I have a dataset with visitors and weather variables. I'm trying to forecast visitors based on the weather variables. Since the dataset only consists of visitors in season there is missing values and gaps for every year. When running proc reg in sas it's all okay but the issue comes when i'm using proc VARMAX. I cannot run the regression due to missing values. How can i tackle this?
proc varmax data=tivoli4 printall plots=forecast(all);
id obs interval=day;
model lvisitors = rain sunshine averagetemp
dfebruary dmarch dmay djune djuly daugust doctober dnovember ddecember
dwednesday dthursday dfriday dsaturday dsunday
d_24Dec2016 d_05Dec2013 d_24Dec2017 d_24Dec2014 d_24Dec2015 d_24Dec2019
d_24Dec2018 d_24Sep2012 d_06Jul2015
d_08feb2019 d_16oct2014 d_15oct2019 d_20oct2016 d_15oct2015 d_22sep2017 d_08jul2015
d_20Sep2019 d_08jul2016 d_16oct2013 d_01aug2012 d_18oct2012 d_23dec2012 d_30nov2013 d_20sep2014 d_17oct2012 d_17jun2014
dFrock2012 dFrock2013 dFrock2014 dFrock2015 dFrock2016 dFrock2017 dFrock2018 dFrock2019
dYear2015 dYear2016 dYear2017
/p=7 q=2 Method=ml dftest;
garch p=1 q=1 form=ccc OUTHT=CONDITIONAL;
restrict
ar(3,1,1)=0, ar(4,1,1)=0, ar(5,1,1)=0,
XL(0,1,13)=0, XL(0,1,14)=0, XL(0,1,13)=0, XL(0,1,27)=0, XL(0,1,38)=0, XL(0,1,42)=0;
output lead=10 out=forecast;
run;
As with any forecast, you will first need to prepare your time-series. You should first run through your data through PROC TIMESERIES to fill-in or impute missing values. The impute choice that is most appropriate is dependent on your variables. The below code will:
Sum lvisitors by day and set missing values to 0
Set missing values of averagetemp to average
Set missing values of rain, sunshine, and your variables starting with d to 0 (assuming these are indicators)
Code:
proc timeseries data=have out=want;
id obs interval = day
setmissing = 0
notsorted
;
var lvisitors / accumulate=total;
crossvar averagetemp / accumulate=none setmissing=average;
crossvar rain sunshine d: / accumulate=none;
run;
Important Time Interval Consideration
Depending on your data, this could bias your error rate and estimates since you always know no one will be around in the off-season. If you have many missing values for off-season data, you will want to remove those rows.
Since PROC VARMAX does not support custom time intervals, you can instead create a simple time identifier. You can alternatively turn this into a format for proc format and converttime_id at the end.
data want;
set have;
time_id+1;
run;
proc varmax data=want;
id time_id interval=day;
...
output lead=10 out=myforecast;
run;
data myforecast;
merge myforecast
want(keep=time_id date)
;
by time_id;
run;
Or, if you made a format:
data myforecast;
set myforecast;
date = put(time_id, timeid.);
drop time_id;
run;
Related
For an if-query I would like to create a macro varibale giving the respective frequency of the underlying time
series. I tried to get some descriptive statistics from proc time series. However, they unfortunately do not include the figure for the frequency.
The underlying times series does not necessarily conclude all periods of the frequency. That excludes a selected count by proc sql from my point of view.
Does anyone know an efficient procedure to determine the frequency without computing the frequency on my own (in a data step or a proc sql code)?
You can use the outspectra statement to help learn what kind of seasonality it has. Based on the data, give PROC TIMESERIES your best guess of day, month, etc. In the example below, we know we want to forecast by month but we do not know what seasonality it has.
proc timeseries data=sashelp.air outspectra=spectra;
id date interval=month;
var air;
run;
Plot this spectra dataset in proc sgplot and you'll see something that looks like this:
proc sgplot data=spectra;
where NOT missing(period);
series x=period y=p;
run;
This line will naturally increase over time, but we're looking for a bumps in the line. Notice the large bump somewhere between 0 and 24 months and the several smaller bumps before it. Let's zoom in on that by filtering out the longer periods.
proc sgplot data=spectra;
where period < 24 and NOT missing(period);
series x=period y=p;
run;
It's pretty clear that there is a strong seasonality of 12, with potentially smaller cycles at 3 and 6 months. From this plot, we can conclude that our seasonality should be 12 based on our spectra plot.
