Modelling multiple occurrences of activities during a day: an extension of the MDCEV model

Abstract

The increased interest in time use among transport researchers has led to a search for flexible but tractable models of time use, such as Bhat's Multiple Discrete Continuous Extreme Value (MDCEV) model. MDCEV formulations typically model aggregate time allocation into different activity types during a given period, such as the amount of time spent working and shopping in a day. While these applications provide valuable insights into activity participation, they ignore disaggregate activity-episodes, that is the fact that people might split their total time spent working in multiple separate blocks, with breaks or other activities in between. Insights into this splitting into episodes are necessary for predicting trips and understanding time use satiation. We propose a modified MDCEV model where an activity-episode, rather than an activity type, is the basic choice alternative, using a modified utility function to capture the reduced likelihood of individuals performing a very large number of episodes of the same activity. Results from two large revealed preference datasets exhibit equivalent forecast accuracy between the traditional and proposed approach at an aggregate level, but the latter also provides insights on the number and duration of activity-episodes with significant accuracy.

1. Introduction

Travelling is a necessity that arises due to individuals' activity patterns (Bhat and Koppelman 1999). People choose how to spend their time and then travel to different locations to carry out their chosen activities. This perspective has gained momentum among transport researchers, who have then been developing models to accurately understand and predict time use decisions.

Time use decisions can be thought of as choosing the activity type (purpose, e.g. work, education, shopping, etc.), number (count by purpose, i.e. number of episodes of a given activity) and duration. In the last decade, the multiple discrete continuous (MDC) structure pioneered by Hanemann (1984) has evolved into an elegant framework to model activity participation and time allocation decisions subject to a budget constraint (Bhat 2008; Bhat, Castro, and Khan 2013; Liu, Susilo, and Karlström 2017; Wang and Li 2011). However, the state-of-the-art MDC models focus on predicting the aggregate duration for an activity type rather than accommodating the time allocation at the episode level (Bhat and Misra 1999; Calastri et al. 2017; Enam et al. 2018). Hence, the time allocation information obtained from the state-of-the-art MDC models can at best act as a constraint (Bhat et al. 2004), but will seldom be (immediately) useful for the representation of downstream travel choices such as number of trips, mode, destination and route, which rely on episode-level activity participation and time allocation decisions (Auld and Mohammadian 2009).

Splitting the time invested in each activity into multiple episodes is relevant from a behavioural perspective, as engaging in an extended episode of an activity is different from engaging in multiple episodes of the same type for the same combined duration. For example, working for 4 hours, having a lunch break and then working for another 4 hours is not behaviourally equivalent to working for 8 hours continuously. For the purpose of travel behaviour analysis, capturing the episode-level activity participation and time allocation decisions allows us to construct the trips that tie together consecutive activity episodes and subsequently model the associated travel decisions such as mode, destination or route choice (Gärling 1998). Our example above involving two episodes of work separated by a lunch break will lead to four trips (home-work, work-restaurant, restaurant-work, work-home), but simply knowing that an individual works for 8 hours a day does not provide information on the particular number of trips performed on top of the first and last (home-work and work-home). This is a crucial shortcoming of the existing modelling approaches, as trip information is necessary to generate the actual demand for a transport system, which is often the end goal of transport planning operations and management.

Our approach, even though not directly applicable to agent-based simulation models (ABM), can be a useful input if combined with a scheduling algorithm. Understanding time use behaviour is an integral part of any ABM. ABMs are an increasingly popular tool and can be broadly categorised into two groups, namely (i) tour-based models where the home-based tours (trip chains) are considered as the building blocks of the day, and (ii) activity-based models, where single activities are considered the building blocks of the day. In both the tour-based and activity-based models, the analyst needs to estimate the amount of time a person invests in a stop (trip-end) or in an activity episode. Given the potential for the time spent in an activity to be split across multiple non-adjacent episodes (e.g. shopping in the morning and again in the afternoon), it is not sufficient to only know the aggregate amount of time a person spends in an activity type. Therefore, the approach undertaken in the current paper is a step towards evaluating the amount of time spent in an activity episode rather than in an activity type (which might involve multiple episodes per day). Both Garikapati et al. (2014) and Enam and Konduri (2017) tried to model time allocation in such a way that their prediction was suitable for ABMs. However, the frameworks they propose are more aligned with the tour-based approach of ABM and not the activity-based approach of ABM, and their methods are not applicable for finding the time allocation at the activity-episode level which is the focus of the current paper. While providing a step forward, the approach proposed in this paper only predicts the number and duration of activity-episodes, but not their sequence, and it would take an additional scheduling algorithm to make the output of the proposed approach directly applicable for ABM.

Other approaches to deal with the ‘episodic consumption of time’ have been proposed in the literature. Pinjari and Bhat (2010) suggest splitting the day into periods (e.g. night, morning, afternoon and evening) and estimating an MDCEV model for each. While having the benefit of providing a rough schedule or at least a time window inside which each episode is performed, this method imposes fairly arbitrary definitions of the time periods.

