Biostatistics PDF

BIOSTATISTICS CHAPTER 14 BIOSTATISTICS PURPOSE OF BIOSTATISTICS Statistics involves the collection and analysis of all types of data, from the average number of cars on a freeway to the blood pressure reduction expected from a calcium channel blocker. When statistics on people and animals, the stat...

BIOSTATISTICS CHAPTER 14 BIOSTATISTICS PURPOSE OF BIOSTATISTICS Statistics involves the collection and analysis of all types of data, from the average number of cars on a freeway to the blood pressure reduction expected from a calcium channel blocker. When statistics on people and animals, the statistical analysis is called biostatistical analysis, or simply, biostatistics. A basic understanding of biostatistics is required to interpret studies in medical and pharmacy journals, such as the New England Journal of Medicine or Pharmacotherapy. Simple formulas and definitions, described here, prepare the reader to interpret most journal articles and feel confident tackling common practice-based situations, such as: A physician asks if a patient should be switched from standard of care treatment to a new drug based on the relative risk reduction reported in a clinical trial. A patient taking warfarin wants to know if he should switch to Xarelto because he saw a commercial claiming that "it prevents DVT or PE in 98% of patients." STEPS TO JOURNAL PUBLICATION ""':·~. 208 The path to publication for the classic type of research study is shown in the figure on the next page. A study manuscript (description of the research, with results) can be submitted for publication in a professional, peer-reviewed journal. The editor of the journal selects potential publications and sends them to experts in the topic area for peer review. Peer review is intended to assess the research design and methods, the value of the results and conclusions to the field of study, how well the manuscript is written, and whether it is appropriate for the readership of the journal. The reviewers make a recommendation to the editor to either accept the article (usually with revisions) or reject it. Data that contradicts a previous recommendation, or presents new information, can change treatment guidelines. RxPREP 2022 COURSE BOOK\ RxPR EP ©2021, ©2022 ih' ,.. ,(°' BEGIN with ;1 • "··• · ' \I)/' . "" RESEARCH QUESTION Write a null hypothesis to answer the research question, such as: New drug is not as effecti ve as current drug. DESIGN COLLECT the DATA , ~-~ ) ENROLL the the STUDY ·" Q,,::2., SUBJECTS Assign to a treatment group or control group, or identify subjects belonging to a cohort or other group. Is it randomized, placebo-controlled, a case-control or other type of study? Prospectively (going into the future for a set period of time) or retrospectively (looking back in time using medical records). ANALYZE the DATA Enter the data into statistical software; assess the results (e.g., risk reductions, confidence intervals). iKl.1 (,lRxPrep Rl'J iStock.com/dilyanah ORGANIZATION OF A PUBLISHED CLINICAL TRIAL A published clinical trial begins with an abstract that provides a brief summary of the article. The introduction to the study comes next, which includes background information, such as disease history and prevalence, and the research hypothesis. This is followed by the study methods, which describe the variables and outcomes, and the statistical methods used to analyze the data. The results section includes figures, tables and graphs. A reader needs to interpret basic statistics and common graphs in order to understand the study results. The researchers conclude the article with an interpretation of the results and the implications for current practice. TYPES OF STUDY DATA When data points, or values, are collected during a study, they can be analyzed to determine the degree of difference between groups, or some other type of association. The statistical tests used to perform the analysis depend on the type of data. CONTINUOUS DATA Continuous data has a logical order with values that continuously increase (or decrease) by the same amount (e.g., a HR of 120 BPM is twice as fast as a HR of 60 BPM). The two types of continuous data are interval data and ratio data. The difference between them is that interval data has no meaningful zero (zero does not equal none) and ratio data has a meaningful zero (zero equals none) . The Celsius temperature scale is an example of interval data because it has no meaningful zero (0°C does not mean no temperature; it is the freezing point of water). Heart rate is an example of ratio data; a HR of O BPM is cardiac arrest (zero equals none; the heart is not beating). DISCRETE (CATEGORICAL) DATA The two types of discrete data, nominal and ordinal, have categories, and are sometimes called categorical data. Nominal and name are derived from the same word; with nominal data, subjects are sorted into arbitrary categories (names), such as male or female (o = male, 1 = female or O = female, 1 = male). It is sometimes described as "yes/no" data. Ordinal comes from the word order; ordinal data is ranked and has a logical order, such as a pain scale. In contrast to continuous data, ordinal scale categories do not increase by the same amount; a pain scale rating of 4 is worse than a pain scale rating of 2, but it does not mean that there is twice as much pain. CONTINUOUS DATA DISCRETE (CATEGORICAL) DATA Data is provided by some type of measurement which has unlimited options (theoretically) of continuous values Data fits into a limited number of categories RAT IO DATA Equal difference between values, with a true, meaningful zero (0= NONE) Examples: age, height, weight, time, blood pressure INTERVAL DATA NOMINAL DATA ORDINAL DATA Equal difference between values, but without a meaningful zero Categories are in an arbitrary order Categories are ranked in a logical order, but the difference between categories is not equal (OtcNONE) Examples: Celsius and Fahrenheit temperature scales Order of categories does not matter Examples: gender, ethnicity, marital status, mortality Order of categories matters Examples: NYHA Functional Class I-IV; 0-10 pain scale r(l'; (i'!RxPrep 209 I BIOSTATI STI CS 14 SUMMARIZING THE DATA MEASURES OF CENTRAL TENDENCY • Descriptive statistics provide simple summaries of the data. The typical descriptive values are called the measures of central tendency, and include the mean, the median and the mode (see Study Tip Gal below for mean, median and mode calculation examples) . • Mean: the average value; it is calculated by adding up the values and dividing the sum by the number of values. The mean is preferred for continuous data that is normally distributed (described below}. Median: the value in the middle when the values are arranged from lowest to h ighest. When there are two center values (as with an even number of values), take the average of the two center values. The median is preferred for ordinal data or continuous data that is skewed (not normally distributed}. • Mode: the value that occurs most frequently. The mode is preferred for nominal data. SPREAD (VARIABILITY) OF DATA Two common methods of describing the variability, or spread, in data are the range and the standard deviation (SD) . Range: the difference between the highest and lowest values. Standard deviation (SD}: indicates how spread out the data is, and to what degree the data is dispersed away from the