is spent by the other armed services together. These advertis- ... How effective is advertising in producing enlistments? Which ... the youth's qualifications for service. 2. Take a ..... months; Capps, Seo, and Nichols (1997) employ a polynomial.
Summer 2006 111
EFFECTIVENESS OF ADVERTISING IN DIFFERENT MEDIA The Case of U.S. Army Recruiting James N. Dertouzos and Steven Garber ABSTRACT: The U.S. military spends roughly $100 million annually to place advertisements to promote enlistments. Army data for 1981 to 1984 exhibit substantial cross-sectional and times series variation in advertising intensities, which enables estimation of advertising effects separately for four media (television, radio, magazines, and newspapers) using a functional form allowing media-specific, S-shaped response functions and relatively flexible, media-specific lag structures. We find that army advertising was very productive in producing enlistments; response functions for television, radio, and magazines are consistent with widespread advertising practice; lag patterns differ substantially over media; and the army’s allocation of spending across media was nearly optimal.
The U.S. Army spends about $100 million per year on advertising to promote its recruiting efforts, and a similar amount is spent by the other armed services together. These advertising activities raise several questions, including the following: How effective is advertising in producing enlistments? Which media are most productive over what ranges of expenditures? What are the dynamic or delayed effects of advertising? How well are military advertising budgets allocated across media? In this paper, we analyze such questions using data for the U.S. Army during the early 1980s. These data, which are relatively disaggregated over time, geography, and media, allow us to implement an econometric specification of the advertising–sales relationship that is unusually rich in terms of functional form, dynamics, and distinguishing effects of advertising through different media. Our results indicate that army advertising was quite effective during the period studied; different media are most effective over ranges of spending that accord well with advertising practice; patterns of dynamic effects differ substantially across media; and the allocation of spending across media was not far from optimal, but did leave room for improvements substantial enough to be of interest to policymakers. The next section presents background information on the recruitment and enlistment processes. We then describe the data and present our econometric specifications and estimation methods. Related literature is then discussed. We then present parameter estimates, depict media-specific advertising– sales relationships, consider the potential for improvement by James N. Dertouzos (Ph.D., Stanford University) is a senior economist at the RAND Corporation, Santa Monica, CA. Steven Garber (Ph.D., University of Wisconsin) is a senior economist at the RAND Corporation, Santa Monica, CA.
reallocating spending across media and increasing total spending, and consider implications of sampling uncertainty. The concluding section summarizes our main findings. THE ENLISTMENT (SALES) PROCESS In the context of army recruiting, a sale is an enlistment or a contractual commitment by a youth to enter the army. Our econometric analysis focuses on effects of advertising on potential numbers of contracts signed. The key salesperson is a recruiter. Modeling how advertising may affect sales requires consideration of steps involved in the enlistment process. The description here, which relies heavily on Polich, Dertouzos, and Press (1986, pp. 9–10), pertains to the early 1980s, the time period during which our data were generated. During the recruiting and enlistment process, a youth will go through the following steps, assuming he or she is qualified and sufficiently motivated to do so: 1. Meet with a recruiter to obtain general information about army opportunities; the recruiter begins to assess the youth’s qualifications for service. 2. Take a series of tests, including the Armed Forces Qualification Test (AFQT) and tests gauging the applicant’s aptitudes for various military occupations. 3. Travel to a Military Entrance Processing Station (MEPS) for a physical examination and to meet with a job counselor. 4. Sign an enlistment contract that specifies an army occupation and a service entry date. Based on monthly “missions,” which are the army’s version of sales quotas, army recruiters have strong incentives to locate, persuade, and sign potential enlistees. During the peJournal of Advertising, vol. 35, no. 2 (Summer 2006), pp. 111–122. © 2006 American Academy of Advertising. All rights reserved. ISSN 0091-3367 / 2006 $9.50 + 0.00.
112 The Journal of Advertising
riod covered by our data, typical monthly quotas for individual recruiters were one “high-quality” and one “low-quality” contract per month. A high-quality recruit is defined as a high-school graduate whose AFQT score is above the median for the U.S. youth population. Recruiting high-quality youths is typically much more difficult than recruiting low-quality youths, and the army places a high priority on increasing numbers of high-quality enlistees. Accordingly, our econometric analysis focuses on how advertising affects the opportunities of the army to enlist high-quality youth. THE DATA The unit of observation for our econometric analysis is a MEPS area during one month. Data are available for the U.S. Army for the 36 months from July 1981 through June 1984, during which time there were 66 MEPS areas in the United States. The outcome (sales) measure is the number of high-quality contracts signed. Data were also constructed at the MEPSarea level for high- and low-quality missions and numbers of active recruiters. To control for alternative employment opportunities of potential enlistees, we also constructed measures at the MEPS-area level of the average hourly wage rate in manufacturing (from the Bureau of Labor Statistics [BLS], Employment and Earnings reports) and the state unemployment rate (from the Current Population Survey, conducted by the U.S. Bureau of the Census for the BLS). We also include indicator variables for two sets of MEPS areas, each including about 15% of the national youth population, in which enlistmentbonus experiments were in effect during the time period studied (Polich, Dertouzos, and Press 1986). Our advertising measures were constructed using extensive information provided by N. W. Ayer, the army’s advertising agency during the time period covered by the data (Dertouzos [1985] provides a more extensive discussion). During that time, about 85% of army spending on television, radio, magazine, and newspaper advertisements was purchased at the national level. For advertising purchased nationally, we obtained media-specific data on monthly, national spending to place ads that ran during that month. We also obtained media-specific data on monthly impressions in each of 210 Areas of Dominant Influence (ADI), which are sets of counties. The impressions data were derived from geographically detailed information on ratings of individual television and radio programs on which the army advertised and readership of individual magazines in which the army advertised. We then allocated total dollars spent monthly on each medium nationally to ADIs in proportion to impressions. These media-specific spending levels for ADIs were then allocated to the 66 MEPS areas using 1980 county-level Census population figures. The significant cross-section variation in both the level and
