Notes for SAS programming
Econ424
Fall 2009
Why SAS?
•
Able to process large data set(s)
•
Easy to cope with multiple variables
•
Able to track all the operations on the data set(s)
•
Generate systematic output
• Summary statistics
• Graphs
• Regression results
•
Most government agencies and private sectors use
SAS
Where to find SAS?
•
•
The MS Window version of SAS is only available on campus
– class website, “computer resources”, “On-campus computer lab”
– list computer lab location, hours and software
You can access SAS remotely via glue.umd.edu, but you cannot use the
interactive windows
– use a secured telnet (e.g. ssh) to remotely login glue.umd.edu with your
directory ID and password
– type “tap sas” to tell the system that you want to access SAS
– use any text editor (say pico) to edit your sas program (say myprog.sas).
You can also create the text-only .sas file in your PC and (securely) ftp it
into glue.
– In glue, type “sas myprog.sas &” will send the sas program to run in the
background.
– The computer will automatically generate myprog.log to tell you how
each command runs in sas. If your program produces any output, the
output will be automatically saved in myprog.lst. All in the same
directory as your .sas file.
Roadmap
•
•
•
•
•
•
•
•
Thinking in “SAS”
Basic rules
Read in data
Data cleaning commands
Summary statistics
Combine two or more datasets
Hypothesis testing
Regression
Thinking in “SAS”
• What is a program?
– Algorithm, recipe, set of instructions
• How is programming done in SAS?
– SAS is like programming in any language:
• Step by step instructions
• Can create your own routines to process data
• Most instructions are set up in a logical manner
– SAS is NOT like other languages:
• Some syntax is peculiar to SAS
• Written specifically for statistics so it isn’t all-purpose
• Canned processes that you cannot edit nor can you see the
code
Thinking in “SAS”
• Creating a program
– What is your problem? (take project 3 as an example)
– How can you find a solution?
– What steps need to be taken to find an answer?
• Do I need to read in data?
– What variables do I need?
– Where is the data?
– What format is the data in?
• How do I need to clean the data?
– Are there outliers?
– Are there any unexpected values in the data?
• How do I need to transform the data?
– Are the variables in the form that I need?
Basic rules (1) – organize files
•
•
•
•
•
.sas – program file
.log – notes, errors, warnings
.lst – output
.sas7bdat – data file
library – a cabinet to put data in
– Default: Work library
• temporary, erased after you close the session
– Permanent library
• libname mylib “m:\”;
• mylib.mydata
= a sas data file named “mydata” in library “mylib”
• run and recall .sas
Basic rules (2) -- program
• every command ends with ;
• format does not matter
if x=1 then y=1; else y=2; is the same as
if x=1 then y=1;
else y=2;
• case insensitive
• comment
* this is comment;
/* this is comment */;
Basic rule (3) – variable
• Type
– numeric (default, 8 digit, . stands for missing value)
– character ($, default 8 digit, blank stands for missing)
• Variable names
– <=32 characters if SAS 9.0 or above
– <=8 characters if SAS 8 or below
– case insensitive
• Must start with letter or “_”
_name, my_name, zip5, u_and_me
-name, my-name, 5zip, per%, u&me, [email protected], my$sign
Basic rules (4) – data step
• Data step
create a new data set called “newdata” in the
temporary library
DATA newdata;
use the data set called “proj3rawdata” in
set proj3rawdata;
the temporary library
fracuninsured=uninsured/total;
percentuninsured=fracuninsured*100;
run;
Define new variables
input data
“proj3rawdata”
obs1
obs2
…
obs n
define
“fracuninsured”
define
“percentuninsured”
output data
“newdata”
obs1
…
obs n
Basic rules (5) – proc step
• PROC step
Action
PROC PRINT data=newdata;
var fracuninsured percentuninsured;
title “print out new data”;
run;
Signal the end of PROC step, could
be ignored if this is followed by a
Data or Proc step
Data
source
Read in data (1) – table editor
• Tools – table editor
• choose an existing library or define a new
library
• rename variable
• save as
Read in data (2) – by program
• Data format
–
–
–
–
Datalines (enter the data in the program)
Existing data text, comma or tab delimited
Existing data text, fixed width
From an existing excel file
• The following several slides won’t be
covered in class. But you are welcome to
use them by yourselves.
