What is attribute non-attendance?

In discrete choice experiments (DCEs), respondents are asked to choose amoung different options which vary across different attributes. For instance, a DCE on mobile phone preferences could have processor speed, battery life, screen size and cost as attributes. A DCE looking at different treatments could have expected survival, anticipated side effects and cost as attributes.

DCEs assume that respondents are rational and have complete, monotonic, and continuous preferences. However, “Continuity of preferences implies that individuals use compensatory decision-making processes, meaning that they take into account all the available information to make their decisions.” In practice, however, this may not be the case. Instead, respondents may use simple strategies or heuristics to make their decisions. Under one common heuristic, known as attribute non-attendance (ANA), respondents ignore one or more attributes when deciding across DCE attributes.

Is ANA common? In short, yes.

In a recent study, Hole (2011) used a two-step process to account for endogenous ANA of respondents and found that ‘a substantial share of the respondents ignored one or more attributes when making their choices’.

So how do you fix this problem?

A paper by Legarde (2013) proposes using a latent class model (LCM), where each class represents a specific non-attendance decision rule, where the attributes ignored have a parameter set to zero.

In the standard logit model, with individual i choosing among J alternatives, for a specific class q, the choice probability function is:

The probability that individual i belongs to class q is:

where θ_Q is normalized to 0. The log likelihood function is:

ln(L) = Σ_{i=1 to N} ln(P_{ii=1 to N} ln{Σ_{q=1 to Q} H_iq Π _{t=1 to T}P_it|q}

the class specific probabilities for each individual H_q|i can be calculated with Bayes rule.

In practice, how is this done? This simplest approach is to assume that respondents take into account all attributes or ignore only one. One could continue to the case where multiple attributes are ignored simultaneously. However, there is a risk that the number of people per class is small, and thus not very informative for research purposes. To address this, the authors propose a stepwise approach:

…the first specification includes eight classes: one class that allows respondents to have not ignored any attribute (class 1), and seven others where only one attribute at a time can be ignored (classes 2–7). …Based on the results of this first model, another LCM is estimated, which includes two types of classes: ‘old’ and ‘new’ ones. Old classes are those that were found relevant in the first model (i.e. classes that did not have a class probability equal to zero), whereas those that were not found relevant are dropped. ‘New’ classes define new ANA patterns, not tested in the previous model estimated.

Another approach to solving the ANA is the endogenous attribute attendance (EAA) proposed by Hole, Norman and Viney (2014). The authors apply EAA to a model trading off longevity against morbidity. Assuming a utility of 0 at death, the authors assume the following utility function for individual i, for option j, in choice set s:

• U_isj = α(TIME_isj) + Σ_k TIME_isj[x_isjβ]+ε_isj

if we assume some attributes are ignored, then:

• U_isj = α(TIME_isj) + Σ_{k ∈ Cq} TIME_isj[x_isjβ]+ε_isj

Thus, the probability a person selects choice j is given by:

What are the tradeoffs between the two models?

To be specific, while the number of parameters in the [LCM] logit model increases exponentially with the number of attributes in the model, in the EAA model, the number of parameters increases linearly. However, the EAA approach involves additional assumptions regarding patterns of attendance that may not be realistic in some contexts

Specifically,

The main advantage of the EAA model is that it is parsimonious; while in the LC model, one vector of parameters is estimated per attribute subset, there is only one vector per attribute in the EAA model.