mean (i.e., spread out over a smaller or larger range} . A large number of data values close to the mean has a smaller SD. Data that is highly dispersed has a larger SD. Example ____ _______ The diastolic blood pressure (DBP mmHg) reduction for 9 patients in a trial 3, 2,3,8,6, 3,4,4,3 ::;:::::_ Put the ~umbers in order: 2+3+3+3+ 3 +4+4+6+8 l The MODE is 3 the value that occurs most frequently _ ' j The MEAN is 4 (36 + 9 = 4). The MEDIAN is 3, the value in the middle of the ranked (ordered) list. The RANGE is the highest value (8; minus the smallest value (2). The range in DBP reduction is 6. GAUSSIAN (NORMAL) DISTRIBUTIONS Lal'ge sample sets of continuous data tend to form a Gaussian, or "normal" (bell-shaped), distribution (see the figure at the top of the next page) . For example, if a researcher collects 5,000 blood pressure measurements (continuous data) from Idaho residents and plots the values, the graph would form a normal distribution. Characteristics of a Gaussian Distribution When the distribution of data is normal, the curve is symmetrical (even on both sides), with most of the values closer to the middle. Half of the values are on the left side of the curve, and half of the values are on the right side. A small number of values are in the tails. When data is normally distributed: A The mean, median and mode are the same value, and are at the center point of the curve. !\ 68% of the values fall within 1 SD of the mean and 95% of the values fall within 2 SDs of the mean. Normal Distribution Shapes The examples to the right show how the curve of normally distributed data changes based on the spread (or range) of the data. The curve gets taller and skinnier as the range of data narrows. The curve gets shorter and wider as the range of data widens (or is more spread out}. 210 ,J A A ,j, Range of Range of data narrows data widens Rx PREP 20 22 CO UR SE BOOK I RxPREP ©2 0 2 1, © 2022 GAUSSIAN (NORMAL) DISTRIBUTION ; I l l ! i I i ! I I I 2.5% Outliers- very low values .. I I I 35D i 25D 1SD mean median mode 1 SD 2 SD 3 ~D Outliers- : very high values ©RxPreµ SKEWED DISTRIBUTIONS Data that are skewed do not have the characteristics of a normal distribution; the curve is not symmetrical, 68% of the values do not fall within 1 SD from the mean and the mean, median and mode are not the same value. This usually occurs when the number of values (sample size) is small and/or there are outliers in the data. Outliers (Extreme Values) An outlier is an extreme value, either very low or very high, compared to the norm. For example, if a study reports the mean weight of included adult patients as 90 kg, then a patient in the same study with a weight of 40 kg or 186 kg is an outlier. When there are a small number of values, an outlier has a large impact on the mean and the data becomes skewed. In this case, the median is a better measure of central tendency. In the examples to the right, the median is right in the middle of the data and is not affected by outliers. The distortion of the central tendency caused by outliers is decreased by collecting more values; as the number of values increases, the effect of outliers on the mean decreases. More low Negative (left) skew Positive (right) skew More high values Skew Refers to the Direction of the Tail Data is skewed towards outliers. When there are more low values in a data set and the outliers are the high values, data is skewed to the right (positive skew). When there are more high values in the data set and the outliers are the low values, the data is skewed to the left (negative skew). DEPENDENT AND INDEPENDENT VARIABLES A variable in a study is any data point or characteristic that can be measured or counted. Examples include age, gender, blood pressure or pain. Variables can be clinical endpoints such as death, stroke, hospitalization or an adverse event, or they can be intermediate (or surrogate) endpoints used to assess an outcome, such as measuring serum creatinine to assess the degree of renal impairment. Independent variables are changed by the researcher The dependent variables can be affected by the Independent variables Examples: drugs, drug dose/s, placebos, patients included (e.g., age, gender, comorbid conditions) Examples: HF progression , hemoglobin A1C, blood pressure, cholesterol values, mortality An independent variable is changed (manipulated) by the researcher in order to determine whether it has an effect on the dependent variable (the outcome). Independent variables are the characteristics of the subject groups (treatment and control) selected for inclusion (e.g., age, gender, presence or absence of hypertension, diabetes or other comorbid conditions), or any other characteristic that could have an effect on the dependent variable. 211 14 I BIOSTAT IS TI CS TESTING THE HYPOTHESIS FOR SIGNIFICANCE If a drug or device manufacturer wants to sell their product and make money, they will want research data that demonstrates that their product is significantly better than (or superior to) the current treatment or a placebo (no treatment). To show significance, the trial needs to demonstrate that the null hypothesis is not true and should be rejected, and the alternative hypothesis can be accepted. The null hypothesis and alternative hypothesis are always complementary; when one is accepted, the other is rejected. THE NULL HYPOTHESIS AND ALTERNATIVE HYPOTHESIS Null means none or no; a null hypothesis (H 0 ) states that there is no statistically significant difference between groups. A researcher who is studying a drug versus a placebo would write a null hypothesis that states that there is no difference in efficacy between the drug and the placebo (drug efficacy= placebo efficacy). The null hypothesis is what the researcher tries to disprove or reject. The alternati ve hypothesis (HA) states that there is a statistically significant difference between the groups (drug efficacy-:;:. placebo efficacy) . The altemative hypothesis is what the researcher hopes to prove or accept. ALPHA LEVEL: THE STANDARD FOR SIGNIFICANCE When investigators design a study, they select a maximum permissible error margin, called alpha (a.). Alpha is the threshold for rejecting the null hypothesis. In medical research, alpha is commonly set at 5% (or 0.05). A smaller alpha value can be chosen (e.g., 1%, or 0.01), but this requires more data, more subjects (which means more expense) and/or a larger treatment effect. ALPHA CORRELATES WITH THE VALUES IN THE TAILS WHEN DATA HAS A NORMAL DISTRIBUTION 99.7% of all values are within 3 SDs on each side of the mean. 95% of all values are within 2 SDs on each side of the 111e;u1. Y X 2.5% on eacn side= 5%. ,,th w ith alpha ; 0 .05 = 0.01 Ulf Comparing the P-Value to Alpha Once the alpha value is determined, statistical tests are performed to compare the data, and a p-value is calculated. The p-value is compared to alpha. If alpha is set at 0 .05 and the p-value is less than 0 .05, the nuU hypothesis is rejected, and the result is termed stat istically significant. If the p-value is greater than or equal to alpha (p 0.05), the study has failed to reject the null hypothesis, and the result is not statistically significant. The Null Hypothesis: Reject or Accept? Write a Determine the Null Hypothesis Alpha Value (such as 5%) Run study, co11ect data, analyze My product =other product/placebo. Goal: reject the null hypothesis. data with statistical tests to calculate p-values. Compare p-value to alpha value. < p-vatue < alpha REJECT the Null Hypothesis Alternative Hypothesis Accepted (e.g.. p < 0.05) $$$$$ p•value alpha ACCEPT the Null Hypothesis Alternative Hypothesis Rejected DARN! (e.g.. p O.OS) (~)RxPrep 212 RxPREP 2022 COURSE BOOK I RxPREP ©2021, ©2022 CONFIDENCE INTERVALS A confidence interval (Cl) provides the same information about significance as the p-value, plus the precision of the result. Alpha and the CI in a study will correlate with each other. 