Read in data (2) -- datalines
data testdata1;
infile datalines;
input id height weight gender $ age;
datalines;
1 68 144 M 23
2 78 . M 34
No ; until you finish all
3 62 99 F 37
the data lines
;
/* you only need one semicolon at the end of all data lines, but the
semicolon must stand alone in one line */
proc contents data=testdata1; run;
proc print data=testdata1; run;
Read in data (3) – more datalines
/* read in data in fixed columns */
data testdata1;
infile datalines;
input id 1 height 2-3 weight 4-6 gender $7 age 8-9;
datalines;
168144M23
278 M34
36299 F37
;
Read in data (4) – more datalines
data testdata1;
infile datalines;
input id : 1. height : 2. weight : 3. gender : $1. age : 2.;
datalines;
1 68 144 M 23
2 78 . M 34
3 62 99 F 37
;
Read in data (4) – more datalines
*alternatively;
data testdata1;
infile datalines;
informat id 1. height 2. weight 3. gender $1. age 2.;
input id height weight gender age;
datalines;
1 68 144 M 23
2 78 . M 34
3 62 99 F 37
;
Read in data (5) – .csv
data testdata1;
infile datalines dlm=‘,’ dsd missover;
/* what if you do not have dsd and/or missover */
input id height weight gender $ age ;
datalines;
1, 68, 144, M, 23
2, 78, , M, 34
3, 62, 99, F, 37
; /* what if you forget to type 23 */
run;
Read in data (6) – .csv file
/* save the project 3 cleaned raw data in a comma delimited
file (.csv) in M:\ before running the following codes */
Restrict your filename <=8 characters
libname mylib "M:\";
filename mycsv "M:\project3-rawdata-cleaned.csv";
data mylib.proj3rawdata;
Be consistent with yourself
infile mycsv firstobs=2 dlm=',' dsd missover lrecl=900;
input year: 4. state: $2. total insured uninsured;
Be consistent with yourself
run;
proc contents data=mylib.proj3rawdata;
run;
proc print data=mylib.proj3rawdata;
run;
Note the options in the infile command!
Read in data (7) – from excel
filename myexcel “M:\project3-rawdata-cleaned.xls”;
proc import datafile=myexcel out=proj3rawdata2
DBMS=excel replace;
run;
No ; until
data proj3rawdata3;
Be consistent with yourselfthe end of
set proj3rawdata2;
the whole
fracuninsured=uninsured/total;
sentence
percentuninsured=fracuninsured*100;
run;
proc contents data=proj3rawdata2; run;
proc print data=proj3rawdata2; run;
proc contents data=proj3rawdata3; run;
proc print data=proj3rawdata3; run;
21
Read in data (7) – from excel
Be careful !
SAS will read the first line as variable names, and assume
the raw data start from the second row.
SAS assigns numeric and character type automatically.
Sometime it does make mistake.
22
Data cleaning (1) – if then
Format:
IF condition THEN action;
ELSE IF condition THEN action;
ELSE action;
Note:
(1) the if-then-else can be nested as many as you
want
(2) if you need multiple actions instead of one
action, use “DO; action1; action2; END; ”
Data cleaning (1) – if then
•
•
•
•
•
•
•
=
~=
>
<
>=
<=
in
or
or
or
or
or
or
EQ
NE
GT
LT
GE
LE
means equals
means not equal
means greater than
means less than
means greater than or equal
means less than or equal
means subset
– if gender in (‘M’, ‘F’) then ..;
• Multiple conditions: AND (&), OR(|)
Data cleaning (1) – if then
*reading in program of proj3rawdata3 is on page 21;
data proj3rawdata3;
set proj3rawdata3;
IF fracuninsured<0.15 THEN uninsuregrp=0;
ELSE uninsuregrp=1;
run;
proc contents data=proj3rawdata3; run;
proc print data=proj3rawdata3; run;
Note: (1) the code is less efficient if you replace ELSE ..; with
IF fracuninsured>=0.15 THEN ..;
(2) missing value is always counted as the smallest negative, so
fracuninsured=. will satisfy the condition fracuinsured<0.15. If
you want to ignore the missing obs set the condition as
0<=fracuninsured<0.15.