mix of advertising spending greatly limits any potential biases associated with the endogeneity of advertising resources. To illustrate this, we regressed local allocations advertising dollars per 1,000 male youths on monthly and local MEPS dummies (101 in all). Controlling for the overall budget level and systematic area differences in this way accounted for only 28% of the variance. In other words, national media choices and unsystematic patterns in the viewing behavior of local audiences created significant cross-section variation that was unrelated to any policy choices made at the national level. The fraction of national-level spending on a particular advertising medium allocated to a particular MEPS varies considerably across MEPS and over months. This variation is due to geographic variation in the audience penetration levels of the programs and magazines in which the army advertised and changes over time in such programs and magazines. Monthly data on local media purchases from local radio stations and daily and weekly newspapers were obtained at the recruiting battalion level. (Army recruiting battalions, of which there were 54 during our study period, are recruiting territories.) Monthly spending levels by counties were computed and converted to MEPS areas by aggregating over counties. Data on daily and weekly newspapers were combined, as were data on national and local radio. This provided us with monthly advertising data at the MEPS-area level for four media—television, radio, magazines, and newspapers—measured in dollar terms. In the econometric model, these spending data are expressed in per capita terms (specifically, per 1,000 young males residing in the MEPS area) because the extent to which the total impressions generated by any level of spending repeatedly reach the same individuals (frequency) will be lower the larger the audience is. ECONOMETRIC SPECIFICATIONS Following Dertouzos (1985) and Polich, Dertouzos, and Press (1986), a fundamental premise of our analysis is that if advertising improves recruiting opportunities, recruiters facing fixed mission levels may respond by reducing their effort levels. Thus, to assess the effectiveness of advertising, it is necessary to hold recruiter effort constant. Accordingly, as in Dertouzos and Polich (1989), we aim to estimate how advertising affects potential high-quality enlistments, which are defined as the number of such enlistments that would occur holding recruiter effort constant. Our analysis, however, extends that of Dertouzos and Polich (1989), which used simpler (constant elasticity or log-linear) functional forms and dynamic specifications (infinite, geometrically declining lags) to estimate effects of advertising. We estimate a reduced-form equation that enables assessment of effects of advertising on high-quality enlistments,
Summer 2006 113
holding recruiter effort constant, while allowing the number of low-quality enlistments to adjust to changes in recruiting opportunities.
ln H t + ψ ln Lt = Et* + β0 + ∑ βi Xit + ∑ κ j f j (a jt ) i
j
∞
+∑ κ j1 f j (a jt−1 ) + ∑ θ j ∑ ρs f j (at−s ) + εt , j
j
(1)
s=2
Recruiting Possibility Frontier
where
The conceptual foundation for our analysis is the “recruiting possibility frontier” (RPF) facing recruiters in a MEPS area in a particular month. An RPF depicts the trade-offs between potential numbers of high-quality (Ht*) and low-quality recruits (Lt*), holding constant recruiter effort, which is endogenous, and all exogenous determinants of recruiting opportunities. In a graphical representation, the height of the RPF for a given value of Lt (measured on the horizontal axis) is the maximum number of high-quality prospects who can be enlisted, given recruiter effort levels and the exogenous determinants of recruiting opportunities, if the given number of low-quality prospects are enlisted. Three sets of factors exogenous to recruiters determine MEPS-area recruiting opportunities, and hence the height of the RPF: (1) local economic conditions, which affect the willingness of youths to enlist; (2) resources allocated to recruiting, such as numbers of recruiters and enlistment incentives; and (3) recruitment advertising messages received by youths in the area.
Xit = observable, nonadvertising variables that affect recruiting possibilities, with X1, X2, and X3 equaling the logarithms of numbers of recruiters, civilian wage rate, and local unemployment rate, respectively, and the variables X4 and X5 taking the value of 1 if the corresponding bonus test was in operation in the MEPS area in month t;
Observed Numbers of Contracts Empirical relationships between advertising and actual contracts may give a misleading picture of advertising effectiveness if, for example, advertising enhances recruiting opportunities but recruiters are not sufficiently motivated to capitalize on enhanced opportunities by increasing enlistments. Our interest centers on how advertising affects recruiting opportunities because it is such effects that are most relevant to understanding advertising’s potential and how policy can best exploit that potential. Suppose, for example, that advertising substantially enhances recruiting opportunities, but that much of the potential enlistment gains go unexploited because recruiter effort declines as it becomes easier to achieve assigned missions. Then a direct empirical relationship between advertising and enlistments will suggest that advertising is largely ineffective when, in fact, advertising is effective in increasing opportunities, and from a policy point of view, increases in advertising might best be accompanied by actions to maintain recruiter effort, such as higher missions, or by increases in numbers of recruiters. Let the numbers of contracts actually signed in a MEPS area in month t, which are observed in the data, be denoted by (Lt, Ht). (For economy of notation, we suppress the subscript indicating the MEPS area.) The trade-off between actual levels of high- and low-quality contracts is specified from the RPF by introducing an unobserved level of recruiter effort, Et*:
αjt = spending per thousand young males in the MEPS area on advertising that ran in month t through medium j = 1 (television), 2 (radio), 3 (magazines), and 4 (newspapers); (ψ, βι, κj, κj1, θj, ρ)= parameters to be estimated; fj(ajt) = functions (introduced below) of advertising in medium j; and εt = a disturbance term assumed to include fixed effects for all combinations of MEPS and calendar months. The fixed effects are included to capture various factors affecting MEPS-level recruiting opportunities that are constant over the sample period, such as the size of the military-eligible youth population and youth attitudes toward military service and MEPS-specific seasonal patterns in recruiting opportunities. Table 1 reports descriptive statistics for selected variables. In equation (1), Et is the MEPS-area recruiters’ unobservable effort level, which is unobserved and has no natural scale. As can be seen from equation (1), a unit increase in effort is assumed to increase potential high-quality enlistments (holding low-quality enlistments constant) by 1%: this restriction fixes the scale on which effort is implicitly measured. Thus, the RPF summarizes the trade-offs between ln Ht and ln Lt as determined by equation (1), with all exogenous determinants of recruiting opportunities given. The parameter ψ corresponds to minus the slope of the RPF in (ln Lt, l ln Ht ) space, and the other parameters of equation (1) can be interpreted as effects of observable righthand-side variables on the trade-off between actual high- and low-quality contracts, holding effort levels constant. Specification for Advertising Effects Advertising effects depend on parameters explicit in equation (1) and parameters of the functions fj (ajt). The parameters κ j are coefficients of these functions evaluated at current-month (per capita) spending levels, and thus relate to effects of advertising in a particular month on enlistments during the same month. The parameters κj1 pertain to effects
114 The Journal of Advertising TABLE 1 Descriptive Statistics for Army Data, July 1981 Through June 1984 Variable
Mean
SD
Number of recruiters Civilian wage rate (hourly) Unemployment rate (percent) Television spending, per thousand young males Radio spending, per thousand young males Magazine spending, per thousand young males Newspaper spending, per thousand young males High-quality contracts Low-quality contracts High-quality quota Low-quality quota
73.8 8.62 8.85 123.2 41.3 20.3 7.7 71.0 78.9 64.0 72.5
48.3 1.19 2.30 85.4 33.2 11.8 9.40 49.3 55.5 44.5 48.8
of spending in a particular month on enlistments in the immediately following month. Since the κ ’s are free parameters, equation (1) allows completely flexible dynamic patterns of advertising effects over the first two months. (The data seem incapable of supporting usefully precise estimation of specifications involving flexible dynamic patterns over more months.) Thus, as can be seen from equation (1), beginning with lags of two months (t-2), the dynamic pattern of effects for medium j is assumed to be an infinite distributed lag with geometrically declining media-specific coefficients that depend on the values of θj. The rate of monthly decline depends on the parameter ρ, the monthly rate of carryover of advertising effects beginning with month t-2. (Estimating separate, media-specific carryover rates appears to be infeasible.) Effects of advertising on ln Ht (holding ln Lt constant) also depend on the functions fj (ajt), which we specify in a way that adds only one parameter per medium. Depending on the data, these functions allow, but do not impose, “S-shaped” (sigmoid or logistic) advertising–sales relationships widely discussed in the advertising literature (e.g., Hanssens, Parsons, and Schultz 1990; Mesak 1992; Simon and Arndt 1980). A response function of this form embodies a “threshold” level of spending below which advertising has essentially no effect (i.e., a convex segment for low levels of spending). Such thresholds formalize the belief that to have a discernable impact, advertising must produce a requisite level of frequency (number of impressions) for several members of the target audience. An S-shaped response function also embodies a “saturation” level of spending above which additional advertising has essentially no impact (i.e., a concave segment for high levels of spending). Such saturation levels formalize the belief that additional advertising will have no impact once the requisite frequency has been attained for virtually all members of the target audience. Specifically, we assume: f j (a jt ) = 1 /[1+ exp(5 − µ j a jt )] , with the parameters µj to be estimated.
(2)
The value 5 in equation (2) is imposed to facilitate estimation while imposing the logical requirement that zero advertising spending has essentially no effect on enlistments. (In view of estimates reported below, this imposed value implies that the effects of zero spending would be predicted to be substantially less than one-tenth of 1% of actual enlistments.) In addition, if µj > 0, the function increases monotonically and approaches a limit of 1 as a jt increases. Thus, in view of equation (1), the parameters (κj , κj1, θj)—which are multiplied by the fj (ajt) during the current month, previous month, and all earlier months, respectively—determine the saturation or maximum possible effects on ln Ht of advertising spending in any medium in any month. The µj ’s determine how rapidly these saturation effects are approached as per capita spending levels increase. The parameters (κj, κ j1, θj , µj , ρ) determine whether the effects of advertising spending through medium j increase or decrease with spending in that medium and whether the effects exhibit the hypothesized S-curve shape over the observed range of advertising spending. To control for variation in recruiter effort levels across observations and to develop a specification involving observable variables only, we assume that Et* = γ H ln( H t / HQt ) + γ L ln( Lt / LQt ) ,
(3)
where (HQt , LQt) are the recruiting unit’s quotas (missions) for high- and low-quality contracts, respectively, in month t. Since recruiters have strong incentives to meet their missions, higher missions should elicit additional effort, and we would expect both γH and γL to be negative. Developing an Estimable Reduced-Form Equation Substituting equation (2) and equation (3) into equation (1) gives an expression involving only observable variables and parameters to be estimated, but involves an infinite number
Summer 2006 115