1-a J If alpha is 0.05, the study reports 95% Cis; an alpha of 0.01 corresponds to a CI of 99%. The relationship between alpha, the p-value and the CI is described in the table here and in the figure on the previous page. P-VALUE ALPHA :. MEANING 0.05 I 2: o.o5 0.05 <0.05 95% probability (confidence) that the conclusion is correct; less than 5% chance it's not. 0.01 <0.01 99% probability (confidence) that the conclusion is correct; less than 1% chance it's not. 0.001 <0.001 99.9% probability {confidence) that the conclusion is correct; less than 0.1% chance it's not. Not statistically significant -------------- Statistically _ _ _ _ _ __ _Si!niflca!!!_ The values In the Cl range are used to determine whether significance has been reached Determining statistical significance using the Cl alone (without a p-value) Is required for the exam COMPARING DIFFERENCE DATA(MEANS) Difference data is based on subtraction [e.g., the difference in t,. FEV1 between roflumilast and placebo (below) was 38 (46 - 8 = 38)) The result is statistically significant if the Cl range does not include zero (e.g., zero is not present in the range of values); for example: D The 95% Cl for the difference int,. FEV1 (18-58 ml) does not include the result is statistically significant D The 95% Cl for the difference int,. FEV1/FVC (-0.26-0.89%) includes the result is not statistically significant DRUG" (N = 745) LUNG FUNCTION . 46 t,.fEV1 (ml ) t,. FEV1/FVC (%) 0.314 PLACEBO (N = 745) DIFFERENCE (95% Cl) 8 38 (18- 58) 0.001 0.313 (-0.26-0.89) 'Roflumilast 95% Cl for LI FEVl does NOT include ("cross") zero STATISTICALLY SIGNIFICANT 95% Cl for LI FEVl/FVC DOES include ("cross") zero NOT STATISTICALLY SIGNIFICANT -0.26 1---+--------i 0.89 18 1--- - - - - - 58 -20 0 40 20 -1 60 -0.5 1 0.5 0 COMPARING RATIO DATA (RELATIVE RISK, ODDS RATIO, HAZARD RATIO) Ratio data is based on division [e.g., the ratio of severe exacerbations between roflumilast and placebo (below) was 0.92 (0.11/0.12 = 0.92)] The result is statistically significant if the Cl range does not include one (e.g., one is not present in the range of values); for example: D The 95% Cl for the relative risk of severe exacerbations (0.61-1.29) includes the result is not statistically significant D The 95% Cl for the relative risk of moderate exacerbations (0.72-0.99) does not include EXACERBATIONS" DRUG'" (N = 745) PLACEBO (N = 745) Severe 0.1 1 - - - - - - - - - , - 0.12 Moderate 0.94 1.11 ----- the result is statistically significant RELATIVE RISK (95% Cl) 0.92 (0.61-1.~ I o.a5 {o.72-0.99) _ _ __, 'Mean rate, per patient per year "Roffumi/ast I l 95% Cl for severe exacerbation DOES include C'cross'1) one NOT STATISTICALLY SIGNIFICANT 0.75 1 1.25 does NOT include ("cross") one STAT!STICALLY SIGNIFICANT 0.72 1 - - - - - - - - - - - - - - - , 0.99 0.61 1 - - - - - - - + - - - - - i 1.29 0.5 95% Cl for moderate exacerbation 1.5 0.7 0.8 0.9 1.1 213 14 I BIOS TATIST ICS Confidence Intervals and Estimation (Extent and Variability in the Data) The goal of the majority of medical research is to use the study results to promote the procedure or drug for use in the general population of patients with the same medical condition. Clinicians need to understand how their patients would benefit. The CI includes the treatment effect and the range; both are helpful in estimating the effect on others. A CI can be written in slightly different formats. For example, a study comparing metoprolol to placebo finds a 12% absolute risk reduction (ARR) in heart failure progression, with a 95% CI range of 6 - 35%. This can be written as ARR 12% (95% CI 6% - 35%) or as decimals, with commas in the range, such as ARR 0.12 (0.95 CI 0.06, 0.35). The CI indicates that you are 95% confident that the true value of the ARR for the general (or true) population lies somewhere within the range of 6%35%, with some values as low as 6% and others as high as 35%. A narrow CI range implies high precision, and a wide CI range implies poor precision. If the reported CI range was 4% - 68%, the true value would still be within the range, but where? The range is wider, and therefore less precise. Cardiologists who interpret the results for their patients would not know whether to expect a result closer to 4% or 68%. A large range correlates to a large dispersion in the data. A narrower range is preferable. In some studies, specific patient types will cause a wider distribution in data. For example, fibrates are used to lower triglyceride levels; they cause a greater reduction in patients with higher triglycerides. The consideration of where the patient is likely to fall within the range will become part of the assessment of the individual's baseline risk. TYPE I AND TYPE II ERRORS Consider what would happen if a drug manufacturer developed and marketed a new drug as better for heart failure than the standard of care, when in fact the new, expensive drug has similar benefits to the old drug (it is not better at all) . The null hypothesis stated that the new drug and the old drug are equal. The statistical tests found a significant benefit with the new drug, and the null hypothesis was rejected when it should have been accepted. Type I Errors: False-Positives In the scenario described above, the conclusion was wrong and a type I error was made. The alternative hypothesis was accepted and the null hypothesis was rejected in error. The probability, or risk, of making a type I error is determined by alpha and it relates to the confidence interval. r 7 Cl- ,.- When alpha is 0.05 and a study result is reported with~ 0.05, it is statistically significant and the probability of a ~ I error (making the wrong conclusion) is < 5%. You are 95% confident (0.95 = 1- 0.05) that your result is correct and not due to chance. Type II Errors: False-Negatives The probability of a type II error, denoted as beta(~), occurs when the null hypothesis is accepted when it should have been rejected. Beta is set by the investigators during the design of a study. It is typically set at 0.1 or 0.2, meaning the risk of a type II error is 10% or 20%. The risk of a type II error increases if the sample size is too small. To decrease this risk, a power analysis is performed to determine the sample size needed to detect a true difference between groups. Study Power Power is the probability that a test will reject the null hypothesis correctly (i.e., the power to avoid a type II error) . Power = 1 - ~- As the power increases, the chance of a type II error decreases. Power is determined by the number of outcome values collected, the difference in outcome rates between the groups, and the significance (alpha) level. If beta is set at 0.2, the study has 80% power (there is a 20% chance of missing a true difference and making a type II error). If beta is set at 0.1, the study has 90% power. A larger sample size is needed to increase