Measures of Locatio ad Variability Sprig, 2009
Skill set: You should kow the defiitios of the major measures of locatio (mea, media, mode, geometric mea) ad variability (stadard deviatio, variace, stadard error of the mea, skewess ad kurtosis). You should kow: Set of Observatios Descriptive Statistic x i xi + c cx i Mea x x + c cx Variace s 2 s 2 cs Stadard deviatio s s cs c Meas the absolute value of c. You should be able to use Stata to graph histograms ad box plots. You should kow how to use the help meu.
Outlie Scales of measuremet Page 1 Measures of Locatio Mea Page 2 Media Page 7 Mode Page 9 Geometric Mea Page 10 Properties of Meas Page 15 Stata commads used: Dropdow meus Page 25 log usig describe (des) summarize (sum) geerate (ge) codebook label display (di) list ameas Measures of spread or variability Rage Page 30 Percetiles Page 30 Iterquartile rage Page 32 Variace Page 33 Stadard deviatio Page 34 Stadard error of the mea Page 34 Kurtosis Page 35 Skewess Page 35 Defiitio of whiskers Page 36 Drop dow meus Box Plots Page 38 Dataset used: weight.dta
Scales used with data: Four scales are used with variables: omial, ordial, iterval ad ratio. omial - the variable has o order, just category ames Geder (male, female) ad hypertesive (yes, o) are examples ordial - the variable ca be rak ordered but there is o cosistet distace betwee the categories Icome scaled as low, medium ad high is a example. We kow that someoe i the category low has a smaller icome tha someoe i the category high but we do t kow how much smaller. Is the distace betwee low ad medium the same as the distace betwee medium ad high? We just kow the order ot the differece or distace betwee categories. iterval ad ratio - both of these are scales of equally spaced uits (i.e. cosistet distaces) like height i iches. A differece betwee the two scales is that variables o the ratio scale have a zero poit that ca be iterpreted as there is oe of the quatity beig measured but variables o the iterval scale do ot have such a zero poit. Height is o the ratio scale ad 0 iches tall meas there is o height. The Celsius scale is o the iterval scale but ot the ratio scale. Zero degrees Celsius does ot mea there is o heat. I order to be o the ratio scale, the ratio of two umbers has to make sese. A perso 140 cm tall is twice as tall as oe 70 cm tall. A ove at 300 degrees Celsius is ot twice as hot as oe at 150 degrees Celsius. Measures of locatio: We will cosider several measures of locatio. The mea, which we cosider first, is the most commoly used measure of locatio. Page -1-
x, x, x,..., x x Mea: If the sample cosists of poits 1 2 3, the the mea ( ) is defied as x xi i= = = x + x + x + + x 1 1 2 3... This is just the arithmetic mea of the values. I order to calculate a mea, the variable has to be at least o the iterval scale. We will create ad use the small data set smalldbp.dta with the diastolic blood pressures of 10 people to illustrate meas. We will follow the steps i the picture below. 1) We click o the log butto which opes the Begi loggig Stata output meu. 2) We select the folder i which we wish to save our log file (i.e. Chapter2"). 3) We tell Stata we wat a log type of log file rather tha the smcl type of log file. 4) We give our log file a ame (smalldbp.log) 5) We save our log file to Chapter2" 6) The results of 1-5. Page -2-
6). log usig "W:\WP51\Biometry\AAAABiostatFall2007\Data\Chapter2\smalldbp.log" ------------------------------------------------------------------------------ log: W:\WP51\Biometry\AAAABiostatFall2007\Data\Chapter2\smalldbp.log log type: text opeed o: 29 Aug 2007, 18:49:36 log o (text) tells you that you have a log file ruig ad that it is text as opposed to smcl We are goig to eter our data usig the data editor. Eterig data here is just like eterig data i Excel. (1) I click o the data editor butto (the highlighted butto below) ad that brigs up the Data Editor meu. I the just type i a ID variable ad 10 diastolic blood pressures (DBP). (2) I preserve the data so I wo t lose it ad (3) close the data editor because Stata wo t let me type o the commad lie if the data editor is ope I the Itroductio to Stata hadout I show you how to use the dropdow meus to give the variables ames other tha var1 ad var2 ad to give the variables descriptive Page -3-
