13.1 Introduction
Parameters and estimates
This chapter discusses the traditional methods used to address random error in quantitative research. Before discussing these methods, we should recall the distinction between statistical parameters and estimates as discussed in Chapter 9. The term parameter refers to an error-free numerical constant that statistically describes a numerical characteristic of a population. We cannot in fact know the value of the parameter exactly but can estimate it statistically. However, the calculated statistical estimate will be imperfect due to random and systematic error. Methods presented in this chapter address the problem of random error in statistical estimates but have no influence on the problem of systematic error.
Let us introduce new notation in this chapter that preserves the distinction between parameters and estimates by using Greek letters to denote parameters, while using the same Greek letter with an overhead hat (^)] to denote analogous estimators. For example, φ is used to denote a risk ratio parameter, while is used to denote a risk ratio estimator. The only exception to the convention of using Greek characters for parameter is seen in using the Roman letter p to denote the binomial proportion parameter and the symbol to denote its estimator. This exception is adopted to avoid confusion with the constant π and maintain conventions of accepted biostatistical usage.
Population and sample
Understanding the population that is the source of the parameter we want to learn about is fundamental to statistical inference. This topic was introduced in Section 9.2 and bears further comment here. Statistical populations may be real or hypothetical, depending on the object of the research. Real populations are composed of a finite number of potential observations, as describes the situation when conducting a prevalence survey. In contrast, hypothetical populations represent an infinite number of potential observations from which to sample. While imagining a real population is relatively undemanding, conceiving of a hypothetical population is not always self-evident. The difference between a real population and hypothetical population can be made more tangible by considering the Framingham Heart Study. While it is true that the Framingham cohort is real and exists in eastern Massachusetts (United States), most inferences based on these data are used to learn about general physiologic relations between various risk factors (e.g., hypercholesterolemia) and heart disease. Since these generalizations go beyond the present Framingham population, data are seen as representing a sample from a hypothetical “superpopulation” of general causal phenomena.
Statistical inference
Regardless of whether a study is based in a real population or hypothetical population, the two traditional methods of inferring statistical parameters are estimation and hypothesis (significance) testing. Examples will introduce their use. It is common for epidemiologists to want to learn about the prevalence of a condition (e.g., smoking) in a population based on the prevalence of the condition in a sample. In a given study, the inference may be “25% of the population smokes” (point estimation). In addition, estimation takes the form of an interval, such as “we are 95% confident that between 20% and 30% of the population smokes” (interval estimation). Finally, the epidemiologist might want to test the claim that the prevalence of smoking has changed. Under such instances, a categorical “yes” or “no” conclusion is sought (hypothesis testing). Our first order of business is to present methods of calculating confidence intervals.
13.2 Confidence intervals
Estimation
Estimation—the process of using data to “locate” population parameters—comes in two forms: point estimation and interval estimation. For example, when in the past we had calculated a risk difference of, say, 10%, this was the point estimate for risk difference parameter δ. Interval estimation surrounds the point estimate with a margin of error, thus creating a confidence interval (Figure 13.1). For example, a 95% confidence interval for a risk difference might be 0.10 ± .02. This is written (0.08, 0.12), where 0.08 is the lower confidence limit (LCL) of the interval and 0.12 is the upper confidence limit (UCL). The width of this confidence interval—simply its upper limit minus its lower limit—is a measure of the estimate’s precision. (For risk ratios and rate ratios, the ratio of the upper and lower confidence limits is a comparable measure of precision.) Wide confidence intervals indicate low precision, and narrow confidence intervals indicate high precision. Large studies tend to derive narrow confidence intervals (precise estimates). Small studies tend to derive wide confidence intervals (imprecise estimates). Formulas and illustrations follow.
Confidence intervals for proportions (incidence proportion and prevalence)
Both incidence proportions (risks) and prevalences are mathematical proportions. Assuming observations are independent, the number of cases in a given sample will follow a binomial distribution with parameters n and p, where n represents the sample size and p represents the incidence proportion or prevalence parameter. The estimator of this parameter, denoted (“p hat”), is simply the sample proportion:
13.1
where A represents the number of cases in the sample and n represents the sample size. For example, a sample of 57 people in which 17 smoke demonstrates = 15/57 = 0.298 for the characteristic of smoking.
Let = 1 − . When ≥ 10, a normal approximation to the binomial can be used to calculate a confidence 95% interval for p as follows:
13.2
where represents the estimated standard error of the proportion:
Other levels of confidence for this and other confidence intervals in this chapter based on normal error distributions are calculated by replacing the 1.96 in formulas with 1.645 for 90% confidence intervals and 2.58 for 99% confidence intervals. For example, a 90% confidence interval for p is given by ± (1.645)(). A 99% confidence interval for p is given by ± (2.58)().
For the random sample of n = 57 with A = 17, = 17/57 = 0.2982, = 1 − 0.2982 = 0.7018, and = (57)(0.2982)(1 − 0.2982) = 11.9. Therefore, the normal approximation to the binomial (large sample method) can be applied. The estimated standard error of the proportion is and a 95% confidence interval for p = 0.2982 ± (1.96)(.0606) = 0.2982 ± 0.1188 = (0.179, 0.417).