study power and decrease the risk of a type II error. HO ACCEPTED H0 is TRUE (There IS a difference between groups) _I Correct Conclusion (NO difference between groups) H0 is FALSE Type 11 Error Committed FALSE NEGATIVE 214 l · a (type I error) H 0 REJECTED Type I Error Committed FALSE POSITIVE Correct Conclusion RxPREP 2022 CO UR SE BOO K I Rx PR EP ©2 0 2 1, ©2 0 22 RISK In healthcare, risk refers to the probability of an event {how likely it is to occur) when an intervention, such as a drug, is given. The lack of intervention is measured as the effect in the placebo {or control) group. RELATIVE RISK (OR RISK RATIO) The relative risk (RR) is the ratio of risk in the exposed group {treatment) divided by risk in the control group. RR Formula __ N_u_m_b_er_o_f_s_ub _i_ec_t_s _in_g_r_ ou_p_w ~_ i~ t h=an_-_u=nf=a_ v-o~r_ab_l_ e _ev_e_n_t __ j Risk L Total number of subjects in group - ~-- F Risk in trea~m~ nt group ~ 7 in control group _ l J RR Calculation A placebo-controlled study was performed to evaluate whether metoprolol reduces disease progression in patients with heart failure (HF). A total of 10,lll patients were enrolled and followed for 12 months. What is the relative risk of HF progression in the metoprolol-treated group versus the placebo group? Calculate the risk of HF progression in each group. Then calculate RR. METOPROLOL CONTROL N = 5,123 N = 4,988 HF progression 823 Metoprolol Risk 823 5,123 1,397 "'0. 16 I_ I Control Risk 1,397 - - - - - =0.28 4,988 0.16 R R = - - - - = 0.57 0.28 X 100 = 57% Answer can be expressed as a decimal or a percentage; the exam question will specify with instructions RR Interpretation RR= 1 (or 100%) implies no difference in risk of the outcome between the groups. RR> 1 (or 100%) implies greater risk of the outcome in the treatment group. RR< 1 (or 100%) implies lower risk {reduced risk) of the outcome in the treatment group. In the metoprolol study, the RR of HF progression was 57%. Patients treated with metoprolol were 57% as likely to have progression of disease as placebo-treated patients. INTERPRETING THE RELATIVE RISK (RR) RR= 1 Equal risk between intervention (treatment) & control groups Intervention had no effect RR< 1 The treatment .J, the risk of the outcome (endpoint) (e.g., less HF progression in the treatment group) A RR of 0.5 indicates there is 50% (0.5 is 50% less than 1) risk in the treatment group as compared to the risk in the control group \ RR> 1 The treatment i the risk of the outcome (endpoint) (e.g., more HF progression in the treatment group) A RR of 1.5 indicates there is 50% (1.5 is 50% greater than 1) increased risk in the treatment group as compared to the risk in the control group [m ~l RxPrep 215 14 I BIO STATI STI CS RELATIVE RISK REDUCTION The RR calculation determines whether there is less risk (RR< 1) or more risk (RR> 1). The r elative risk r eduction (RRR) is calculated after the RR and indicates how much the risk is reduced in the treatment group compared to the control group. RRRFormula I (% risk in control group - % risk in treatment group) % risk in the control group ------- Decimals ar percentages may be used for risks I J 1- RR* or l J 'Must use decimal form of RR RRR Calculation Using the risks previously calculated for HF progression in the treatment and control groups (metoprolol: 16% and placebo: 28%), calculate the RRR of HF progression. (28% - 16%) RRR or 0.43 RRR 1 - 0.57 0.43 28% Answer can be expressed as a decimal or percentage; the exam question will specify with instructions RRR Interpretation The RRR is 43%. Metoprolol-treated patients were 43% less likely to have HF progression than placebo-treated patients. INTERPRETING THE RELATIVE RISK REDUCTION (RRR) RR 0.57 + RRR 0.43 = 1 RRR rnr,toprolr,I p,1twnts were 43% ~l!t.£J Y (than the control group) to suifer from HF - - progr,i~STon . ·· - RR RRR Therefore ·- ..___ AS likely (vs. the control) LESS likely (vs. the control) RR+ RRR = 100% _ _ _ __ __ _ _ _ _ _ ___, fll1 ©RxPrep ABSOLUTE RISK REDUCTION A clinician is listening to a presentation on a drug. The drug manufacturer representative reports that the drug causes 48% less nausea than the standard treatment. The result sounds great; the clinician asks the pharmaceutical representative: what is the absolute risk reduction (ARR)? The RR and RRR provide relative (proportional) differences in risk between the treatment group and the control group; they have no meaning in terms of absolute risk. Absolute risk reduction is more useful because it includes the reduction in risk and the incidence rate of the outcome. If the risk of nausea is reduced, but the risk was small to begin with (perhaps the drug caused very little nausea}, the large risk reduction has little practical benefit. It is best if a study reports both ARR and RRR, and for clinicians to understand how to interpret the risk for their patients. If the ARR is not reported, it is possible that the risk reduction, in terms of a decrease in absolute risk, is minimal. ARR Formula ARR 216 --------(% risk in control group) - (% risk in treatment group) 7 ) RxPREP 20 22 COUR SE BOOK I Rx PREP ©202 1, ©2 0 22 ARR Calculation Using the risks previously calculated for HF progression in the metoprolol study, calculate the ARR of HF progression. Metoprolol Risk 823 Control Risk 1,397 = 0.16 5,123 4,988 = 0.28 • ARR= 0.28 - 0.16 = 0.12 x 100 = 12% Answer can be expressed as a decimal or a percentage; the exam question will specify with instructions ARR Interpretation The ARR is 12%, meaning 12 out of every 100 patients benefit from the treatment. Said another way, for every 100 patients treated with metoprolol, 12 fewer patients will have HF progression. An additional benefit of calculating the ARR is to be able to use the inverse of the ARR to determine the number needed to treat (NNT) and number needed to harm (NNH). These concepts are discussed next. INTERPRETING THE ABSOLUTE RISK REDUCTION (ARR) Placebo risk minus the Treatment risk = ARR The absolute risk reduction is the true difference in risk between the treatment and the placebo groups. Said another way, the ARR is the net effect (benefit) beyond the effect obtained from a placebo. fffi © RxPrep NUMBER NEEDED TO TREAT OR HARM NNT and NNH help clinicians answer the question: how many patients need to receive the drug for one patient to get benefit (NNT) or harm (NNH)? This information, taken into consideration with the patient's individual risk, helps guide decisions. NUMBER NEEDED TO TREAT NNT is the number of patients who need to be treated for a certain period of time (e.g., one year) in order for one patient to benefit (e.g., avoid HF progression). NNTFormula r 1 NNT ,, - - - -- -- - or I __!__-l (risk in control group) - (risk in treatment group)* *Risk and ARR are expressed as decimals NNT Calculation The ARR in the metoprolol study was 12%. The duration of the study period was one year. Calculate the number of patients that need to be treated with metoprolol for one year in order to prevent one case of HF progression. NNT = 8.3, rounded up to 9* 0.12 *Numbers greater thon a whole number are rounded up NNT Interpretation For every 9 patients who receive metoprolol for one year, HF progression is prevented in one patient. 