labels. Here I am just goig to type i the appropriate commads o the commad lie. - preserve. reame var1 id. label variable id "Uique Idetifier". reame var2 dbp. label variable dbp "Diastolic Blood Pressure i mm Hg". des Cotais data obs: 10 vars: 2 size: 60 (99.9% of memory free) ------------------------------------------------------------------------------- storage display value variable ame type format label variable label ------------------------------------------------------------------------------- id byte %8.0g Uique Idetifier dbp byte %8.0g Diastolic Blood Pressure i mm Hg ------------------------------------------------------------------------------- Sorted by: Note: dataset has chaged sice last saved des is short for describe. The mea diastolic pressure of these 10 people is: x 10 x i i= 1 = 10 90 + 85 + 100 + 87 + 92 + 78 + 80 + 96 + 93 + 99 = 10 900 = = 90. 0 10 It is customary to write the value for the mea to oe more decimal place tha the origial data. The origial DBP s are itegers so I report the mea of the DBP s as 90.0. We usually report the stadard deviatio to two decimal places beyod the origial data (7.51). Page -4-
The easy way to get the mea is to just type i sum dbp or for more iformatio type sum dbp, det where sum is short for summarize ad det is short for detail. The results are below.. sum dbp Variable Obs Mea Std. Dev. Mi Max -------------+-------------------------------------------------------- dbp 10 90 7.512952 78 100. sum dbp,det Diastolic Blood Pressure i mm Hg ------------------------------------------------------------- Percetiles Smallest 1% 78 78 5% 78 80 10% 79 85 Obs 10 25% 85 87 Sum of Wgt. 10 50% 91 Mea 90 Largest Std. Dev. 7.512952 75% 96 93 90% 99.5 96 Variace 56.44444 95% 100 99 Skewess -.248569 99% 100 100 Kurtosis 1.914099 To use dropdow meus to do the same thig see the back of this hadout. Graph #1 based o origial set of 10 DBP values. Page -5-
The mea ca be thought of as the ceter of gravity (if you have weights of equal size hagig off each sample poit, the mea would be the balace poit.). Advatages of usig the mea: it uses all the observatios i the sample each sample has a uique mea A disadvatage of usig the mea is that it is sesitive to extreme values (ad the smaller the sample, the more impact the extreme values have). Below I create a ew variable which is equal to the old variable dbp except the value 99 is chaged to 130 (we ll call this set of 10 values the ewdbp). Note that this chages the mea of the sample from 90.0 to 93.1 (see graph below to uderstad how the ceter of gravity has chaged just by chagig oe value).. ge ewdbp = dbp ge is short for geerate. replace ewdbp = 130 if dbp == 99 (1 real chage made). sum ewdbp Variable Obs Mea Std. Dev. Mi Max -------------+-------------------------------------------------------- ewdbp 10 93.1 14.64734 78 130 Graph #2 is based o the set of 10 DBP values with 99 replaced by 130. Page -6-
Notice that the mea is pulled from 90.0 to 93.1 (i.e. the mea is pulled toward the outlyig value).. save smalldbp.dta file smalldbp.dta saved. log close log: W:\WP51\Biometry\AAAABiostatFall2007\Data\Chapter2\smalldbp.log log type: text closed o: 29 Aug 2007, 20:29:53 The largest value for baselie cholesterol i the dataset weight.dta is 412. Try chagig that to 1500 ad comparig the mea of the origial sample with the mea of the chaged sample. Notice that there are 10,273 participats with baselie cholesterol values but there are 10,355 participats i the dataset. The way to create the ew DBP variable with dropdow meus is give at the back of the hadout. Whe we study the Cetral Limit Theorem, we will fid that the mea has some ice properties that allow us to get cofidece itervals ad do hypothesis testig. The type of data eeded to calculate a mea is iterval (i.e. you have to have the ability to divide ad still have a legitimate observatio). So we calculate meas for variables such as age ad diastolic blood pressure (i.e. cotiuous variables). Media: If the sample cotais a odd umber of observatios, the media is the middle observatio provided the sample is ordered from smallest to largest. If the sample cotais a eve umber of observatios, the media is the average of the two middle observatios give that the sample is ordered from smallest to largest. You ca see that this defiitio makes the media such that a equal umber of poits are greater tha or equal to ad less tha or equal to the media. A advatage for the media over the mea is that the media is ot sesitive to extreme values. Notice that both the variable dbp ad the variable ewdbp have the same media, but ot the same mea. The media is the 50 th percetile. Media Mea dbp 91 90.0 ewdbp 91 93.1 Page -7-
