The dataset comes from an online retailer of luxury clothing brands. Data included 16,236 shoe purchases over five years (2010–2015) from 2,007 women who had changed primary residence locations (i.e., US Zip Code) at least once. Average heel-size of shoes purchased by participants in the destination location was the key dependent variable (i.e., new sartorial behavior). Key predictor variables were the average heel-size purchased by participants in the origin location (past sartorial behavior); the average heel-sizes purchased by others in the participant’s origin location (origin norms) and destination location (destination norms); and the median income level in the US Census Region of origin (origin SES) and destination (destination SES), as given by 2010 US Census data.

Our data come from an online retailer of luxury clothing brands. The retailer’s name cannot be disclosed due to a non-disclosure agreement prohibiting the naming of the company, though use of the data for research purposes and disclosure of summary statistic is permitted. Of primary interest are individuals who we can identify as having moved from one geographic location to another during our observation window. This identification task was accomplished by defining a “mover” as someone in United States who had at least five transactions in one location followed by at least five transactions in a new location. Though this specification is imperfect, it is reasonable to assume that an individual who purchased clothing a number of times in one area and then systematically switched their purchases to a new area has, at minimum, a strong physical presence in the two geographies, and, at best, permanently moved. Moreover, because our measure of interest is heel size, we further restricted our data set to those individuals who purchased at least one shoe in either of the two locations. Looking only at shoe transactions (which is what we have direct access to), women remain in a location, 115 days on average. This likely underestimates how long they actually live in a location, because our data are both left and right truncated and we only examine shoe purchases, not all clothing purchases. This resulted in data from 2,007 individuals who placed 16,236 orders for shoes between the dates 6/14/2010 and 2/3/2015. Of those, 1,865 had data from areas where we were able to ascertain sufficient information (more on this below). Of those, 15 were outliers in terms of their number of moves (more than two standard deviations from the mean) and so were excluded, resulting in a final data set of 1,850 customers who placed 14,496 transactions (see Table 1 for summary statistics and S1 Fig for histogram of number of transactions per customer and S1 Table for correlation matrix of all relevant variables). Of note, the online retailer does not customize offerings by region within the United States, and therefore all shoe options were available to all individuals regardless of their geographic location.

The dataset contains panel data at the transaction level. For each transaction, we observe unique user id, date of transaction, census region of transaction, MSRP (not the price paid) of shoe, number of shoes purchased in that location on that date, and heel size of shoe purchased. In addition to the transaction data, we constructed norm indices at the transaction level that allow us to identify the norms of the region that an individual purchased their shoes in. This was accomplished by averaging across all transactions of other customers in the region of interest for the 14 days prior to the transaction in question. That is, for each transaction, we constructed unique norm indices by averaging across all other transactions from other customers in the census region where the transaction took place for the previous 14 days. This allowed us to construct norm indices for heel size and MSRP. These norm indices are particularly useful as they naturally account for seasonal variations in norms (e.g. higher heels purchased in the summer time) because they only look at recent transactions.

In order to ascertain the impact of location on heel size preferences, we transform the data into individual level transaction data averaging across sales within a geographic region within an individual. We classify the data into two groups: data from the origin location and data from the destination location. Origin location data is the average of all transaction data from a given individual within the first observed location in the dataset. Destination location data is the average of all transaction data from a given individual within a subsequent location in the dataset. For each individual, there may be more than one such origin-destination pair (if someone moved multiple times) and so we include multiple rows of data as needed and control for both individual level variation and move number with random effects. See S2 Fig for a histogram of number of moves.

Because some transactions were either 1) at the very beginning of our data collection window (left censored data) or 2) in geographic regions with no other transactions for the 14 days prior to the transaction in question, we were unable to construct norm data for them. There were 398 (2.7%) observations that had any such missing data. Because the type of data missing varied across customers, we do not exclude customers merely for having one missing data value. Rather, we exclude customers when the data missing applies to the model being tested (see bottom of Table 2 for Customer and Observation sample sizes in each model).

The socioeconomic status of a geographic region was determined with 2010 ZIP code level US census data and create census region averages. Specifically, for each region, we observe the population size, % of population that is female, the median age of the individuals, average years of education of individuals, and, most critically, the median household income of individuals.

The first author consulted the institutional review board of his university prior to the start of research. The review board deemed that this research does not require approval as the data being used are archival in nature and, critically, lack any identifiable information making them anonymous.

The firm providing the data used in the subsequent analyses has requested that the data not be made publicly available. All authors signed non-disclosure agreements with the firm to such an effect. The firm has allowed aggregate data such as that reported in the Tables of this paper and any results from data analyses that do not inadvertently reveal individual level data to be made publicly available. For any additional aggregate analyses, the corresponding author can be contacted and aggregate results can be shared on an ad hoc basis.