of lagged terms. To move toward an estimable form, we employ the well-known Koyck method to eliminate the infinite sequence of geometrically declining lagged terms in advertising. This process eliminates the lagged advertising terms beyond month t-2, but results in an equation with a serially correlated disturbance term for each MEPS area of the form vt ≡ εt– ρεt–1. To eliminate this serial correlation, we lag the equation involving vt, multiply by λ =– ρ/(1 + ρ2), and subtract, which results in the following equation with a serially uncorrelated disturbance ut: ln H t = ( λ + ρ) ln H t−1 − λρ ln H t−2 + ( γ L − ψ) ln Lt − ( λ + ρ) ln Lt−1 + 1− γ H +λρ ln Lt−2 1 −γ H ( ln HQt − ( λ + ρ) ln HQt−1 + λρ ln HQt−2 ) + 1− γ H −γ L ( ln LQt − ( λ + ρ) ln LQt−1 + λρ ln LQt−2 ) 1 ∑ i βi ( Xit − ( λ + ρ) Xit−1 + λρXit−2 ) + 1− γ H 1 ∑ j κ j f j (a jt ) + ∑ ( κ j1 − j + 1 − γ H ( ρ + λ) κ j ) f j (a jt−1 ) 2 + λ) κ j1 ) f j (a jt−2 ) 1 ∑ j ( ρ θ j + λρκ j − ( ρ+ + ut . (4) 2 γ 1 − H −∑ λ( ρ θ j − ρκ j1 ) f j ( a jt−3 ) j
The reduced-form equation we estimate is obtained by eliminating from equation (4) the terms involving ln Lt and its first and second monthly lags. This equation is estimated by nonlinear least squares regression (in annual-difference form to eliminate MEPS-level, seasonal fixed effects). The interpretation of the reduced-form equation is that it determines ln Ht—holding effort constant—allowing the values of ln Lt to adjust to differences in recruiting opportunities. The equation to be estimated is nonlinear in the parameters of equations (1), (2), and (3) and λ, the parameter introduced by the second round of lagging and differencing. The number of observations available for estimation, after constructing lagged variables and performing the annual differencing, is 1,452. RELATED LITERATURE The literature on the advertising–sales relationship is massive, and the empirical methods employed vary widely with the substantive context and available data. Literature reviews include Assmus, Farley, and Lehmann (1984); Berndt (1991); Clarke (1976); Dertouzos and Polich (1989); Farley, Lehmann, and Sawyer (1995); Hanssens, Parsons, and Schultz (1990);
and Lodish et al. (1995). We comment on relatively recent literature to place our data and methods in context. Aggregation of Advertising Data Our advertising data are much less aggregated over media, time, and geography than are advertising data typically available to researchers. Moreover, our advertising data exhibit considerable variation over both MEPS areas and months. As we discuss below, estimates based on aggregated data are subject to bias. Aggregation over Geographic Areas Aggregation of advertising data over broad geographic areas, which is typical of the literature, implicitly assumes that advertising intensities are constant over space. If this is not the case—and it is not the case in our data—the aggregated advertising measure can be viewed as error-laden for all locations. As is well known, measurement error in independent variables causes regression estimates to be biased and inconsistent. Aggregation over Time Studies of effects of time aggregation conclude that estimates of advertising effectiveness can be greatly distorted if the advertising data are aggregated over longer time periods than the periods for which decisions about advertising spending levels are made (Clarke 1976; Leone 1995; Russell 1988; and Tellis and Weiss 1995). Fortunately, our data are reported monthly, which appears to be the same time interval the army uses for making advertising spending decisions. Aggregation over Media Most studies use advertising data that are aggregated over all media, and those studies that do distinguish media raise concerns that our data enable us to avoid. Recent studies estimating separate effects for different media include: (1) Seldon and Jung (1993), which distinguishes advertising in four media, but whose sales and advertising measures are aggregated nationally over all advertised products; (2) Giannakas and Tzouvelekas (1998), which studies effects of advertising in three media on market shares of Greek meat processing firms, but whose data are annual and national, and wherein lagged effects are not considered; (3) Pritchett, Liu, and Kaiser (1998), which estimates effects of generic milk advertising in four media on milk demand using quarterly data and fairly flexible lag specifications, but wherein advertising effects are assumed to be characterized by constant elasticities; (4) Notta and Oustapassidis (2001), which studies effects of advertising in four media on profits of Greek food manufacturers, and in
116 The Journal of Advertising
some cases separately considers categories of food, but whose data are annual and national, and wherein lagged effects are not considered; and (5) Yiannaka, Giannakas, and Tran (2002), which studies effects of advertising in three media on sales of Greek meat processors and presents evidence of biases due to aggregating over media, but whose data are annual and national, and wherein the only dynamic feature of specification is inclusion of lagged sales as an independent variable. Dynamic Specifications for Advertising Effects In principle, advertising can affect sales quite quickly after target audiences are exposed to the messages, as well as for some time into the future. Developing tractable econometric specifications allowing potentially long-lived effects to enter in plausible, relatively unrestrictive ways is very challenging. A common approach is to specify an infinite, geometrically declining distributed lag, but this rules out plausible dynamic patterns (e.g., current-period effects that are smaller than effects a few months in the future). Some recent studies adopt less restrictive lag structures. For example, Reberte et al. (1996) use a quadratic exponential function to determine lag coefficients over six months; Capps, Seo, and Nichols (1997) employ a polynomial inverse-lag specification; and Pritchett, Liu, and Kaiser (1998) allow for second-order polynomial distributed lags over three quarters, in some cases imposing endpoint restrictions. Forms of Advertising–Sales Relationships We are aware of no other studies estimating media-specific, S-shaped advertising–response functions. There is considerable empirical evidence for concavity of advertising–sales relationships, which suggests saturation, but little direct evidence of thresholds (Hanssens, Parsons, and Schultz 1990, pp. 178–182; Simon and Arndt 1980). The lack of evidence of thresholds may be largely attributable to advertisers typically choosing not to advertise at levels with increasing marginal returns (Hanssens, Parsons, and Schultz 1990; Simon and Arndt 1980). As it happens, our data have considerable fractions of observed (per capita) spending levels on both the convex and concave segments of the S-curves that we estimate.