More recently, Saxena, Pinjari, and Paleti (2019) suggested an approach which has many similarities with our own. Both approaches predict the number and length of episodes from each activity during a day (or any other period of time), but do not predict their chronological order, i.e. they do not generate a schedule or a sequence. Both models are based on the MDCEV formulation, using activity-episodes as the basic alternative. The main difference is that Saxena, Pinjari, and Paleti (2019) enforce the ordering of the episodes in the model formulation, therefore ensuring that a forecast will never allocate time to episode 2 of an activity unless episode 1 of the same activity has been allocated time to (i.e. episode 2 cannot be performed unless episode 1 has, see Section 2.4 for more details). While desirable, this property comes at the cost of a different, and more complex, likelihood function, requiring specialised estimation and forecasting software (e.g. a custom R or Matlab script). Our approach, on the other hand, does not enforce the order of the episodes, instead assuming independence between all episodes (implications are discussed in Section 2.4). In exchange for this simplifying assumption, our method retains the simpler MDCEV formulation, which is readily available in multiple software packages (Hess and Palma 2019; Lloyd-Smith 2020a).

The objective of the current research is to expand on the activity participation and time allocation research based on MDC formulations with an episode-based framework. The proposed framework can produce time allocation choices at the activity-episode level, shedding light on time use behaviour at a more granular level. Its results can be readily used for trip generation models, as each change from one activity-episode to the next requires the individual to travel. This approach considers an activity-episode to be the basic choice alternative of the MDCEV model, in contrast with an activity type. Additionally, the proposed formulation accounts for the increasingly lower likelihood of performing later episodes of an activity type compared to the first. We demonstrate its potential using two large-scale household travel survey datasets, one from Leeds, UK and the other from the Puget Sound Region (PSR), USA.

The remainder of the paper is organised as follows: Section 2 presents the estimation and forecasting methodology. Section 3 describes the data sources used for our empirical examples and Section 4 presents the estimation and forecasting results. Section 5 provides a summary of the work and concludes the paper.

2. Estimation and forecasting methodology

In this section, we introduce the MDCEV model (Bhat 2008) and describe two modelling approaches to time use data based on this model. First, we discuss the traditional or aggregate approach, used by most time use applications. Second, we describe the episode-based approach which we propose in the present paper.

2.1. The MDCEV model

The MDCEV model is derived from a classical individual utility maximisation problem (individual subscript n has been omitted for clarity), as follows:

$$\begin{array}{rl}{\mathrm{Max}}_{x_n}& \sum _{k=1}^K\frac{{\gamma }_k}{\alpha }{\psi }_k({(\frac{x_k}{{\gamma }_k}+1)}^{\alpha }-1)\\ \mathrm{s}.\mathrm{t}.& \sum _{k=1}^K{x}_k{p}_k=B\end{array}$$ (1)

where $x_k$ is the amount of alternative k consumed (i.e. the time allocated to activity k). The utility function in (1) fulfils the requirements of additive separability and is driven by two different sets of parameters, ${\psi }_k$ and ${\gamma }_k$. The ${\psi }_k$ parameters (one for each choice alternative) represent the marginal base utility of alternative k, while ${\gamma }_k$ relates to the associated level of satiation, with a larger value implying a lower satiation for alternative k, i.e. a higher consumption when chosen. The α parameter is also related to satiation, but is the same across all alternatives. We use this particular parameterisation of the MDCEV model (generic α with alternative-specific ${\gamma }_k$) as it leads to the most efficient forecasting algorithm (Pinjari and Bhat 2011). Consumption is subject to a budget constraint B, expressed either in money or time unit (24 h in the present case), with $p_k$ representing the price per unit of alternative k.

Stochasticity is included in the model through a random error term ${ϵ}_k$ in the base utility of each alternative. Both the base utility ${\psi }_k$ and the satiation parameter ${\gamma }_k$ need to be positive, and can be further parameterised as follows:

$$\begin{array}{rl}{\psi }_k& ={\mathrm{e}}^{{\delta }_k+{\beta }_k{z}_k+{ϵ}_k}\end{array}$$ (2)

$$\begin{array}{rl}{\gamma }_k& ={\theta }_k+{\lambda }_k{z}_k\end{array}$$ (3)

where ${\gamma }_k$ and ${\theta }_k$ are constants for alternative k for the baseline utility and satiation parameters, $z_k$ is a vector of attributes of the alternative and/or characteristics of the individual (e.g. a measure of the activity attractiveness, age of the individual, whether this observation was during a weekend, etc.), and ${\beta }_k$ and ${\lambda }_k$ are estimated parameters capturing the impact of $z_k$. Many implementations of MDCEV use an exponential transformation in the definition of ${\gamma }_k$ to ensure positivity, but we have found that this often leads to slow model convergence and inferior solutions, while unconstrained estimation generally still yields positive estimates. If ${ϵ}_k$ is assumed to follow an independent and identical $Gumbel(0,\sigma )$ distribution across individuals and alternatives, then the following closed form expression for the likelihood of a choice (i.e. time allocation throughout a day) can be derived.

$$L(x_1,x_2,\dots ,x_K)=\frac{1}{p_1{\sigma }^{M-1}}\left(\prod _{k=1}^M{f}_k\right)\left(\sum _{k=1}^M\frac{p_k}{f_k}\right)\frac{\prod _{k=1}^M{\mathrm{e}}^{\frac{V_k}{\sigma }}}{{\left(\sum _{k=1}^K{\mathrm{e}}^{\frac{V_k}{\sigma }}\right)}^M}(M-1)!$$ (4)

where $f_k=\frac{1}{x_k+{\gamma }_k}$ and $V_k={\gamma }_k+{\beta }_k{z}_k+(\alpha -1)\ln (\frac{x_k}{{\gamma }_k}+1)-\ln (p_k)$. Alternatives are ordered in such a way that the first M are consumed. This formulation corresponds to the MDCEV model without an outside good (Bhat 2008). In the case of time use applications, the cost attribute $p_k$ for each alternative is expressed as a single time unit. The fact that the cost attribute is thus the same across alternatives simplifies the equations and makes (4) independent of which alternative is labelled as the first one. In the context of this paper, a time unit corresponds to an hour.

