/*
Attempt to replicate:
Sugata Roychowdhury, 2006. Earnings management through real activities manipulation.
Journal of Accounting and Economics (42) pp. 350-370
Section 4.1 on page 343 describes sample selection:
- all firms in Compustat between 1987 and 2001
- dropping regulated firms (drop between 4400 and 5000, and between 6000 and 6500)
- at least 15 observations for each 2-digit SIC-fyear
- data availability to construct variables
“Imposing all the data-availability requirements yields 21,758 firm-years over the
period 1987–2001, including 36 industries and 4,252 individual firms.”
Issues replicating
- The sample size is much larger (64,257 observations vs 21,758 in the paper; I can reduce it to 37,768 when filtering on exchanges)
- Table 2: Coefficients for models 1-3 are (somewhat) similar (but much less significant compared to paper)
- Table 2: Coefficient for model 4 (accruals) in table 2 are very different
- Table 4: Coefficients are similar, standard deviations are larger so no significant findings
(maybe this has to do with not correcting for auto correlation)
*/
%let wrds=wrds-cloud.wharton.upenn.edu 4016;
options comamid=TCP remote=WRDS;
signon username=prompt;
libname local ‘C:\Users\Xiaoqin(Patrick) Wei\Desktop\Chapter3 data collection\Real earning management’;
options nolabel nocenter;
rsubmit;
/* Key variables */
proc sql;
create table a_comp as
select gvkey, datadate, fyear, sich, sale, at, xsga, xrd, xad, cogs, invt, oancf, ib, ppegt, exchg,
prcc_f * csho as mcap, calculated mcap / ceq as mtb, log(calculated mcap) as size, ceq / at as lev
from comp.funda
where 1992 <= fyear <= 2022
and indfmt=’INDL’ and datafmt=’STD’ and popsrc=’D’ and consol=’C’;
quit;
/* lagged values for assets (at), sales (sale), inventory (invt), ppe (ppegt)/ proc sql; create table b_comp as select a., b.at as at_lag, b.sale as sale_lag, b.invt as invt_lag, b.ppegt as ppegt_lag,
a.sale / b.sale -1 as sale_gr /* percent sales growth */
from a_comp a, comp.funda b
where a.gvkey = b.gvkey and a.fyear – 1 = b.fyear
and b.indfmt=’INDL’ and b.datafmt=’STD’ and b.popsrc=’D’ and b.consol=’C’;
quit;
/* 2-year lagged sales (needed for lagged change in sales) / proc sql; create table b_comp2 as select a., b.sale as sale_lag2
from b_comp a, comp.funda b
where a.gvkey = b.gvkey and a.fyear – 2 = b.fyear
and b.indfmt=’INDL’ and b.datafmt=’STD’ and b.popsrc=’D’ and b.consol=’C’;
quit;
/* header SIC (from comp.company) / proc sql; create table c_comp as select a., b.sic from b_comp2 a, comp.company b where a.gvkey = b.gvkey;
quit;
proc download data=c_comp out = c_comp;run;
/* key variables to be winsorized; also used to drop observations that have a missing value for any of these */
%let keyVars = one_at sale_ch_at sale_ch2_at sale_lag_at sale_at ppe_at accruals_at prod_at disexp_at cfo_at mcap mtb size lev sale_gr;
/* create main variables / data c_comp2; set c_comp; / missing R&D or advertising replace with 0 */
if xrd eq . then xrd = 0;
if xad eq . then xad = 0;
/* drop SIC 4400-5000, 6000-6500 */
if sic >= 4400 and sic <5000 then delete; if sic >= 6000 and sic <6500 then delete;
/* 2-digit SIC */
sic2 = floor(sic/100);
/* require positive lagged assets (used as scalar) */
if at_lag > 0;
/* net income scaled by lagged assets */
ni = ib / at_lag;
/* indicator, 1 if ib / lagged assets is between 0 and 0.005 */
