Anda di halaman 1dari 2

Smith, A. H., Handley, M. A., & Wood, R. ple.

For some of the effects reported in our


(1990). Epidemiological evidence indicates Figure 2 (Bushman & Anderson, 2001, Table 1
asbestos causes laryngeal cancer. Journal of p. 481), we converted odds ratios to corre- Percentages for Amount of Smoking
Occupational Medicine, 32, 499 –507. lations. In their critique of our article, Among 605 Male Patients With
Weller, S. C. (1993). A meta-analysis of condom
effectiveness in reducing sexually transmitted
Block and Crain (2007, this issue) claimed, Lung Cancer and 780 Male
HIV. Social Science and Medicine, 36, 1635– “There is no data transformation that con- Patients Without Lung Cancer
1644. verts an odds ratio or relative risk into a
Cancer No cancer
Wells, A. J. (1998). Lung cancer from passive correlation” (p. 252). They further stated
Amount of
smoking at work. American Journal of Public that our transformation method is “wrong” smoking % n % n
Health, 88, 1025–1029. and “faulty” (p. xxx). They are incorrect. In
Wynder, E. L., & Graham, E. A. (1950). To- fact, there are many different methods for None 1.3 8 14.6 114
bacco smoking as a possible etiological factor transforming an odds ratio or relative risk Light 2.3 14 11.5 90
in brochiogenic carcinoma. Journal of the into a correlation. We used the following Moderate 10.1 61 19.0 148
American Medical Association, 143, 329 –
transformation proposed by Digby (1983): Heavy 35.2 213 35.6 278
336. Excessive 30.9 187 11.5 90
(OR3/4 ⫺ 1) / (OR3/4 ⫹ 1), where OR is the
odds ratio. Bonett (2007, this issue) pro- Chain 20.3 123 7.6 59
Correspondence concerning this comment vides a brief review of methods that trans- Note. Data are from Figure 3 in Wynder and Gra-
should be addressed to Jerald J. Block, 1314 NW form a relative risk or odds ratio into a ham (1950). The sum of counts of men with cancer
Irving Street #508, Portland, OR 97209. E-mail: correlation. equals 606 rather than 605 owing to rounding er-
jblock@aracnet.com ror. The sum of counts of men without cancer equals
Block and Crain (2007) also stated 779 rather than 780 owing to rounding error.
that at least 6 of the 9 comparison correla- None ⫽ less than 1 cigarette per day for more than
tions shown in Figure 2 of our 2001 article 20 years; light ⫽ 1–9 cigarettes per day for more
DOI:10.1037/0003-066X.62.3.253
were calculated incorrectly. Of course, dif- than 20 years; moderate ⫽ 10 –15 cigarettes per
day for more than 20 years; heavy ⫽ 16 –20 ciga-
ferent transformations will likely yield dif- rettes per day for more than 20 years; excessive ⫽
Measuring the Strength of the
ferent estimates, but in the present case the 21–34 cigarettes per day for more than 20 years;
Effect of Violent Media on different estimates generally do not vary chain ⫽ 35 or more cigarettes per day for more than
20 years.
Aggression that much. Block and Crain cited one ex-
ample where their estimate varies greatly
Brad J. Bushman from our estimate. To create the bar in our
University of Michigan Figure 2 that gives the correlation between
and Vrije Universiteit cigarette smoking and lung cancer, we used “Wynder and Graham’s (1950) Figure 1 (p.
data from Figure 1 (based on 100 male 332) is a subset of their Figure 3 (p. 333).
Craig A. Anderson patients with lung cancer and 186 male It is wrong to pool the data from these two
Iowa State University patients with other chest diseases; total figures; cases are counted twice.” This