Forecasting with the MDCEV can be done efficiently by using the algorithm proposed by Pinjari and Bhat (2011). Even though this method is proposed for MDCEV models with an outside good, it is easy to generalise it to the case without an outside good, by taking the alternative with the highest base utility (given ${ϵ}_k$) as the first consumed alternative.

2.2. Approach 1: aggregate

The traditional approach to using MDCEV in a time use context disregards the number of episodes of each activity, focusing only on the total amount of time spent performing each activity during a given day. One observation corresponds to 1 day of data, the number of available alternatives in the model is equal to the number of different activities, and M is equal to the number of activities that have a non-zero consumption on a given day. The duration of each activity is the sum of the time spent in all episodes of that activity.

2.3. Approach 2: episode-based

Like in the aggregate approach, one observation of the episode-based approach corresponds to a day of data, but episodes of each activity are no longer aggregated. Instead, multiple episodes of each activity are available. In contrast with the aggregate approach, M is now the number of activity-episodes that are conducted by a given individual on a given day. While in the aggregate approach, the number of available alternatives K is simply equal to the number of activity types, K now depends on the maximum number of episodes that the analyst defines a priori, which must be at least as large as the maximum number of episodes observed in the data for each activity. This leads to a substantial increase in the number of alternatives – now equal to $\sum _k{E}_k$, where $E_k$ is the maximum number of instances in which activity k is performed by anyone in the data (see Figure 1 for an example on how the alternatives are coded). To avoid an excessive number of alternatives, it is theoretically possible to define a maximum number of episodes smaller than the one observed in the database. For example, even though there might be some outliers in the database who perform seven episodes of drop-off/pick-up, an analyst may decide to consider only up to five episodes. In such a situation, a pragmatic approach is for the analyst to aggregate the time spent in episodes five, six and seven, into a single episode. Alternatively, if there is only a small amount of observations with more than five episodes, the analyst could simply drop those observations. If too many observations are in this situation, then the analysts should consider a higher number of episodes in their modelling.

Figure 1.

Example of alternative coding for traditional and episode approach.

The resulting high number of alternatives poses a problem in terms of parameterisation. As time use studies generally use large datasets where participants engage with a potentially large number of episodes of each activity, it quickly becomes infeasible to use different alternative specific constants (ASC) parameters δ for each alternative (activity-episode). However, it is clearly important to allow for variability in the base utility across the episodes of the same activity type, for two key reasons. First, engaging in too many episodes of the same activity is likely to be undesirable as the amount allocated to each episode would become too small to be productive. Second, a large number of non-adjacent episodes would also imply more travel between activities that are geographically not in the same place, where this would then affect the amount of time left for non-travel activities. Additionally, different activities may have different average number of episodes. For example, if we assume that a day starts at 3 AM as in Figure 1, being at home will probably have at least two episodes during the day (morning and evening), while getting petrol would probably be performed no more than once a day. This phenomena requires episode penalties to be different across activities.

The situation is similar when considering the duration (and satiation) of different episodes. For most activities, later episodes will likely be shorter than earlier ones. For example, a third episode of education in the evening will likely be shorter than the previous ones due to fatigue. But behaviour can change across activities. For example, the second at home episode might be longer than the first one, depending on commuting time. Using different θ parameters for each activity-episode could help capture these effects, but again, this is unpractical given the potentially high number of activity-episodes. This point about satiation also explains why, despite there being a disutility associated with engaging in an additional episode, splitting the time for major activities across events might still be beneficial. For example, engaging in two 4-hour episodes of work with a rest in between is probably more desirable than engaging in one 8-hour episode. However, the disutility from satiation needs to be offset against the disutility of conducting additional episodes, and two 4-hour episodes are likely to be preferred to eight 1-hour episodes.

To capture these effects, we use a generic baseline constant for each activity type (δ and θ parameters) and add a polynomial episode penalty to the base utility ψ and satiation γ of each alternative, where the value of this depends on the number of the episode. To avoid identification issues, the first episode of each activity does not have any penalty; instead, the penalties apply from the second episode onwards. Penalties can be used inside both ${\psi }_{ki}$ and ${\gamma }_{ki}$. The analyst needs to decide on what degree of polynomials to include, being mindful that a high degree will provide more flexibility to the penalty, at the cost of a higher number of parameters to estimate, and potential issues with multicollinearity. For example, a second degree polynomial will allow for a parabola-shaped penalty, which can have a single maximum or minimum. A fourth degree polynomial would allow for two local maxima or minima, and so forth. The analyst may decide what is the best degree based on the histogram of episode consumption for each activity (see, e.g., Figures 3 and 4). The equations of the baseline utility and satiation parameters result as follows:

$$\begin{array}{rl}{\psi }_{ki}& =\exp ({\delta }_k+{\beta }_k{z}_k+\sum _{p=1}^{P_{\psi k}}{\pi }_{\psi kp}(i-1{)}^p+{\epsilon }_{ki})\end{array}$$ (5)

$$\begin{array}{rl}{\gamma }_{ki}& ={\theta }_k+{\lambda }_k{z}_k+\sum _{p=1}^{P_{\gamma k}}{\pi }_{\gamma kp}(i-1{)}^p\end{array}$$ (6)

where i enumerates the episode of activity k, $P_{\psi k}$ and $P_{\gamma k}$ represent the number of polynomial terms used for ${\psi }_k$ and ${\gamma }_k$, and ${\pi }_{\psi kp}$ ( $p=1,\dots ,P_{\psi k}$) and ${\pi }_{\gamma kp}$ ( $p=1,\dots ,P_{\gamma k}$) are the associated penalty parameters to be estimated. The penalty is only within activities, i.e. there is no penalty at the day level in terms of episode counts across all activities. Indeed, such a penalty term would require us to model the order between activities as well, which we are not doing.