linear functional form. Warner’s data, however, are much more aggregated over both media and time than those used here. Hogan et al. (1996) study Navy enlistments using unusually strong advertising data, namely, monthly data for 4 years, 31 recruiting districts, and 4 media. The authors also consider alternative functional forms and alternative lag specifications. Their geometric-decay specification involves imposition (rather than estimation) of the carryover rate. They also consider a polynomial distributed lag specification. Warner, Simon, and Payne (2003) analyzed determinants of high-quality enlistments for four military services using state-level, monthly data for fiscal years 1988 through 1997. The focus of the study is effects of enlistment incentives, but the authors use rich advertising data. In modeling advertising effects, however, per capita advertising variables are entered in levels, advertising in a particular month is not allowed to affect recruiting outcomes in that same month, and advertising in each of the past 11 months is assumed to have the same marginal effect on enlistments in the current month. In sum, the studies of military recruitment advertising (of which we are aware) are undermined by one or more of the following weaknesses in the data or the models employed: (1) too much aggregation of advertising data over time, space, or media; (2) functional forms that force estimates to conform to implausible patterns; (3) overly restrictive dynamic specifications; and (4) failure to take into account that if advertising makes a recruiter’s task less difficult, levels of recruiter effort are likely to decline. ESTIMATES AND THEIR IMPLICATIONS In this section, we begin by reporting estimates of the parameters of the econometric model. Then we use graphical methods to explore implications for the functional form and dynamic patterns of advertising effects on potential high-quality enlistments. Finally, we consider the extent to which reallocating advertising spending across media and the extent to which increasing the advertising budget might have enhanced recruiting opportunities. Parameter Estimates
Studies of Military Recruitment Advertising Dertouzos and Polich (1989) used the data employed in the present study and allowed recruiter effort to respond to changes in recruiting opportunities, but they used simpler (log-linear or constant elasticity) functional forms and (geometric-decay) dynamic specifications. Warner (1991) estimated a model of high-quality Navy enlistments that allowed recruiter effort to adjust according to the difficulty of meeting recruiting goals and used a log-
Estimates are reported in Table 2. First consider the estimated effects of the nonadvertising variables in equation (1) reported in the lower portion of the table. The elasticity of high-quality contracts with respect to the number of recruiters is estimated to be rather small (.150) and not statistically significant. The estimated elasticity of high-quality enlistments with respect to the civilian wage rate is much higher (around 1.0) and statistically significant. The estimated elasticity with respect to the local unemployment rate is roughly .7 and highly
Summer 2006 117 TABLE 2 Estimated Determinants of the Logarithm of High-Quality Enlistments and Recruiter Effort for the U.S. Army in the Early 1980s Variable description Advertising effects on potential recruits Television spending
Radio spending
Magazine spending
Newspaper spending
Ad carryover rate Other determinants of enlistments Intercept Log (recruiters) Log (civilian wage) Log (unemployment rate) Bonus B Bonus C Determinants of recruiter effort High-quality mission Low-quality mission R2 Sample size
Parameter
Estimate
t ratio
κ1 κ11 θ1 µ1 κ2 κ21 θ2 µ2 κ3 κ31 θ3 µ3 κ4 κ41 θ4 µ4 ρ
.122 .00758 .215 .0505 –.000968 .0622 .210 .0936 –.0125 .0549 .125 .443 .0119 –.0138 .0537 3.11 .489
4.90 .34 2.48 8.96 –.05 3.29 2.50 6.01 –.71 3.15 1.61 4.37 .84 –.97 1.01 1.32 14.98
β0 β1 β2 β3 β4 β5
–.00872 .150 1.06 .703 –.0107 .0204
–.45 1.45 1.94 11.08 –.36 .70
γH γL
significant. Neither of the experimental bonus programs appears to have had much effect. The estimated coefficients of equation (3) suggest that higher missions (quotas) for high-quality contracts increase recruiter effort, with a statistically significant estimated elasticity of about .17. The estimated effect of low-quality quotas on effort is opposite from the direction expected, but it is not statistically significant. Turning to the estimated effects of advertising, recall that effects for each of the four media depend on four media-specific parameters (κj , κj1, θj , µj ) and ρ, the monthly carryover rate for effects two or more months after the advertising runs. None of these parameters is expected to be negative a priori, and, in fact, 14 of the 17 point estimates are positive, and none of the negative estimates is statistically significant. The monthly carryover rate is estimated to be a bit less than 50%, an estimate that is highly statistically significant. Regarding the media-specific parameters, we first review qualitative patterns from Table 2, and then use graphical methods to gauge quantitative implications.
–.178 .0457 .5931 1,452
–3.04 1.17
The estimate of κ1 indicates that spending on television advertising increases high-quality enlistments during the month in which the ads are run. The size of this estimated coefficient suggests that saturation spending levels on television advertising can increase high-quality contracts by about 15% during the same month. The estimate of κ11 (which is not significant) suggests no substantial effect one month later. Moreover, the estimate of θ1 suggests that substantial effects persist for two and more months. Evidence of substantial effects during the month in which ads run suggests that a major role for television advertising is to encourage potential enlistees who have already negotiated the early stages of the recruiting process to make the enlistment commitment. The indication of long-lived effects suggests that television ads might also lead potential recruits to contact recruiters and initiate the process. The estimated effects of radio and magazine advertising on potential enlistments have similar time patterns. More specifically, there is no evidence of effects during the month that the
118 The Journal of Advertising
FIGURE 1 Effects of One-Tie Levels in Spending over Six Months Additional high-quality contracts
20 15 10 5 0 5
20
35
50
65
80
95
110
125
140
155
170
185
200
Extra spending per capita
Television
Radio
Magazines
advertising messages are transmitted, an indication of moderate-size effects one month later, and evidence of relatively large, long-lived effects. These estimates suggest that effects of radio and magazine advertising are largely to induce youths to contact a recruiter or take subsequent steps in the enlistment process, but that it is only relatively rarely that they induce youths to quickly sign a contract. Finally, none of the four medium-specific parameters determining effects of newspaper advertising is estimated to be statistically significant. Potential Enlistments and Advertising in Different Media The effects of spending on any medium are spread over several months and depend on several parameters. Figure 1 uses the point estimates to illustrate the effects of spending on television, radio, and magazine advertising on monthly high-quality contracts if recruiter effort levels do not adjust to changes in recruiting opportunities. (Newspaper advertising is ignored because the corresponding curves would not be visible using the spending scales used in the figures.) The figures plot predicted additional contracts (relative to zero spending) over a six-month time horizon—which captures virtually all of the total effects—against per capita spending, assuming a baseline level of contracts of 70 per month (roughly the sample mean). As can be seen from the figure, the estimated advertising– sales relationships are S-shaped for the three media. Table 3 reports estimated, media-specific saturation spending levels and additional contracts from spending at these levels for television, radio, and magazine advertising. (For numerical purposes, a medium-specific saturation level is defined as the level for which an additional five dollars per 1,000 young males would increase potential high-quality contracts by less than .02 over a six-month period.) Figure 1 strongly suggests that the estimated S-curves for the three media differ across media. In particular, as we move from magazine to radio to television, both saturation spending levels and saturation effects on contracts increase. If only a small budget were available—$20 per capita, for example—
