8.2 – The controversy over proper hypothesis testing

Introduction

Over the next several chapters we will introduce and develop an approach to statistical inference, which has been given the title “Null Hypothesis Significance Testing” or NHST.

In outline, NHST proceeds with

• statements of two hypotheses, a null hypothesis, H_O, and an alternative hypothesis, H_A
• calculate a test statistic comparison of the null hypothesis (assuming some characteristic of data).
• The value of the test statistic is to be compared to a critical value for the test, identified for the assumed probability distribution at associated degrees of freedom for the statistical function, and assigned Type I error rate.

We will expand on these statements later in this chapter, so stay with me here. Basically, the null hypothesis is often a statement like the responses of subjects from the treatment and control groups are the same, eg, no treatment effect. Note that the alternative hypothesis, eg, hypertensive patients receiving hydalazine for six weeks have lower systolic blood pressure than patients receiving a placebo (Campbell et al 2011), would be the scientific hypothesis we are most interested in. But in the Frequentist NHST approach we test the null hypothesis, not the alternative hypothesis. This framework over proper hypothesis testing is the basis of the Bayesian vs Frequentist controversy.

Consider the independent sample t-test (see Chapter 8.5 and 8.6), our first example of a parametric test.

After plugging in the sample means and the standard error for the difference between the means, we calculate t, the test statistic of the t-test. The critical value is treated as a cut-off value in the NHST approach. We have to set our Type I error rate before we start the experiment, and we have available the degrees of freedom for the test, which follows from the sample size. With these in hand, the critical value is found by looking in the t-table of probabilities (or better, use R).

For example, what is the critical value of t-test with 10 degrees of freedom and Type I error of 5%?

In Rcmdr, choose Distributions → Continuous distributions → t distribution → t quantiles… (Fig 1).

Figure 1. Screenshot t-quantiles Rcmdr menu.

Note we want Type I equal to 5%. Since there are two tails for our test, we divide 5% by two and enter 0.025 and select the Upper tail.

R output

> qt(c(.025), df=10, lower.tail=FALSE)
[1] 2.228139

which is the same thing we would get if we look up on the t-distribution table (Fig 2).

Figure 2. Screenshot of portion of t-table with highlighted (red) critical value for 10 degrees of freedom.

If the test statistic is greater than the critical value, then the conclusion is that the null hypothesis is to be provisionally rejected. We would like to conclude that the alternative hypothesis should favored as best description of the results. However, we cannot — the p-value simply tells us how likely our results would be obtained and if the null hypothesis was true. Confusingly, however, you cannot interpret the p-value as telling you the probability (how likely) that the null hypothesis is true. If however the test statistic is less than the critical value, then the conclusion is that the null hypothesis is to be provisionally accepted.

The test statistic can be assigned a probability or p-value. This p-value is judged to be large or small relative to an a priori error probability level cut off called the Type I error rate. Thus, NHST as presented in this way may be thought of as a decision path — if the test statistic is greater than the critical value, which will necessarily mean that the p value is less than the Type I error rate, then we make one type of conclusion (reject H_O). In contrast, if the test statistic is less than the critical value, which will mean that the p-value associated with the test statistic will be greater than the Type I error rate, then we conclude something else about the null hypothesis. The various terms used in this description of NHST were defined in Chapter 8.3.

Sounds confusing, but, you say, OK, what exactly is the controversy? The controversy has to do whether the probability or p-value can be interpreted as evidence for a hypothesis. In one sense, the smaller the p-value, the stronger the case to reject the null hypothesis, right? However, just because the p-value is small — the event is rare — how much evidence do we have that the null hypothesis is true? Not necessarily, and so we can only conclude that the p-value is one part of what we may need for evidence for or against a hypothesis (hint: part of the solution is to consider effect size — introduced in Chapter 9.2 — and the statistical power of the test, see Ch 11). What follows was covered by Goodman (1988) and others. Here’s the problem. Consider tossing a fair coin ten times, with the resulting trial yielding nine out of ten heads (eg, a value of one, with tails equal to zero).

R code

set.seed(938291156)
rbinom(10,1,0.5)
[1] 1 1 1 1 1 1 0 1 1 1

While rare (binomial probability 0.0098), do we take this as evidence that the coin is not fair? By itself, the p-value provides no information about the alternative hypothesis. More about p-value follows below in sections What’s wrong with the p-value from NHST? and The real meaning and interpretation of P-values.

Statisticians have been aware of limitations of the NHST approach for years (see editorial by Wasserstein et al 2019), but only now is the message getting attention of researchers in the biosciences and other fields. In fact, the New York Times recently had a nice piece by F.D. Flam (“The Odds, Continually Updated,” 29 Sep 2014) on the controversy and the Bayesian alternative. Like most controversies there are strong voices on either side, and it can be difficult as an outsider to know which position to side with (Fig 3).

Figure 3. “Frequentists vs. Bayesians,” https://xkcd.com/1132/.

The short answer is — as you go forward do realize that there is a limitation to the frequentist approach and to be on the correct side of the controversy, you need to understand what you can conclude from statistical results. NHST is by far the most commonly used approached in biosciences (eg, out of 49 research articles I checked from four randomly selected issues of 2015 PLoS Biology, 43 used NHST, 2 used a likelihood approach, none used Bayesian statistics). The NHST is also the overwhelming manner in which we teach introductory statistics courses (eg, checking out the various MOOC courses at www.coursera.org, all of the courses related to Basic Statistics or Inferential Statistics are taught primarily from the NHST perspective). However, right from the start I want to emphasize the limits of the NHST approach.