. sum(dbp),det (origial set of 10 values for DBP) Diastolic Blood Pressure (dbp) ------------------------------------------------------------- Percetiles Smallest 1% 78 78 5% 78 80 10% 79 85 Obs 10 25% 85 87 Sum of Wgt. 10 50% 91 Mea 90 Largest Std. Dev. 7.512952 75% 96 93 90% 99.5 96 Variace 56.44444 95% 100 99 Skewess -.248569 99% 100 100 Kurtosis 1.914099 Note that i the Stata output below the 50th percetile is the media ad that although the largest value chages from 100 to 130 the media remais the same.. sum(ewdbp),det New versio of DBP with 99 chaged to 130 ------------------------------------------------------------- Percetiles Smallest 1% 78 78 5% 78 80 10% 79 85 Obs 10 25% 85 87 Sum of Wgt. 10 50% 91 Mea 93.1 Largest Std. Dev. 14.64734 75% 96 93 90% 115 96 Variace 214.5444 95% 130 100 Skewess 1.644196 99% 130 130 Kurtosis 5.212837 Aother advatage for the media is that each sample has a uique media. A disadvatage for the media is that it does ot utilize all the data i the sample. I order to obtai a media, the data has to be o at least the ordial scale (i.e. you ca order the observatios). Whe should we use the mea ad whe should we use the media? The cartoo below sort of gives the correct aswer. Page -8-
Mode: The mode is the most frequetly occurrig value i a set of observatios. A disadvatage for the mode is that ot all samples have a mode ad some samples have multiple modes. Sample 1 = {1,2,3,4,5,6,7,8,9,10} has o mode. Sample 2 = {1,1,1,2,3,4,4,4,5} has modes 1 ad 4. Sample 3 = {M, F, F, F, M, M, M, F, F, F} has mode F where M = male ad F = female. The mode ca be calculated with data o the omial scale (i.e. all you have to be able to do is categorize each observatio). The mode will ot come up agai i this course uless it is i a discussio of a bimodal distributio because it is ot ameable to mathematical maipulatio. Thigs about logs you have probably log sice forgotte. log here ca be to ay base (i.e., ) log e log 10 1) log(a) is defied oly if a > 0. 2) log(ab) = log(a) + log(b) 3) log(a/b) = log(a) - log(b) 4) log( a k ) = k log( a ) Page -9-
Geometric mea: x, x, x,..., x x g If the sample is 1 2 3 the the geometric mea ( ) is defied as x = x x x x g 1 2 3 (This is the th root of the product of sample elemets) 1 g 1 2 3 This ca also be writte as x = ( x x x x ) or as log( x ) g i = = 1 log( x ) i The geometric mea turs up whe doig such thigs as dilutio assays. So usig our ewly remembered facts about logs we have the followig: log( x ) = log ( x x x x ) g 1 2 3 1 = log( x x x x ) 1 2 3 1 = log( x ) + log( x ) + log( x ) + log( x ) 1 2 3 log( x ) i i = = 1 So we have that the mea of the logs is the log of the mea. Roser gives a good example of the use of the geometric mea o pages 14 ad 15, Table 2.4. Page -10-
The geometric mea is more appropriate tha the arithmetic mea i the followig circumstaces: 1) Whe losses/gais ca best be expressed as a percetage rather tha a fixed value. 2) Whe rapid growth is ivolved, as i the developmet of a bacterial or viral populatio. 3) Whe the data spa several orders of magitude as with a cocetratio of pollutats. Take from Commo Errors i Statistics 2 d editio by Good ad Hardi. The most commoly used of the above measures of locatio is the mea with the media secod because it is used i o-parametric aalyses. Questio: Why would the CMS (Ceter for Medicare ad Medicaid Services) preset the geometric mea to summarize the legth of hospital stay. Note that this does t fit ay of the reasos give above. It has to do with trasformed data. Below is a small study of the legth of hospital stay for 25 patiets. The dataset used is hospital.dta which is a file that is also used i the Itroductio to Stata. hospital.dta is o the class website. Page -11-
The distributio of a variable is said to be symmetric if the pieces o either side of the ceter poit are mirror images. Otherwise the distributio is described as skewed. If the distributio is symmetric the skewess value give i the detailed versio of the commad summarize is zero. The variable legth of hospital stay is skewed to the right (also described as positively skewed). Notice that the skewess value is 2.2. A positive skewess value (i.e. value > 0) idicates that the skewess is to the right (see the histogram of hospital stay above). A egative skewess value idicates the distributio is skewed to the left. Idividuals who have much loger hospital stays tha most of the other patiets is very commo for legth of stay data.. sum stay,det Legth of hospital stay i days ------------------------------------------------------------- Percetiles Smallest 1% 3 3 5% 3 3 10% 3 3 Obs 25 25% 5 4 Sum of Wgt. 25 50% 8 Mea 8.6 Largest Std. Dev. 5.715476 75% 11 11 90% 14 14 Variace 32.66667 95% 17 17 Skewess 2.203535 99% 30 30 Kurtosis 8.959067 This is a case where the value 30 days is probably correct so we ca t just set it to missig. Oe thig that we ca do is trasform the data to brig the 30 days closer to the rest of the data. Oe of the trasformatios which will brig i the larger values is the atural (i.e. base e) logarithmic trasformatio (log to base 10 will also brig i the more distat data). To get the log trasformatio we simply geerate a ew variable that is equal to log base e of the variable stay.. ge logofstay = log(stay). label variable logofstay "The atural logarithm of the variable legth of hospital stay" You ca also use l(stay) to get the log base e of stay. To get the log base 10 you use log10(stay). The thigs about logs that we ve probably log sice forgotte are true regardless of the base. Notice i the histogram below that the log trasformatio has pulled the largest value i earer the other values. Page -12-
