It has been said that people learn more from mistakes and failure
Contents
1. WHAT SAS SOFTWARE DOES AND DOESN T DO THINGS YOUR MOTHER NEVER TOLD YOU Calvin D Croy National Demographics amp Lifestyles It has been said that people learn more from mistakes and failure than from success Yet learning by self discovery can cost time composure and sometimes money Because of these costs most people have probably sometime in their lives had occasion to reflect Gosh I wish someone had told me The purpose of this paper is to h lp prevent such after thoughts concerning SAS code Listed in this paper are a few of the SAS software version 82 3 peculiarities I or my colleagues have encoun tered while submitting batch jobs within a V51 operating system We ve all discovered instances where SAS software handles data in unexpected ways The few here might take more than a little digging through the user s manuals to unearth They re listed below according to decreasing likelihood of being encountered Surprise 1 Little elementary level documen tation in Basics User s Manual User familiarity with SAS software insufficient to select the best Procedure Though perhaps analysts above the novice stage won t consider this a SAS software peculiarity many beginning SAS software users do have problems deciding which Proc to select For example the Basics User s Manual doesn t directly help identify which Proc a user should select if all the user wants to do is compare2. Keep Y The resulting error message Explicit subscript ing of array variables is not supported in this release of SAS means little No array has been referenced though SAS software has mistaken Second for one Surprise 3 The logical comparison SAS soft ware makes may not be the one wanted The user should take care to understand the Boolean logic behind SAS software logical comparisons Example The statement if X or Y gt 5 will always be true regardless of the value of X and Y The Boolean interpretation of this statement is if X X or gt 5 Since KX X is always true all records will pass this test Similarly the test if X 10 or 20 will always be true because 20 is always equal to 20 The correct syntax for these tests are if X gt 5 or Y gt 5 and if X 10 or X 20 Problems such as these with Boolean logic frequently snare unwary SAS software users new to computer languages in general Surprise 4 Because of the way SAS software represents numbers internally using floating point decimal tests that are actually true can be taken as false Consequences can be disastrous Example Data Two Set One If Area 5 5 then go to Codel Else go to Codez If the variable Area s value is 25 where will the flow of execuation be carried To label Codel or to label Codez Answer Code2 Floating point decimal repre sentations can cause minute errors in the right most bits SAS software
3. than or equal to 1 Using the code if SUM of M1 M5 gt 5 will merely insure that at least one of the variables has a value of 5 or more Note that if M1 5 and M2 M5 all have missing values then SUM M1 M5 will equal 5 and the test is satisfied even though 4 out of 5 variables had missing values Or perhaps one wants to make sure retained records have no missing values for variables M1 M5 The test if SUM of Mi M5 then delete will delete records only where all 5 variables have missing values If Mi 1 and M2 M5 all have missing values then SUM of Mi M5 will return a value of 1 not a missing value Of interest too is that if A 10 B 20 C 30 andX 20 Y Z then MEAN of A B C MRAN of X Y Z 20 Surprise 8 Using the trim function to remove trailing blanks from character values doesn t work unless you include a length statement Unexpected though well documented in 82 3 Basics User s Manual on p 194 The trim function removes trailing blanks from a character variable s value and the resulting trimmed value has the length specified in an accompanying length statement If no length statement is specified the trim function in effect does nothing since any trimmed blanks are reinserted Example A XYZ 3 B TRIM A Here the value of B will be XYZ To geta value returned of XYZ the user needs to supply the statement Length B 3 Surprise 9 Variables created from the s
4. value for variable Y A is read using list input which will search for a blank to separate the two variables values Since no blank separates A s value from Y s value Y s value is not found and is set to missing Note however that it is safe to use Input Y 1 A with data of 8abcdefghij Since Y is read using an informat no blank in the data is searched for and A s value will be sought beginning in byte 2 Surprise 6 SAS software may interpret arguments to functions differently than you do Example Xi 1 X2 100 X3 1 X4 100 X5 1 Total SUM of X1 X2 X5 The value of Total above is 1 not 401 The reason here is that because of the preced ing comma SAS software considers X2 X5 to be a subtraction to be performed prior to the operation of the SUM function Since the value of X2 subtract X5 is 0 X1 O 1 0 1 The solution here is to remove the comma SUM of X1 X2 X5 401 Surprise 7 Functions may not generate missing values or may return unexpected values compared to doing the calculation by hand Example Given X 1 Y T missing value SUM of K Y T 3 but X Y T missing value if il Many SAS software functions will exclude variables with missing values and operate on 896 the remaining variables Hence one should try to avoid using functions to critically screen data For example suppose one would like to keep records where variables M1 through M5 have values greater
