EBM-Types of data
Range
Range: either the difference between the highest and lowest value or the smallest interval containing all the data Interquartile range: a limited range index Example: data between the 25th and 75th percentiles
Interval data
A measurement that's assigned to one of an unlimited number of equally-spaced values The order is important, the interval between values is important, and the interval between successive values is equal. However, there is no absolute/true zero. Example: temperature in °F or °C, wt
AV fistula
don't take BP over
Abnormality definition
A determination of abnormality should be in the context of those features known to be associated with disease, disability, or death; in other words, it should be a clinically-recognizable entity, and ideally one that can be txed to yield *clinically-meaningful* outcomes Be careful of labeling and especially careful of overdx, which often -> overtx -> :(
Bias due to flaw in measurement
A faulty measuring instrument may yield values that are consistently higher or lower, or sometimes altogether meaningless If the measurement methods are not consistent across groups, this is obviously riddled with problems as well (remember the BP example?) Surrogate outcome measures do not necessarily translate into clinically-meaningful outcomes This is a hugely important point.
Intraobserver variation
A radiologist reviewing the same radiograph at two different times may give a different interpretation on each occasion
Systematic measurement bias
A variation of clinical measurement(s) resulting from any number of various possible failures in the research process that can skew the results Bias is bad, can be sneaky/hard to detect, and can be avoided (more on this later) Example - BP fluctuation occurs if it is taken after smoking a cigarette or drinking a latte, with the position of the patient (and his or her arm), etc.
How should abnormality be confirmed?
Abnormality should generally be confirmed through repetition Regression to the mean, variability, etc. With respect to markers/values, the cutoff points for normal/abnormal are often arbitrary or not as firmly established as some might think (e.g. they might vary based on particulars, they might only be loosely evidence-based, etc.), and for many conditions, the cutoff has changed over the years (take, for instance, DM, HTN, osteoporosis)
Validity
Accuracy The degree to which the data measure what they are intended to measure Very important testing or measuring concept Relates to the inferences we can make based on test results for a particular use Construct validity=are you really testing what you say you are testing Surrogate=protein in urine indicates kidney health; marker for something else
Variable
All medical research involves the study of relationships among variables A variable can be defined as a quality or characteristic of a person or thing that can be measured By definition, variables may change
Ordinal Data
Also known as ranked scale Categorical specification with inherent ordering The ordering is important, but the intervals between categories are not necessarily equal; additionally, the intervals may not even be clearly defined A symbol is assigned to represent the category, but typically is not important as long as the order is preserved Can see relationship between ordinal variables and know whether one is more, less, or equally desirable Examples: Grading edema, reflexes, or strength Pregnancy risk factor for pharmaceuticals
Bias due to confounding
Arises from inadequate control over a study or inadequate consideration of the myriad things that could be related to a condition/outcome/etc. Occurs when an extraneous variable is associated with both the dependent and independent variables, thus confounding (confusing, distorting, etc.) the true effects of the independent variable on the dependent variable This can even make it look like there's an association between two variables when in fact there isn't any direct relationship at all; instead, the apparent relationship is merely an artifact of each variable's relationship with the confounder Frequent stumbling block with studies of association/correlation Correlation ≠ causation
Intrasubject variation
Blood pressure will vary based on biological variation, the time of day it is taken, position, stress status, consumption (e.g. caffeine, nicotine), etc.
Bias!
Channeling bias Detection/surveillance bias Expectation bias Incorporation bias Interviewer bias Lead-time bias Length-time bias Partial verification bias, differential verification bias Recall bias Observer bias Social desirability bias (related: Hawthorne effect) Reporting bias, selective outcome reporting bias, publication bias (a devastatingly destructive bias at the heart of all science-based/evidence-based practices) Hawthorn effect: people behave a certain way because they are being watched
Construct validity
Construct validity addresses whether the measurement relates in consistent, understandable, and meaningful ways to the construct This implies consistency with other measurements of the same phenomena Example: Whether a physician assistant interviewing a patient for depression in a clinical setting would make the same determination as the questionnaire regarding whether the patient had depression
Content validity
Content validity addresses representativeness (should include all aspects of the element of interest, but nothing extraneous) For example, the degree to which a sample of test items represents the content that the test is designed to measure Usually derived by an objective comparison of the test items with established criteria Good example from the text: For a pain scale, it would be appropriate to include painful symptoms (e.g. aching, throbbing, burning, and stinging), but not other symptoms (e.g. itching, numbness, tingling, nausea)
Continuous data
Continuous data can take on any value in the continuum Examples: Temperature (in °F or °C) Height or weight (in inches, centimeters, feet, meters, etc.) Many lab values (e.g. glucose, Cr, K+)
Criterion validity
Criterion validity addresses how well the measurement predicts or correlates with a directly observable phenomenon (criterion) that is considered to provide a direct measurement of the characteristic or behavior in question Example: Does a patient who scores high on a DSS have the affect of a clinically-depressed person? Example: Does a person with a chronically-high cholesterol have xanthelasma?