Histogram 2 above is the graph of the atural logarithm of the variable stay, so the log of the geometric mea of stay will equal the arithmetic mea of the variable logofstay.. ameas stay Variable Type Obs Mea [95% Cof. Iterval] -------------+---------------------------------------------------------- stay Arithmetic 25 8.6 6.240767 10.95923 Geometric 25 7.303239 5.774765 9.236272 Harmoic 25 6.308454 5.148257 8.143695 ------------------------------------------------------------------------. ameas logstay Variable Type Obs Mea [95% Cof. Iterval] -------------+---------------------------------------------------------- logstay Arithmetic 25 1.988318 1.753498 2.223138 Geometric 25 1.907722 1.685849 2.158796 Harmoic 25 1.8248 1.613525 2.09974 ------------------------------------------------------------------------. di log(7.303239) 1.9883179 Or the atilog of the arithmetic mea of the variable logstay is the geometric mea of the variable stay.. di exp(1.988318) 7.3032394 The atilog i this case is the iverse of the log fuctio which is the expoetial e x e = 2. 7182818 fuctio (i.e. where ). Page -13-
So what does the log trasformatio do? If the ratios of two pairs of poits are equal the o the log scale the distace betwee the two members of a pair is the same for both pairs. 10 100 = 1 10 so 10 1 log = log 100 10 but 10 1 log( 10) log( 100) = log = log = log( 1) log( 10) 100 10 So we have. di log(10/100) -2.3025851. di log(1/10) -2.3025851. di log(1) - log(10) -2.3025851. di log(10) - log(100) -2.3025851 So istead of havig 1 ad 10, 9 uits apart while 10 ad 100 are 90 uits apart both are 2.3 uits apart o the atural log scale. So the short aswer to why CMS presets the geometric mea is to lesse the ifluece of outlyig values. Page -14-
Properties of meas: Property 1: Sometimes we wish to rescale the elemets of our sample. For example, we may have collected the weight of our participats i pouds ad ow we are goig to publish our paper i a joural that requires the weight to be reported i grams. The data file we are usig is weight.dta. I double (left) clicked o the data set weight which was stored o the W drive ad the file opeed i Stata. I the use statemet below from the W to weight.dta gives the path to fid the data set. Whe we ope a data set i this fashio, Stata will store ay log file we create i the same folder where the dataset was stored. Page -15-
There are several properties that I would like you to otice about the file above: 1) The file is sorted by the variable weight. This meas if I list the variable weight, the smallest weight will be listed first ad the largest weight will be listed last. 2) Each variable has a variable label describig the data the variable cotais. 3) The categorical variables have value labels. Notice i the descriptio above that the umber of observatios is give as 10,355 but the summary of weight below says there are 10,341 values for weight.. sum weight Variable Obs Mea Std. Dev. Mi Max -------------+-------------------------------------------------------- weight 10341 183.1275 39.37125 54 392 If I use the commad codebook, we ca see that there are 14 missig values for weight.. codebook weight ------------------------------------------------------------------------------ weight Weight (lbs) at Baselie ------------------------------------------------------------------------------ type: umeric (float) rage: [54,392] uits: 1 uique values: 262 missig.: 14/10355 mea: 183.127 std. dev: 39.3713 percetiles: 10% 25% 50% 75% 90% 136 156 180 206 234 We kow that 1 poud = 453.26 grams. So let us create a ew variable called wtigms that is the baselie weight i grams.. ge wtigms = weight*453.26 (14 missig values geerated). label variable wtigms Weight i grams Note that wtigms is missig 14 values because weight is missig 14 values (i.e. missig 453.26 = missig). Stata uses the period to represet missig data. Page -16-