217 14 I BIOS TATI STI CS NUMBER NEEDED TO HARM NNH is the numbe1· of patients who need to be treated for a certain period of time in order for one patient to experience harm. NNT and NNH are calculated with the same formula (see the NNT formula on the previous page}. There are two differences: • 1. NNT is rounded !:!P, and NNH is rounded down (see Study Tip Gal). 2. The absolute value of the ARR is used with NNH, as shown in the following example. NNH Calculation A study evaluated the efficacy of clopidogrel versus placebo, both given in addition to aspirin, in reducing the risk of cardiovascular death, MI and stroke. The study reported a 3.9% risk of major bleeding in the treatment group and a 2.8% risk of major bleeding in the control group. Normal rounding rules do not apply: For NNT: O Anything greater than a whole number, round up to the next whole number. This avoids overstating the potential benefit of an intervention. O Example: NNT of 52.1 to 53 n- For NNH: O Anything greater than a whole number, round down to the nearest whole number. This avoids understating the potential harm of an intervention. O Example: NNH of 41.9 round down to 41 ARR== 2.8%- 3.9% = -1.1%; the absolute value is the difference between the two groups. There is a 1.1% higher risk of major bleeding in the treatment group. 90.9, rounded down to 90* NNH O.Dll ' Numbers greater than a whole number are rounded do wn NNH Interpretation One additional case of major bleeding is expected to occur for every 90 patients taking clopidogrel instead of placebo. ODDS RATIO AND HAZARD RATIO ODDS RATIO Odds are the probability that an event will occur versus the probability that it will not occur. Case-control studies, described in the Types of Medical Studies section, are not suitable for relative risk calculations. In case-control studies, the odds ratio is used to estimate the risk of unfavorable events associated with a treatment or intervention. Case-control studies enroll patients who have a clinical outcome or disease that has already occurred (e.g., lung cancer}. The patient medical charts are reviewed retrospectively (in the past) to search for possible exposures (e.g., smoking) that increased the risk of the clinical outcome or disease. In this case, the odds ratio (OR) is used to calculate the odds ofan outcome occurring with an exposure, compared to the odds of the outcome occurring without t h e exposure. ORs are used most commonly with case-control studies but can be used in cohort and cross-sectional studies. OR Formula EXPOSURE/TREATMENT Present Absent OUTCOME PRESENT OUTCOME ABSENT A B t A = # that have the outcome, with exposure B = # without the outcome, with exposure C = # that have the outcome, without exposure D = # without the outcome, without exposure 218 C T D ---, ~ I RxPREP 2022 COURSE BOOK I RxPREP ©2021, ©2022 OR Calculation A case-control study was conducted to assess the risk of falls with fracture (outcome) associated with serotonergic antidepressant (AD) use (exposure) among a cohort of Chinese females 65 years old. Cases were matched with 33,000 controls (1:4, by age, sex and cohort entry date). FALLS W/ FRACTURE (CASES) EXPOSURE/ TREATMENT Serotonergic AD-YES 18,270 ------ 73,517,430 3,259 59,541,930 X 59,541,930 14,730 3,259 14,730 73,517,430 OR= Serotonergic AD-NO X BC = 18,270 -----,- 4,991 4,991 AD FALLS W/O FRACTURE (CONTROLS) = 1.23 Conclusion: serotonergic ADs are associated with a 23% increased risk of falls with fracture (see OR and HR Interpretation below). HAZARD RATIO In a survival analysis (e.g. analysis of death or disease progression), instead of using "risk," a hazard rate is used. A hazard rate is the rate at which an unfavorable event occurs within a short period of time. Similar to RR, the hazard ratio (HR) is the ratio between the hazard rate in the treatment group and the hazard rate in the control group. HR Formula l Hazard rate in the treat~ HR _:_ t g~ Hazard rate in the control group -- - - -1 -- HR Calculation A placebo-controlled study was performed to evaluate whether niacin, when added to intensive statin therapy, reduces cardiovascular risk in patients with established cardiovascular disease. The primary endpoint was the first event of the composite endpoint (death from coronary heart disease, nonfatal myocardial infarction, ischemic stroke, hospitalization for an acute coronary syndrome or coronary or cerebral revascularization). A total of 3,414 patients were enrolled and followed for three years. Calculate the hazard ratio. Primary endpoint Niacin Hazard Rat~ NIACIN PLACEBO N = 1,718 N = 1,696 282 274 1~18 282 J Control Hazard Rate 274 = 0.16 0.16 HA = - - - = l 0.16 = 0.16 1~96 x 100 : 100% Answer can be expressed as a decimal or a percentage; the exam question will specify with instructions Conclusion: there is no benefit to cardiovascular risk when adding niacin to intensive statin therapy (see OR and HR Interpretation below). OR AND HR INTERPRETATION OR and HR are interpreted in a similar way to RR: OR or HR= 1: the event rate is the same in the treatment and control arms. There is no advantage to the treatment. OR or HR> 1: the event rate in the treatment group is higher than the event rate in the control group; for example, a HR of 2 for an outcome of death indicates that there are twice as many deaths in the treatment group. OR or HR< 1: the event rate in the treatment group is lower than the event rate in the control group; for example, a HR of 0.5 for an outcome of death indicates that there are half as many deaths in the treatment group. 219 14 I BIOSTATISTIC S PRIMARY AND COMPOSITE ENDPOINTS The primary endpoint is the main (primary) result that is measured to see if the treatment had a significant benefit. In the metoprolol trial, the primary endpoint was HF progression. Primary Endpoints (distinct and separate) Composite Endpoint (combined into one) A composite endpoint combines multiple individual endpoints into one measurement. This is attractive to researchers, as combining several endpoints increases the likelihood of reaching a statistically significant benefit with a smaller, less costly trial. Death from cardiovascular causes Death from cardiovascular causes or and Nonfatal stroke Nonfatal stroke or and Nonfatal Ml Nonfatal Ml When a composite endpoint is used, each individual endpoint gets counted toward the same (composite) outcome. COMPOSITE ENDPOINTS: CAUTION All endpoints in a composite must be similar in magnitude and have similar, meaningful importance to the patient. For example, the composite endpoint of blood pressure reduction should not be included with heart attack and stroke reduction. The FDA requires each individual endpoint to be measured and reported when a composite endpoint is used. When assessing a composite measurement, it is important to use the composite endpoint value, rather than adding together the values for the individual endpoints. The value of the sum of the individual endpoints may not equal the value of the composite endpoint, since a patient can have more than one non-fatal endpoint during a trial. TYPES OF STATISTICAL TESTS The next step following data collection (and calculation of risks, RR, ARR, HR, etc.) is to analyze the data to check if differences between the treatment and control groups are statistically significant or if there is an association or relationship in the data. Selecting the correct test to analyze the data depends on the type of data and the outcomes measured. CONTINUOUS DATA With continuous data, the type of test used to determine statistical significance