SUSPECT_NI = ( 0 <= ni <= 0.005 );
/* change in sales (and change in lagged sales)*/
sale_ch = sale – sale_lag;
sale_ch2 = sale_lag – sale_lag2;
/* dependent variables, all scaled by lagged assets */
cfo_at = oancf / at_lag;
disexp_at = (xrd + xad + xsga) / at_lag;
prod_at = (cogs + invt – invt_lag ) / at_lag;
accruals_at = (ib – oancf) / at_lag;
/* control variables */
ppe_at = ppegt_lag / at_lag;
sale_at = sale / at_lag;
sale_lag_at = sale_lag / at_lag;
sale_ch_at = sale_ch / at_lag;
sale_ch2_at = sale_ch2 / at_lag;
one_at = 1 / at_lag;
/* key variables may not be missing (cmiss function counts missing values of a list of variables)*/
if cmiss (of &keyVars) eq 0;
/* dummy for main exchanges (not mentioned in paper, but brings sample size more in line with #obs in paper)*/
exch_main = ( exchg in (11,12,14) );
run;
/* winsorize */
filename mwins url ‘http://www.koopsom.com/macros/winsorize.sas’;
%include mwins;
/* winsorize &keyVars */
%winsor(dsetin=c_comp2, dsetout=c_comp2_wins, vars=&keyVars, type=winsor, pctl=1 99);
/* keep 15 obs per industry-year (35,342 obs remain)*/
proc sql;
create table d_15obs as
select * from c_comp2_wins
group by sic2, fyear
/* at least 15 obs in each sic2-fyear/ having count() >= 15;
quit;
/* Calculate relative net income (ni_rel) as net income (ni) minus industry-year mean */
proc sql;
create table e_main as
select *,
/* take ni and subtract the mean net income (calculated for each sic2-fyear) */
ni – mean(ni) as ni_rel
from d_15obs
group by sic2, fyear;
quit;
/* 5,245 unique gvkeys vs 4,252 mentioned in the paper */
proc sql; create table gvkey as select distinct gvkey from e_main; quit;
/* 38 unique industries vs 36 mentioned in the paper */
proc sql; create table sic2 as select distinct sic2 from e_main; quit;
/* 474 unique industry-years vs 416 mentioned in the paper (table 2 footnote) */
proc sql; create table sic2_yrs as select distinct sic2, fyear from e_main; quit;
/* some descriptives, these are similar as those in paper
Median market cap: 154.6 vs 137.3 inpaper
Median assets: 150.8 vs 137.3 in paper
Median sales: 166.4 vs 221.0 in paper
Median CFO: 0.088 vs 0.082 in paper
Median Production costs: 0.750 vs 0.788 in paper
*/
proc means data=e_main n mean median; var mcap at sale oancf ib ni cfo_at disexp_at prod_at accruals_at; run;
/* Table 2
——-
Four models: cfo, disexp, prod, accruals
Regressions by industry-year: table presents mean and t-values (mean/standard deviation)
Some of the coefficients are similar, but overall not significant (i.e., higher standard errors)
I may have overlooked something.*/
/* First regression: cfo */
proc sort data=e_main; by fyear sic2;run;
proc reg data=e_main noprint edf outest=e_parms_cfo;
model cfo_at = one_at sale_at sale_ch_at ;
output out=f_fitted1 p=yhat r=yresid ;
by fyear sic2;
run;
/* sale_at: 0.0476839 vs 0.0516 in paper
sale_ch_at: -0.0161828 vs 0.0173 in paper
*/
proc means data=e_parms_cfo n mean stderr;
var Intercept one_at sale_at sale_ch_at RMSE;
run;
/* Second regression: disexp */
proc reg data=e_main noprint edf outest=e_parms_disexp;
model disexp_at = one_at sale_lag_at ;
output out=f_fitted2 p=yhat r=yresid ;
by fyear sic2;
run;
/* sale_lag_at is 0.1506, vs 0.1596 in paper
by the way: not clear if there is a typo in the paper, St listed twice, assuming second one should be St-1
*/
proc means data=e_parms_disexp n mean stderr;