If the purpose of science is to increase knowledge, then the NHST approach by itself is an inadequate framework at best, and in the eyes of some, worthless! Now, I think this latter sentiment is way over the top, but there is a need for us to stop before we begin, in effect, to set the ground rules for what can be interpreted from the NHST approach. The critics of NHST have a very important point, and that needs to be emphasized, but we will also defend use and teaching of this approach so that you are not left with the feeling that somehow this is a waste of time or that you are being cheated from learning the latest knowledge on the subject of statistical inference. The controversy hinges on what probability means.

P-values, statistical power, and replicability of research findings

Science, as a way of knowing how the world works, is the only approach that humans have developed that has been empirically demonstrated to work. Note how I narrowed what science is good for — if we are asking questions about the material world, then science should be your toolkit. Some (eg, Platt 1964), may further argue that there are disciplines in science that have been more successful (eg, molecular biology) than others (eg, evolutionary psychology, cf discussion in Ryle 2006) at advancing our knowledge about the material world. However, to the extent research findings are based solely on statistical results there is reason to believe that many studies in fact have not recovered truth (Ioannidis 2005).

In a review of genomics, it was reported that findings of gene expression differences by many microarray studies were not reproducible (Allison et al 2006). The consensus is that confidence in the findings should hold only for the most abundant gene transcripts of many microarray gene expression profiling studies, a conclusion that undercuts the perceived power of the technology to discover new causes of disease and the basis for individual differences for complex phenotypes. Note that when we write about failure of research reproducibility we are not including cases of alleged fraud (Carlson 2012 on Duke University oncogenomics case), we are instead highlighting that these kinds of studies often lack statistical power; hence, when repeated, the experiments yield different results.

Frequentist and Bayesian Probabilities

Turns out there is a lot of philosophical problems around the idea of “probability,” and three schools of thought. In the Fisherian approach to testing, the researcher devises a null hypothesis, H_O, collects the data, then computes a probability (p-value) of the result or outcome of the experiment. If the p-value is small, then this is inferred as little evidence in support of the null hypothesis.

In the Frequentists’ approach, the one we are calling NHST, the researcher devises two hypotheses, the null hypothesis, H_O, and an alternative hypothesis, H_A. The results are collected from the experiment and, prior to testing, a Type I error rate (α, chance) is defined. The Type I error rate is set to some probability and refers to the chance of rejecting the null hypothesis purely due to random chance. The Frequentist then computes a p-value of result of the experiment and applies a decision criterion: If P-value greater than Type I error rate, then provisionally accept null hypothesis. In both the Fisherian and Frequentist approaches, the probability, again defined at the relative frequency of an event over time, is viewed as a physical, objective and well-defined set of values.

Bayesian approach: based on Bayes conditional probability, one identifies the prior (subjective) probability of an hypothesis, then, adjusts the prior probability (down or up) as new results come in. The adjusted probability is known as posterior probability and it is equal to the likelihood function for the problem. The posterior probability is related to the prior probability and this function can be summarized by the Bayes factor as evidence the evidence against the null hypothesis. And that’s what we want, a metric of our evidence for or against the null hypothesis.

We can calibrate the Bayesian probability to the frequentist p-value (Selke et al 2001; Goodman 2008; Held 2010; Greenland and Poole 2012). Methods to achieve this calibration vary, but the Fagan nomogram Held (2010) proposed is a good tool for us as we go forward. We can calculate our NHST p-value, but then convert the p-value to a Bayes factor by looking at the nomogram. I mention this here not as part of your to do list, but rather as a way past the controversy: the NHST p-value can be interpreted as a Bayesian conditional probability, but they do not test the same hypothesis.

Likelihood

Before we move on there is one more concept to introduce, that of likelihood. We describe a model (an equation) we believe can generate the data we observe. By constructing different models with different parameters (hypotheses), you generate a statistic that yields a likelihood value. If the model fits the data, then the likelihood function has a small value. The basic idea then is to compare related, but different models to see which fits the data better. We will use this approach when comparing linear models when we introduce multiple regression models in Chapter 18.

What’s wrong with the p-value from NHST?

Well, really nothing is “wrong” with the p-value.

Where we tend to get into trouble with the p-value concept is when we try and interpret it. See below, Why is this important to me as a beginning student? The p-value is not evidence for a position, it is a statement about error rates. The p-value from NHST can be viewed as the culmination of a process that is intended to minimize the chance that the statistician makes an error.

In Bayesian terms, the p-value from NHST is the probability that we observe the data (eg, the differences between two sample means), assuming the null hypothesis is true. If we want to interpret the p-value in terms of evidence for a proposition, then we want the conditional error probability.

Sellke et al (2001) provided a calibration of p-values and, assuming that the prior probabilities of the null hypothesis and the alternative hypothesis are equal (that is, that each have a prior probability of 0.5), by using a formula provided by them (equation 3), we can correct our NHST p-value into a probability that can be interpreted as evidence in favor of the interpretation that the null hypothesis is true. In Bayesian terms this is called the posterior probability of the null hypothesis. The formula is

where e is Euler’s number, the base of the natural logarithm (ln), and p is the p-value from the NHST. This calibration works as long as   (Sellke et al 2001).

By convention we set the Type I error at 5% (cf Cohen 1994). How strong of evidence is a p-value near 5% against the null hypothesis being true, again, under the assumption that the prior probability of the null hypothesis being true is 50%? Using the above formula I constructed a plot of the calculated conditional error probability values against p-values (Fig 4).

Figure 4. Conditional error probability values plotted against p-values.