Figure 3.

Frequency of episode engagement (%) and size of base utility penalty ( $\sum _{p=1}^{p=3}{\pi }_{\psi kp}(i-1{)}^p$ where k, i index activity and episode, respectively) in the Leeds dataset.

Figure 4.

Frequency of episode engagement (%) and size of base utility penalty ( $\sum _{p=1}^{p=3}{\pi }_{\psi kp}(i-1{)}^p$ where k, i index activity and episode, respectively) in the PSR dataset.

To understand the effect of penalties more clearly, consider the case of two activities: at home and get petrol. As the first activity is usually performed twice a day, while the second one is usually performed only once, we expect penalties for getting petrol to be much more negative than for at home. Such values would make a second episode of the getting petrol activity much less likely than a second episode of the at home activity.

Revisiting the comparison between approaches, and using the notation in (5), Saxena, Pinjari, and Paleti (2019) assume that ${\psi }_{ki}>{\psi }_{k(i+1)}$, where ${\psi }_{ki}$ is the marginal utility of episode i of activity k when consumption is zero. These assumptions lead to a closed-form likelihood function that is conditional on the order of episodes. This likelihood is different to the one from a traditional MDCEV, therefore requiring specialised software for model estimation.

2.4. Forecasting

In both approaches, the forecast for each observation is calculated by solving the optimisation problem in (1) multiple times, each time with different values of ${\epsilon }_{ki}$ drawn from the corresponding distribution. The final forecast is the average across solutions for all different ${\epsilon }_{ki}$.

Pinjari and Bhat (2010) propose an efficient algorithm to solve the optimisation problem based on an iterative process. First, the price-normalised baseline marginal utilities ( $\frac{{\psi }_k}{p_k}$) are sorted in descending order of magnitude and one alternative is incrementally added to the consumption set, until the choice set is exhausted or the magnitude of the baseline marginal utility of the next alternative in line becomes less than the Lagrangian multiplier. While the original algorithm is proposed for models with an outside good, it is easy to generalise to a model without it, by assuming that the alternative with the highest price-normalised base utility is consumed, and then proceed to calculate the optimal consumption of the remaining alternatives. More formally, for each observation in the dataset:

1. Draw a complete set of ${\epsilon }_k$ ( $k=1,\dots ,K$) from the appropriate distribution.

2. Sort alternatives in decreasing order of magnitude $r_k=\frac{{\psi }_k}{p_k}$ Let this new ordering be indexed by m, and set M = 1.

3. Compute $\lambda ={\left(\frac{B+\sum _{m=1}^M{p}_m{\gamma }_m}{\sum _{m=1}^M{p}_m{\gamma }_m{r}_k^{\frac{1}{1-\alpha }}}\right)}^{\alpha -1}$

4. If $\lambda \le r_{M+1}$ and M is smaller than the total number of alternatives, then make M = M + 1 and go back to step (3).

5. Calculate optimal consumption as $x_k=({\left(\frac{\lambda }{r_k}\right)}^{\frac{1}{{\alpha }_k-1}}-1){\gamma }_k$.

Both the aggregate and episode-based approaches use exactly the same forecasting algorithm. In the aggregate approach, each activity type constitutes an alternative, while in the episode-based approach an event of an activity type is considered to be an alternative. The only change in the described algorithm is that the index k should be replaced by the composite index ki.

When forecasting with the episode-based approach, nothing a priori forces an individual to choose event i before event i + 1 for a given activity k. For example, episode 2 will be consumed before episode 1 of activity k if ${\epsilon }_{k1}<{\epsilon }_{k2}+\sum _{p=1}^{P_{\psi k}}{\pi }_{\psi kp}$. This is a consequence of all ${\epsilon }_{ki}$ being independent from each other. However, despite the ordering of episode consumption not being guaranteed for a given set of ${\epsilon }_{ki}$ draws, this is respected when averaging across a sufficiently large number of draws. If a large number of ${\epsilon }_{ki}$ sets are used (i.e. if the forecasting algorithm is applied numerous times to each individual using different draws for each occasion) then the mean forecast across these draws will show a decreasing probability of engaging in a higher number of episodes. This is caused by the penalty terms ${\pi }_{\psi ki}$, which make each episode less (more) likely to be consumed if ${\pi }_{\psi ki}$ is negative (positive). Similarly, negative (positive) values of ${\pi }_{\gamma ki}$ will make later episodes more likely to be shorter (longer) in the average forecast. As usual when working with random draws, our advice is to calculate the forecast multiple times, with an increasing number of draws each time, and stop when further increases do not yield a significant change in prediction. The particular number of draws will depend on the dataset being analysed, meaning it is not possible to recommend a specific number.

A limitation when forecasting with the episode-based approach is that the maximum number of possible episodes is defined a priori by the analyst, preventing the model from predicting more than that number of episodes. Despite these limitations, the forecasting algorithm is efficient, and no adaptations are required for it to be applied to the episode-based approach. This allows an analyst to use standard MDCEV software to implement the episode-based approach, such as Apollo (Hess and Palma 2019) or rmdcev (Lloyd-Smith 2020a).

