For a number of reasons discussed below, we decided that populations of restaurant establishments were nearly ideal candidates for analysis of niche width and mortality. Because our study of restaurants was designed to test hypotheses from one kind of model, it is the most narrowly focused of the data sets. It is also the only design that uses a prospective observation plan.
Although the restaurant industry is huge in terms of employment and sales, individual restaurants are neither large nor powerful. Why study restaurants rather than the large firms and government bureaus that have preoccupied organizational research? The choice of restaurants offers several advantages for testing our version of niche width theory. First, restaurants are small enough that most cities have many of them. Hence it is feasible to estimate the distributions of effects of environmental variations on restaurants. And restaurants have diverse forms. For example, fast- food forms, many types of ethnic restaurants, coffee shops, luxury dinner houses, and natural food restaurants can readily be identified. Moreover an obvious spectrum of specialism/generalism underlies the variety of forms. Some restaurants, such as sandwich shops and night clubs specialize to a narrow price range; others span much of the range. Similarly some restaurants, such as doughnut shops and luncheonettes, specialize to one part of the day; others operate 24 hours a day. Finally, some restaurants specialize in a narrowly defined cuisine while others provide a broad array of menu items. Overall, there appears to be great variability in terms of specialism/generalism.
Restaurants supposedly have much shorter life expectancies than many other kinds of firms. If disbanding rates are sufficiently high, observation over short periods can yield enough failures to permit meaningful survival analysis. We chose to study restaurants partly for the reason that geneticists study drosophila: both turn over rapidly enough that one is not restricted to retrospective analysis.
Finally, it is easy to find reasonably complete listings of restaurants in urban areas. Since restaurants depend on local consumer environments they usually purchase listings in the Yellow Pages. This publication provides a convenient universe from which to sample restaurants.
1. Research Design
We conducted prospective analysis because following establishments for- ward in time increases the likelihood of recording events for short-lived organizations. As we have noted repeatedly, we wanted to avoid the com- mon practice of sampling only the most successful members of an organi- zational population.
The unit of observation here is the individual establishment. Our design goal was to maximize variation in grain and variability of sales among establishments. However, we knew nothing about the environments of particular establishments before sampling. Therefore, we used information about variation in sales among cities in designing a plan for sampling estab- lishments. We analyzed time series of aggregate restaurant sales by quarter for the 200 largest cities in California for the period 1974-1977 and chose 18 cities that fell toward the extremes of the joint distribution of seasonality and variability of aggregate sales. Sampling establishments within these cities should maximize variation in grain and variability, and minimize the correlation between them.
In 1977 we sampled up to 50 establishments in each city. In cities with more than 50 restaurants we took a random sample; in smaller cities we selected the whole population. Establishments that had already closed or that refused to participate were replaced with a randomly chosen alternative when the local population had not been exhausted.
We conducted telephone interviews with owners or managers to obtain information about each restaurant’s history and current operations. The interviews were repeated in 1978 and 1979. In later interviews, we added establishments to replace those that had failed. In an effort to obtain more variation in life-cycle characteristics, we oversampled newly listed restaurants in the second and third waves. In all, we sampled 985 establishments.78
The period of observation ranges from one to three years because res- taurants were added in the second and third years of the study. This feature of the design requires use of a method of analysis that adjusts properly for differential length of exposure to the risk of having an observed disbanding, as we discussed in Chapter 8.
Although most establishments in the sample were independently owned local establishments, 128 were branches of large chains. We suspect that the contingencies of selection differ for local establishments and members of chains. For one thing, chains operate in many local markets. Since we have information only on local environments, we restrict attention to organizations that operate completely within the city; this restriction leaves us with 810 observations. Eighty-one of these cases had missing data on seasonality or age. Excluding these cases, the effective sample size is 729. Of these, 102 (14 percent) closed during the observation period and 91 (12 percent) were sold. We treat closure as the event of interest and sale of the establishment as an independent competing risk. Cases in which a restaurant is sold are treated as censored at that time with respect to the risk of closure.
A restaurant can specialize in many ways: type of cuisine, style of service, hours of operation, price range, diversity of menu items, or range of ser-vices. We use a single measure of specialism that reflects our understanding of the main strategic choices involved in attempting to establish a restaurant. Some attempt to appeal to a narrow range of the population of potential customers. Others try to appeal to the “average” consumer who occupies the middle of the market range reflecting quality and price. After studying the completed interviews and advertisements in the Yellow Pages, we classified establishments into 33 forms. This coding relied heavily on menu specialization; it also took into account the hours of operation, number of entrees, and staff composition. We distinguished cafés, steak houses, coffee shops, health food restaurants, taco stands, and Basque restaurants, to name only a few.