You can turn this into a macro to help identify the season if you'd like. Simply search for the largest bump within a reasonable timeframe. In our case we'll choose 36 because we do not suspect that we have any seasonality > 36 months.
proc sort data=spectra;
by period;
run;
data identify_period;
set spectra;
by period;
where NOT missing(period) AND period LE 36;
delta = abs(p - lag(p) );
run;
proc sql;
select period, max(delta) as max_delta
from identify_period
having delta = max(delta)
;
quit;
Output:
PERIOD max_delta
12 163712
I don't know how to do this without data step logic, but you could wrap the data step in a macro as follows:
%macro get_frequency(data,date_variable,output_variable);
proc sort data=&data (keep=&date_variable) out=__tempsorted;
by &date_variable;
run;
data _null_;
set __tempsorted end=lastobs;
prevdate=lag(&date_variable);
if _n_ > 1 then do;
interval_number+1;
interval_total + (&date_variable - prevdate);
end;
if lastobs then do;
average_interval = interval_total/interval_number;
frequency = round(365.25/average_interval);
call symput ("&output_variable",left(put(frequency,best32.)));
end;
run;
proc datasets nolist;
delete __tempsorted;
run;
quit;
%mend get_frequency;
Then you can call the macro on your original data set timeseries to examine the variable date and create a new macro variable frequency1 with the required frequency.
data work.timeseries;
input date date. value;
format date date9.;
datalines;
01Oct18 3000
01Nov18 4000
01Dec18 6500
01Jan19 7000
01Feb19 4000
01Mar19 5000
01Apr19 7500
01May19 4800
01Jun19 4500
;
run;
%get_frequency(timeseries,date,freqency1)
%put &=frequency1;
This seems to work on your sample data where each date is the first of the month. If your dates are evenly distributed (e.g. always near month start/end, or always near mid-month etc.) then this macro should work ok. Obviously if you have multiple observations per date then it will give the completely incorrect frequency.
When I run a proc glimmix in SAS, sometimes it drops observations.
How do I get the set of dropped/excluded observations or maybe the set of included observations so that I can identify the dropped set?
My current Proc GLIMMX code is as follows-
%LET EST=inputf.aarefestimates;
%LET MODEL_VAR3 = age Male Yearc2010 HOSPST
Hx_CTSURG Cardiogenic_Shock COPD MCANCER DIABETES;
data work.refmodel;
set inputf.readmref;
Yearc2010 = YEAR - 2010;
run;
PROC GLIMMIX DATA = work.refmodel NOCLPRINT MAXLMMUPDATE=100;
CLASS hospid HOSPST(ref="xx");
ODS OUTPUT PARAMETERESTIMATES = &est (KEEP=EFFECT ESTIMATE STDERR);
MODEL RADM30 = &MODEL_VAR3 /Dist=b LINK=LOGIT SOLUTION;
XBETA=_XBETA_;
LINP=_LINP_;
RANDOM INTERCEPT/SUBJECT= hospid SOLUTION;
OUTPUT OUT = inputf.aar
PRED(BLUP ILINK)=PREDPROB PRED(NOBLUP ILINK)=EXPPROB;
ID XBETA LINP hospst hospid Visitlink Key RADM30;
NLOPTIONS TECH=NRRIDG;
run;
Thank you in advance!
It drops records with missing values in any variable you're using in the model, in a CLASS, BY, MODEL, RANDOM statement. So you can check for missing among those variables to see what you get. Usually the output data set will also indicate this by not having predictions for the records that are not used.
You can run the code below.