Data cleaning (1) – if then
* Multiple actions in each branch;
data proj3rawdata3;
set proj3rawdata3;
IF fracuninsured<0.15 AND uninsured>1000000 THEN DO;
uninsuredgrp=0; uninsuredpop=‘over 1 million';
END;
the do-end pair
ELSE DO;
acts as brackets
uninsuredgrp=1; uninsuredpop=‘less than 1 million';
END;
run;
proc print data=proj3rawdata3; run;
Data cleaning (1) – if then
*Use if commands to choose a subsample;
data proj3subsample; /* note here we generate a new data set */
set proj3rawdata3;
IF fracuninsured=. Then delete;
If fracuninsured<=0.1;
run;
proc print data=proj3subsample; run;
Data cleaning (1) – exercise
still use proj3rawdata.
define newgrp = 1 if fracuninsured <0.1 (low)
2 if 0.1<=fracuninsured<0.15 (mid-low)
3 if 0.15<=fracuninsured<0.2 (mid-high)
4 if fracuninsured>=0.2 (high).
Data cleaning (1) – exercise answer
data proj3rawdata3;
set proj3rawdata3;
if fracuninsured<0.1 then newgrp=1;
else if fracuninsured<0.15 then newgrp=2;
else if fracuninsured<0.2 then newgrp=3;
else newgrp=4;
run;
proc contents data=proj3rawdata3; run;
proc print data=proj3rawdata3; run;
Question: What if one observation has fracuninsured=.?
Save data
* Save in sas format;
libname mylib “M:\”;
data mylib,proj3rawdata3;
set proj3rawdata3;
run;
* Export data to excel;
Proc export data=proj3rawdata3
outfile=“M:\proj3data-fromsas.xls”
dbms=excel replace;
Run;
No ; here
You can also export a sas data file into a comma delimited text file if you write
dbms=csv.
Pages 31-34 are optional material for
data cleaning.
They are not required, but you may
find them useful in the future.
We skip them in the regular class.
Data cleaning (2)
– convert variable type
Numeric to character:
age1=put(age, $2.);
age and age1 have the same contents but different formats
Character to numeric:
age2=input(age1, 1.);
now age and age2 are both numeric, but age2 is chopped at the first
digit
Take a sub string of a character
age3=substr(age1,2,1);
now age3 is a sub string of age1, starting from the second digit of
age1 (the meaning of “2”) and having one digit in total (the meaning
of “1”).
Data cleaning (2) - example
* we want to convert studid 012345678 to 012-345-678;
data testdata2;
infile datalines;
input studid : 9. studname : $1.;
datalines;
012345678 A
135792468 B
009876543 C
;
proc print; run;
data testdata2;
set testdata2;
if studid LT 1E+7 then studid1= '00’||compress(put(studid, $9.));
else if 1E+7 LE studid LT 1E+8 then studid1='0'||compress(put(studid,
$9.));
else studid1= put(studid, $9.);
studid2=substr(studid1,1,3)||'-'||substr(studid1,4,3)||''||substr(studid1,7,3);
proc print; run;
Data cleaning (2) - exercise
You have the following data, variables in sequence are
SSN, score1, score2, score3, score4, score5:
123-45-6789 100 98 96 95 92
344-56-7234 69 79 82 65 88
898-23-1234 80 80 82 86 92
Calculate the average and standard deviation of the five
scores for each individual. Use if-then command to find
out who has the highest average score, and report his SSN
without dashes.
Data summary roadmap
•
•
•
•
•
•
•
•
•
Proc contents – variable definitions
Proc print – raw data
Proc format – make your print look nicer
Proc sort – sort the data
Proc means – basic summary statistics
Proc univariate – detailed summary stat
Proc freq – frequency count
Proc chart – histogram
proc plot – scatter plot
proc format (1)
*Continue the data cleaning exercise on page 29;
Defining “group” as a
data proj3rawdata3;
numeric variable will
save space
set proj3rawdata3;
if fracuninsured<0.1then newgrp=1;
else if fracuninsured<0.15 then newgrp=2;
else if fracuninsured<0.2 then newgrp=3;
else newgrp=4;
run;
proc format (2)
proc format;
value newgroup 1=‘low’
2=‘mid-low’
no ; here until
3=‘mid-high’
the end of the
4=‘high’;
value command
run;
proc print data=proj3rawdata3;
format newgrp newgroup. fracuninsured 4.2;
title ‘report new group in words’;
no ; here until
label fracuninsured=‘fraction uninsured’
the end of the
newgrp=‘new group 1-4’;
var fracuninsured newgrp;
label command
run;
proc sort
proc sort data=proj3rawdata3;
by year state;
run;
proc sort data=proj3rawdata3
out=proj3rawdata3_sorted;
by year descending fracuninsured;
run;
* note that missing value is always counted as the
smallest;
proc means and proc univariate
proc means data=proj3rawdata3;
class newgrp;
var insured uninsured fracuninsured;
run;
By default, proc means report
mean, stdev, min, max
Could choose what to report:
proc means data=proj3rawdata3
n mean median;
proc sort data=proj3rawdata3; by newgrp; run;
proc univariate data=proj3rawdata3;
by newgrp;
var insured uninsured fracuninsured;
run;
By default, proc univariate
report median, and many
other statistics
Notes on proc means and proc univariate
*if you do not use class or by command, the
statistics are based on the full sample. If you use
class or by var x, the statistics are based on the
subsample defined by each value of var x.