depends on the distribution of data (discussed previously). If it is normally distributed, parametric methods are appropriate. If the data is not normally distributed, nonparametric methods are appropriate. T-Tests This is a parametric method used when the endpoint has continuous data and the data is normally distributed. When data from a single sample group is compared with known data from the general population, a one-sample t-test is performed. If a single sample group is used for a pre-/post-measurement (i.e., the patient serves as their own control), a paired t-test is appropriate. A student t-test is used when the study has two independent samples: the treatment and the control groups. For example, a study comparing the reduction in hemoglobin AlC values between metformin and placebo would use an independent or unpaired student t-test. Analysis of Variance Analysis of variance (ANOVA), or the F-test, is used to test for statistical significance when using continuous data with 3 or more samples, or groups. DISCRETE (CATEGORICAL) DATA Chi-Square Test For nominal 01· 01•dinal data, a chi-square test is used to determine statistical significance between treatment groups. For example, if a study assesses the difference in mortality (nominal data) between two groups, or pain scores based on a pain scale (ordinal data), a chi-square test could be used. 220 RxPREP 2022 COURSE BOOK I RxPREP ©2021, ©2022 SELECTING A TEST TO ANALYZE THE DATA Numerical/Continuous Data Discrete/Categorical Data PARAMETRIC TESTS (Data has a normal distribution) NON-PARAMETRIC TESTS (Data does not have a normal distribution) One group One-sample t-test One group Sign test One group Chi-square test Dependent/paired t-test (if one group has before and after measures) Wilcoxon Signed-Rank test (if one group has before and after measures) Wilcoxon Signed-Rank test (if one group has before and after measures) Two groups (e.g. treatment and control groups) Two groups (e.g. treatment and control groups} Independent/unpaired student t-test Mann-Whitney {Wilcoxon Rank-Sum) test Two groups (e.g. treatment and control groups) Chi-square test or Fisher's exact test Mann-Whitney {Wilcoxon Rank-Sum) test (may be preferred for ordinal data) Three or more groups ANOVA (or F-test) Three or more groups Kruskal-Wallis test Three or more groups Kruskal-Wallis test EXAMPLES OF TEST TYPE SELECTION Example 1 A study is performed to assess the safety and efficacy of ketamine-dexmedetomidine (KD) versus ketamine-propofol (KP) for sedation in patients after coronary artery bypass graft surgery. ENDPOINT (MEAN VALUES) Fentanyl dose, mcg Weaning/extubation time, min KP KO 41.94 ± 20.43 152.8 ± 51.2 374.05 ± 20.25 445.23 ± 21.7 Measurements of dose and time are both continuous data. The trial has two independent samples, or groups (KD and KP). An appropriate test is an independent/ unpaired student t-test. If the trial included a third group, ANOVA would be used. Example2 An emergency medical team wants to see if there is a statistically significant difference in death due to multiple drug overdose (OD) with at least one opioid taken, versus no opioid taken. Which test can determine a statistically significant difference in death? ENDPOINT Death (n, %) YES OPIOID NO OPIOID N =250 N = 150 52 (20.8%) T The variable (dead or alive) is nominal. The chi-square test is used to test for significance when there are two groups. 35 (23.3%) CORRELATION AND REGRESSION CORRELATION Correlation is a statistical technique that is used to determine if one variable (such as number of days hospitalized) changes, or is related to, another variable (such as incidence of hospital-acquired infection). When the independent variable (number of hospital days) causes the dependent variable (infections) to increase, the direction of the correlation is positive (increases to the right). When the independent variable causes the dependent variable to decrease, the direction of the correlation is negative (decreases to the right). Different types of data require different tests for correlation. Spearman's rank-order correlation, referred to as Rho, is used to test correlation with ordinal, ranked data. The primary test used for continuous data is the Pearson's correlation coefficient, denoted as r, which is a calculated score that indicates the strength and direction of the relationship between two variables. The values range from -1 to +1, and are described in the figure on the next page. It is not possible to conclude from a correlation analysis that the change in a variable causes the change in another variable. A correlation, whether positive or negative, does not prove a causal t·elationship. 221 14 I BIOSTATISTICS TESTING FOR CORRELATION WITH THE PEARSON CORRELATION COEFFICIENT y • ••• No Correlation The correlation coefficient measures the STRENGTH & DIRECTION of a linear relationship between two variables. shown here as X and Y, on a SCATTERPLOT • r= 0 stronger weak -1 y -0.5 stronger weak Negative correlation Positive correlation t y A perfect downhill (negative) I ' ""' :aeoo•'> 1 0.5 0 t A perfect uphill (positive) I-?~ X 161 ©RxPreµ REGRESSION Regression is used to describe the relationship between a dependent variable and one or more independent (or explanatory) variables, or how much the value of the dependent variable changes when the independent variables changes. Regression is common in observational studies where researchers need to assess multiple independent va1·iables or need to control for many confounding factors. There are three typical types of regressions: 1) linear, for continuous data, 2) logistic, for categorical data, and 3) Cox regression, for categorical data in a survival analysis. SENSITIVITY AND SPECIFICITY Lab and diagnostic tests are used to screen for and diagnose medical conditions. Interpreting sensitivity and specificity correctly is required to answer these two questions concerning the validity of lab or diagnostic test results: • If the result is positive, what is the likelihood of having the disease? • If the result is negative, what is the likelihood of not having the disease? SENSITIVITY, THE TRUE POSITIVE Sensitivity describes how effectively a test identifies patients with the condition. The higher the sensitivity, the better; a test with 100% sensitivity will be positive in all patients with the condition. Sensitivity is calculated from the number of patients who test positive, out of those who actually have the condition (sensitivity is the percentage of "true-positive" results). SPECIFICITY, THE TRUE NEGATIVE Specificity describes how effectively a test identifies patients without the condition. The higher the specificity the better; a test with 100% specificity will be negative in all patients without the condition. Specificity is calculated from the number who test negative, out of those who actually do not have the condition (specificity is the percentage of "true-negative" results). Sensitivity and Specificity Formula TEST RESULT HAVE CONDITION NO CONDITION Positive A B Negative C D Total A+C A = # that have the condition, with a positive test result B = # without the condition, with a positive test result C = # that have the condition, with a negative test result D = # without the condition, with a negative test result 222 B+D A rI Sensitivity A+C I - ---- 100 , D ~ pec: city X B+D J 100 I I ---' RxPREP 2022 COURSE BOOK I