the estimates suggest that magazine would be the medium of choice because of the three media, it is the only one that is estimated to offer appreciable enlistment effects for such small spending levels. For larger budgets—$75 per capita, for example—a mixture of magazine advertising (which achieves its maximum effect for roughly $25 per capita) and radio advertising appears to be the best choice. Only when larger budgets are available does television advertising appear to be worthwhile, and television can achieve larger enlistment effects—if enough is expended—than can radio or magazine advertising. We interpret these patterns as being consistent with widespread views of practitioners and researchers. For example, magazines and radio are attractive media for relatively small advertising budgets because they offer more impressions per dollar than does television, and they therefore offer the opportunity to achieve required levels of frequency for relatively low levels of spending (albeit, perhaps, for smaller segments of the target audience). Time Patterns of Advertising Effects for Different Media The estimated effects presented in Figure 1 and Table 3 are spread out over six months. Figure 2 illustrates the time patterns of effects implied by our point estimates for television, radio, and magazine advertising. Specifically, this figure decomposes the effects of spending an additional five dollars per capita on each medium starting at the mean levels for each medium (see Table 1). As before, effects are calculated assuming a baseline of 70 high-quality contracts per month. Note first the geometrically declining effects for all media starting two months after ads run, which is referred to as month 3; these shapes are imposed by the econometric specification in equation (1). Next, consider the qualitative patterns described in reviewing the estimates in Table 2. In particular, the curves for radio and magazine are inverted-U shaped and indicate that current-period effects for these media are unimportant. In contrast, the most important effects for television appear to be during the same month that the advertising messages are broadcast. Optimal Allocation of Spending Across Media Table 4 reports the optimal (i.e., potential-contract maximizing) spending levels for television, radio, and magazines, assuming that total monthly, per capita spending is the observed MEPS-area average of $192.50 (Table 1). (Optima were located using grid-search methods.) To calculate these values, we set spending on newspaper advertising at $3 per capita, which corresponds to the saturation spending level for that medium suggested by our estimates. Since average spending
Summer 2006 119
FIGURE 2 Effects of $5 per Capita Extra Spending from Actuals on Additional High-Quality Contracts, by Month
TABLE 3 Estimated Effects of Actual and Saturation Spending Levels for Different Media on High-Quality Contacts Spending level ($ per capita)
Additional high-quality contracts
Television Actual Saturation
123.20 195.00
12.5 16.3
Radio Actual Saturation
41.30 110.00
2.6 11.0
Magazines Actual Saturation
20.30 25.00
6.8 7.0
Newspaper Actual Saturation
7.70 3.00
1.5 1.5
Additional contracts
0.50 0.40 0.30 0.20 0.10 0.00
– -0.10 1
2
3
4
5
6
Month (1 = Current) Television
Radio
Magazines
on newspapers was $7.70 per capita, we allocate $4.70 (the difference between average spending and the $3.00 saturation point) to the other three media. First, the actual allocation of spending across media accords reasonably well with the estimated optimal allocations. Second, the estimates suggest nonetheless that there was room for improvement through reallocation of spending across media. In particular, during the early 1980s, given the army’s actual advertising budget, monthly spending was roughly $16 per capita too high for television, about $25 too low for radio, and perhaps a few dollars too high for magazines. How much might such reallocations of spending across media have yielded in terms of additional high-quality enlistments? The second-to-last row of Table 4 reports increases in monthly levels of high-quality contracts predicted to result from optimally reallocating spending across media, assuming (again) a baseline level of 70 contracts per MEPS area per month. The estimates suggest an additional 2.4 contracts per month—or a 3.5% improvement on 70 contracts—was attainable. This figure is interpreted as the effect holding recruiter effort constant, but allowing the allocation of effort between high- and low-quality youth to adjust to changing recruiting opportunities. In sum, the estimates suggest considerable potential to improve enlistment outcomes without increasing total spending. The last column of Table 4 relates to an optimal allocation of a budget of $10 per capita (per MEPS area per month) higher than the actual budget, which represents a bit more than a 5% budget increase. As can be seen from the table, relative to the optimal allocation of the actual budget, most of the $10 budget increment would be allocated to television, with most of the remainder allocated to radio. In addition, the budget increment would have been productive. Specifically, relative to optimal allocation of the actual budget, an optimal allocation of the higher budget is predicted to increase potential high-quality enlistments by 1.9 per MEPS area per month.
Total Actual Saturation
192.50 333.00
23 36
TABLE 4 Optimal Versus Actual Per Capita Spending Levels: Potential High-Quality Contributions from Reallocating and Increasing Budget
Television ($/capita) Radio ($/capita) Magazine ($/capita) Newspaper ($/capita) Total ($/capita) Total high-quality contracts from ads Additional high-quality contracts, reallocation Additional high-quality contracts, budget increase
Actual allocation
Optimal allocation
Optimal allocation for actual budget + $10/cap
123.20 41.30 20.30 7.70 192.50
107.40 65.10 17.10 3.0a 192.50
115.10 66.80 17.60 3.0a 202.50
23.56
25.93
27.83
n.a.
2.37
n.a.
n.a.
n.a.
1.90
Note: n.a. = not applicable. a
Newspaper spending constrained to $3 per thousand young males.
Sampling Variability and Its Implications So far, we have ignored uncertainty due to sampling variability in our estimates or predictions that involve more than one parameter. Once we consider sampling variability, questions
120 The Journal of Advertising TABLE 5 Optimal Constant-Budget Spending Levels Conditional on Various Parameter Estimates (U.S. Army, Early 1980s) Optimal allocation over samples
Actual allocation Spending Television Magazines Newspaper
123.2 41.3 20.3 7.7
Optimal allocation for point estimates 107.4 65.1 17.1 3a
Mean
SD
125.6 46.2 17.7 3a
32.0 34.4 7.87 0
a
Spending on other media optimized, given newspaper spending constrained to $3 per capita.