As you can see, a p-value of 5% is not strong evidence at just 0.289. Not until p-values are smaller than 0.004 does the conditional error probability value dip below 0.05, suggesting strong evidence against the null hypothesis being true.

R note: For those of you keeping up with the R work, here’s the code for generating this plot. Text after “#” are comments and are not interpreted by R.

At the R prompt type each line

NHSTp = seq(0.00001,0.37,by=0.01)      #create a sequence of numbers between 0.0001 and 0.37 with a step of 0.01
CEP = (1+(-1*exp(1)*NHSTp*log(NHSTp))^-1)^-1    #equation 3 from Sellke et al 2001
plot(NHSTp,CEP,xlab="P-value", ylab="Conditional error probability",type="l",col="blue")

Why is this important to me as a beginning student?

As we go forward I will be making statements about p-values and Type I error rates and null hypotheses and even such things as false positives and false negative. We need to start to grapple with what exactly can be said by p-values in the context of statistical inference, and to recognize that we will sometimes state conclusions that cut some corners when it comes to interpreting p-values. And yet, you (and all consumers of statistics!) are expected to recognize what p-values mean. Always.

The real meaning and interpretation of P-values

This is as good of a time as any to make some clarification about the meaning of p-value and the whole inference concept. Fisher indeed came up with the concept of the p-value, but its use as a decision criterion owes to others and Fisher disagreed strongly with use of the p-value in this way (Fisher 1955; Lehmann 1993).

Here are some common p-value corner-cutting statements to avoid using (after Goodman 2008; Held 2010). P-values are sometimes interpreted, incorrectly, as

Table. Incorrect interpretations of NHST p-values

1. the probability of obtaining the observed data under the assumption of no real effect
2. an observed type-I error rate
3. the false discovery rate, ie, the probability that a significant finding is a “false positive”
4. the (posterior) probability of the null hypothesis.

So, if p-values don’t mean any of these things, what does a p-value mean? It means that we begin by assuming that there is no effect of our treatments — the p-value is then the chance we will get as large of a result (our test statistic) and the null hypothesis is true. Note that this definition does not include a statement about evidence of the null hypothesis being true. To get evidence of “truth” we need additional tools, like the Bayes Factor and the correction of the p-value to the conditional error probability (see above). Why not dump all of the NHST and go directly to a Bayesian perspective, as some advise? The single best explanation was embedded in the assumption we made about the prior probability in order to calculate the conditional error probability. We assumed the prior probability was 50%. For many, many experiments, that is simply a guess. The truth is we generally don’t know what the prior probability is. Thus, if this assumption is incorrect, then the justification for the formula by Sellke et al (2001) is weakened, and we are no closer to establishing evidence than before.

The take home message is that it is unlikely that a single experiment will provide strong evidence for the truth. Thus the message is repeat your experiments — and you already knew that!  And the Bayesians can tell us that with the addition of more and more data reduces the effect of the particular value of the prior probability on our calculation of the conditional error probability. So, that’s the key to this controversy over the p-value.

Reporting p-values

Estimated p-values can never be zero. Students may come to use software that may return p-values like “0” — I’m looking at you Google Sheets re: default results from CHISQ.TEST() — but again, this does not mean the probability of the result is zero. The software simply reports values to two significant figures and failed to round. Some journals may recommend that 0 should be replaced by p < 0.01 or even < 0.05 inequalities, but the former lacks precision and the latter over-emphasizes the 5% Type I error rate threshold, the “statistical significance” of the result. In general, report p-value to three significant figures and four digits. If a p-value is small, use scientific notation and maintain significant digits. Thus, a p-value of 0.004955794 should be reported as 0.00496 and a p-value of 0.0679 should be reported as 0.0679. Use R’s signif() function, for example p-value reported as 6.334e-05, then

signif(6.334e-05,3)
[1] 6.33e-05

Rounding and significant figures were discussed in Chapter 3.5. See Land and Altman (2015) for guidelines on reporting p-values and other statistical results.

Questions

1. Revisit Figure 1 again and consider the following hypothesis — the sun will rise tomorrow.
   • If we take the Frequentist position, what would the null hypothesis be?
   • If we take the Bayesian approach, identify the prior probability.
   • Which approach, Bayesian or Frequentist, is a better approach for testing this hypothesis?
2. Consider the pediatrician who, upon receiving a chest X-ray for a child notes the left lung has a large irregular opaque area in the lower quadrant. Based on the X-ray and other patient symptoms, the doctor diagnoses pneumonia and prescribes a broad-spectrum antibiotic. Is the doctor behaving as a Frequentist or a Bayesian?
3. With the p-value interpretations listed in the table above in hand, select an article from PLoS Biology, or any of your other favorite research journals, and read how the authors report results of significance testing. Compare the precise wording in the results section against the interpretative phrasing in the discussion section. Do the authors fall into any of the p-value corner-cutting traps?

Quiz Chapter 8.2

The controversy over proper hypothesis testing

Start Quiz

Question

✕

✓

Your answer:

Correct answer:

Next

Please wait..