Below I used the commad list to list the values of weight ad wtigms for the last 19 participats (whe the data is ordered by weight) which icludes the 14 people with missig values for wtigms. oobs asks that Stata ot to umber the rows.. list id weight wtigms if weight >= 364,oobs +---------------------------+ id weight wtigms --------------------------- 10337 364.00 164986.6 10338 370.00 167706.2 10339 382.00 173145.3 10340 392.00 177677.9 10341 392.00 177677.9 --------------------------- 10342.. 10343.. 10344.. 10345.. 10346.. --------------------------- 10347.. 10348.. 10349.. 10350.. 10351.. --------------------------- 10352.. 10353.. 10354.. 10355.. +---------------------------+ I have listed the last 19 observatios for weight. The periods represet missig data. Sice the missig data is listed last, we kow that Stata cosiders missig values to be larger tha ay other values. The other thig to otice is that 164986.6 = 453.26 364 167706.2 = 453.26 370 etc. Below we see that the mea of the wtigms variable is 453.26 times the mea of the weight variable.. sum weight Variable Obs Mea Std. Dev. Mi Max -------------+-------------------------------------------------------- weight 10341 183.12745 39.37125 54.00000 392.00000. sum wtigms Variable Obs Mea Std. Dev. Mi Max -------------+-------------------------------------------------------- wtigms 10341 83004.35 17845.41 24476.04 177677.9. di 453.26*183.12745 83004.348 The di above stads for display. The * says multiply 183.12745 times 453.26. That is, I m usig Stata like it is a calculator. Page -17-
This shows that if c is a costat (here 453.26), the samplecx 1, cx 2, cx 3,..., cx cx x x 1, x 2, x 3,... x (wtigms) has mea where is the mea of the sample (weight). That is, you ca obtai the mea of a sample ad the multiply by the costat or you ca multiply each elemet by the costat ad the get the mea. Property 2: Sample x1+ c, x2 + c, x3 + c,..., x + c has mea x + c if the sample has mea ad is a costat. x 1, x 2, x 3,..., x x c This says you ca add (or subtract) a fixed value to each of the origial values ad the get the mea or you ca get the mea of the origial values ad the add (or subtract) the fixed value. You will fid later whe doig regressio that people sometimes ceter their data by subtractig the mea of the variable from each of the origial observatios. So istead of puttig the origial variable i the regressio equatio, the variable they use is the origial variable mius its mea. So let s take a look at what happes whe you add a fixed value to each elemet of a sample. Let us take the variable chol (this is the baselie cholesterol from the dataset weight.dta) ad add 50 to the baselie value for each of the10273 people who have a baselie value (i.e. 82 people have missig listed for the baselie value of cholesterol ad missig + 50 = missig).. sum chol,det Lipid BL Cholesterol ------------------------------------------------------------- Percetiles Smallest 1% 167 130 5% 181 134.5 10% 189.5 142.5 Obs 10273 25% 205 144 Sum of Wgt. 10273 50% 223 Mea 223.7146 Largest Std. Dev. 26.80037 75% 241.5 320.5 90% 259 322 Variace 718.2601 95% 269 345 Skewess.2067261 99% 288.5 412 Kurtosis 3.099006. ge cholplus = chol + 50 (82 missig values geerated). label variable cholplus50 "Baselie cholesterol + 50 mg/dl" Soapbox momet: I recommed always labelig your variables. You thik you ll remember how the variable is defied, but whe you come back to the data six moths later you may fid that you ve forgotte. Page -18-
. sum cholplus50,det Baselie cholesterol + 50 mg/dl ------------------------------------------------------------- Percetiles Smallest 1% 217 180 5% 231 184.5 10% 239.5 192.5 Obs 10273 25% 255 194 Sum of Wgt. 10273 50% 273 Mea 273.7146 Largest Std. Dev. 26.80037 75% 291.5 370.5 90% 309 372 Variace 718.2601 95% 319 395 Skewess.2067261 99% 338.5 462 Kurtosis 3.099006 So we ca see that addig 50 to each baselie value shifts all of the percetiles, the mea, the miimum ad the maximum up by 50 poits. Notice that the stadard deviatio ad the variace (which we will defie o later) remai uchaged (this is because they refer to shape, while the mea ad percetiles etc. refer to positio). The skewess ad kurtosis (to be defied later) also remai the same because the oly thig we ve doe is to shift the curve up 50 poits. See the graphs o the ext 2 pages. Below is the codebook for both chol ad cholplus50.. codebook chol cholplus50 ------------------------------------------------------------------------------ chol Lipid BL Cholesterol ------------------------------------------------------------------------------ type: umeric (float) rage: [130,412] uits:.1 uique values: 326 missig.: 82/10355 mea: 223.715 std. dev: 26.8004 percetiles: 10% 25% 50% 75% 90% 189.5 205 223 241.5 259 ------------------------------------------------------------------------------ cholplus50 Baselie cholesterol + 50 mg/dl ------------------------------------------------------------------------------ type: umeric (float) rage: [180,462] uits:.1 uique values: 326 missig.: 82/10355 mea: 273.715 std. dev: 26.8004 percetiles: 10% 25% 50% 75% 90% 239.5 255 273 291.5 309 Page -19-