such as the following arise. How much sampling uncertainty is involved in our estimated optimal spending levels? Does consideration of uncertainty lead to different estimates of optimal spending levels? How does uncertainty affect inferences about the effectiveness of advertising generally, contributions of advertising through different media, and potential improvements from reallocating or increasing spending? Tables 5 and 6 report results bearing on these questions. Analyzing sampling uncertainty for estimates of optimal spending levels and predicted effects of budget reallocations or increases in spending is complicated because these estimates depend on estimates of all 17 parameters determining advertising effects. To estimate sampling standard deviations for point estimates that depend on multiple parameter estimates, we simulated their sampling distributions. More specifically, we generated samples of 500 realizations of a multinormal distribution with mean (17 × 1) vector equal to the associated parameter point estimates and variance– covariance matrix equal to corresponding elements of the estimated variance–covariance matrix from the nonlinear regression. (Use of the normal distribution relies on the asymptotic normality of the estimators; with roughly 1,400 degrees of freedom, the normal approximation should be quite accurate.) We then recomputed various quantities of interest for each realization and computed their means and standard deviations over realizations. The first two columns of Table 5 repeat information from Table 4 to aid comparison with the new information contained in the last column of Table 5. To develop the estimates in the last column, for each of the 500 realizations of the sampling distribution of parameters we calculated optimal levels of spending on television, radio, and magazines, assuming news-
paper spending equals $3 per capita and a remaining monthly budget of $189.50 per capita. Over 500 draws from the sampling distribution of the estimator, the average optimal levels of per capita spending on television, radio, and magazines are 125.6, 46.2, and 17.7, respectively (Table 5, last column). As the standard deviations associated with these optimal allocations indicate, however, there is considerable sampling variation associated with these estimates. For example, a confidence interval constructed for optimal level of radio spending (mean plus and minus roughly two standard deviations) would both contain zero. In sum, these results indicate that—even assuming the econometric model is correct—sampling variability in the parameters leads to considerable uncertainty about the optimal allocation of spending across the three media. Of more practical interest, perhaps, are implications of sampling uncertainty for questions such as the following. How effective is advertising? What are the contributions of different media? How much improvement could have been achieved if spending had been optimally allocated across media or if total spending had been higher? Table 6 reports results bearing on these issues. Specifically, these results pertain to the performance over sampling realizations of three spending policies or budget allocations. In subjective probability or Bayesian decision-theoretic terms (i.e., interpreting the sampling distributions as posterior distributions), given a subjective distribution over parameter values, the questions explored presently can be interpreted along the lines of expected performance of three spending policies. As detailed in the upper panel of Table 6, the three spending policies are: (1) the actual (average) monthly spending levels observed in the data; (2) the optimal spending levels assuming that the point estimates in Table 2 are correct (and assuming spending of $3 per capita on newspapers); and (3) the averages over the realizations of the sampling distribution of the optimal spending levels, given the realizations. It is unclear a priori which of the latter two spending policies should perform better over repeated samples. In the bottom panel of Table 6 we report estimated contributions to monthly high-quality contracts (assuming a base level of 70 contracts per month)—averaged over 500 repetitions of a sampling distribution—of the spending level for each medium indicated in the upper panel, the total contribution of advertising, and the standard deviations associated with these estimates. For example, the actual spending levels are estimated to have contributed 23.0 additional potential contracts per month. Moreover, the standard deviation of this estimate is less than one-quarter of the corresponding mean, indicating that actual spending levels did indeed increase potential high-quality enlistments. The corresponding estimates reported in the last row of columns 2 and 3 provide an indication of how much better the army might have done by reallocating spending across
Summer 2006 121 TABLE 6 Performance of Constant-Budget Spending Policies over 500 Draws from Sampling Distributions of Parameters (U.S. Army, Early 1980s) Optimal for point estimates
Actual
Spending Television Radio Magazines Newspaper
Average over samples
Mean
SD
Mean
SD
Mean
SD
123.2 41.3 20.3 7.7
0 0 0 0
107.4 65.1 17.1 3
0 0 0 0
125.6 46.2 17.7 3
0 0 0 0
Additional high-quality contracts Television Radio Magazines Newspaper
12.3 2.74 6.8 1.23
3.56 1.34 3.73 2.11
Total
23.0
5.32
media, as well as which of the two spending policies is more advantageous. The spending levels determined as optimal given the point estimates (column 2) appear to outperform the spending levels determined by averaging optimal spending levels over repetitions of the sampling experiment (column 3). Specifically, the former reallocation is estimated to increase monthly high-quality contracts, relative to actual spending levels, by about twice as much as the latter policy would. In view of the associated standard deviations, however, we cannot reject the hypothesis that the two spending policies are equally effective. What do the estimates in Table 6 suggest about our degree of confidence in the contributions of the individual media? The mean estimates of the contributions of television and radio advertising are considerably more than twice its standard deviation for all three spending policies, but those for magazines and newspapers never exceed twice their standard deviations. Thus, the hypotheses that television and spending are unproductive can be confidently rejected for all spending allocations considered. We cannot, however, reject the hypotheses that magazine and newspaper advertising are unproductive. Finally, how much sampling variability is there in our estimates of the benefits of increasing monthly budgets by $10 per thousand young males in each MEPS area (representing about a 5% increase in the overall budget)? Recall that in Table 4 (last row and column) we reported a point estimate of 1.90 additional potential contracts per MEPS area per month from optimal use of this budget increment. To examine sampling uncertainty for these predictions, we recalculated them for each of 500 realizations of the parameters. (Results are not tabulated.) The result is an average increment in potential contracts
9.70 7.66 6.30 1.03 24.7
3.26 2.80 3.68 1.96
12.6 3.73 6.43 1.03
3.58 1.77 3.70 1.96