↺

You got {{SCORE_CORRECT}} out of {{SCORE_TOTAL}}

Your Answers

Chapter 8 contents

• Introduction
• The null and alternative hypotheses
• The controversy over proper hypothesis testing<.span>
• Sampling distribution and hypothesis testing
• Tails of a test
• One sample t-test
• Confidence limits for the estimate of population mean
• References and suggested readings

6.7 – Normal distribution and the normal deviate (Z)

Introduction

In chapter 3.3 we introduced the normal distribution and the Z score, aka normal deviate, as part of a discussion about how some knowledge about characteristics of the dispersion of our data sampled from a population could be used to calculate how many samples we need (the empirical rule). We introduced Chebyshev’s inequality as a general approach to this problem, where little is known about the distribution of the population and contrasted it with the Z score, for cases where the distribution is known to be Gaussian or the normal distribution. The normal distribution is one of the most important distributions in classical statistics. All normal distributions are bell-shaped and symmetric about the mean. To describe a normal distribution only two parameters are needed: the population mean, μ, and the population standard deviation, σ. The normal distribution with mean equal to zero and standard deviation equal to one is called the standard normal, or Z distribution. With use of the Z score any normal distribution can be quickly converted to the standard normal distribution.

Proportions of a Normal Distribution

This concept will become increasingly important for the many statistical tests we will learn over the next few weeks. What is the proportion of the populations that is greater than some specific value? Below (Fig 1), again, I have generated a large data set, now with population mean µ = 5 and σ = 2. The red line corresponds to the equation of the normal curve using our values of µ = 5 and σ = 2.

Figure 1. Frequency of observations expected to be greater than 7 (red bars) from a large population with mean µ = 5 and σ = 2.

R code for Fig 1.

set.seed(123)
myData <- rnorm(1000, mean = 5, sd = 2)
h <- hist(myData, breaks = 20, plot = FALSE) 
threshold <- 7
# Bars with midpoints greater than threshold will be "red", 
# and others will be "white."
bar_colors <- ifelse(h$mids > threshold, "red", "white")
plot(h, col = bar_colors, xlab = "Value", ylab = "Frequency")

Note that this is a crucial step! We assume that our sample distribution is really a sample from a population density (= “area under the curve”) function (= “an equation”) for a normal random (= “population”) variable.

Once I (you) make this assumption, then we have powerful and easy to use tools at our command to answer questions like

Question: What proportion of the population is greater than 7 (Fig 1, colored in red)?

This gets to the heart of the often-asked question, How many samples should I measure? If we know something about the mean and the variability, then we can predict how many samples will be of a particular kind. Let’s solve the problem.

The Z score

We could use the formula for the normal curve (and a lot of repetitions), but fortunately, some folks have provided tables that short-cut this procedure. R and other programs also can find these numbers because the formulas are “built in” to the base packages. First, let’s introduce a simple formula that lets us standardize our population numbers so that we can use established tables of probabilities for the normal distribution.

Below, we will see how to use Rcmdr for these kinds of problems.

However, it’s one of the basic tasks in statistics that you should be able to do by hand. We’ll use the Z score as a way to take advantage of known properties of the standard normal curve.

Z = 1 (with the mean and SD). Z (say “Z-score”) is called the normal deviate (aka “standard normal score”; it is also called the “Z-score”); it gives us a shortcut for finding the proportion of data greater than 7 in this case).

We use the normal deviate to do a couple of things; one use is to standardize a sample of observations so that they have a mean of zero and a standard deviation of one (the Z distribution). The data would then said to have been normalized.

The second use is to make predictions about how often a particular observation is likely to be encountered. As you can imagine, this last use is very helpful for designing an experiment — if we need to see a specified difference, we can conduct a pilot study (or refer to the literature) to determine a mean and level of variability for our observation of choice, plug these back into the normal equation and predict how likely we can expect to see a particular difference. In other words, this is one way to answer that question — how many observations need I make for my experiment to be valid?

Table of normal distribution

A portion of the table of the normal curve is provided at our web site and in your workbook. For our discussions, here’s another copy to look at (Fig 2).

Figure 2. Portion of the table of the normal distribution. Only values equal to or greater than Z = 0 are visible.

See Table 1 in the Appendix for a full version of the normal table.

We read values of Z from the first column and the first row. For Z = 0.23 we would scan the top row, scoot over to fourth column, then trace to where the row and column intersect (Fig 3); the frequency of occurrence of values at Z = 0.23 is 0.409046 or 40.9% (Fig 3).

Figure 3. Highlight Z = 0.23, frequency is 0.409046.

Z on the standard normal table is going to range between -4 and +4, with Z = 0 corresponding to 0.500. The Normal table values are symmetrical about the mean of zero.

What to make of the values of Z, from -4, -3, … +2, +3, up to +4 and beyond?  These are the standard deviations! Recall that using the Z score you corrected to mean of zero (got it!), and a standard deviation of one! A Z of 2 is twice the standard deviation; a Z = 3 is therefore three times the standard deviation, and so forth. The distribution is symmetrical: you get the same frequency for negative as for positive values. So on the “X” axis on a standard normal distribution, we have units of standard deviation plus (greater) or minus (less) than the mean. In Figure 4, area less than -1 standard deviations is highlighted (Fig 4).

Figure 4. Plot of standard normal distribution; area less than -1 σ.

R code for Fig 4.

set.seed(123)
myData2 <- rnorm(100000, mean = 0, sd = 1)
threshold <- -1
bar_colors <- ifelse(h$mids < threshold, "red", "white")
# Additional code used, same as Figure 1.

Question. How many multiples of standard deviations would you have for a Z score of Z = 1.75?

Answer = 1.75 times

Examples

See Table 1 in the Appendix for a full version of the normal table as you read this section.

What proportion of the data set will have values greater than 7? After applying our Z score equation, I get Z = 1.0 and for a Z of 1.0 = 0.1587 or 15.87% of the observations are greater than 7.