RxPREP ©2021, ©2022 Sensitivity and Specificity Calculation and Interpretation The tables below show how sensitivity and specificity is calculated for two lab tests used in the diagnosis of rheumatoid arthritis (RA), cyclic citrulline peptide (CCP) and rheumatoid factor (RF). Based on study data, CCP has a sensitivity of 98% and a specificity of 98% for RA, while RF has a sensitivity of 28% and a specificity of 87%. Using the RF lab test as an example, a sensitivity of 28% means that only 28% of patients with the condition will have a positive RF result; the test is negative in 72% of patients with the disease (and the diagnosis can be missed). A specificity of 87% means that the test is negative in 87% of patients without the disease; but 13% of patients without the disease can test positive (potentially causing an incorrect diagnosis). CCP RESULTS Positive HAVE CONDITION A= 147 Negative C=3 Total A+ C = 150 Sensitivity Specificity NO CONDITION -- RF RESULTS B=9 Positive HAVE CONDITION NO CONDITION A= 21 B = 26 ·r D =441 Negative C = 54 D = 174 B + D = 450 Total A+ C = 75 B+D=200 Sensitivity 21/75 X 100 = 28% 441/450 X 100 = 98% Sped fi city .! 147/ 150 X 100 = 98% 174/200 X 100 = 87% Sensitivity and Specificity Application If an elderly female patient with swollen finger joints is referred to a rheumatologist and lab tests reveal a positive CCP and a positive RF, the positive CCP indicates a very strong likelihood that the patient has RA because it has high sensitivity and specificity (98%). If the RF is positive and the CCP is negative, the rheumatologist would consider the possibility of other autoimmune/inflammatory conditions that could be contributing to swollen joints because of the low sensitivity of RF (28%). INTENTION-TO-TREAT AND PER PROTOCOL ANALYSIS Data from clinical trials can be analyzed in two different ways; intention-to-treat or per protocol. Intention-to-treat analysis includes data for all patients originally allocated to each treatment group (active and control) even if the patient did not complete the trial according to the study protocol (e.g., due to non-compliance, protocol violations or study withdrawal). This method provides a conservative (real-world) estimate of the treatment effect. A per protocol analysis is conducted for the subset of the trial population who completed the study according to the protocol (or at least without any major protocol violations). This method can provide an optimistic estimate of treatment effect since it is limited to the subset of patients who were adherent to the protocol. NONINFERIORITY AND EQUIVALENCE TRIAL DESIGNS The standard design of most trials is to establish that a treatment is superior to another treatment; the researcher wishes to show that the new drug is better than the old drug or a placebo. Perhaps a new treatment is developed that is less expensive or less toxic than the standard of care. Researchers would hope to demonstrate that the new drug is roughly equivalent, or at least not inferior, to the standard of care. Two types of trials are used for this purpose: equivalence and non-inferiority trials. Equivalence trials attempt to demonstrate that the new treatment has roughly the same effect as the old (or reference) treatment. These trials test for effect in two directions, for higher or lower effectiveness, which is called a two-way margin. Non-inferio1•ity trials attempt to demonstrate that the new treatment is no worse than the current standard based on the predefined non-inferiority (delta) margin. The delta margin is the minimal difference in effect between the two groups that is considered clinically acceptable based on previous research. 223 14 I BIOSTATISTICS FOREST PLOTS AND CONFIDENCE INTERVALS Forest plots are graphs that have a "forest" of lines. Forest plots can be used for a single study in which individual endpoints are pooled (gathered together) into a composite endpoint (see figure labeled Pogue, et al. below). More commonly, forest plots are used when the results from multiple studies are pooled into a single study, such as with a meta-analysis (see figure labeled Miller, et al. below). Forest plots provide Cis for difference data or ratio data. Interpreting forest plots correctly can help identify whether a statistically significant benefit has been reached. When interpreting statistical significance using a forest plot: The boxes show the effect estimate. In a meta-analysis, the size of the box correlates with the size of the effect from the single study shown. Diamonds (at the bottom of the forest plot) represent pooled results from multiple studies. The horizontal lines through the boxes illustrate the length of the confidence interval for that particular endpoint (in a single study) or for the particular study (in a meta-analysis). The longer the line, the wider the interval, and the less reliable the study results. The width of the diamond in a meta-analysis serves the same purpose. The vertical solid llne is the line of no effect; a significant benefit has been reached when data falls to the left of the line; data to the right of the line indicates significant harm. The vertical line is set at zero for difference data and at one for ratio data. COMPARING DIFFERENCE DATA The study shown to the right (a meta-analysis by Miller, et al.) uses a forest plot to test for significance with difference data. Recall for difference data, a result is not statistically significant if the confidence interval crosses zero, so the vertical line (line of no difference) is set at zero. Examples (for high-dosage vitamin E): Sludy, YIIH (Reference) All-Cause Mortallly Risk Difference (95% Cl) Low-dosase vitamin E MIN,VIT.AOX, 1999 (35) -a-l Llnxlan A, 1993{36) - ·' SU. VI MAX, 2004 (37) ATBC, 1994 (38, 39) Llnxlan B, 1993 (40) Llnqu, 2001 (41) OISSI, 1999(42) PPP,2001 (43) Totiil(lowd~e) 3rd study (PPS): shows a statistically significant benefit; the data point, plus the entire confidence interval, is to the left of the vertical line and does not cross zero. Hlgh-do588e vitamin E HOPE, 2000 (44) AREDS, 2001 (45) PPS, 19941 (46) VECAT, 200'! (-47) CHAOS, 1996 (8, 9} REACT, 2002 MB) MRC/BHF HPS, 2002 (49) 5th study (CHAOS): the result is not statistically significant; the confidence interval crosses zero. SPACE, 2000 {50) WAVE, 2002 (51) - AOCS, 1997 (51) DAJATO,. ·IJU t5!U 9th study (WAVE): shows a statistically significant harmful outcome; the data point, plus the entire confidence interval, is all to the right of the vertical line and does not cross zero. 1tiU.1tot £1)1 tu1llet Vitamin E Bem1flcJal Vllamln E Hannful Miller, et al. Ann Intern Med. 2005 Jan 4;142/1):37-46. COMPARING RATIO DATA The study shown to the right (by Pogue, et al.) uses a forest plot to test for significance of a composite endpoint reported as ratio data, in this case hazard ratio. Recall for ratio data, the result is not sta.tistically significant if the confidence interval crosses one, so the vertical line (line of no difference) is set at one. Examples: Primary composite endpoint: a statistically significant benefit was shown with treatment; the CI (0.7- 0.99) does not cross one (and the horizontal line representing the CI does not touch or cross the vertical line at one). CV death: shows no statistically significant benefit (or harm); the CI (0.92-1.83) crosses one (and the horizontal line representing the CI crosses the vertical line at one). 