Below I have created a histogram for each of chol ad cholplus50. You ca see that the two histograms below are the same shape. The lower oe is just shifted 50 mg/dl to the right. Origial Baselie Cholesterol Frequecy 0 200 400 600 800 1000 224 100 150 200 250 300 350 400 450 Baselie Cholesterol mg/dl Baselie Cholesterol + 50 Frequecy 0 200 400 600 800 1000 273.7 100 150 200 250 300 350 400 450 Baselie Cholesterol mg/dl + 50 mg/dl Page -20-
Box ad whisker plots: The lie i the middle of the box is the media or 50th percetile. The bottom of the box is the 25 th percetile ad the top of the box is the 75 th percetile. The height of the box (i.e. from 25 th to 75 th percetile) is called the iterquartile rage ad it is a measure of variability. Lipid BL Cholesterol 100 200 300 400 Box ad Whisker Plot Upper whisker Lower whisker 75 th percetile 50 th percetile 25 th percetile Addig a costat chages locatio but ot variability Cholesterol for baselie ad baselie + 50 100 200 300 400 500 Lipid BL Cholesterol Baselie cholesterol + 50 mg/dl The box plot above shows eve more clearly that the distributio is just shifted up without chagig the relatioship of the various pieces. So what I ve worked hard to show is that addig a fixed umber to each uit of a sample chages the locatio of the distributio but leaves the shape uchaged. We will discover that multiplyig each uit of a sample by a fixed umber chages the shape of the distributio. Page -21-
Now go back to multiplyig the origial values by some costat We ll geerate a ew variable which we obtai by multiplyig each of the origial baselie cholesterol values by 2.. ge cholx2 = 2*chol (82 missig values geerated). label variable cholx2 "Baselie cholesterol times 2 mg/dl" Notice below that almost all of the values produced by the summarize commad are multiplied by 2. There are three exceptios. The variace is multiplied by 4 = 2 2 (we will later lear the variace = SD 2, where SD = stadard deviatio) ad the skewess ad kurtosis are the same as they were for baselie cholesterol (as opposed to beig multiplied by 2). We ll discuss skewess ad kurtosis later.. sum cholx2,det Baselie cholesterol times 2 mg/dl ------------------------------------------------------------- Percetiles Smallest 1% 334 260 5% 362 269 10% 379 285 Obs 10273 25% 410 288 Sum of Wgt. 10273 50% 446 Mea 447.4292 Largest Std. Dev. 53.60075 75% 483 641 90% 518 644 Variace 2873.04 95% 538 690 Skewess.2067261 99% 577 824 Kurtosis 3.099006. sum chol,det Lipid BL Cholesterol ------------------------------------------------------------- Percetiles Smallest 1% 167 130 5% 181 134.5 10% 189.5 142.5 Obs 10273 25% 205 144 Sum of Wgt. 10273 50% 223 Mea 223.7146 Largest Std. Dev. 26.80037 75% 241.5 320.5 90% 259 322 Variace 718.2601 95% 269 345 Skewess.2067261 99% 288.5 412 Kurtosis 3.099006 I have created a histogram for each of baselie cholesterol ad baselie cholesterol times 2. I order to compare the 2 graphs they eed to be o the same scale. Notice that the smallest value for cholesterol is 130 mg/dl ad the largest for cholesterol times Page -22-
2 is 824 mg/dl. So I will select the x-axis scale as 125(100)825 for both versios of cholesterol. 125(100)825 says label the x-axis startig with the smallest value (i.e. 125) ad the goig up by uits of 100 util you reach 825. Frequecy 0 200 400 600 800 1000 125 225 325 425 525 625 725 825 Baselie cholesterol mg/dl Frequecy 0 200 400 600 800 1000 125 225 325 425 525 625 725 825 Baselie cholesterol mg/dl times 2 Page -23-
Baselie cholesterol ad baselie cholesterol times 2 mg/dl 200 400 600 800 Lipid BL Cholesterol Baselie cholesterol times 2 mg/dl Lookig at the graphs o the previous page ad above we see that multiplyig by 2 has chaged ot oly the locatio (mea) but also the shape. The cholesterol times 2 is much more spread out (we ll come back to these graphs whe we discuss measures of variability). So we ve leared that addig to the elemets of a sample chages oly the locatio but multiplyig chages both the locatio ad the shape. We kow that we ca measure locatio usig the mea ad media, but we do t yet kow how to idicate (other tha graphically) that the shape has chaged. Page -24-
Meus to get meas: Click o Submit to ru the commad but leave the meu up so you ca make chages as eeded. Click OK just to ru the commad. Click o? to brig up the help meu for summarize. Click o R to clear the etries i the meu. Page -25-