What proportion of the data set will have values less than -7? After applying our Z score equation, I get Z = -1.0. Taking advantage of symmetry argument, I just take my Z = -1.0 and make it positive — instead of values smaller than -1.0 we now have values greater than +1.0. And for a Z of 1.0 = 0.1587 or 15.87% of the observations are greater than 7, which means that 15.87% will be -7 or smaller.

What proportion of the data set will have values greater than 8? Again, apply the Z score equation I get Z of 1.5 = 0.0668 or 6.68% of the observations are greater than Z.

What proportion of the population is between 5 and 7 ? (colored in red). Draw the problem (example shown in Fig 5).

Figure 5. Proportion of the population is between 5 and 7.

R code for Fig 5.

# Used myData from Fig 1 code.

highlight_start <- 5
highlight_end <- 7
bar_colors <- rep("white", length(h$counts)) 
bar_colors[h$breaks[-length(h$breaks)] >= highlight_start & h$breaks[-1] <= highlight_end] <- "red"
# Additional code used, same as Figure 1.

Worked problem

1 – (proportion beyond 7) – (proportion less than 5)

1 – (0.1587) – (proportion less than 5)

And the proportion less than 5?

Use Z-score equation again. Now we find that

Z = 0

And look up Z of 0 from the table  = 0.5 proportion or 50.0%

Therefore, the proportion between 5 & 7 equals

1 -0.1587 – 0.50 = 0.3413

Answer = 34.13 % of the observations are between 5 and 7 when the mean µ = 5 and s = 2.

Questions

Several questions were asked in the text.

1. Proportion of observations greater than 7  from a large population with mean µ = 5 and σ = 2.

2. How many multiples of standard deviations would you have for a Z score of Z = 1.75?

3.

Quiz Chapter 6.7

Normal distribution questions

Start Quiz

Question

✕

✓

Your answer:

Correct answer:

Next

Please wait..

↺

You got {{SCORE_CORRECT}} out of {{SCORE_TOTAL}}

Your Answers

Chapter 6 contents

• Introduction
• Some preliminaries
• Ratios and probabilities
• Combinations and permutations
• Types of probability
• Discrete probability distributions
• Continuous distributions
• Normal distribution and the normal deviate (Z)
• Moments
• Chi-square (Χ²) distribution
• t distribution
• F distribution
• References and suggested readings

5.6 – Sampling from Populations

Introduction

Researchers generally can’t study an entire population. More generally, striving to study each member of a population is not necessary to arrive at answers about the population. For example, consider the question, does taking a multivitamin daily improve health? What are our options? Do we really need to follow every single individual in the United States of America, monitoring their health and noting whether or not the person takes vitamins daily in order to test (inference) this hypothesis? Or, can we get at the same answer by careful experimental design (see Schatzkin et al 2002; Dawsey et al 2014)? Supplement use is wide-spread in the United States, but both health and vitamin use differ by demographics. Young people tend to be healthier than older people and older people tend to take supplements more than younger people.

A subset of the population is measured for some trait or characteristic. From the sample, we hope to refer back to the population. We want to move from anecdote (case histories) to possible generalizations of use to the reference population (all patients with these symptoms). How we sample from the reference population limits our ability to generalize. We need a representative sample: simple to define, hard to achieve.

Statistics becomes necessary if we want to infer something about the entire populations. (Which is usually the point of doing a study!!) Typically tens to thousands of individuals are measured. But in addition, HOW we obtain the sample of individuals from the reference population is CRITICAL.

Kinds of sampling include (adapted from Box 1, Tyrer and Heyman 2016; see also McGarvey et al 2016)

Probability sampling

• random
• stratified
• blocking
• clustered
• systematic

Nonprobability sampling

• convenience, haphazard
• judgement
• quota
• snowball

For an exhaustive, authoritative look at sampling, see Thompson (2012).

How can samples be obtained?

Sampling from a population may be convenient. For one famous example, consider the Bumpus data set. (We introduced this data set in Question 5, Chapter 5.) So the legend goes, Professor Bumpus was walking across the campus of Brown University, the day after a severe winter storm, and came across a number of motionless house sparrows on the ground. Bumpus collected the birds and brought them to his lab. Seventy-two birds recovered; 64 did not (Table1).

Table 1. A subset of Bumpus data set, summarized by sex of birds.

Bumpus reported differences in body size that correlated with survival (Bumpus 1899), and this report is often taken as an example of Natural Selection (cf. Johnston et al 1972). The Bumpus dataset is clearly convenience sampling. It’s also a case study: a report of a single incident. But given that is is a large sample (n = 136), it is tempting to use the data to inform about about possible characteristics of the birds that survived compared with those that perished.

Another way we collect samples from populations is best termed haphazard. In graduate school I got the opportunity to study locomotor performance of whiptail lizards (Aspidoscelis tigris, A. marmoratus, genus formerly Cnemidophorus**) across a hybrid zone in the Southwest United States (Dohm et al. 1998). During the day we would walk in areas where the lizards were known to occur and capture any individual we saw by hand. (This would sometimes mean sticking our hands down into burrows, which was always exciting — you never really knew if you were going to find your lizard or if you were going to find a scorpion, venomous spider, or …) Lizards collected were returned to the lab for subsequent measures. Clearly, this was not convenience sampling; it involved a lot of work under the hot sun. But just as clearly, we could only catch what we could see and even the best of us would occasionally lose a lizard that had been spotted. Moreover, one suspects we missed many lizards that were present, but not in our view. Lizards that were underground at the time we visited a spot would not be seen nor captured by us; individual lizards that were especially wary of people (Bulova 1994) would also escape us. In other words, we caught the lizards that were catchable and can only assume (hope?) that they were representative of all of these lizards. Applying a grid or quadrant system to the area and then randomly visiting plots within the grid or quadrant might help, but still would not eliminate the potential for biased sampling we faced in this study.