Does purchase behavior after moving reflect conformity to the new location or consistency with their past behavior? To address this question, we ran a series of mixed model regressions with random intercepts for individual, move number, origin location, and destination location predicting own average heel size purchased in the destination location (Table 2). Of note, origin location and destination location random effects allow us to control for any variation in the general nature of the locations that we could not observe with region level data (e.g. New York City’s cutting edge fashion scene vs. Topeka, Kansas’ more conservative approach to fashion). Heel size was log transformed in order to normalize the data.

When predicting the average heel size purchased in the destination location as a function of the average heel size purchased in the origin location (a measure of personal preferences), origin norms, and destination norms (Model 1), own preferences from the origin location (B = .20, SE = .02, t = 10.97, p < .001) and destination norms (B = .30, SE = .04, t = 7.23, p < .001) significantly predict own preferences in the destination location, while origin norms have no predictive power (B = .05, SE = .03, t = 1.58, p = .12). In other words, behavior in a new location demonstrates both consistency and conformity, balancing past behavior with the influence of new norms. Interestingly, origin norms, though not significant direct predictors of destination behavior, do operate indirectly via established individual preferences (see S1 Appendix for an explanation).

People who move may demonstrate some level of conformity, but it may differ between moves to higher SES locations (i.e., upward relocations) versus moves to lower SES locations (i.e., downward relocations). To test this, we computed a new variable—change in median income between origin and destination behavior—and tested whether this interacted with conformity level in the same mixed model (Model 3). As can be seen by the crossing lines in Fig 1, the analysis revealed a significant interaction (B = .005, SE = .002, t = 2.52, p = .01). Of note, to assure that the origin norms and origin behavior covariates do not bias our interaction estimate [20] we also test the interaction in Model 4 without these statistical controls. Doing so does not change our conclusion as the interaction term is of the same magnitude and still statistically significant (B = .005, SE = .002, t = 2.66, p = .007).

The influence of destination norms is largest when the change in median income between locations is large and positive, and smallest when the same change is large and negative. When an individual moves to a higher status location, their behavior strongly reflects the norms of that destination. However, when an individual moves to a lower (or similar) status location, they mostly do not assimilate to those new norms and instead remain relatively consistent in their preferences.

To give a sense of the size of this effect, we can consider an individual moving to a new location that has a median income level that is $22,723 higher (1 standard deviation, in terms of difference between destination and origin median income levels) than their origin location’s income. All else being equal, if the norms of that new destination are that shoes are purchased with an average heel height of 1”, then the behavior of this individual will be to purchase shoes with heel height of about .95”. Likewise, if the norms of that new destination are that shoes are purchased with an average heel height of 3”, then the behavior of this individual will be to purchase shoes with heels of about 2.11” in height. In other words, this person will largely assimilate to the preferences of their new location. In contrast, if the same person moved to a new location that has a median income level that is $22,723 lower than their origin location, all else being equal, their decision about heel height is far less influenced by the norms of that new location. For instance, if the average heel height preference of others in that new location is 1”, then the behavior of this individual will be to purchase shoes with heels that are about 1.14” in height. However, unlike when moving to a higher income level location, when the average heel height preference of others in this new location is 3”, then the behavior of this individual is largely unaffected by these norms and they purchase shoes that are about 1.56” in height. In other words, though there is some assimilation to this new higher heel norm, there is far less assimilation than when the new location is of a higher SES. Put another way, when moving to a new location with a median income that is $22,723 higher than the origin location, the unstandardized beta coefficient between destination norms and destination behavior is .54, whereas when moving to a new location with a median income that is $22,723 lower than the origin location, the same beta coefficient is only .21, a sign of far less conformity to new norms.

To test the robustness of the SES change by conformity interaction, we preform three checks. First (Model 5), we include a series of fixed effects controls: change in average MSRP of shoes across locations, change in population size across locations, change in % of women across locations, change in median age across locations, and change in number of years of education across locations. Even when including these controls, we observe a significant interaction between destination norms and change in median income between locations (B = .005, SE = .002, t = 2.61, p = .009).

Second, we change the nature of the norm variables to include averages across all previous transactions, not just those made in the previous 14 days (Model 6). Because this specification removes the benefit of seasonality being controlled for by the nature of how the norms variables were computed, we include a new random effect of Month. This ensures that any variations we observe cannot solely be attributed to seasonal differences. In Model 6, we still observe a significant interaction between destination norms and change in median income between locations (B = .011, SE = .004, t = 2.67, p = .007).

Finally, it is possible that customers who move often exert excessive influence on our model and so even with the customer and move number random effects controls, our results are driven by these unusual individuals (Model 7). We reran our analysis with only the first observed move from each individual (removing the customer and move number random effects as they are no longer needed). Doing so once again yields a significant interaction between destination norms and change in median income between locations (B = .006, SE = .002, t = 2.45, p = .01). In sum, regardless of model specification, upward relocations involve greater conformity to destination norms than downward relocations, which involve greater consistency with origin behavior. This is true even when controlling for location factors including MSRP of shoes purchased, population size, average age, gender ratio, and education level. To make this complex relationship a bit clearer, we include three illustrative examples in Fig 2.