3. Data

In order to demonstrate the proposed approach and compare it to traditional time use modelling using the MDC framework, we use two Revealed Preferences data sources, one collected in Leeds, UK, and the other in the Puget Sound Region (PGS), USA.

3.1. Leeds dataset

The Leeds dataset was collected in 2017 as part of the ERC-funded project ‘DECISIONS’. Time use was only one of the aspects on which the data collection effort was focused, see Calastri and Hess (2020) for more details. The study participants first completed a background survey providing data on their socio-demographics, commuting behaviour, and attitudes. At a later stage, they were asked to install the mobility tracking application rMove (Resource Systems Group 2017) on their smartphones. rMove recorded participants' location for 2 weeks through their phone's GPS. Every time the application detected the end of a trip, it would prompt a short survey asking the participant for the trip purpose, mode, cost (if any), and who else was part of the trip. At the end of each day of tracking, participants saw a summary of all their daily trips, giving them the opportunity to correct or complete the information if there was any error.

A total of 449 respondents successfully completed 2 weeks of tracking, providing full information for at least 95% of all their trips. Most participants lived in the greater Leeds area, yet the sample is not representative of this area's population. Women (58%) and University graduates (69%) are over-represented. Most participants (30%) are between 30 and 39 years old, with under-25 participants representing only 15% of the sample. About 25% reports an income between £20 K and £30 K a year (see Table 1).

Table 1.

Summary of Leeds database: sample socio-demographics.

		Female	Male	Total
Participants		260	189	449
Bachelor degree		182	126	308
Age	18–24	41	26	67
	25–29	29	14	43
	30–39	81	53	134
	40–49	57	36	93
	50–59	41	40	81
	60–65	7	12	19
	66–75	3	8	11
	>75	1	0	1
Personal	Missing	12	11	23
income	<10	43	21	64
(thousands	10–20	65	24	89
of $\text{£}$)	20–30	70	44	114
	30–40	51	44	95
	40–50	12	27	39
	50–75	4	16	20
	>75	3	2	5

On the basis of the recorded trips and their stated purposes, it is possible to construct a daily activity schedule for each participant, which we use to model time use. We aggregated trip purposes into 11 activities: home (i.e. being at home), work (either at main location or elsewhere), leisure or social (e.g. meeting friends, going to the cinema, eating out, etc.), drop-off/pick-up (i.e. driving someone else to their activity location, e.g. taking children to school), exercise (e.g. spending time at a gym), shopping (both maintenance, such as grocery shopping, and non-maintenance, such as leisure shopping), private business (e.g. doctor's appointment), petrol (i.e. buying petrol for a vehicle), and education (e.g. school or university classes). We considered two additional activities: travelling, i.e. travelling to an activity location, and an other/unknown activity, used in the presence of errors in the tracking (e.g. participant did not provide the purpose of a trip or the end location of a trip did not match – within a tolerance – the beginning of the next trip).

Key to our approach is the observation that people can engage in the same activity across multiple episodes throughout a day. The Leeds data contains 28,839 episodes in total for all activities. Among these, at home is the one participants engage with more often and for longer, and the only one in the dataset with an average of more than two daily episodes. Travelling follows as the second activity in terms of number of episodes, but work is the second highest in terms of time spent. Table 2 presents a summary of average daily time use in the Leeds dataset. As the data was collected using geographical tracking, travelling is a pre-requisite to record a new activity. For this reason, travelling is a very common activity in our sample, and we did not split it into episodes as their number would have been perfectly correlated with the total number of episodes in a day (given that a new activity only starts after travelling). We did not disaggregate other/unknown into episodes either, as this activity mostly represents errors in data collection. We simply decided to retain it in the model to make sure that the 24-hour daily budget would be satisfied.

Table 2.

Summary of daily activity engagement and time consumption in the Leeds sample.

		Time spent when		Number of episodes (#)				Length of episodes
		Engaged (h)				When engaged		When engaged (h)
	Fraction of sample who Engage (%)	Average	s.d.	Min.	Max.	Average	s.d.	Average	s.d.
At home	98	15.38	5.57	2	13	2.21	0.91	6.96	5.22
Work	46	6.66	3.30	0	18	1.78	1.12	3.74	3.38
Exercise	17	3.84	4.76	0	7	1.40	0.80	2.74	3.72
Education	4	3.55	2.93	0	5	1.43	0.79	2.49	2.59
Leisure	40	3.25	3.35	0	11	1.66	1.01	1.96	2.56
Other/unknown*	4	3.06	4.71	0	1
Travelling*	91	2.34	2.49	0	1
Drop-off/Pick-up	20	2.20	3.92	0	10	1.68	0.98	1.31	2.77
Private Business	25	1.98	3.28	0	10	1.49	0.88	1.33	2.58
Shopping	34	1.54	2.94	0	9	1.55	0.93	0.99	2.25
Getting petrol	3	0.95	2.78	0	2	1.02	0.14	0.94	2.76

*Engagement not split across episodes.

We limited the number of episodes per activity to five in the Leeds dataset by aggregating subsequent episodes into the fifth one. We chose not to remove observations with more than five episodes per activity, as this would have implied dropping more than 5% of the sample.

Exercise and Drop-off/pick-up exhibit unusually high average daily time allocations in the Leeds dataset, as compared to the PSR dataset. This is probably due to limitations in the data collection, as several participants recorded leisure activities such as hiking and cycling (quite popular around Leeds) as exercise. On the other hand, drop-off/pick-up episodes often include the time of the following activity because if the time taken to drop-off or pick-up was short, the tracking app may have confused it with a short stop on a longer trip.