The distinction used in this chapter collapses the 33 forms into three: generalist, fast-food establishment, and other specialist. To be coded a generalist, a restaurant had to meet three criteria: (1) offer a fairly general menu, not limited to items such as pizza or hamburgers and not dominated by a specific ethnic cuisine; (2) offer seating; and (3) employ at least one person in the differentiated role of chef or cook.80 Thirty-one percent of the restaurants in the sample are generalists. The fast-food form contains es- tablishments that specialize in one of the following: pizza, hamburgers, hot dogs, fried chicken, tacos, doughnuts, or ice cream. Thirty-four percent of the establishments fall in this category. The remaining thirty-five percent, the specialists, tend to feature ethnic cuisine or have a very limited menu.
In our first analyses of these data (Hannan and Freeman 1982; Freeman and Hannan 1983) we established that the dynamics for the fast-food and for the specialist forms do not differ substantially. So we collapse this distinction here. With this revision, roughly two-thirds of the population are specialists, and a third are generalists.
In preliminary analysis we learned that specialists tend to be slightly older than generalists but have slightly lower average sales (Freeman and Hannan 1983, table 1). The fact that these differences are small suggests that we are not comparing generally successful organizations with generally unsuccessful ones when we compare specialists with generalists.
We assume that these broad forms respond differently to changing demand that they have different fitness functions. Demand for restaurant services has many distinguishable components, such as the business lunch trade, family dining, tourism, meetings, and banquets. We assume that the various components do not change proportionately each time aggregate sales rise or fall. Generalists are presumably less sensitive to fluctuations in aggregate demand because they offer a greater range of services to more diversified populations of consumers. They achieve this reduced sensitivity at least partly by maintaining the excess capacity of employing a chef.
Since many respondents were reluctant to reveal detailed financial infor- mation, we could not measure variability in sales over time for individual establishments. It is not clear, however, that this would have been appro- priate. Extreme fluctuations in sales for a particular establishment may be confounded with disbanding. As business failure threatens, restaurants may manipulate their prices, hours of operation, quality of food or service, and other features of their operations in ways that temporarily influence sales. Therefore, for both practical and theoretical reasons, we do not measure variability at the establishment level. Instead we give each establishment the variability score for the city in which it is located. We used the time series of quarterly gross sales by city mentioned earlier. Our measure of variability (denoted by V) is the coefficient of variation of each city’s gross restaurant sales by quarter over the period extending from the fourth quarter of 1976 to the first quarter of 1980. This measure ranges from a low of .058 to a high of
The second important dimension is grain. It is not obvious a priori whether a particular patch size is large or small for a class of organizations. Should we analyze weekly variations in demand? Quarterly variations? Yearly variations? Our understanding of the economics of the industry is that restaurants can average over variations from one week to the next, but that several months of low sales create serious problems. This view suggests that seasonal variations in sales ought to be considered coarsegrained for restaurants. Therefore, for our purposes a restaurant has a coarse-grained environment when its sales variations have a strong seasonal component.
We have tried two approaches to measuring coarseness of grain. The first is the strength of the seasonal component in the time series of aggre-gate sales for the city in which the establishment is located. This is the same method used to stratify the sample. The second method uses a self- report from each establishment about the seasonality of its sales. We obtain sharper results with the self-reports. The patterns of estimates are similar with the two measures, but the standard errors of estimates are considerably smaller with the self-report. This is not surprising because there are likely to be important differences among subenvironments in each city. For example, restaurants located near highway interchanges probably face opposite seasonal patterns from those in central business districts. Moreover, the calendar quarters used in official statistics do not square well with the seasonal patterns actually faced by restaurants in some cities.82 Therefore, we use self-reported seasonality as our measure of coarseness of grain (denoted by C) in the analysis that follows. Twenty- three percent of the establishments report a highly seasonal pattern of sales.
It turns out that generalism and coarseness of grain are not correlated the X2 statistic for independence in this four-fold classification is only 0.02. Generalism has a weak correlation with variability (r = 0.02) and the correlation between variability and coarseness of grain is 0.33. Restaurants in coarse-grained environments tend to be slightly older than fine-grained ones and to have slightly higher sales.