How to chage the values of a variable.. replace chol = 1500 if chol == 412 (1 real chage made) Page -26-
How to get geometric, arithmetic ad harmoic meas. Page -27-
How to get a histogram. Page -28-
Frequecy 0 500 1000 1500 2000 200 250 300 350 400 450 cholplus50 Page -29-
Measures of spread or variability: Rage: rage = largest value - smallest value Note that codebook gives the rage as a iterval. Statisticias ted to use the defiitio as give so that the rage is a sigle umber Advatage: This is the simplest measure of spread. Disadvatage Very sesitive to extreme values The rage for the baselie cholesterol is 412-130 = 282. If we chage the largest value (412) to 550, the the rage becomes 550-130 = 420 Oe of the problems with the rage is there is a tedecy for larger samples, to have larger rages. How does addig 50 to the variable cholesterol or multiplyig by 2 chage the rage. The rage for the baselie cholesterol is 412-130 = 282. The rage for the cholesterol + 50 = 462-180 = 282. So these two variables with the same shape also have the same rage. The rage for cholesterol times 2 = 824-260 = 564 = 2 times the rage of baselie cholesterol. The rage for cholesterol times 2 is twice that of the origial cholesterol. We ca see that i the histograms ad the box-ad-whisker plots i the Chapter 2 Part 1 hadout. Percetiles: Roser says that ituitively, the p th percetile is the value V p such that p percet of the sample poits are less tha or equal to You will also see percetiles called quatiles. Quartiles are the Quitiles are the Deciles are the th th th 25, 50, 75 th th th th 20, 40, 60, 80 10, 20, 30, 40, K, 90. The media is the 50 th percetile. V p percetiles percetiles percetiles th th th th th Page -30-
Below we ca see the chage i the 25 th, 50 th ad 75 th percetiles as you add a costat (here 50) to the origial cholesterol or multiply the origial cholesterol by a costat (here 2). Percet Cholesterol Cholesterol + 50 Cholesterol x 2 25% 205 205 + 50 = 255 205 x 2 = 410 50% 223 223 + 50 = 273 223 x 2 = 446 75% 241.5 241.5 + 50 = 291.5 241.5 x 2 = 483 Page -31-
Iterquartile rage: Iterquartile rage = value of the 75 th percetile - value of the 25 th percetile As we saw i the last hadout, the iterquartile rage is the height of the box i the box plot graph. Notice below that the values of baselie cholesterol cluster together whereas the values of baselie cholesterol times 2 are much more spread out. We would like to be able to describe this variability i a way that uses all of the data as opposed to the rage ad iterquartile rage which use oly 2 of the values i the dataset. We ll call this ew statistic the variace. Page -32-
Variace: A first guess at a defiitio for variace might be guess() 1 = ( x x) i= 1 This defiitio uses all of the observatios i the sample. It also seems reasoable to use the distace of each observatio from the mea as a measure of how spread out the values are. The problem is that this sum is always equal to zero. i A secod guess might be guess( 2) = x x i= 1 This secod guess solves the problem of the sum addig to zero ad it is scaled the same as the origial data. However, this secod guess has two problems: (1) is that the absolute value is mathematically itractable ad (2) this sum gets larger as the sample size gets larger. The secod problem could be dealt with by dividig the sum by the size of the sample, amely. Guess umber 3 is to square the differece because the square is easier to deal with mathematically tha the absolute value ad it prevets the sum from beig zero as the absolute value did. If we also divide by, the we have provided a correctio for the sample size (i.e. we adjusted the sum of squares so that the sum does t icrease just because the sample size icreases). i guess() 3 = i= 1 ( x x) i 2 The problem with this estimate, which we wo t uderstad util we lear about biased ad ubiased estimators, is that o the average it is too small (this meas if we took a large umber of repeated samples of size from a give populatio ad averaged all of the variaces from these samples, the average would be smaller tha the true variace 1 of the populatio). To solve this problem we divide by rather tha. What we have t stated before is that the sample estimate for the variace is iteded to Page -33-