To compare the different approaches tested in this study, we set aside 20% of the sample and estimate the models with the remaining 80%. This led to 4429 days of data used for estimation, and 1101 days used for forecasting comparison (i.e. validation). We randomly split the full dataset at the individual level, meaning that all observations from a single individual belong to either the estimation or validation sets, but are never spread across both.

3.2. Puget sound region dataset

The Puget Sound Region (PSR) dataset was collected through a household travel survey from four counties (King, Kitsap, Pierce and Snohomish) located in the Puget Sound Region (PSR), Washington State. The survey collected information in a trip diary format, and was collected using two modes: a proprietary smartphone app (Resource Systems Group 2017) that automatically recorded participants' trips; or a more traditional travel diary filled by participants. Assignment to each group depended on participants owning a smartphone capable of running the tracking app. Additionally, the households completed a telephone or online survey recording their socio-demographic characteristics. The survey collected travel patterns (e.g. trip start and end time, origin and destination purpose, transport mode) of the household members on a randomly selected weekday (Tuesday, Wednesday and Thursday) in Spring 2014. A total of 4786 participants from 2419 households participated in the survey. The current study uses information from 3618 participants after filtering based on age (>18).

Table 3 provides a summary of the socio-economic characteristics of the 3618 survey respondents. The sample is slightly skewed in terms of gender – with 54% being male. The sample over-represents highly educated people with 65% of the sample having a bachelor or higher degree. One-third of the sample belong to the 35–54 age group, while 43% of the respondents are older than 55. In terms of household income, 37% of the households have an income of more than $100,000 per year, while 46% of households' income falls between $25,000 and $100,000 per year.

Table 3.

Summary of Puget Sound Region (PSR) database: sample socio-demographics.

		Male	Female	Total
Gender		1951	1667	3618
Bachelor degree		1249	1118	2367
Age	18–24	91	78	169
	25–34	359	316	675
	35–44	314	308	622
	45–54	316	264	580
	55–64	415	343	758
	65–74	309	242	551
	75–84	121	95	216
	>85	26	21	47
Household	Missing			231
income	<25			375
(thousands	25–50			593
of $ per	50–75			543
year)	75–100			541
	>100			1335

The next task in terms of data preparation was to create an activity diary from the trip data collected in the survey. The 16 trip purposes were first re-coded into 13 broad categories – home (i.e. being at home), work (either at main location or elsewhere), shopping (both maintenance shopping, such as grocery shopping, and non-maintenance, such as leisure shopping), education (e.g. day-care, school or university classes), medical (e.g. doctor's appointment), personal-business (e.g. bank, post office), drop-off/pick-up (i.e. driving someone else to their activity location, e.g. taking children to school), exercise (e.g. gym, walk, jog, bike ride), eat-out (e.g. go to restaurant to eat/get take-out), leisure (e.g. attend social event such as visit with friends, family, co-workers, attend recreational event such as movies, sporting event), religious (go to religious/community/volunteer activity), travel (e.g. transfer to another mode of transport such as changing from ferry to bus) and other. Like in the case of the Leeds dataset, we only disaggregated the first 11 activity types into episodes. All the travel and other activities undertaken during the day were aggregated into a single episode. Table 4 presents a summary of time engagement from the PSR sample.

Table 4.

Summary of activity engagement and time consumption in the PSR sample.

		Time spent when		Number of episodes (#)				Length of episodes
		Engaged (h)				When engaged		When engaged (h)
	Fraction of sample who Engage (%)	Average	s.d.	Min.	Max.	Average	s.d.	Average	s.d.
Home	99	15.79	4.65	0	5	2.48	0.82	6.36	3.92
Work	54	8.04	2.55	0	5	1.43	0.76	5.63	3.36
Other*	6	5.68	5.37	0	1
Education	3	4.89	3.18	0	3	1.12	0.36	4.36	2.97
Leisure	17	3.36	3.58	0	5	1.25	0.60	2.69	2.54
Religion	4	2.58	2.01	0	5	1.25	0.71	2.06	1.62
Travel*	100	1.70	1.40	0	1
Medical	10	1.47	1.72	0	3	1.08	0.29	1.36	1.54
Exercise	16	1.20	1.02	0	4	1.12	0.36	1.08	0.96
Eat out	23	0.91	0.85	0	4	1.22	0.50	0.75	0.70
Personal business	22	0.77	1.32	0	5	1.32	0.68	0.58	1.02
Shopping	39	0.71	0.67	0	5	1.49	0.82	0.48	0.46
Drop-off/pick-up	13	0.41	0.56	0	5	1.51	0.71	0.27	0.43

*Engagement not split across episodes.

As in the case of the Leeds dataset, we limited the number of episodes per activity to five, this time by dropping observations with more than five episodes of a single activity, as these cases constituted less than 5% of the sample.

Like in the Leeds sample, almost 100% of respondents participate in home and travel on the survey day, while 53% and 4% of the sample participate in work and in education activity, meaning increases by 7 and 1 percentage points, respectively, compared to the Leeds sample. Another considerable difference noted in the PSR sample is that while in Leeds only 33% of respondents engage in shopping on the survey day, 40% do so in the PSR sample.