5. a value of 1 less than the true value Suggestion User Row 10 d instead of Row 0 4 SAS and SAS GRAPH are registered trademarks of SAS Institute Inc Cary NC USA 1 Doyle Bishop Chris Hamlin Don Hinman Bernie Schneider Bill Schneider Duane Schulte and Frank Scoviak For further information contacts Dr Calvin D Croy Senior Statistical Analyst National Demographics amp Lifestyles 1621 Eighteenth Street Denver Colorado 80202 USA 303 292 5000
6. aving GROUP values A through E and an integer weight value for variable X were run twice through the following code Proc Freq Tables GROUP Weight X Results Group Frequency Cumulative Frequenc Run 1 E 1 430 393 CH 294 606 Run 2 E 1 430 393 64 161 260 In the above example the final total frequency calculated was first 64 284 606 and when rerun 64 161 260 a difference of 123 346 or 0 5976 for each record processed Neither of these totals was correct 898 The true count total was 64 517 076 as calcu lated using counters in a data step and also by using Proc Summary with a Freq statement The inaccuracy of Proc Freq in this use remains a mystery Moral Use Proc Summary with a Freq statement to process very large sets of data Surprise 16 SAS software writes a missing value using packed decimal format as 0 However when such 0 values are read using the same packed decimal format with which they were written the value input is 0 The negative sign is lost In datasets written using packed decimal format the distinction between a missing value and 0 is lost when the data are read using a packed decimal informat Surprise 17 The column binary informats gt 5 Row 0 6 and Row 10 6 are not equal despite what the 82 3 Basics User s Manual shows i Consider a puch occurs in row 10 The informat Row 0 6 will return a value of 0 whereas Row 10 6 will return the value of 1 In general using Row 0 6 returns
7. e since its name will consist entirely of blanks A variable name made of nothing but blanks makes such scores very hard to later reference Example 1 Without use of optional label Proc Reg data Readin Outest Coefs Model SALES PRICE j Output out Regout Predicted YHAT Proc Score data New score Coefs out Scored type OLS Var PRICE Proc Print data Scored Data Scored PRICE 10 2000 50 500 30 1000 Example 2 With use of optional label Proc Reg data Readin Outest Coefs Score Model SALES PRICE Output out Regout Predicted YHAT Proc Score data New score Coefs out Scored type OLS Var PRICE Proc Print data Scored Data Scored PRICE SCORE 10 2000 50 500 30 1000 Moral If the regression coefficients created by Proc Reg are to be used later with Proc Score always use alabel in the Proc Reg s Model statement ae i i x 7 f y 4 Surprise 13 Using a permanent file for work space by giving it a dd name of Work in the JCL can cause headaches When one creates or uses a permanent external file by having a dd name of Work in the JCL the last line number of source code is stored in the external file regardless of the number or position of the other internal pointers This facilitates using statements such as include When the physical external space is reused for work space SAS software begins counting lines of source code beginning with the line
8. ervations 2 and 3 the contributing dataset mentioned latest in the merge statement is the only contributing dataset Data A For observations 2 and 3 the value of M is taken from Data A the only contributing dataset Why the value 20 for TOTAL is kept for all records is succinctly stated as well on p 105 of the 82 3 Basics User s Manual When a dataset runs out of observations for a by value values from the last observation contributed by the dataset are retained for the remaining observations in the by group Surprise 11 The sign of the t value gener ated by Proc T test depends upon the spellings of the class variable s values Let s say you always want to compare another company s data to your company s data If the other company s mean is significantly larger than your company s mean the t statistic should have a positive value If the other company s mean is significantly less than yours you expect the t value to be negative But in practice it doesn t work this way with Proc T test Proc T test orders the values of the class variable alphabetically on the printout and the mean of the data represented lowest 1s subtracted from the mean printed above Depending on the name you furnish to designate the other company s data their mean will appear either above or beneath yours regard less of the direction of difference in means Hence comparing the same sets of data ordered the same way can lead to diffe
9. rent signs for the t value depending upon the spelling selected Example Other company s variable TIME has mean of 5 Class variable GROUP given value of other Your company s variable TIME has mean of 2 Class variable GROUP given value of ours Proc T test Glass GROUP Var TIME 897 The listing from this program will show a positive t value since other will physically appear above ours Alphabetically ot comes before ou If the same data are run again but the class value for your company s data is changed from ours to mine the same program will generate a negative t value of the sam magni tude This is because mine will physically appear above other since alphabetically m precedes o Surprise 12 To use Proc Score with coef ficients from Proc Reg a label must have been present on the Model statement of Proc Reg This is true even though the label is optional i Proc Score uses the value of the _MODEL_ variable in the scoring coefficient dataset from Proc Reg When used with regression coefficients Proc Score takes the value of the MODEL variable to name the scores it creates The value of the MODEL variable is the label the user supplied with the original Model statement of Proc Reg If the user supplies no label then the value of _MODEL_ in the coef ficient dataset is blank and Proc Score will create scores for a variable without a useable nam