*You can use class or by in proc means, but only by
in proc univariate;
*whenever you use “by var x”, the data set should
be sorted by var x beforehand;
proc means and proc univariate
allow multiple groups
data proj3rawdata3;
set proj3rawdata3;
if totalpop<6000000 then popgrp=“low”;
else popgrp=“high”;
run;
proc means data=proj3rawdata3;
class newgrp popgrp;
var fracuninsured;
run;
proc sort data=proj3rawdata3;
by newgrp popgrp;
run;
proc univariate data=proj3rawdata3;
by newgrp popgrp;
var fracuninsured;
run;
proc freq
* Remember we already generate a variable called newgrp
to indicate categories of fraction uninsured and a variable
called popgrp to indicate categories of population size;
proc freq data=proj3rawdata3;
tables newgrp
One dimension frequency table
popgrp
newgrp*popgrp;
run;
Two-dimension frequency table
proc chart – histogram for
categorical variables
proc chart data=proj3rawdata3;
title ‘histogram for newgrp’;
vbar newgrp;
run;
proc chart data=proj3rawdata3;
title ‘frequency by two variables’;
vbar newgrp / group=popgrp;
run;
proc chart – histogram for
continuous variable
proc chart data=proj3rawdata3;
title “histogram for continuous variable’;
vbar fracuninsured;
run;
proc chart data=proj3rawdata3;
title ‘histogram with specific midpoints’;
vbar fracuninsured / midpoints=0 to 1 by 0.05;
run;
proc plot – scatter plot
proc plot data=proj3rawdata3;
title ‘scatter plot of fracuninsured and
totalpop’;
plot fracuninsured*totalpop;
run;
scatter plot is less informative for
categorical variables
proc plot data=proj3rawdata3;
title ‘scatter plot of newgrp and popgrp’;
plot newgrp*popgrp;
run;
fancy proc means
proc means data=proj3rawdata3;
class newgrp popgrp;
var uninsured fracuninsured;
output
out = summary1
mean = avguninsured avgfracuninsured;
run;
proc print data=summary1;
run;
The following page may be
useful in practice, but I am not
going to cover it in class.
some summary stat. in proc print
* Assume we have already defined newgrp and
popgrp in proj3rawdata3;
proc sort data=proj3rawdata3; by popgrp; run;
proc print data=proj3rawdata3 n;
where fracuninsured>=0.1;
by popgrp;
sum totalpop;
var totalpop insured uninsured fracuninsured;
run;
How to handle multiple data sets?
• Add more observations to an existing data
and the new observations follow the same
data structure as the old one  append
• Add more variables to an existing data and
the new variables refer to the same subjects
as in the old data  merge
• Sometimes we may need to change data
structure to fit in append or merge ….
merge and append
proj3rawdata3: year state totalpop … fracuninsured newgrp popgrp
2009 MA 6420947 … 0.0548
summary1:
newgrp
1
1
high
popgrp avguninsured avgfracuninsured
high
7500000
0.073
merged:
year state totalpop ….fracuninsured newgrp popgrp avguninsure avgfracuninsured
2009 MA 6420947 … 0.0548
1
high
7500000
0.073
appended:
year state totalpop ….fracuninsured newgrp popgrp avguninsure avgfracuninsured
2009 MA 6420947 … 0.0548
1
high
.
.
.
.
.
1
high
7500000
0.073
merge two datasets
proc sort data=proj3rawdata3;
by newgrp popgrp;
run;
proc sort data=summary1;
by newgrp popgrp;
run;
data merged;
merge proj3rawdata3 (in=one) summary1 (in=two);
by newgrp popgrp;
if one=1 & two=1;
What if this line is
run;
“if one=1 OR two=1;”?