Quota sampling implies selection of subjects by some specific criteria, weighted by the proportions represented in the population. It’s different from stratified sampling because there is no random selection scheme: subjects are selected based on matching some criteria, and collection for that group stops when the sample number matches the proportion in the population. Consider our vitamin supplement survey. If student population at Chaminade University was the reference population, and we have enough money to survey 100 students, then we would want a sample of 70 women students and 30 male students, representing the proportions of the student population.

Snowball sampling implies that you rely on word-of-mouth to complete sampling. After initial recruitment of subjects, sample size for the study increases because early participants refer others to the researchers. This can be a powerful tool for reaching underrepresented communities (eg, Valerio et al 2016).

Types of Probability sampling

Random sampling is an example of probability sampling. As we defined earlier, simple random sampling requires that you know how many subjects are in the population (N) and then each subject has an equal chance of being selected

Examples of nonprobability sampling include

• convenience sampling
• volunteer sampling
• judgement sampling

Convenience sampling (the first 20 people you meet at the library lanai); volunteer sampling (you stand in front of a room of strangers and ask for any ten people to come forward and take your survey — or more seriously, persons with a terminal disease calling a clinic reportedly known to cure the disease with a radical new, experimental treatment), and judgement sampling (to study tastes in fashion, you decide that only persons over six feet tall should be included because .…).

Random in statistics has a very important, strict meaning. As opposed to our day-to-day usage, random sampling from a population means that the probability that any one individual is chosen to be included in a sample is equal. Formally, this is called simple random sampling to distinguish it from more complex schemes. For a sampling procedure to be random requires a formal procedure for sampling a population with known size (N).

For example, at the end of the semester, I may select the order for student presentations at random. Thus, students in the classroom would be considered the population (groups of students are my sample unit, not individual students!). What is the probability that your group will be called first? Second? We need to know how many groups there are to conduct simple random sampling. Let’s take an extreme and say that all groups have a size of one; there are 26 students in this room, so

or

of being selected first.

Now to determine the probability of your group being selected second, we need to distinguish between two kinds of sampling:

• Sampling with replacement — after I select the first group, the first group is returned to the pool of groups that have not been selected. In other words, with replacement, your group could be selected first and selected second! The probability of being selected second then remains

     

  .

• Sampling without replacement — after I select the first group, then I have

     

  groups left to select the second group, so probability that your group will be second is

     

  . The first group has already been selected and is not available.

Random sampling refers to how subjects are selected from the target, reference population. Random assignment, however, describes the process by which subjects are assigned to treatment groups of an experiment. Random sampling applies to the external validity of the experiment: to the extent that a truly random sample was drawn, then results may be generalized to the study population. Random assignment of subjects to treatment however makes the experiment internally valid: results from the experiment may be interpreted in terms of causality.

R code, sample without replacement

sample(26)
[1] 1 10 11 23 17 13 8 6 19 20 21 16 14 15 5 2 12 4 26 22 9 7 25 18 3 24

R code, sample with replacement

mySample <- sample(26, replace = TRUE); mySample
[1] 9 26 24 9 3 1 23 14 26 6 5 19 17 23 11 12 3 8 7 19 2 16 3 11 1 1

How many elements duplicated?

out <- duplicated(mySample); table(out)
out
FALSE TRUE 
17 9

More on sampling in R discussed in Sampling with computers below.

Additional sampling schemes

Simple random sampling is not the only option, but in many cases it is the most desirable. Consider our multivitamin study again. Perhaps studying the entire USA population is a bit extreme. How about working from a list of AARP members, sending out questionnaires to millions of persons on this list, getting back about 20% of the questionnaires, sorting through the responses and identifying the respondent to diet categories and multivitamin use. The researchers had nearly 500,000 persons willing and able to participate in their prospective study (Schatzkin et al 2001; Dawsey et al 2014; Lim et al 2021). It’s an enormous study. But is this really much better than our described lizard experiment? Let’s count the ways: Not all older people are members of AARP (that 500,000? That’s less than 1% of the 50 and older persons in the USA); A large majority of AARP members did not return surveys; Some fraction of the returned surveys were not usable; How representative of diverse aged populations in the United States is AARP?

Simple random sampling may not be practical, particularly if sub-populations, or strata, are present and members of the different sub-populations are not available to the researcher in the same numbers. In ecology, habitat-use studies routinely need to account for the patchiness of the environment in space and time (Wiens 1976; Summerville and Crisp 2021). Thus, samples are drawn in such a way to represent the frequency within each sub-population.

For a simple example, researchers conducting a controlled breeding program of mice don’t use simple random sampling to choose pairs of mates; after all, random sampling without regard of sex of the mice would lead to some pairings of males only, or females only (Fig 1).

Figure 1. Sixteen mice, eight red and eight blue. Image © 2024 Mia D Graphics

R code

pairs <- c(rep("red", 8), rep("blue", 8))
replicate(2, sample(pairs, 8), simplify = FALSE)

[[1]]
[1] "blue" "blue" "blue" "red" "blue" "red" "blue" "red"

[[2]]
[1] "blue" "blue" "red" "red" "red" "red" "red" "blue"

Oops. If we interpret the output blue = male and red = female, then four out of the eight pairs were same sex pairings.

Thus, the breeding strategy is to random sample from female mice and from male mice and the stratification is sex of the mouse. Alternatively, breeders may select mice to form breeding pairs systematically: From a large colony with dozens of cages, the breeder may select one mouse from every third cage.