In the PSR sample, respondents spend around 16 hours at home per day on average. Average aggregate duration for work is around 8 hours, which is 1.3 hours higher than in the Leeds sample. Similarly, the average aggregate duration for education is around 1.3 hours higher in the PSR sample (5 hours) compared to the Leeds sample (3.7 hours). Discretionary activities are generally shorter compared to the Leeds sample. For example, the average (aggregated across all episode) shopping activity duration in the PSR sample is only about 40 minutes, while the average duration in the Leeds sample is almost 50 minutes higher. Similarly, the average aggregate travel duration in the PSR sample is around 1.7 hours, which is almost 40 minutes shorter than the average in the Leeds sample.

In terms of episodes, people tend to engage in an average 2.5 episodes of the home activity during the day, which is consistent with the out-of-home activities splitting the home activity into at least two episodes. Other than home, an average of around 1.5 episodes per day is noted for work, shopping, pick-up/drop-off and personal-business activity – for the rest of the activities an average closer to 1 episode is more probable.

As with the Leeds data, 80% of the PSR sample (3000 observations) is used for model estimation and 20% ( 750 observations) was set aside for the validation of the model estimation and forecasting routines.

Figure 2 shows the average length of each episode (when conducted) in both datasets. The average length of episodes decreases monotonically only among a few activities: exercise, private business and work in the Leeds dataset, and exercise, religious and work in the PSR dataset. The Leeds dataset exhibits peaks in the fifth episode for several activities, due to the aggregation of later episodes into the fifth one.

Figure 2.

Average duration of episodes when engaged in the Leeds (left) and PSR (right) datasets, in hours. Episodes are ordered in a clockwise fashion.

4. Results

In this section, we present results from the proposed episode-based approach for both the Leeds and PSR datasets. We compare them against MDCEV models using the traditional aggregate approach. We begin by comparing the model parameters, followed by model fit (using the aggregate Root Mean Squared Error, RMSE), and finish with an analysis of the predicted episode numbers as compared to the observed ones.

4.1. Model parameters

The detailed parameter estimates for the Leeds models are shown in Table 5, while those for the PR models are shown in Table 6.

Table 5.

Parameter estimates of aggregate and episode-based approach of the Leeds models (robust t-ratios).

		Aggregate approach				Episode-based approach
		Base utility (ψ)		Satiation (γ)		Base utility (ψ)		Satiation (γ)
		Coeff.	t-ratio	Coeff.	t-ratio	Coeff.	t-ratio	Coeff.	t-ratio
Drop-off/Pick-up	Intercept	−3.116	(−19.17)	0.282	(7.19)	−3.062	(−28.37)	0.136	(7.88)
	Weekend			0.242	(2.00)			0.116	(2.60)
	${\pi }_1$							−0.955	(−16.40)
Work	Intercept	−1.773	(−11.93)	4.456	(20.01)	−1.766	(−22.38)	2.076	(13.04)
	Weekend	−2.175	(−13.65)			−2.057	(−11.61)
	${\pi }_1$					−1.108	(−20.39)	−0.947	(−5.51)
	${\pi }_2$					0.070	(4.76)
Education	Intercept	0.147	(0.20)	4.456	(20.01)	0.639	(0.87)	1.549	(6.86)
	Weekend	−2.254	(−5.09)			−2.536	(−5.69)
	Age	−1.076	(−5.01)			−1.248	(−5.43)
	Income					−1.015	(−5.04)
	${\pi }_1$					−1.274	(−7.18)
Shopping	Intercept	−2.888	(−18.89)	0.364	(17.38)	−2.900	(−30.15)	0.230	(20.62)
	Weekend	0.672	(9.18)			0.786	(9.64)
	Female	0.293	(3.19)			0.312	(2.93)
	${\pi }_1$					−1.311	(−24.29)
	${\pi }_2$					0.068	(3.69)
Private business	Intercept	−3.154	(−19.96)	0.511	(12.04)	−3.102	(−30.45)	0.282	(13.02)
	Weekend	0.463	(5.82)			0.493	(6.08)
	Female	0.246	(2.16)			0.229	(1.67)
	${\pi }_1$					−1.308	(−18.34)
	${\pi }_2$					0.074	(3.33)
Get petrol	Intercept	−5.376	(−24.00)	0.091	(4.04)	−5.321	(−31.08)	0.093	(3.94)
	Weekend	0.527	(2.76)			0.518	(2.69)
	${\pi }_2$					−3.755	(−6.93)
Leisure	Intercept	−2.618	(−17.04)	1.455	(20.10)	−2.611	(−30.06)	0.678	(16.29)
	Weekend	0.973	(16.14)			1.126	(17.16)
	Female	0.143	(1.69)			0.109	(1.17)
	${\pi }_1$					−1.215	(−28.71)	0.116	(2.29)
	${\pi }_2$					0.069	(5.17)
Exercise	Intercept	−3.642	(−22.31)	1.736	(9.49)	−3.656	(−32.26)	1.332	(8.97)
	Weekend	0.712	(8.18)	0.464	(1.50)	0.889	(8.60)	−0.203	(−1.43)
	${\pi }_1$					−1.495	(−14.20)	−0.696	(−5.57)
	${\pi }_2$					0.123	(3.93)
At home	Intercept	0.000	(fixed)	2.030	(6.40)	0.000	(fixed)	0.845	(15.99)
	Weekend	−0.146	(−1.04)	2.839	(3.87)	0.204	(3.94)	0.684	(8.04)
	Female	0.118	(1.68)			0.136	(1.90)
	${\pi }_1$					−1.367	(−38.91)	1.577	(15.53)
Travel	Intercept	0.666	(4.36)	0.115	(13.40)	0.869	(9.26)	0.098	(14.32)
	Full time			0.024	(3.56)			0.022	(4.02)
Other	Intercept	−4.840	(−25.98)	0.612	(2.83)	−4.792	(−34.54)	0.612	(2.83)
Loglikelihood		−41,499.86				−79,633.72
Number of parameters		40				57

Table 6.

