How Quickly We Forget: The Duration of Persuasion Effects From Mass Communication

Abstract

Scholars do not usually test for the duration of the effects of mass communication, but when they do, they typically find rapid decay. Persuasive impact may end almost as soon as communication ends. Why so much decay? Does mass communication produce any long-term effects? How should this decay color our understanding of the effects of mass communication? We examine these questions with data from the effects of advertising in the 2000 presidential election and 2006 subnational elections, but argue that our model and results are broadly applicable within the field of political communication. We find that the bulk of the persuasive impact of advertising decays quickly, but that some effect in the presidential campaign endures for at least 6 weeks. These results, which are similar in rolling cross-section survey data and county-level data on actual presidential vote, appear to reflect a mix of memory-based processing (whose effects last only as long as short-term memory lasts) and online processing (whose effects are more durable). Finally, we find that immediate effects of advertising are larger in subnational than presidential elections, but decay more quickly and more completely. [Supplementary material is available for this article. Go to the publisher's online edition of Political Communication for the following free supplemental resource(s): discussion of methodological issues; results for a alternative specifications of key models; full reports of model results.]

Keywords:

• elections
• advertising
• mass communication
• decay

Notes

1. For reviews, see Iyengar and Simon (2000) and Kinder (2003).

2. In a review of evidence on this point, Petty, Haugtvedt, and Smith argue that strong attitudes are more likely to persist than weak attitudes (1995, pp. 100–108). But the number of studies reviewed is small and, as the authors note, some have validity problems. Even in psychology, duration of opinion change appears to be a peripheral concern.

3. Lodge, Steenbergen, and Brau find a small effect from recall for one of the two candidates, but the effect of the full set of positions was dominant (see Table 2). The study included an experimental manipulation, but it did not figure in these results, which we take to be their key results.

4. Mitchell (2012)reports that her subjects did maintain an online tally summarizing reaction to previous stories; they didn't, however, use it in making weekly evaluations, relying instead on information from the current week's story.

5. In studies that do not distinguish online from memory-based processing, the typical pattern is rapid decay of persuasion effects (Gaines et al., 2007). A study following this pattern but not cited is de Vreese (2004,p. 203). However, an exception is an experiment in News That Matters that checks for and finds duration over a 1-week period (Iyengar & Kinder, 1987, pp. 25–26, 44).

6. Late deciders in elections are likely to care less, know less, and be less educated, all of which would incline them toward less effortful processing (Campbell, Converse, Miller, & Stokes, 1960).

7. This is our eyeball estimate from Figure 4a. Althaus et al. report effect sizes at lags of 14, 28, 42, and 56 days, plus an estimate for all days, but do not report an initial impact.

8. We rescaled the data to make ad volumes comparable to those reported by Shaw (2006); that is, we divided by 10.

9. The quality of Internet samples has been challenged and defended (Malhotra & Krosnick, 2007; but also see Hill, Lo, Vavreck, & Zaller, 2007; Sanders, Clarke, Stewart, & Whiteley, 2007).

10. In this case, “reach” is the percentage of the target population that sees an ad at least once.

11. If we code advertising exposure strictly by calendar day, results degrade somewhat, but the main patterns remain.

12. Controls are gender, race, age, income, education, political information, party identification, a PID and political information interaction, Republican vote share in the respondent's county from 1996, Perot vote share in respondent's county in 1996, church attendance, fixed effects for interview week, media market, and state separately, and indicators for “don't know” responses on the income and church attendance variables.

13. Results are unchanged if instead we use fixed effects for the last 4, 7, or 10 days.

14. Full model results are available.

15. The basic pattern of ad effects is similar within each type of race.

16. We are skeptical that the long-term effect of ads in these races is negative or zero, as the data weakly suggest. Although we are using ad data from up to 42 days prior to each interview, few races were competitive enough to deploy long-term advertising. This means that, in contrast with presidential elections, we have little leverage for identifying long-term effects and their decay; the previous coefficients should be viewed in this light.

17. They actually identify five top models, but two are close cousins, so we dropped the weaker of this pair.

18. Rubin and Wenzel refer to this function as the exponential power, but it is better known to political scientists as a variant of the Weibull function.

19. Estimation of the hyperbolic power model would not converge in either data set.

20. No likelihood ratio test is necessary for this comparison because all decay models have the same number of parameters.

21. The logarithmic function finds that ad effects survive about 10 days, but it fits the data poorly, pushing survival rates into negative territory after about 12 days.

22. A statistical test involving this difference is reported below.

23. From the derivative of the best fitting Weibull function, one can see that if the impact and decay coefficients are positive (as they are), f’(x) is always negative and f’’(x) is always positive. This means that the functions are strictly decreasing and convex, which is to say that the rate of decay is always increasing, but at an increasingly slow rate.

24. The simulation calculates the marginal effect for a respondent whose baseline probability of voting for the first candidate is .5.

25. A significance test of differences in the coefficients in Tables 4 and 5 does not strictly bear on the difference in effect sizes because of differences in the cut-points of the models.

26. The non-parametric bootstrap estimates of the standard errors are 1.14 and 1.30.

27. We note that our estimate of short-term ad impact in gubernatorial, Senate, and House elections is close to the estimate from the state gubernatorial primary contest of Gerber et al. They tested the impact of ad buys of 1,000 GRPs per week or, in our units, 1.4 ad viewings per day on average (10 ads a week divided by 7 days). Their estimated impact was about 5 percentage points, which they took to be constant over the week-long period in which they surveyed voters. We calculate the comparable effect in our modeling framework as the effect of 7 days of exposure to 1.4 ad viewings per day from one candidate and none from the other, as measured on the seventh day. This effect, as calculated from the best fitting model in Table 5, is 8.07 points (with a standard error of 2.41). About half of this 8-point effect would carry over to the eighth day.

28. We use the replication data set from Wand et al. (2001).

29. Coefficients themselves cannot be directly compared because the forms of the response models are different, OLS and probit.

30. We obtain the following next-day estimate for a county in which one candidate runs two ad viewings and the other runs one: a .34-point impact (with a standard error of .09). The individual-level estimate (from the Weibull model) gave an impact of .22 points (with a standard error of .07). The county data suggest a bigger next-day effect than the survey data; however, given the uncertainty of the estimates, they are best regarded as comparable.

31. Johnston et al. report that Bush's ad advantage in the last week of the campaign netted 4 points. Their analysis, however, differs from ours in important ways. Inter alia, they do not appear to have added either previous-week or next-day decay.

32. The effects of advertising in the 2012 presidential election appear to have been similar to those in 2000. John Sides and Lynn Vavreck (2013) demonstrate sizable impacts of same-day ads and a rapid rate of decay, but their model uses a different functional form and different controls. Sides and Vavreck have made their data available to us for a robustness test of our models. The survey data are from the 2012 Cooperative Campaign Analysis Project, which is broadly similar in design to the 2006 CCES described above. The advertising data are from Nielsen under license to Lynn Vavreck and John Geer. With fairly minor exceptions, these data have the same form as our data from the 2000 election and permitted tests that are very similar to the tests reported in Table 5. We found that the initial impact of ads in the 2012 election was greater than in the 2000 election, but decay was also greater. The half-life of an ad's impact was about 3 days (compared to 4 in 2000). In both initial impact and decay, results fall between those obtained for the 2006 subnational elections and the 2000 presidential election. Coefficients, however, are imprecisely estimated because the 2012 sample was only 60% as large as the 2000 sample. In view of this lack of precision, we do not offer an interpretation for such differences as appear to exist from 2000.