Keep track of matched and
unmatched records
data allrecords;
merge proj3rawdata3 (in=one) summary1 (in=two);
by newgrp popgrp;
myone=one;
SAS will drop variables
mytwo=two;
“one” and “two”
if one=1 or two=1;
automatically at the end of
run;
the DATA step. If you want
proc freq data=allrecords;
to keep them, you can copy
tables myone*mytwo;
them into new variables
“myone” and “mytwo”
run;
be careful about merge!
• always put the merged data into a new data set
• must sort by the key variables before merge
• ok for one-to-one, multi-to-one, one-to-multi, but
no good for multi-to-multi
• be careful of what records you want to keep, and
what records you want to delete
• what if variable x appears in both datasets, but x
is not in the “by” statement?
– after the merge x takes the value defined in the last
dataset of the “merge” statement
append
data appended;
set proj3rawdata3 summary1;
run;
proc print data=appended;
run;
proc print data=merged;
run;
Class example of merge and
append: reshape and summarize
Task1: reshape proj3rawdata3 from long to wide
Task2: generate average fracuninsured per state and merge it
back to the main data
Source format of Proj3rawdata3 (long):
year state totalpop fracuninsured ….
2009 MA 6420947 0.0548
….
2009 HI 1257622 0.078
….
….
2008 MA 6339513 0.0536
….
2008 HI
1267409 0.075
…..
Target format (wide)
state totalpop2009 fracuninsured2009 . .. Totalpop2008
MA 6420947
0.0548 ….
6339513
HI 1257622
0.078
….
1267409
fracuninsured2008 …..
0.0536 ……
0.075 ……..
Main data issues
• From long to wide, variable names are
different except for the key variable (state)
• We can generate average fracuninsured per
state either before or after the reshape, but
the merge code will be different depending
on when we compute the average
fracuninsured
Step 1 to reshape:
generate a sub-sample for each year,
so that we have:
subsample2009
subsample2008
….
subsample2003
focus on 2009
data subsample2009;
set proj3rawdata3;
if year=2009;
run;
data subsample2009;
set subsample2009;
rename totalpop=totalpop2009;
rename insured=insured2009;
rename uninsured=uninsured2009;
rename fracuninsured=fracuninsured2009;
drop year;
run;
Same for 2008
data subsample2008;
set proj3rawdata3;
if year=2008;
run;
data subsample2008;
set subsample2008;
rename totalpop=totalpop2008;
rename insured=insured2008;
rename uninsured=uninsured2008;
rename fracuninsured=fracuninsured2008;
drop year;
run;
Step 2 to reshape: merge each year’s
subsample by state
proc sort data=subsample2009; by state; run;
…
proc sort data=subsample2003; by state; run;
data reshaped;
merge subsample2009 subsample2008 subsample2007
subsample2006 subsample2005 subsample2004
subsample2003;
by state;
run;
proc print data=reshaped; run;
Generate average fracuninsured
by state
proc means data=proj3rawdata3;
class state;
var fracuninsured;
output out=avgperstate
mean=avgfrac_bystate;
run;
Merge the data file “avgperstate”
back to reshaped
proc sort data=reshaped;
by state;
run;
proc sort data=avgperstate;
by state;
run;
data reshaped_withavg;
merge reshaped (in=one) avgperstate (in=two);
by state;
myone=one;
mytwo=two;
if one=1 | two=1;
run;
proc print data=reshaped_withavg; run;
Check observations in
reshaped_withavg
proc freq data=reshaped_withavg;
tables myone*mytwo;
run;
mean comparison: two groups
Example: does the average fracuninsured differ between 2008 and 2009?
H0: mean of fracuninsured2008 = mean of fracuninsured2009.
H1: mean of fracuninsured2008 not equal to mean of fracuninsured2009.
This is a two-tail mean-comparison test between the 2008 sample and
the 2009 sample.
The test result will be different if
(1)
(2)
we treat 2008 and 2009 as two independent samples; or
We treat 2008 and 2009 as matched pairs (matched by state).
mean comparison: two groups
SAS performs mean comparison in a regression framework.
H0: mean of fracuninsured2008 = mean of fracuninsured2009.
H1: mean of fracuninsured2008 not equal to mean of fracuninsured2009.
Step 1: focus on the subsample that has 2008 and 2009 data only.
Step 2: create a binary variable dummy2009=1 if year=2009, 0 if
year=2008.