5.32
23.8
5.27
of 1.70 with a standard deviation of .36. Moreover, the mean estimate for the contribution of the additional television spending easily exceeds twice its standard deviation. (The means [standard deviations] for the predicted contributions of television, radio, and magazine are 1.35 [.33], .28 [.13], and .11 [.10].) In sum, sampling variability in our parameter estimates is not nearly large enough to undermine the conclusions suggested by the point estimates that a budget increase of roughly 5% would have increased opportunities for the army to recruit highquality youths and that the associated increments in spending on television advertising would have been productive. CONCLUSION The availability of unusually detailed advertising data has enabled us to specify and estimate an econometric model of advertising–sales relationships that is considerably less restrictive than is typical in the literature. In particular, the models we estimated allow: (1) advertising–sales relationships to differ across four media; (2) each relationship to take on the Sshape often discussed, but rarely estimated, in the literature; and (3) completely flexible time patterns of advertising effects over the month ads run and the following month. Our estimated advertising–sales relationships are, in fact, S-shaped, and the estimated timing of effects of advertising on sales also differ greatly across media. Broadly, our substantive findings support advertising practice generally and army advertising decisions during the early 1980s. First, the advertising–sales relationships we estimate are consistent with conventional wisdom and advertising practice. More specifically, depending on the budget, advertising
122 The Journal of Advertising
through television, radio, magazines, and (tentatively) newspapers can help to increase sales. Second, if only a rather small budget is available, it seems best to advertise only in print media; as budgets expand, it becomes optimal to add radio advertising to the mix; and as budgets expand further, television should also be used. Due to significant changes in the media landscape, most notably the growth of cable, satellite, and Internet services, as well as evolving perceptions about military service, empirical results from the early 1980s should be used only with caution in the design of policies today. Still, it remains interesting that U.S. Army advertising during the early 1980s appears to have been very productive. In particular, our estimates suggest that the army advertising program increased potential high-quality enlistments by roughly 32% (relative to no advertising at all). Moreover, the allocation of spending across media was impressive; our estimates suggest that an optimal reallocation across media would have yielded only a 3.5% improvement in recruiting opportunities. Finally, our estimates indicate that a 5% increase in the advertising budget, most of which would have been best allocated to television, could have increased recruiting opportunities by another 3%. REFERENCES Assmus, Gert, John U. Farley, and Donald R. Lehmann (1984), “How Advertising Affects Sales: Meta-Analysis of Econometric Results,” Journal of Marketing Research, 21 (February), 65–74. Berndt, Ernst R. (1991), “Causality and Simultaneity Between Advertising and Sales,” in The Practice of Econometrics, Classic and Contemporary, Reading, MA: Addison-Wesley. Capps, Oral, Jr., Seong-Cheon Seo, and John P. Nichols (1997), “On the Estimation of Advertising Effects from Branded Products: An Application to Spaghetti Sauces,” Journal of Agricultural and Applied Economics, 29 (December), 291–302. Clarke, Darral G. (1976), “Econometric Measurement of the Duration of Advertising Effect on Sales,” Journal of Marketing Research, 13 (November), 345–357. Dertouzos, James N. (1985), Recruiter Incentives and Enlistment Supply, Santa Monica: RAND (R-3065-MIL). ———, and J. Michael Polich (1989), Recruiting Effects of Army Advertising, Santa Monica: RAND (R-3577-FMP). Farley, John U., Donald R. Lehmann, and Alan Sawyer (1995), “Empirical Marketing Generalizations Using Meta-Analysis,” Marketing Science, 14 (special issue), G36–G46. Giannakas, Konstantinos, and Vangelis Tzouvelekas (1998), “Strategic Behavior and Competition in Dynamic Industry: Greek Processed Meats,” Agribusiness, 14 (March/April), 157–169. Hanssens, Dominique M., Leonard J. Parsons, and Randall L. Schultz (1990), Market Response Models: Econometric and Time Series Analysis, Boston: Kluwer Academic.
Hogan, Paul F., Timothy M. Dall, Pat Mackin, and Chris Mackie (1996), “An Econometric Analysis of Navy Television Advertising Effectiveness,” Research Report (draft), United States Navy Recruiting Command (December). Leone, Robert (1995), “Generalizing What Is Known About Temporal Aggregation and Advertising Carryover,” Marketing Science, 14 (special issue), G141–G150. Lodish, Leonard M., Magid Abraham, Stuart Kamelson, Jeanne Livelsberger, Beth Lubetkin, Bruce Richardson, and Mary Ellen Stevens (1995), “How T.V. Advertising Works: A Meta-Analysis of 389 Real World Split Cable T.V. Advertising Experiments,” Journal of Marketing Research, 32 (May), 125–139. Mesak, Hani I. (1992), “An Aggregate Advertising Pulsing Model with Wearout Effects,” Marketing Science, 11 (Summer), 310–326. Notta, Ourania, and Kostas Oustapassidis (2001), “Profitability and Media Advertising in Greek Food Manufacturing Industries,” Review of Industrial Organization, 18 (February), 115–126. Polich, J. Michael, James N. Dertouzos, and S. James Press (1986), The Enlistment Bonus Experiment, Santa Monica: RAND (R-3353-FMP). Pritchett, James G., Donald J. Liu, and Harry M. Kaiser (1998), “Optimal Choice of Generic Milk Advertising Expenditures by Media Outlet,” Journal of Agricultural and Resource Economics, 23 ( July), 155–169. Reberte, Carlos, Harry M. Kaiser, John E. Lenz, and Olan Forker (1996), “Generic Advertising Wearout: The Case of the New York City Fluid Mild Campaign,” Journal of Agricultural and Resource Economics, 21 (December), 199–209. Russell, Gary J. (1988), “Recovering Measures of Advertising Carryover from Aggregate Data: The Role of the Firm’s Decision Behavior,” Marketing Science, 7 (Summer), 252– 270. Seldon, Barry J., and Chulho Jung (1993), “Derived Demand for Advertising Messages and Substitutability Among the Media,” Quarterly Review of Economics and Finance, 33 (Spring), 71–86. Simon, Julian L., and Johan Arndt (1980), “The Shape of the Advertising Response Function,” Journal of Advertising Research, 20 (August), 11–28. Tellis, Gerald J., and Doyle L. Weiss (1995), “Does TV Advertising Really Affect Sales? The Roles of Measures, Models and Data Aggregation,” Journal of Advertising, 24 (Fall), 1–12. Warner, John T. (1991), “Navy Recruiting Incentive Models” (manuscript), San Diego: Navy Personnel Research and Development Center. ———, Curtis J. Simon, and Deborah M. Payne (2003), “The Military Recruiting Productivity Slowdown: The Roles of Resources, Opportunity Cost and the Tastes of Youth,” Defence and Peace Economics, 14 (October), 329–342. Yiannaka, Amalia, Konstantinos Giannakas, and Kien C. Tran (2002), “Medium, Message and Advertising Effectiveness in the Greek Processed Meats Industry,” Applied Economics, 14 (September), 1757–1763.