Stratified Random Sampling: Individuals in the population are classified into strata or sub-groups: age, economic class, ethnicity, gender, sex …, as many as needed. Then choose a simple random sample from each group, typically in proportion to the size of the sub-groups in the population. Combine those into the overall sample. For example, when I wanted a random sample of mice for my work, I called the supplier and requested that a total of 100 male and female mice be randomly selected from the five colonies they maintained. The reference population is the entire supply of mice at that company (at that time), but I wished to make sure that I got unrelated mice, so I needed to divide the population into groups (the five colonies) before my sample was constructed.

For a simple example of the need for stratified sampling, consider our 16 blue and red mice again. If mice were randomly assigned to treatment group, then by chance unequal samples within strata may be assigned to a treatment group, potentially biasing our conclusions (Fig 2).

Figure 2. Sixteen mice, randomly assigned to treatment groups C and T; by chance, 75% blue in C and just 25% in T. If color was a confounding factor then our conclusions about the effectiveness of the treatment would be associated with color. Image © 2024 Mia D Graphics

Block sampling: superficially resembles stratified sampling, but refers to how experimental units are assigned to treatment groups. Blocks in experimental design refers to potentially confounding factors for which we wish to control, but are otherwise not interested in. For our example described in Figure 2, mouse color may not be of specific interest for our study and therefore would be considered a blocking effect.

Agricultural research provides a host of other blocking situations: plots distributed across an elevation or hydration gradient (Fig 3), soil types, and many others.

Figure 3. Agricultural fields in Central Poland. NASA https://visibleearth.nasa.gov/images/146593/agriculture-fields-in-central-poland/146595w

For a microbiology example, edge effects of heated culture plates. For a growth experiment it’s unlikely we are specifically interested in the plate layout on the growth characteristics of treated cells. However, it’s also unlikely we can assume placement has no influence on growth (eg,  Mansoury et al 2021). A 96-well microplate has 60 inner wells and 36 edge wells (A1 – H1, A12-H12, plus 1^st and 12^th wells of columns B – H, for a total of 36 wells, Fig 4). Thus, analysis of growth effects should include a block effect for well position.

Figure 4. Format of 96-well plate. Red cells = “edge” wells; blue cells = “inner” wells. Plotted with ggplate package.

For long incubation periods it’s not unexpected that edge wells and inner wells may evaporate at different rates, potentially confounding interpretation of treatment effects if, by chance, treatment samples are more commonly assigned to edge wells. Block sampling implies subjects from treatment groups are randomly assigned to each block, in equal numbers if possible.

Cluster Sampling: In many situations, the population is far too large or too dispersed and scattered for a list of the entire population to be known. And, a random approach ignores that there is a natural grouping — people live next to each other, so there is going to be things in common. A multi-stage approach to sampling will be better than simply taking a random sample approach. Most surveys of opinion (when conducted reputably) use a multi-stage method. For example, if a senator wishes to poll his constituents about an issue, he may his pollster will randomly select a few of the counties from his state (first stage), then randomly select among towns or cities (second stage), to obtain a list of 1000 people to call. In some instances, they might use even more stages. At each stage, they might do a stratified random sample on sex, race, income level, or any other useful variable on which they could get information before sampling. If you are interested in this kind of work, for starters, see Couper and Miller (2009).

There are more types of sampling, and entire books written about the best way to conduct sampling. One important thing to keep in mind is that as long as the sample is large relative to the size of the population, each of the above methods generally will get the same answers (= the statistics generated from the samples will be representative of the population).

As long as some attempt is made to randomize, then you can say that the procedure is probability sampling. Nonprobability, or haphazard sampling, describes the other possibility, that is, each element is selected arbitrarily by a non-formal selecting of individual. … all the fish or birds that you catch may not be a random sample of those present in a population. For example,

• wild Pacific salmon do not feed on the surface, hatchery salmon feed on the surface. …
• all the individuals who respond to a survey.
• Phone surveys, web surveys, person on the street surveys… how random, how representative?

Sampling with computers

Now to determine the probability of your group being selected second, we need to distinguish between two kinds of sampling:

Sampling with replacement — after I select the first group, the first group is returned to the pool of groups that have not been selected. In other words, with replacement, your group could be selected first and selected second! The probability of being selected second then remains

   

Sampling without replacement — after I select the first group, then I have

   

groups left to select the second group, so probability that your group will be second is

   

. The first group has already been selected and is not available, and so on.

Sampling is usually easiest if a computer is used. Computers use algorithms to generate pseudo-random numbers. We call the resulting numbers pseudo-random to distinguish them from truly random physical processes (eg, radioactive decay). For more information about random numbers, please see random.org.

If all you wish to do is select a few observations or you need to use a random procedure to select subjects prior to observations, then these websites can provide a very quick, useful tool.

Sampling in Microsoft Excel or LibreOffice Calc

Microsoft’s Excel is pretty good at sampling, but requires knowledge of included functions. Here are the steps to generate random numbers and select with and without replacement in Excel. I’ll give you two cases.

1. For random numbers, enter the function =rand() in a cell, then drag the cell handle to fill in cells to N (in our case = 26, so A1 to A26). This function generates a random (more or less!) number between 0 and 1. We want digits between 1 and 26, not fractions between 0 and 1, so combine INT function with RAND function

=INT(27*RAND())

Note 2: To get between 1 and 9, multiply by 10 instead of 27; to get between 1 and 100, multiply by 101, etc.