Step 3: depend on whether 2008 and 2009 are independent samples or
matched pairs.
If independent samples, regress fracuninsured as:
fracuninsured = a + b* dummy2009 + error
If matched pairs, regress fracuninsured as:
fracuninsured = a + b * dummy2009
+ c1* dummy_AL + c2* dummy_AK
+ …+c51*dummy_WY + error
mean comparison: two groups as
two independent samples
* Focus on 2008 and 2009 data only;
Data subsample0809;
Set proj3rawdata3;
If year=2008 or year=2009;
If year=2009 then dummy2009=1; else dummy2009=0;
Run;
* Treat 2008 and 2009 as two independent samples;
Proc reg data=subsample0809;
Model fracuninsured=dummy2009;
Run;
The difference between 2008 and 2009 is captured in the coefficient of dummy2009.
So the hypothesis test is equivalent to:
H0: coefficient of dummy2009=0.
H1: coefficient of dummy2009 not equal to 0.
mean comparison: two groups as
matched pairs (matched by state)
* Focus on 2008 and 2009 data only;
data subsample0809;
set proj3rawdata3;
if year=2008 or year=2009;
if year=2009 then dummy2009=1; else dummy2009=0;
run;
* Treat 2008 and 2009 as matched pairs (matched by state);
proc glm data=subsample0809;
class state;
model fracuninsured =dummy2009 state/solution;
run;
The within-state difference between 2008 and 2009 is captured in the coefficient of
dummy2009.
So the hypothesis test is equivalent to:
H0: coefficient of dummy2009=0.
H1: coefficient of dummy2009 not equal to 0.
mean comparison: more than two groups
Example: does the average fracuninsured differ between any two years
in our data? (here year = 2003, 2004, … 2009)
The test result will be different if
(1)
(2)
we treat year t and year t’ as two independent samples; or
We treat year t and year t’ as matched pairs (matched by state).
mean comparison: every year as an
independent sample
* Treat every year as an independent sample;
* Now we need to use the whole data of proj3rawdata3;
proc glm data=proj3rawdata3;
class year;
model fracuninsured=year;
means year/waller;
means year/lsd cldiff;
run;
mean comparison: every year as
matched pairs (matched by state)
* Treat every year as matched by state;
* First define dummies for each year;
data proj3rawdata3;
set proj3rawdata3;
if year=2004 then dummy2004=1; else dummy2004=0;
if year=2005 then dummy2005=1; else dummy2005=0;
…
if year=2009 then dummy2009=1; else dummy2009=0;
run;
proc glm data=proj3rawdata3;
class state;
model fracuninsured=dummy2004 dummy2005 dummy2006 dummy2007
dummy2008 dummy2009 state/solution;
run;
The coefficient of dummy2004 captures the difference between 2003 and 2004.
The coefficient of dummy2005 captures the difference between 2003 and 2005.
notes on mean comparison
1.
The logic of mean comparison is the same as in Excel
2.
Be careful about one-tail and two-tail tests.
The standard SAS output of coefficient t-stat and p-value are based on a two-tail test
of H0: coeff=0, but could be used for a one-tail test if we compare 1-alpha vs. pvalue/2 instead of p-value.
3.
Comparison across more than two groups (as independent samples)
H0: all groups have the same mean
 F-test of whole regression
OR
H0: group x and group y has the same mean
 waller or lsd statistics
4.
Comparison across more than two groups (as matched pairs) requires specific test
on regression coefficients.
H0: 2003 = 2004  test the coefficient of dummy2004=0 because 2003 is set
as the benchmark
H0: 2004=2005  test the coefficient of dummy2004 = coefficient of
dummy2005.
Explicit hypothesis tests
in proc reg and proc glm
* Treating each year as independent samples;
Proc reg data=proj3rawdata3;
Model fracuninsured=dummy2004 dummy2005 dummy2006 dummy2007
dummy2008 dummy2009;
Test fracuninsured2003=fracuninsured2008;
Test dummy2008=0;
Test dummy2008=dummy2009;
Test fracuninsured2008=fracuninsured2009;
Run;
* Treating each year as matched by state;
Proc glm data=proj3rawdata3;
Class year state;
Model fracuninsured=year state/solution;
Contrast ‘test 2003 vs. 2008’
year 1 0 0 0 -1 0;
Test fracuninsured2003=fracuninsured2008;
Contrast ‘test 2008 vs. 2009’
year 0 0 0 0 1 -1;
Test fracuninsured2008=fracuninsured2009;
Run;
In class exercise
for mean comparison
Main question:
compare fracuninsured in west, midatlantic and everywhere else, where
west = CA, WA, OR
midatlantic = DE, DC, MD, VA
step 0: define west, midatlantic and everywhereelse
exercise 1: compare west and midatlantic
1(a). as two independent samples
1(b). consider the match by year
exercise 2: compare west, midatlantic, and everywhere else
2(a). as three independent samples;
2(b). consider the match by year;
regression in SAS
Question: how do fracuninsured vary by total population of a state?