estimate the variace of the populatio from which the sample was draw. So the variace ( s 2 ) is defied as follows: s ( xi x) = 1 2 i= 1 2 The variace of each of the baselie cholesterol ad the baselie cholesterol + 50 is 718.26. The variace of the cholesterol times 2 = 2873.04 (i.e. 2 2 baselie cholesterol variace). Notice that the variace is ot i the same uits as the origial data (i.e. mg 2 /dl 2 versus mg/dl). See the Stata output o page 2. Stadard deviatio: The oly problem left with the above defiitio is that the variace is ot i the same uits as the origial data. This ca be solved by takig the square root of the variace. The square root of the variace is called the stadard deviatio ad is deoted by s. We take the o-egative square root so $ 0. s s = ( xi x) 2 i= 1 1 Stadard Error of the Mea: The stadard error of the mea, deoted either SEM or SE is the stadard deviatio divided by the square root of or SE = s The SE is goig to come i hady whe we get to cofidece itervals ad the Cetral Limit Theorem. Small preview: The stadard deviatio ( ) tells us about the spread for s Page -34-
a sigle sample. The stadard error (SE) is actually the stadard deviatio of the distributio of all sample meas from samples of size. Notice that the size of the SE is depedet upo the size of the sample. Kurtosis: The kurtosis of a distributio describes its peakedess relative to the legth ad size of its tails. The kurtosis of the ormal distributio is 3. Distributios with values of kurtosis higher tha 3 ted to have sharp peaks ad log taperig tails (see the histogram of triglycerides ). Values lower tha 3 idicate distributios that are relatively flat with short tails. Users of SAS eed to be aware that the value that SAS gives for kurtosis is Stata s value mius 3 (i.e. the ormal distributio will have a kurtosis of 3 accordig to Stata ad 0 accordig to SAS). There are at least two differet defiitios of kurtosis ad SAS ad Stata have just selected differet defiitios. Kurtosis = 17.6 Skewess = 1.8 Skewess: A symmetric distributio is oe that you ca fold over at the mea ad the two halves will coicide. A symmetric distributio (e.g. the ormal distributio) will have a skewess of zero. Those distributios that are skewed to the right, like triglycerides, have a positive umber for skewess. Those skewed to the left will have a egative umber for skewess. Page -35-
The directio of the skewess goes with the side the loger tail is o. So the triglycerides graph above is said to be skewed to the right. Lipid BL Triglycerides 0 500 1,000 1,500 How to tell the graph is skewed whe usig a box plot: The 50 th percetile lie is ot i the ceter of the box. This is hard to see but the media lie is a little below the middle if the box. The whiskers are ot the same legth. Ad, of course, that log strig of poits outside the upper whisker with o similar strig outside the lower whisker. Defiitio of the whiskers. First order the uits of the sample i ascedig order (smallest to largest). xp [ ] x[ 25] Let deote the pth percetile. So is the 25 th percetile. x[ 25 ] x[ 75 ] x[ 50] The box exteds from to. The lie i the middle is. Defie U = x[ 75] + 15. ( x[ 75] x[ 25]) ad L= x[ 25] 15. ( x[ 75] x[ 25]) Page -36-
Notice that if the whiskers were defied by U ad L, the the legth of the upper ad lower whiskers would always be the same. After we ve looked at a buch of examples you ll kow the upper ad lower whiskers are ot always the same legth. The legth depeds o the upper ad lower adjacet values defied below. The otatio x () i idicates that the x' s are ordered from smallest to largest. If there xs ' x ( ) are, the 1 is the smallest ad x ( ) is the largest. The upper adjacet value (i.e. the upper whisker) is defied as the x () i such that x U ad x > U (i.e. is just iside or o U). () i ( i+1) x () i The lower adjacet value (i.e. the lower whisker) is defied as the x () i such that x() i L ad x( i 1) < L (i.e. x () i is just iside or o L). Notice that Roser refers to poits outside the whiskers as outlyig values. The upper ad lower adjacet values (defied above) are a creatio of Joh Tukey (Exploratory Data Aalysis, 1977). Page -37-
Joh Tukey - Statisticia He died at 85 i 2000 Coied the Word 'Software' ad the word bit for biary digit. Tukey used the term software three decades before the foudig of microsoft. Joh Wilder Tukey was oe of the most ifluetial statisticias of the last 50 years ad a wide-ragig thiker. Mr. Tukey developed importat theories about how to aalyze data ad compute series of umbers quickly. He spet decades as both a professor at Priceto Uiversity ad a researcher at AT&T's Bell Laboratories, ad his ideas cotiue to be a part of both doctoral statistics courses ad high school math classes. I 1973, Presidet Richard M. Nixo awarded him the Natioal Medal of Sciece. Take i part from the New York Times Obituary. How to graph a box plot I the meu above click o box plot ad you will get the meu o the right. There are a lot of facy thigs you ca do but just puttig trig i the variables widow gets you the graph a couple of pages up. Page -38-