Small sample methods
In small samples the above method should be voided in preference for either a quadratic method (Formula 13.3) or exact binomial method. The lower confidence limit (pLCL) and upper confidence limit (pUCL) for a 95% confidence interval for p by the quadratic method (Fleiss, 1981, Section 1.4) are
13.3
Coverage of the method used to calculate exact binomial confidence limits is beyond the scope of this book. Fortunately, there exist reliable public domain computer programs that use this method to compute confidence limits. The two that have been emphasized in this book are OpenEpi.com (Dean et al., 2006) and WinPEPI (Abramson, 2011). For example, OpenEpi → Counts → Proportions derives a mid-P exact 95% confidence interval for p of 0.191–0.426.
Confidence intervals for rates
When estimating an incidence rate based on a person-time denominator, the number of cases in a given sample is assumed to follow a Poisson distribution with expectation μ. Let A represent the observed number of cases in a sample comprising T person-years of population experience. The point estimator of rate parameter λ is
13.4
A Fisher’s exact 95% confidence interval for rate parameter λ is calculated in two steps. First, the lower confidence limit (Alcl) and the upper confidence limit (Aucl) for the expected number of cases are determined using the Poisson limits in Appendix 1. Then, these confidence limits are expressed relative to the person-time (T) in the sample:
13.5
Suppose 25 deaths are found in 4054 person-years of observation. The mortality rate,
This may be expressed with a population base of 1000 as 6.2 per 1000 person-years. A 95% confidence interval for the rate is calculated in two steps. The confidence limits for the number of cases is (16.18, 36.90)—see Appendix 1—and a 95% confidence for λ is (16.18/4054 person-years, 36.90/4054 person-years) = (0.0040, 0.0091) year−1. Expressed with a population base of 1000, the 95% confidence interval for λ is (4.0, 9.1) per 1000 person-years. The same results can be derived with OpenEpi.com → Person-time → 1 Rate → Fisher’s exact.
Confidence intervals for proportion ratios (risk ratios and prevalence ratios)
The ratio of two incidence proportions is a risk ratio. The ratio of two prevalences is a prevalence ratio. Let φ represent a proportion ratio parameter (either a risk ratio or prevalence ratio). The point estimator of the proportion ratio parameter is
13.6
where represents the proportion in the exposed group and represents the proportion in the nonexposed group. Table 13.1 displays additional notation. The sampling distribution of the natural logarithm of is approximately normal. Therefore, we transform the risk ratio estimate into a natural logarithmic (ln) scale before calculating its confidence interval. (Use the “ln” key on your calculator.)
The standard error of the natural log of the proportion ratio is
13.7
The 95% confidence interval for ln φ is given by
13.8
Antilogs (exponents) of the confidence limits are taken to convert them into a nonlogarithmic scale. (The antilog key may be labeled e on your calculator or might be accessed by pressing “inverse ln.”)
The drug cytarabine is used for bone marrow ablation in preparation for transplantation. Even under the best of circumstances, this drug is associated with a high risk of cerebellar toxicity. There was a suspicion that the drug produced by a generic manufacturer presented a greater risk than the innovator product to those who used it. Table 13.2 contains data from a study that addressed this question. Based on these data, the risk of toxicity with the generic drug () = 11/25 = 0.440. The risk of cerebellar toxicity with the innovator drug () = 3/34 = 0.088. The risk ratio estimate = 4.99 and ln = ln(4.99) = 1.607. The standard error estimate is
and a 95% confidence interval for ln φ = 1.607 ± (1.96)(0.5964) = (0.438, 2.776). The 95% confidence limits for our risk ratio = e0.438,2.776 = (1.55, 16.05). An identical result is derived with OpenEpi.com → Counts → Two-by-Two Table.
Confidence intervals for rate ratios
Let ω represent the rate ratio parameter. The rate ratio estimator is
13.9
where represents the rate in the exposed group (= A1/T1) and represents the rate in the nonexposed group (= A0/T0). With moderate to large samples, the random error distribution of the natural log of the estimator (ln ) is approximately normal with a standard error of
13.10
The 95% confidence interval for ln ω is calculated in the usual manner:
13.11
The 95% confidence limits for ω is derived by taking the exponents of these limits.
Table 13.3 presents data from a study on physical fitness and mortality (Blair et al., 1995). This study found 25 deaths in 4054 person-years in men who went from the physically unfit to the physically fit category ( = 25/4054 person-years = 0.0062 year−1). It found 32 deaths in 2937 person-years in the men who remained in the unfit category ( = 32/2937 person-years = 0.0109 year−1). Thus, and ln() = ln(0.57) = −0.562. The standard error on a logarithmic scale (base e) is and a 95% confidence interval for ln ω = −0.562 ± (1.96)(0.2669) = −0.562 ± 0.523 = (−1.085, −0.039). The 95% confidence limits for ω is e(−1.085,−0.039) = (0.34, 0.96).
We can use OpenEpi.com → Person-time → Compare 2 rates to derive exact confidence intervals for rate ratios. WinPEPI → Compare2 → D. Rates with person-time denominators can also be used for this purpose. The 95% confidence interval by the Mid-P exact method is (0.33, 0.96).
Confidence intervals for proportion differences (risk differences and prevalence differences)
This section considers differences between proportions (risk differences and prevalence differences). To avoid redundancy, only the risk difference will be discussed.
Let δrisk denote the risk difference parameter. The point estimator of δrisk is
13.12
where represents the incidence proportion in the exposed group and represents the incidence proportion in the nonexposed group.
The standard error of the risk difference is
13.13
A 95% confidence interval for δrisk is calculated with the formula
13.14