Central tendencies
Described by measures of central tendency and dispersion Central tendency Mean (average): typically a very good estimate of central tendency barring any extreme values -Sum of all values divided by number of observations Median: with scores in order by value, the middle score (with half the values above, and half the values below)-Resistant to outliers Mode: the most-frequently-occurring value
Discrete data
Discrete data can only take on specific values; expressed as whole numbers/counts. Examples # of pregnancies a woman has had # of times a patient has had a CXR # of syncopal episodes in a lifetime
Ratio data
Essentially all the qualities of an interval scale, with the addition of having a true zero/absolute zero Example: HR, BP (though the absolute zero for both of these is incompatible with life)
Basis for variation
Fluctuation among clinical measurements reflects the combined effects of several phenomena
Frequency distributions
Frequency distributions portray the quantity (or proportion or percentage) of each category, class, or interval Can be used for nominal (qualitative) data or, as is more commonly done in clinical studies, used on quantitative data to show the number or proportion of people who have different values of the measurement(s), how interventions affect values of the measurement(s), etc.
Variables with qualitative and quantitative values
Heart murmurs due to valvular abnormality can be described qualitatively (harsh, high-pitched, blowing) and/or quantitatively by determining the grade (or may quantify degree of "leakage" by viewing an echocardiogram)
Independent vs dependent variable
Independent - controlled by the researcher (input variable) Dependent - typically relates to the outcome (what occurs because of the treatment or effect) ... or at least that is the idea
Measurement variation
Measurement variation or error is variation among clinical observations attributable to the process May be random (inescapable) or systematic (avoidable) Random measurement error or variation is due to chance. There is, theoretically, an equal probability of being above or below the true value. While we cannot necessarily eliminate this altogether, there are things we can do minimize the effect this can have on the outcome(s) of interest. Determines a range for the true score
Moderator variable
Moderator/confounding/lurking/third - variable potentially moderating the effects of another variable, sometimes "behind the scenes"-need to be randomly distributed between the groups equally
Dichotomous or binary variables
Nominal variables with only two levels Directionality is only implied if one outcome is better than the other (e.g. alive/dead) Directionality may not be important While directionality doesn't affect the statistical analysis, it may be important for the conclusions drawn from the data Type of nominal data
Types of Data
Nominal: name (ex. Yes, no, red, blue, green)-Dichotomous/binary, Other Ordinal: there is an order to values Interval: implies order and clear and meaningful difference between 2 places that can be readily defined Ratio: Closest to real numbers we can get Qualitative vs. quantitative-may try to change qualitative into quantitative with counts Discrete vs. continuous
Problems with normal distribution
Normal distribution is based on mathematical theory and *only takes into account random error*, whereas the distribution of clinical data may have many sources of variation Not everything takes a normal distribution
Evaluating validity
Not an absolute/not binary (not simply present or absent), but rather, validity is assessed by the degree to which one can make a case for it/trust it Obviously, we want to use the most valid or accurate measurements possible; otherwise, the inferences drawn may be detrimentally impacted
Moderating vs confounding
Not the same thing Both effect the relationship of the effect that you see
Systematic measurement bias definition
Occurs as a result of a flaw in the measurement process The values no longer center around the true value, but around a value that is systematically higher or lower than the true value
Interobserver variation
One physician may rate a heart murmur as a "grade 3," while another rates it as "grade 4"
Reliability
Precision The consistency of measurements (reproducibility) Precision/reliability can be evaluated by observing the frequency distribution of the element in question and then calculating a standard deviation With respect to research findings, reproducibility is determined by further/other studies on the matter
Spectrum of Normal vs abnormal
Problem in determination: often there is no clear-cut point as to where pathology or disease exists "Disease" is often acquired by degrees Sucsceptibility...Pre-symptomatic...Clinical disease...Disability...Death Example: fxal classifications of heart disease based on degree of limitation
Qualitative vs quantitative data
Qualitative data Measured at the nominal level Another example: home pregnancy tests - either positive or negative (note that's binary/dichotomous) Quantitative data All others discussed in this lecture Another example: # of MIs in a sample
True biologic variation (random)
Random True biologic variation is due to the sum of many unknown factors, each of which contributes a certain amount of random effect Example: a series of consecutive systolic blood pressures under the same conditions will not be exactly the same due to biologic variability in a given patient
Theoretical frequency distribution
Real frequency distributions are those obtained from actual data Theoretical frequency distributions are assumed to describe the underlying population from which the sample was drawn Normal or Gaussian distribution (bell-curve) Some data distributions (e.g. certain lab tests) may change based on patient characteristics such as age, sex, race, etc.
Bias Due to Sampling Error
Selection Bias Arises when groups differ significantly in ways other than the condition/treatment/topic being studied (i.e. if there are important differences between the groups other than that which is intentionally different by design)
Std Dev
Standard deviation: variability of the scores about the mean
When to treat normal vs abnormal
Statistical or technical abnormality doesn't necessarily translate to practical or clinical abnormality-Na+ of 134 mEq/L? Can have an abnormal finding or state, but not be clinically sick -Homozygous recessive condition with only one recessive allele -Dzs w/ incubation periods or clinically asymptomatic periods Normal test/value does not necessarily rule out disease, either (sensitivity/SnOut, NPV, LR-)
Nominal/categorical data
Used primarily to denote groups If numbers are used, there is no arithmetic (or arithmetical) meaning Binary or dichotomous = nominal variable with only two outcomes Examples of nominal variables: Blood type Eye color Race Sex Birth month
Regression to the mean
the more observations you have, the more they will regress to the mean