*create fake data;
data heart;set sashelp.heart; ;run;
*Logistic Regression model, ageCHDdiag is missing ;
proc logistic data=heart;
class sex / param=ref;
model status(event='Dead') = ageCHDdiag height weight diastolic;
*generate output data;
output out=want p=pred;
run;
*explicitly flag records as included;
data included;
set want;
if missing(pred) then include='N'; else include='Y';
run;
*check that Y equals total obs included above;
proc freq data=included;
table include;
run;
The output will show:
The LOGISTIC Procedure
Model Information
Data Set WORK.HEART
Response Variable Status
Number of Response Levels 2
Model binary logit
Optimization Technique Fisher's scoring
Number of Observations Read 5209
Number of Observations Used 1446
And then the PROC FREQ will show:
The FREQ Procedure
Cumulative Cumulative
include Frequency Percent Frequency Percent
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒ
N 3763 72.24 3763 72.24
Y 1446 27.76 5209 100.00
And 1,446 records are included in both of the data sets.
I think I answered my question.
The code line -
OUTPUT OUT = inputf.aar
gives the output of the model. This table includes all the observations used in the proc statement. So I can match the data in this table to my input table and find the observations that get dropped.
#REEZA - I already looked for missing values for all the columns in the data. Was not able to identify the records there are getting dropped by only identifying the no. of records with missing values. Thanks for the suggestion though.
Is it possible to score a data set with a model created by PROC ARIMA in SAS?
This is the code I have that is not working:
proc arima data=work.data;
identify var=x crosscorr=(y(7) y(30));
estimate outest=work.arima;
run;
proc score data=work.data score=work.arima type=parms predict out=pred;
var x;
run;
When I run this code I get an error from the PROC SCORE portion that says "ERROR: Variable x not found." The x column is in the data set work.data.
proc score does not support autocorrelated variables. The simplest way to get an out-of-sample score is to combine both proc arima and a data step. Here's an example using sashelp.air.
Step 1: Generate historical data
We leave out the year 1960 as our score dataset.
data have;
set sashelp.air;
where year(date) < 1960;
run;
Step 2: Generate a model and forecast
The nooutall option tells proc arima to only produce the 12 future forecasts.
proc arima data=have;
identify var=air(12);
estimate p=1 q=(2) method=ml;
forecast lead=12 id=date interval=month out=forecast nooutall;
run;
Step 3: Score
Merge together your forecast and full historical dataset to see how well the model did. I personally like the update statement because it will not replace anything with missing values.
data want;
update forecast(in=fcst)
sashelp.air(in=historical);
by Date;
/* Generate fit statistics */
Error = Forecast-Air;
PctError = Error/Air;
AbsPctError = abs(PctError);
/* Helpful for bookkeeping */
if(fcst) then Type = 'Score';
else if(historical) then Type = 'Est';
format PctError AbsPctError percent8.2;
run;
You can take this code and convert it into a generalized macro for yourself. That way in the future, if you wanted to score something, you could simply call a macro program to get what you need.
I want to calculate the 95th percentile of a distribution. I think I cannot use proc means because I need the value, while the output of proc means is a table. I have to use the percentile to filter the dataset and create another dataset with only the observations greater than the percentile.
Clearly I don't want to use the numeric value..because I want to use it in a macro.
Don't put summary statistics into macro variables. You risk loss of precision.
This is based on your cryptic description of the problem.
proc means...
output out=pct95 pct95=
run;
data subset;
if _n_ eq 1 then set pct95;
set data;
if value < pct95;
run;
You can suppress proc means from outputting your results in a new tab using the noprint option. Try this:
proc means data = your_data noprint;
var variable_name;
output out = your_data2 p95= / autoname;
run;
I have a data set of patient information where I want to count how many patients (observations) have a given diagnostic code. I have 9 possible variables where it can be, in diag1, diag2... diag9. The code is V271. I cannot figure out how to do this with the "WHERE" clause or proc freq.
Any help would be appreciated!!
Your basic strategy to this is to create a dataset that is not patient level, but one observation is one patient-diagnostic code (so up to 9 observations per patient). Something like this:
data want;
set have;
array diag[9];
do _i = 1 to dim(diag);
if not missing(diag[_i]) then do;
diagnosis_Code = diag[_i];
output;
end;
end;
keep diagnosis_code patient_id [other variables you might want];
run;
You could then run a proc freq on the resulting dataset. You could also change the criteria from not missing to if diag[_i] = 'V271' then do; to get only V271s in the data.
An alternate way to reshape the data that can match Joe's method is to use proc transpose like so.
proc transpose data=have out=want(keep=patient_id col1
rename=(col1=diag)
where=(diag is not missing));
by patient_id;
var diag1-diag9;
run;