10. stored in the external file With each run the stored end line number is added to _ However a SAS program can contain only 32 000 lines of SAS code Thus when large SAS programs run frequently and reuse the external file the 32 000 line limit will eventually be reached and an error generated We have had no success in resetting the line number stored in the permanent external file Neither have we been able to destroy it using the kill option to Proc Datasets The only way we at NDL have been able to remove the offending line pointer has been to run a small PLi program to open and close the external file thus purging all pointers set in the file Moral Don t use Work as a dd name in ICL Surprise 14 Under CMS the device type must be supplied by the user either in the o tiong statement of the program executing SAS GRAPH or in a SAS profile When using SAS GRAPH under 0S SAS software can get the device type from the operating system so the user needn t explicitly declare it But to run SAS GRAPH under CMS the user does need to supply the device type since it can t be obtained from the operating system The real surprise here is that such require ments are listed in the 82 3 SAS Basics User s Manual rather than in the 82 3 SAS Companion for the VM CMS Operating System Surprise 15 Using Proc Freg with the Weight statement on very large datasets can produce incorrect and inconsistent counts Example 206 395 records h
11. two means The advice of a more advanced analyst may not be available and at such an elementary level reference to the Basics User s Manual probably won t help much Should the neophyte select Proc Means Proc Summary Proc Univariate Proc Freq Proc Corr Proc T test Proc GLM Proc ANOVA All these procedures can be used to perform statistical analyses useful in comparing two means For now the only answer is to talk to a more experienced analyst get other manuals read a statistical textbook or all three depending upon the potential user s background Surprise 2 The message printed in the SAS log as a result of an error may not have straightforward connection to the error made Restated error messages in the SAS log do not always help identify an error and can actually misidentify an error Example 1 Consider a title statement imbedded in a 300 line SAS program The user unwittingly places a single apostrophe within the title rather than the double apostrophe required by SAS software The resulting error message character literal has more than 200 characters can be confusing to those i uninitiated into the ways of SAS software This message is especially confusing if the programmer realizes that no character values have been used in the program 895 Example 2 The programmer places a semicolon in the wrong place Consider a data step opened with this accidentally split line of code Data First Keep X Second
12. ubstring function retain the length of the variable from which they were created unless specified otherwise Example Length Y 130 H SUBSTR Y 1 2 As can be revealed with a Proc Contents the length of H will be 130 not 2 In order to have H be of length 2 its length needs to be explicitly declared Length H 2 Surprise 10 How you merge determines what you get But is it what you want Merging can create unexpected holes in your data Consider the merge of Data B with Data A in order to have the variable TOTAL appended to each observation in the resulting Data C Data A Data B OBS ane T ops GROUP SEX TOTAL 2 5 1 M od et oe al abel Data C Merge A B By GROUP Data C will appear as OBS age SEX TO AL 3 i Why does observation 1 have a blank for the value of the character variable SEX when obser vations 2 and 3 have the anticipated value of M Answer The blank value arises because Data A s value for SEX was replaced by Data B s value a blank Also when multiple observa tions occur within a by group such as in Data A the value in the new dataset is the value from the dataset mentioned latest in the merge statement that is still contributing informa tion to the by group p 105 82 3 Basics User s Manual Thus for observation 1 the contributing dataset mentioned latest in the merge statement is Data B and the blank value for SEX is taken from B But for obs
13. uses to store the results of arithmetic operations SAS software unfortunately uses the floating point binary representations of numbers to make comparisons As this example shows the binary representa tion of 25 raised to the power of 5 is not perfectly equal to the binary representation of 5 If the variable Area s value is printed however the value printed will be 5 This occurs because SAS software converts the internal binary code to the nearest base 10 number prior to printing The printed value is base 10 but the tested value is binary floating point The execution of unintended code despite the seeming correctness of values when printed can be almost impossible to debug sus Vert fas in lengthy SAS programs Moral For critical applications round values to a user specified level of precision before testing for perfect matches Surprise 5 SAS software may interpret input statements differently from the way you think it does Consider the following code Data Example Length A 10 Input AY 1 Cards abcdefghi j8 5 A logical interpretation of how the length and input statements interact would be to assume that since character variable A has been declared to have a length of 10 then numeric variable Y would be sought for beginning in byte 11 In this example the anticipated value for A is abcdefghij and Y is 8 However when this code runs SAS software will return a missing
Download Pdf Manuals
Related Search