* model: fracuninsured=a+b*totalpop+error;
proc reg data=proj3rawdata3;
model fracuninsured=totalpop;
run;
* Add year fixed effects;
* Model: fracuninsured=a+b*totalpop+c1*dummy2004
+c2*dummy2005 + …+c51*dummy2009+error;
proc glm data=proj3rawdata3;
class year;
model fracuninsured=totalpop year/solution;
run;
A comprehensive example
A review of
1. readin data
2. summary statistics
3. mean comparison
4. regression
reg-cityreg-simple.sas in N:\share\
A Case Study of Los Angeles Restaurants

Nov. 16-18, 1997 CBS 2 News “Behind the Kitchen Door”

January 16, 1998, LA county inspectors start issuing hygiene
grade cards


A grade if score of 90 to 100

B grade if score of 80 to 89

C grade if score of 70 to 79

score below 70 actual score shown
Grade cards are prominently displayed
in restaurant windows

Score not shown on grade cards
Take the idea to data
Research Question:
Does better information lead to better hygiene quality?
better
Information
better
quality
regulation
hygiene
scores
by county
by city
Data complications
(blue font indicates our final choices)
Unit of analysis:
individual restaurant? city? zipcode? census tract?
Unit of time:
each inspection? per month? per quarter? per year?
Define information:
county regulation? city regulation? the date of passing the regulation?
days since passing the regulation? % of days under regulation?
Define quality:
average hygiene score? the number of A restaurants? % of A
restaurants?
How to test the idea?
• Regression:
quality = α+β*information+error
+something else?
Something else could be:
year trend, seasonality, city specific effects, ….
real test
reg-cityreg-simple.sas in N:\share\
Questions
• How many observations in the sample?
– log of the first data step, or output from proc contents
• How many variables in the sample? How many
are numerical, how many are characters?
– Output from proc contents
• How many percentage of restaurants have A
quality in a typical city-month?
– Output from proc means, on per_A
Questions
• What is the difference between cityreg and ctyreg? We know county
regulation came earlier than city regulation, is that reflected in our
data?
– Yes, cityreg<=ctyreg in every observation
– We can check this in proc means for cityreg and ctyreg, or add a proc
print to eyeball each obs
• What is the difference between cityreg and citymper? What is the
mean of cityreg? What is the mean of citymper? Are they consistent
with their definitions?
– The unit of cityreg is # of days, so it should be a non-negative integer
– The unit of citymper is % of days, so it should be a real number between
0 and 1
– To check this, we can add a proc means for cityreg and citymper
Questions
• Economic theories suggest quality be higher after
the regulation if regulation gives consumers
better information. Is that true?
– The summary statistics reported in proc means
(class citym_g or ctym_g) show the average
percentage of A restaurants in different
regulation environments.
– Rigorous mean comparison tests are done in
proc glm with waller or lsd options.
Questions
Summary statistics often reflect many economic factors, not
only the one in our mind. That is why we need regressions.
Does more regulation lead to higher quality?
– is the coefficient of city regulation positive and
significantly different from zero? (proc reg)
– is the coefficient of county regulation positive and
significantly different from zero? (proc reg)
– Do we omit other sensible explanations for quality
changes? What are they? (proc glm, year, quarter, city)
Count duplicates (not required)
http://support.sas.com/ctx/samples/index.jsp?sid=67&tab=co
de
data dups nodups ;
set clasdata ;
by name class ;
/* If the combination of NAME and CLASS is in the data
set once, output NODUPS, else output DUPS. */
if first.class and last.class then output nodups ;
else output dups ;
run;
course evaluation
• University wide:
– www.CourseEvalUM.umd.edu
• TTclass in particular: (password plstt)
– www.surveyshare.com/survey/take/?sid=81087
Descargar

Notes for SAS programming - University of Maryland