N ⫽ 286) and Figure 3 (based on 605 statement is simply incorrect. In the para-
In our American Psychologist article patients with lung cancer and 780 male graph titled “Comparison of Independent
(June–July 2001), we presented a figure patients in the general hospital population Studies” (p. 332, emphasis added), Wynder
that compared the effect of violent media without lung cancer; total N ⫽ 1,385) in and Graham (1950) stated, “Before the
on aggression with other well-known ef- Wynder and Graham’s (1950) classic arti- smoking habits of the 605 patients with
fects (Bushman & Anderson, 2001, Figure cle. The phi coefficients computed from cancer of the lungs are compared with
2, p. 480). Because the strength of the their Figures 1 and 3 are .35 and .41, re- those of the general hospital population, it
relationship between violent media and ag- spectively. In our 2001 article, we com- might be well to compare the results of two
gression is unclear to many people, it is puted a weighted average of these two co- control studies” (p. 332). The data in
quite useful to compare it with other rela- efficients (i.e., [.35 ⫻ 286 ⫹ .41 ⫻ 1,385] Wynder and Graham’s Figure 1 are from
tionships with which people are more fa- / [286 ⫹ 1,385] ⫽ .40). The same estimates Control Study I.
we reported have also been reported by Second, Block and Crain (2007, p.
miliar (e.g., the relationship between ciga-
252) stated that our estimate is inaccurate:
rette smoking and lung cancer). For other researchers (e.g., Eron, 1996; Hues-
example, Youth Violence: A Report of the mann in Violence on Television, 1993). We repeated the calculation, using just Figure 3
Surgeon General contains a table of corre- Table 1 shows the calculations based from Wynder and Graham (1950). In doing so,
lations showing that TV violence is a larger on Figure 3 of Wynder and Graham’s we defined one variable as the amount of smok-
ing the patients engaged in, as coded in Table 2
risk factor for violence among children (1950) article. The frequency procedure in (p. 331) of Wynder and Graham (1950). The
6 –11 years old than other well-known vi- SAS yields the following correlations: phi second variable was a diagnosis of lung cancer
olence risk factors such as low IQ, being coefficient ⫽ .41, Spearman correlation ⫽ versus other diagnoses. The resulting correlation
from a broken home, having abusive par- .40, Pearson product–moment correlation was .90, not .40 as reported by Bushman and
ents, and having antisocial peers (U.S. De- coefficient ⫽ .40. The phi coefficient as- Anderson (2001, Figure 2, p. 481).
partment of Health and Human Services, sumes dichotomous data, the Spearman We could not replicate the .90 correlation
2001, Table 4-1). correlation assumes ordinal data, and the that Block and Crain (2007) computed, but
It is necessary, of course, to convert Pearson correlation assumes interval data. we suspect it is inaccurate. The correlations
different effect-size indices to a common Block and Crain (2007) criticized our we computed for Figure 3 ranged from .40
index so that people can more easily com- approach to computing the correlation be- to .41. A .90 correlation suggests that 81%
pare effects. We chose the correlation co- tween cigarette smoking and lung cancer (.902) of the variance in lung cancer is
efficient because it is familiar to many peo- on two grounds. First, they stated (p. 252), attributable to smoking cigarettes. This