224 Treatment Benefit Treatment Harm HR (95% Cl) Primary Composite: - - - CV Death/Ml/CA - • -- 0.84 (0,70-0.99) CV Death 1-30 (0.92-1-83) 0.70 (0.57-0.86) Ml --- (0.60-2.06) Cardiac Arrest 0 .5 1.0 1.5 2.0 2.5 HR Pogue, et al. PLoS One 2012; 7/4): e34785. Rx PREP 2 022 COURS E BOOK TYPES OF MEDICAL STUDIES I Rx PREP © 2021 , ©2022 Most Reliable Evidence-based medicine (EBM), which is largely guideline and protocol-driven, is the foundation of practice and most patient care recommendations are based on valid study data. The type of study that a researcher chooses is a major factor in determining the quality of the study data and the clinical value or impact. The pyramid figure to the right depicts the reliability of each of the major study types. Cohort Studies CoStl•Controllo<IStud~ Common types of studies include: Case-control studies: retrospective comparisons of cases (patients with a disease) and controls (patients without a disease). Cohort studies: retrospective or prospective comparisons of patients with an exposure to those without an exposure. Randomized controlled trials: prospective comparison of patients who were randomly assigned to groups. OtM Sf!ries and Ca.ut RGporU E,q,ert Opinion least Reliable Each type of study has benefits and limitations. The table below describes the various study types and provides an example of each study. Meta-analyses: analyzes the results of multiple studies. STUDVTYPE, BENEFITS AND LIMITATIONS STUDY TYPE EXAMPLE CASE-CONTROL STUDY Predictors of surgical site Infection after open lower extremity bypass (LES) revascularlzatlon. Compares patients with a disease (cases) to those without the disease (controls). The outcome of the cases and controls is already known, but the researcher looks back in time (retrospectively) to see if a relationship exists between the disease (outcome) and various risk factors. Benefits Methods Data was pulled from 35 hospitals for all patients who had LEB during a 3-year period. Cases of surgical site infection (551) were identified and compared to those who did not develop an 551 (controls). An odds ratio (OR) was calculated for various risk factors that might increase 551 risk. Data is easy to get from medical records. Good for looking at outcomes when the intervention is unethical (e.g., exposing patients to a pesticide to test an association with cancer; cases that occurred are used instead). Less expensive than a RCT. Renal failure OR: 4.35 (95% Cl 3.45-5.47; p < 0.001). I Hypertension OR: 4.29 (95% Cl 2.74-6.72; p < 0 .001). BMI 2: 25 kg/m 2 OR: 1.78 (95% Cl 1.23-2.57; p = 0.002). Conclusion Limitations Cause and effect cannot reliably be determined (associations may be proven to be non-existent). COHORT STUDY Statistical Data ™ Compares outcomes of a of patients exposed and not exposed to a treatment; the researcher follows both groups prospectively (in the future) or retrospectively (less common) to see if they develop the outcome. Renal failure, hypertension and BMI 2: 25 kg/m 2 were all associated with an increased risk of 551. Statin use and cosnitive function In adults with type 1 diabetes. Methods Patients with type 1 diabetes who were taking statins (exposed) were compared to those not taking statins (not exposed) and followed for 7-12 years to see if statin use was associated with cognitive impairment (outcome). Benefits Good for looking at outcomes when the intervention would be unethical. Statistical Data 5tatin use and odds of cognitive impairment OR: 4.84 (95% Cl 1.63-14.44; p=0.005). Limitations More time-consuming and expensive than a retrospective study. Can be influenced by confounders, which are other factors that affect the outcome (e.g., smoking, lipid levels), Conclusion In type 1 diabetes, patients taking statins were more likely to develop cognitive impairment compared to those who did not take statins. 225 14 I BIOSTAT ISTI CS PHARMACOECONOMICS BACKGROUND Healthcare costs in the United States rank among the highest of all industrialized countries. In 2017, total healthcare expenditures reached $3.5 trillion, which translates to an average of $10,739 per person, or about 17.9% of the national gross domestic product. The increasing costs have highlighted the need to understand how limited resources can be used most effectively and efficiently in the care of individual patients and society as a whole. It is necessary to scientifically evaluate the value (i.e., costs vs. outcomes) of interventions such as medical procedures or drugs. DEFINITIONS Pharmacoeconomics is a collection of descriptive and analytic techniques for evaluating pharmaceutical interventions (e.g., drugs, devices, procedures) in the healthcare system. Pharmacoeconomic research identifies, measures and compares the costs (direct, indirect an.d in.tangible) and the consequences (clinical, economic and humanistic) of pharmaceutical products and services. Various research methods can be used to determine the impact of the pharmaceutical product or service. These methods include cost-effectiveness analysis, cost-minimization analysis, cost-utility analysis and cost-benefit analysis. The term "pharmacoeconomics" is sometimes referred to as "outcomes research," but they are not the same thing. Pharmacoeconomic methods are specific to assessing the costs and consequences of pharmaceutical products and services. Outcomes research represents a broader research discipline that attempts to identify, measure and evaluate the end result of healthcare services. Healthcare providers, payers and other decision makers use these methods to evaluate and compare the total costs and consequences of pharmaceutical products and services. The results of pharmacoeconomic analyses can vary significantly based on the point of view of the analyst; the study perspective is critical for interpretation. What may be viewed as good value for society or for the patient may not be deemed as such from an institutional or provider perspective (e.g., the costs of lost productivity due to illness are critically important to a patient or employer, but perhaps less so to a health plan) . Pharmacoeconomic analyses provide useful supplemental evidence to traditional efficacy and safety endpoints. They help translate important clinical benefits into economic and patient-centered terms, and can assist providers and payers in determining where, or if, a drug fits into the treatment paradigm for a specific condition. Pharmacoeconomic studies serve to guide optimal healthcare resource allocation in a standardized and evidence-based manner. The ECHO model (Economic, Clinical and Humanisti Outcome ) provides a broad evaluative framework to assess the outcomes associated with diseases and treatments. ~conomic outcomes: include direct, indirect and intangible costs of the drug compared to a medical intervention. Clinical outcomes: include medical events that occur as a result of the treatment or intervention. !:!umanistic Qutcomes: include consequences of the disease or treatment as reported by the patient or caregiver (e.g., patient satisfaction, quality oflife). MEDICAL COST CATEGORIES: DIRECT, INDIRECT AND INTANGIBLE DIRECT I MEDICAL I Drug preparation & administration, including home infusion supplies I I Outpatient direct costs: office & clinic visits 228 INTANGIBLE Lost work time Pain, suffering, anxiety, I j "oN-MEDtCAL I Travel and lodging costs for patients & careg