In Excel to sample with replacement, simply pick the first two cells (the algorithm Excel used already has conducted sample with replacement. See next item for method to sample without replacement in Excel. You have to have installed the Data Analysis Tool Pak. Here’s instructions for Office 2010.

If you have a Mac and Office 2008, there is no Data Analysis Tool Pak, so to get this function in your Excel, install a third-party add-in program (eg, StatPlus, a free add-in, really nice, adds a lot of function to your Excel). If you have the 2011 version of Office for Mac, then the Data Analysis Tool Pak is included, but like your Windows counterparts, you have to install it (click here for instructions).

2. Let’s say that we have already given each group a number between 1 and 26 and we enter those numbers in sequence in column A.

To sample without replacement, select Tools → Data Analysis… (if this option is not available, you’ll have to add it — see Excel help for instructions, Fig 5).

Figure 5. Screenshot of Sampling in Data Analysis menu, Microsoft Excel

Enter the cells with the numbers you wish to select from. In our example, column A has the numbers 1 through 26 representing each group in our class. I entered A:A as the Input Range.

Next, select “Random” and enter the number of samples. I want two.

Click OK and the output will be placed into cell B3 (my choice); I could have just as easily had Excel put the answer into a new worksheet.

Figure 6. Screenshot of input required for Sampling in Data Analysis menu, Microsoft Excel

To sample with replacement from column A (with our 1 through 26), type in the formula B1

=INT(27*RAND())

and the formula C1

=INDEX(A:A,RANK(B1,B:B))

then drag the cell handles to fill in the columns (first column B, then column C).

That’ll do it for MS Excel or LibreOffice Calc.

Sampling with R (Rcmdr)

It’s much easier with much more control to get sampling in Rcmdr (R) than in Excel. Sampling in R is based on the function called sample() and sample.int(). I will present just the sample() command here.

sample(x, size, replace = FALSE, prob = NULL)

Example

You want to sample ten integers between 1 and 10

sample(10)

R output

sample(10)
[1]  5  1 10  8  7  3  4  9  2  6

You have a list of subjects, A1 through A10

subjects = c("A1","A2","A3","A4","A5","A6","A7","A8","A9","A10") 
sample(subjects, 3, replace = FALSE, prob = NULL)

R output

sample(subjects, 3, replace = FALSE, prob = NULL)
[1] "A5" "A2" "A9"

Could use this to arrange random order for ten subjects

sample(subjects, 10, replace = FALSE, prob = NULL)
[1] "A9" "A3" "A8" "A2" "A1" "A4" "A10" "A7" "A5" "A6" 

Try with replacement; To sample with replacement, type in TRUE after replace in the sample() function. The R output follows

sample(x, 10, replace = TRUE, prob = NULL)
[1]  5  5  3  5  3  7  3  5 10  6

R’s randomness is based on pseudonumbers and is, therefore, not truly random (actually, this is true of just about all computer-based algorithms unless they are based on some chaotic process). We can use this pseudo part to our advantage: if we want to reproduce our “random” process, we can seed the random number algorithm to a value (eg, 100), with the command in the Script Window

set.seed(100)

For 10 random integers (eg, observations), type in the Script window

sample(10)

R returns the following in the Output Window

sample(10)
[1] 4 5 7 3 10 9 2 1 6 8

Sampling was done without replacement. Here’s another selection round,

First without setting a seed value

sample(10)
[1] 10 4 7 3 8 1 2 9 6 5

Try again to see if get the same 10

sample(10)
[1]  1  4  8  2  9  6  5 10  3  7

Now to demonstrate how setting the seed allows you to draw repeated samples that are the same.

Note 3: If I need to precede the sample command with a set.seed() call, then sampling is repeatable.

set.seed(100)
sample(10)
[1] 4 3 5 1 9 6 10 2 8 7

and try again

set.seed(100)
sample(10)
[1] 4 3 5 1 9 6 10 2 8 7

Additional R packages that help with sampling schemes include: sampling() and spatialsample, which is part of the BiodiversityR package, which is available as a plugin for R Commander.

Quiz Chapter 5.6

Sampling from populations

Start Quiz

Question

✕

✓

Your answer:

Correct answer:

Next

Please wait..

↺

You got {{SCORE_CORRECT}} out of {{SCORE_TOTAL}}

Your Answers

Questions

1. For our two descriptions of experiments in section 6.1 (the sample of patients; the sample of frogs), which sampling technique was used?
2. What purpose is served by set.seed() in a sampling trial?
3. Consider our question, Does taking a multivitamin daily improve health? Imagine you have a grant willing to support a long-term prospective study to follow up to one thousand people for ten years. List at least three concerns with proposed solutions about how sampling of subjects for the study.
4. Imagine you wish to conduct a detailed survey to learn about student preferences. Your survey will include many questions, so you decide to ask just ten students. Student population is 70% female, 30% male.
   1. Assuming you select at random (simple random sampling), what is the chance that no male students will be included in your survey?
   2. You are able to increase the number of surveys to 20, 30, 40, or 50. What is the chance that no male students will be included in your survey for each of these increased sample numbers?
      1. What can you conclude about the effects of increasing survey sample size on representativeness of students for the survey?
5. Discuss how you could apply a stratified sampling scheme to this survey and whether or not this approach improves representativeness.
6. Why are random numbers generated by a computer called pseudo random numbers?

---

Chapter 5 contents

• The basics explained
• Experiments
• Experimental and Sampling units
• Replication, Bias, and Nuisance Variables
• Clinical trials
• Importance of randomization in experimental design
• Sampling from Populations
• References and suggested readings