Basic Biostatistics Statistics For Public Health Practice

Why Most Public Health Workers Get Biostatistics Wrong

I spent eight years doing field epidemiology for state health departments before I stopped trying to find the perfect software. The problem isn't that public health professionals lack statistical training. It's that most programs still teach biostatistics like you're going to be a researcher instead of someone who needs to make decisions under time pressure with messy data. Basic Biostatistics Statistics For Public Health Practice is not about deriving formulas or proving you understand the Central Limit Theorem. It's about taking a dataset that arrived at 2pm with missing values, inconsistent date formats, and at least one column that makes no sense, and extracting a number you can actually report to people who don't care about your p-values.

The Core Methods That Actually Matter

Start with descriptive statistics. I know that sounds obvious, but most people I see struggle here because they rush into inferential tests without checking whether their data distribution justifies them. Your first step should always be to generate frequency tables, cross-tabs, and measures of central tendency stratified by relevant subgroups. Not after the analysis. Before it. Always before it. When you move to inferential statistics, the three tests you will use 90% of the time are the chi-square test for independence, the t-test (independent samples), and the non-parametric alternatives when your data violates normality assumptions. Most beginner guides don't tell you that the independent samples t-test is remarkably robust to violations of normality when your group sizes are roughly equal and larger than 30. I learned this the hard way during a 2019 outbreak investigation where our sample was small and skewed, and I wasted half a day trying to force a parametric approach before switching to Mann-Whitney U. For rates and ratios, you need to understand the difference between prevalence, incidence, and cumulative incidence. This matters more than any statistical test when you're communicating with stakeholders. A colleague of mine once reported an "outbreak spike" using prevalence data when what he actually had was a point-in-time snapshot. The numbers looked dramatic. They were misleading. The correct metric would have been incidence density calculated over person-time, which reduced the apparent magnitude by more than half.

Working With Real Data Instead Of Textbook Examples

Textbook problems give you clean data. Real public health data does not. Here's what I mean. In 2022, I was analyzing influenza surveillance data from five counties. The dataset had 14,000 rows. Approximately 8% of records had missing age values. 12% of the dates were entered as text strings in inconsistent formats. There were duplicate identifiers that needed resolution. Someone had used -999 to code missing values for one variable, which turned out to be a plausible negative test result in another variable. This happened twice. The workaround wasn't fancy. I wrote a simple Stata do-file that first created a codebook of every variable including all unique values, then flagged any suspicious entries. I spent one hour on data cleaning for what should have been a two-hour analysis. The key insight was not to start analyzing until the codebook revealed the -999 problem. Had I run a mean or proportion without catching that, the result would have been garbage masked as precision.

This is the single most important skill in basic biostatistics for public health practice: data auditing before analysis. Not data visualization. Not hypothesis testing. Auditing. You need to know what you have before you try to draw conclusions from it. I check for impossible values, inconsistent categorizations, and coding artifacts in every dataset. It takes less time than fixing errors after peer review catches them.

Confidence Intervals And What They Actually Tell You

Most public health reports I read present point estimates with p-values. Very few present confidence intervals. This is a significant gap. A 95% confidence interval gives you the range of plausible values for your population parameter given the data you observed. It is far more informative than a binary significant-or-not conclusion. Here's a counter-intuitive point that many training programs miss: a non-significant result does not mean "no effect." It means your data is compatible with both a meaningful effect and no effect. When I saw a study claiming no association between a vaccination intervention and disease reduction because the p-value was 0.08, I asked for the confidence interval. The range was 0.02 to 0.89. The intervention could have reduced disease by 89%. That is not a null finding. That is an inconclusive finding driven by sample size. This distinction matters enormously when you're advising policymakers. Saying "there is no evidence of an effect" is qualitatively different from saying "the evidence is inconclusive." The former stops funding. The latter asks for more data.

When Standard Methods Break Down

Biostatistics has assumptions. When those assumptions are violated, standard methods produce biased or misleading results. The most common violation I encounter in public health practice is small expected cell counts in contingency tables. When any cell in a chi-square test has an expected count below 5, the test becomes unreliable. Most software will still output a p-value. That p-value is wrong. The fix is Fisher's exact test for 2x2 tables or combining categories when it makes substantive sense. I once analyzed a table comparing disease rates across four age groups where two groups had very few cases. Combining the two smallest groups was statistically defensible and produced a valid result. Splitting them further would have been meaningless. Another scenario where standard methods fail is clustered data. Public health data is often naturally clustered by geography, clinic, or household. Ignoring this clustering when calculating standard errors produces artificially narrow confidence intervals and inflated significance. The workaround is to use survey-weighted procedures or mixed-effects models. In Stata, the svy prefix handles this cleanly. In R, the survey package does the same. Both require you to specify the cluster identifier and, if applicable, the sampling weight. If you skip this step, your results are not just slightly off. They are systematically biased toward significance.

A Note On Software Choices

I use Stata for routine public health analysis because it handles survey data and complex weighting schemes without requiring extensive programming. For exploratory work and figures, I use R. SPSS is fine if that's what your organization uses, but its default output includes more information than you need and hides important diagnostics unless you dig for them. SAS is overkill for most basic biostatistics work unless you're processing data at scales beyond a few hundred thousand records. The software doesn't matter as much as understanding what each test actually assumes. I've seen analysts switch tools every few years. The underlying statistics didn't change. The bias patterns didn't change. Only the buttons they clicked changed.

Reporting Results Without Misleading Anyone

The way you present biostatistical results is itself a statistical decision. Rounding a rate to one decimal place when the denominator is 47 is different from rounding to two decimal places. A ratio of 1.33 is more precise-sounding than a ratio of 1.3, but it may not be justified by your data. I report effect sizes with 95% confidence intervals and specify the exact test used. I note sample sizes. I state when assumptions were violated and how I addressed it. This takes about 30 extra seconds per result but prevents the kind of misinterpretation that leads to wrong public health actions. There is a persistent pressure to produce "significant" findings in public health. It's real and it's harmful. The solution is not to hide uncertainty. The solution is to quantify it clearly and communicate it plainly. "The intervention reduced incidence by 40% (95% CI: 5% to 62%)" is honest. "The intervention was effective" is not, because it conveys certainty that the data does not support.

Resources That Actually Help

For practical learning, I recommend the CDC's Principles of Epidemiology course. It's free, it covers the core concepts, and it's designed for working public health professionals rather than statistics students. The Johns Hopkins Biostatistics in Public Health specialization on Coursera is also solid, though it moves faster than some practitioners need. For reference, Altman's "Practical Statistics for Medical Research" is dense but exhaustive. Collett's "Modelling Binary Data in Epidemiology" is the book I return to when I need to remember why odds ratios and risk ratios diverge and what that divergence means for interpretation. The downloadable resource most people actually need is not a formula sheet. It's a decision tree: what type of data do I have, what is my comparison group, what is my outcome variable, and which test matches that combination. I keep a one-page version at my desk. It took me three years to develop it through trial and error. The versions online tend to be overly generic or aimed at academic researchers rather than practitioners.

If you want to learn basic biostatistics statistics for public health practice in a way that translates directly to field work, focus on data auditing, proper test selection, confidence interval interpretation, and clear reporting. The rest follows.