequisites: Data Collection and Analysis, Basic probability and statistics
Related Topics: Research Design and Planning, Writing Research Papers
Next Topics: Writing Research Papers, Research Ethics