Use of our adaptation of Levins’ model in analyzing the relative mortality rates of specialist and generalist restaurants requires that we establish the likely shape of the fitness functions of restaurant populations. Are the fitness sets of populations of restaurants convex or concave? We concentrate on the effects of variations in demand for services. Restaurant sales fluctuate with seasons, business cycles, and consumer income cycles. They also respond to purely local events and to fashions and fads. Are typical demand variations large relative to the adaptive capacities of restaurants? Most restaurants seem to operate close to the margin, suggesting that they are sensitive to even small fluctuations in demand. This means that the fitness sets of restaurants are concave with respect to typical variations in demand. The rest of our argument depends on this assumption.
We began our analysis by exploring the effects of various control variables, including measures of the size of the establishment (number of employees, number of seats, number of meals served per week), its age, the logarithm of its gross sales, and characteristics of its local environment such as total restaurant sales, sales per establishment, and measures of growth and decline in aggregate sales. Only the age at time of first observation and the log of gross sales84 had systematic effects on mortality rates. Given the theoretical and empirical importance of age dependence in mortality rates, we concentrate on results of models that include this factor. Readers interested in findings without age (which allows use of a slightly larger number of cases) and with effects of size (which limits the number of cases considerably) may consult Freeman and Hannan (1987).
With these preliminaries, we arrive at our specification:
where A contains the main effects of V and the effect of age.85 The subscripts in equation (12.1) are a reminder that coarseness of grain, general- ism, and age vary among establishments (indicated by the subscript i) but that variability varies only between cities (indicated by c).
Table 12.1 reports ML estimates of this model.86 The first hypothesis is that mortality processes differ by grain. A likelihood ratio test rejects the hypothesis that the two parameters indexing the effects of grain are zero at the .01 level. So the data agree with this broad hypothesis.
The two point estimates are relevant to the specific hypotheses about fine-grained environments. Hypothesis 2a, that β >0, tells that specialists are favored in fine-grained (non-seasonal) environments. This hypothesis receives no support: β does not differ significantly from zero. Hypothesis 2b, that β + γV >0, means that specialism is favored over generalism over the full range of environmental variation in fine-grained environments. This hypothesis is supported in this analysis. The point estimates imply that the mortality rate of generalists exceeds that of specialists over almost the entire range of V. However, the ratio of the two mortality rates is close to unity at the minimum observed level of V. Perhaps more important is the fact that the point estimate of y differs significantly from zero in the direction predicted. So these results suggest that specialists are favored in fine-grained environments that fluctuate greatly, contrary to the conven-tional view noted at the beginning of the chapter (see also Freeman and Hannan 1983).
Next consider the case of coarse grain, that is, seasonal environments. The estimated relative mortality rate of generalists to specialists (see equation 12.3) is
The expression in parentheses is positive for small observed values of V and negative for large observed values of V, as predicted. The expression in parentheses changes sign when V equals 0.08, which falls within the observed range. According to these estimates, the relative mortality rate of generalists to specialists at Vmin equals 1.66. In other words, the mortality rate of generalists exceeds that of specialists by two-thirds in stable, coarse-grained environments. But in extremely variable environments (when V = Vmax), the relative mortality rate is .01, which means that the mortality rate of generalists is only a hundredth as large as that of specialists.
Because coarseness of grain and variability are correlated, estimates of effects at the extremes are unlikely to be precise. Nonetheless, the qualitative pattern implied by the model appears to hold. Not only are the implied differences in mortality rates of specialists and generalists large, but in addition, both δ and ζ differ significantly from zero at the .05 level.
Our model makes several predictions about the effects of environmental variables on the relative mortality rates of specialist and generalist organi- zations. The global prediction is that the patchiness of environmental variation affects the selection process. Analysis of this problem in terms of seasonality of demand confirms the prediction. Establishments that report seasonal sales patterns are affected differently by variability in aggregate sales.
The model also predicts how the relative mortality rates vary by grain. The predictions are confirmed for the case of fine grain: the signs of estimated parameters agree with the predictions. The usual argument is that any kind of environmental instability favors generalist strategies, but our estimates imply that this is not so in fine-grained environments. In particular, our estimates imply that specialists are actually favored in fine-grained environments with high variability. This finding agrees with the prediction of niche width theory.
The second case, involving coarse-grained environments, gives strong support to the model. The implied relative mortality rate shows the pre-dieted reversal over the range of V. The parameters associated with this pattern differ significantly from zero in the predicted direction.
Source: Hannan Michael T., Freeman John (1993), Organizational Ecology, Harvard University Press; Reprint edition.