var Intercept one_at sale_lag_at RMSE;
run;
/* Third regression: prod */
proc reg data=e_main noprint edf outest=e_parms_prod;
model prod_at = one_at sale_at sale_ch_at sale_ch2_at ;
output out=f_fitted3 p=yhat r=yresid ;
by fyear sic2;
run;
/* sale_at: 0.776 vs 0.7874 in paper
sale_ch_at: 0.0165265 vs 0.0404 in paper
sale_ch2_at: -0.0168325 vs -0.0147 in paper*/
proc means data=e_parms_prod n mean stderr;
var Intercept one_at sale_at sale_ch_at sale_ch2_at RMSE;
run;
/* Fourth regression: accruals */
proc reg data=e_main noprint edf outest=e_parms_accr;
model accruals_at = one_at sale_ch_at ppe_at ;
output out=f_fitted4 p=yhat r=yresid ;
by fyear sic2;
run;
/* sale_ch_at: 0.0537103 vs 0.0490 in paper
ppe_at: -0.0492712 vs -0.060 in paper */
proc means data=e_parms_accr n mean stderr;
var Intercept one_at sale_ch_at ppe_at RMSE;
run;
/* Table 4
——-Regressions of abnormal cfo, abnormal discretionary exp and abnormal prod costs
on size, mtb, net income, and suspect_ni
The paper presents Fama McBeth regressions (yearly regressions); below I have pooled regressions
*/
/*
Abnormal CFO:
Intercept -0.04694 0.00213 -22.05 <.0001
size 0.00871 0.00039472 22.06 <.0001
mtb 0.00072744 0.00021366 3.40 0.0007
ni 0.00020496 0.00004853 4.22 <.0001
SUSPECT_NI -0.01856 0.00540 -3.44 0.0006 <<< in line
*/
/* yresid in f_fitted1 has the abnormal cash flow from operations */
proc reg data=f_fitted1 ;
model yresid = size mtb ni suspect_NI;
run;
/*
Abnormal discretionary expenses
Intercept -0.08060 0.00447 -18.03 <.0001
size 0.00645 0.00082902 7.78 <.0001
mtb 0.01609 0.00044874 35.87 <.0001
ni -0.00060751 0.00010193 -5.96 <.0001
SUSPECT_NI -0.03273 0.01135 -2.88 0.0039 <<< in line
*/
/* yresid in f_fitted2 has the abnormal discretionary expenses */
proc reg data=f_fitted2 ;
model yresid = size mtb ni suspect_NI;
run;
/*
Abnormal production costs
Intercept 0.06387 0.00347 18.39 <.0001
size -0.00720 0.00064402 -11.18 <.0001
mtb -0.00918 0.00034860 -26.32 <.0001
ni -0.00006812 0.00007918 -0.86 0.3896
SUSPECT_NI 0.03331 0.00881 3.78 0.0002 <<< in line
*/
/* yresid in f_fitted3 has the abnormal production costs */
proc reg data=f_fitted3 ;
model yresid = size mtb ni suspect_NI;
run;
/* Yearly regressions (Fama McBeth)
——————————-
*/
/* CFO */
proc sort data=f_fitted1; by fyear;
proc reg data=f_fitted1 noprint edf outest=g_parms_cfo;
model yresid = size mtb ni suspect_NI;
by fyear;
run;
/* suspect_NI: -0.000045406 vs -0.020 in paper
stderr is 0.0054624, so not significant (paper: t-value of 3.0)*/
proc means data=g_parms_cfo n mean stderr;
var Intercept size mtb ni suspect_NI RMSE;
run;
/* Discretionary expenses */
proc sort data=f_fitted2; by fyear;
proc reg data=f_fitted2 noprint edf outest=g_parms_de;
model yresid = size mtb ni suspect_NI;
by fyear;
run;
/* suspect_NI: -0.0702146 vs -0.0591 in paper */
proc means data=g_parms_de n mean stderr;
var Intercept size mtb ni suspect_NI RMSE;
run;
/* Production costs */
proc sort data=f_fitted3; by fyear;
proc reg data=f_fitted3 noprint edf outest=g_parms_prod;
model yresid = size mtb ni suspect_NI;
by fyear;
run;
/* suspect_NI: 0.0401730 vs 0.0497 in paper )*/
proc means data=g_parms_prod n mean stderr;
var Intercept size mtb ni suspect_NI RMSE;
run;
/* Download the dataset to the local machine */
proc download data=e_main out=local.e_main;
run;
endrsubmit;
signoff;
Reference