April 2007 ● American Psychologist 253


leaves little room for other factors to influ- U.S. Department of Health and Human Services.
ence lung cancer (e.g., diet, exercise, ge- (2001). Youth violence: A report of the Sur- Table 1
netic predispositions). geon General. Rockville, MD: U.S. Public 2 ⫻ 2 Contingency Table
Regarding our own meta-analysis, Health Service, Office of the Surgeon Gen-
Block and Crain (2007) stated that we “did eral. Retrieved June 9, 2006, from http:// y1 y2
not provide references for the studies in- www.surgeongeneral.gov/library/youthviolence/ x1 p11 p12 p1⫹
images/Table_4-1.gif
cluded in their meta-analysis; therefore it is x2 p21 p22 p2⫹
Violence on television: Hearings before the Sub-
impossible to replicate their study or deter- p⫹1 p⫹2
committee on Telecommunications and Fi-
mine if they again used the faulty transfor- nance of the House Committee on Energy and
mation to convert odds ratios and relative Commerce, 103rd Cong., 1st session, 44
risk to correlation coefficients” (p. 252). (1993) (testimony of L. Rowell Huesmann).
We did not include references for the stud- Wynder, E. L., & Graham, E. A. (1950). To-
ies in our meta-analysis because the editor bacco smoking as a possible etiological factor
of American Psychologist thought it would questions are of a sensitive nature (e.g.,
in brochiogenic carcinoma. Journal of the
be better to have interested individuals con- American Medical Association, 143, 329 –
income, alcohol consumption, body weight),
tact us directly for these references rather 336. the response rate is often higher if the re-
than to use valuable journal space listing spondent is simply asked to check one of
hundreds of references. These references two broad categories (e.g., less than
are readily available from Brad J. Bush- Correspondence concerning this comment $40,000 per year, $40,000 or more per
man. should be addressed to Brad J. Bushman, Insti- year) rather than a specific quantitative
We do agree with Block and Crain’s tute for Social Research, University of Michi- value. In other applications, genetic or psy-
(2007) conclusion that violent media ef- gan, 426 Thompson Street, Ann Arbor, MI chometric theory predicts the existence of a
fects constitute an important and controver-
48106. E-mail: bbushman@umich.edu latent quantitative variable that is observ-
sial topic and that the results from scientific able only on a dichotomous scale as a result
studies on media-related aggression need to of the latent variable exceeding, or not ex-
DOI:10.1037/0003-066X.62.3.254 ceeding, some unknown threshold value.
be accurate and replicable. Our 2001 article
Quantitative variables that are measured on
relied heavily on meta-analytic procedures Transforming Odds Ratios dichotomous scales are referred to as arti-
to integrate the literature on media-related
aggression. Meta-analytic procedures are
Into Correlations for Meta- ficially dichotomous.
Analytic Research When X and Y are naturally or artifi-
more objective, accurate, and replicable
cially dichotomous, data from a sample of
than are traditional narrative procedures
n respondents may be summarized in a 2 ⫻
(e.g., Bushman & Wells, 2001; Cooper & Douglas G. Bonett 2 contingency table as shown in Table 1,
Rosenthal, 1980). Although violence in the Iowa State University where pij are the cell proportions, pi⫹ is a
media is not the only factor that increases
marginal row proportion, and p⫹j is a mar-
aggression, or even the most important fac- Block and Crain (2007, this issue) stated, ginal column proportion. The association
tor, it is not a trivial factor. “There is no data transformation that con- in a 2 ⫻ 2 contingency table is often re-
verts an odds ratio or relative risk into a ported in terms of an odds ratio,
REFERENCES correlation. One needs more data” (p. 252).
The purpose of this comment is to explain OR ⫽ 共p 11 p22兲/共p12 p21兲,
Block, J. J., & Crain, B. R. (2007). Omissions how an odds ratio or relative risk can be
and errors in “Media Violence and the Amer- or a relative risk,
ican Public.” American Psychologist, 62, transformed to approximate a product–mo-
252–253. ment correlation. Such transformations RR ⫽ 共p 11/p 1⫹)/(p 21/p 2⫹),
Bonett, D. G. (2007). Transforming odds ratios have important applications in meta-ana-
into correlations for meta-analytic research. lytic research. where X is the predictor variable and y1 is
American Psychologist, 62, 254 –255. the response category of interest. A relative
Meta-analysis often involves the com-
Bushman, B. J., & Anderson, C. A. (2001). risk may be transformed into an odds ratio
bination of product–moment correlations
Media violence and the American public: Sci- using the following equality:
entific facts versus media misinformation.
obtained from multiple published studies.
American Psychologist, 56, 477– 489. A product–moment correlation between OR ⫽ RR兵共1 ⫺ p21/p2⫹兲/共1 ⫺ p11/p1⫹兲其.
Bushman, B. J., & Wells, G. L. (2001). Narrative two quantitative variables (X and Y) is
impressions of literature: The availability bias commonly referred to as a Pearson corre- In applications where the response cate-
and the corrective properties of meta-analytic lation. If one variable is naturally dichoto- gory (y1) is rare, note that p11 and p21 may
approaches. Personality and Social Psychol- be very small so that p1⫹ ⬇ p12, p2⫹⬇ p22,
mous (male/female, Treatment A/Treat-
ogy Bulletin, 27, 1123–1130. and thus RR ⬇ OR.
ment B, etc.) while the other variable is
Cooper, H. M., & Rosenthal, R. (1980). Statis- The problem of estimating the Pear-
tical versus traditional procedures for summa- quantitative, the product–moment correla- son correlation between two quantitative
rizing research findings. Psychological Bulle- tion between X and Y is called a point- variables using information from a 2 ⫻ 2
tin, 87, 442– 449. biserial correlation. If both X and Y are contingency table is one of the oldest prob-
Digby, P. G. N. (1983). Approximating the tet- naturally dichotomous, the product–mo- lems in statistics (Pearson, 1900) and in-
rachoric correlation coefficient. Biometrics, ment correlation between X and Y is called volves the computation of a tetrachoric
39, 753–757. a phi coefficient.
Eron, L. D. (1996, February 28). TV violence correlation. The computation of the exact
and its effect on kids. Invited address pre-
In contrast to naturally dichotomous tetrachoric correlation is complicated but
sented at the 49th Annual Convention of the variables, quantitative variables are some- may be obtained in the current version of
Ontario Psychological Association, Toronto, times measured on dichotomous scales. For SAS. If a study reports the odds ratio but
Ontario, Canada. instance, in survey research, where certain does not provide enough additional infor-

254 April 2007 ● American Psychologist