Parameter estimates of aggregate and episode-based approach of the PSR models (robust t-ratio).

		Aggregate approach				Episode-based approach
		Base utility (ψ)		Satiation (γ)		Base utility (ψ)		Satiation (γ)
		Coeff.	t-ratio	Coeff.	t-ratio	Coeff.	t-ratio	Coeff.	t-ratio
Home	Intercept	−5.040	(−4.92)	0.530	(3.21)	−13.090	(−116.26)	0.607	(33.7)
	Age 18–34	−0.62	(−6.75)
	Age 35–54			−0.72	(−7.93)
	Age 55–74	−0.500	(−5.34)
	Age >74					0.294	(3.07)
	Income 50–100 k$			−0.05	(−1.18)
	Income >100 k$
	${\pi }_1$					0.118	(1.89)
	${\pi }_2$					−0.716	(−12.64)
	${\pi }_3$					0.099	(9.08)
Work	Intercept	−10.70	(−9.97)	2.04	(55.31)	−14.630	(−118.77)	1.39	(39.98)
	Age 18–34	3.060	(8.39)
	Age 35–54	3.050	(8.4)			0.503	(8.78)
	Age 55–74	2.220	(6.09)
	Age >74					−2.272	(−8.4)
	Male	−0.27	(−5.25)			−0.233	(−4.11)
	Income 50–100 k$	0.110	(1.44)	−0.140	(−2.24)
	${\pi }_1$					−1.586	(−27.55)
	${\pi }_2$					0.090	(3.86)
Shopping	Intercept	−7.710	(−7.57)	−1.110	(−21.29)	−15.120	(−106.27)	−1.307	(−53.65)
	Age 18–34	−1.110	(−7.93)
	Age 35–54	−0.970	(−7.25)	−0.09	(−1.46)	−0.336	(−4.52)
	Age 55–74	−0.440	(−3.47)
	Age >74					0.647	(3.71)
	Male	0.160	(2.33)	0.15	(2.55)	0.258	(3.82)
	Income 50–100 k$	−0.20	(−2.57)	0.09	(1.41)
	${\pi }_1$					−1.300	(−21.35)
	${\pi }_2$					0.038	(1.61)
Education	Intercept	−11.80	(−11.41)	1.44	(7.71)	−17.600	(−91.7)	1.52	(9.96)
	Age 18–34	1.74	(6.93)	0.560	(2.79)
	Age 35–54	0.00				−0.914	(−3.18)
	Age 55–74	−0.86	(−2.06)
	Income 50–100 k$			−0.340	(−1.71)
	${\pi }_1$					−2.26	(−5.18)
Medical	Intercept	−9.11	(−8.87)	−0.05	(−0.77)	−16.91	(−97.18)	−0.03	(−0.5)
	Age 18–34	−1.77	(−7.59)
	Age 35–54	−1.33	(−6.76)	0.20	(1.71)
	Age 55–74	−0.83	(−4.52)
	Age >74					0.93	(4.24)
	Male	0.32	(2.58)			0.28	(1.95)
	Income 50–100 k$	−0.23	(−1.66)
	${\pi }_1$					−2.66	(−12.1)
Personal	Intercept	−8.48	(−8.31)	−1.23	(−23.12)	−15.97	(−117.25)	−1.47	(32.21)
Business	Age 18–34	−1.43	(−8.42)
	Age 35–54	−1.05	(−6.88)			−0.23	(−2.26)
	Age 55–74	−0.55	(−3.76)
	Age >74					0.74	(4.29)
	Male	0.23	(2.74)			0.23	(2.44)
	${\pi }_1$					−2.66	(−12.1)
Pick-up/	Intercept	−10.31	(−10.14)	−1.49	(−12.21)	−17.23	(−98.49)	−1.85	(−47.29)
Drop-off	Age 35–54	0.77	(7.19)			0.89	(7.92)
	Male	0.42	(3.83)	0.12	(1.14)	0.51	(4.11)
	Income 50–100 k$			−0.15	(−1.11)
	Income >100 k$			−0.20	(−1.68)
	${\pi }_1$					−0.71	(−10.78)
	${\pi }_2$					0.11	(6.42)
Exercise	Intercept	−9.25	(−9)	−0.58	(−4.73)	−16.16	(−124.79)	−0.31	(−5.77)
	Age 18–34	−0.39	(−1.94)
	Age 35–54	−0.49	(−2.49)	0.16	(1.32)	−0.11	(−1.02)
	Age 55–74	−0.25	(−1.31)	0.33	(2.9)
	Male	0.10	(1.07)
	Income 50–100 k$			0.26	(2.07)
	Income >100 k$			0.41	(3.52)
	${\pi }_1$					−1.97	(−15.59)
Eat out	Intercept	−9.11	(−9.01)	−0.83	(−7.97)	−15.62	(−127.89)	−0.85	(−19.3)
	Age 35–54			−0.23	(−2.03)	−0.22	(−2.36)
	Income 50–100 k$			0.25	(2.16)
	Income >100 k$			0.36	(3.36)
	${\pi }_1$					−1.79	(−23.34)
Leisure	Intercept	−9.05	(−8.86)	0.63	(7.37)	−16.12	(−111.29)	0.66	(12.65)
	Age 18–34	−0.69	(−4.02)
	Age 35–54	−0.91	(−5.3)			−0.40	(−3.57)
	Age 55–74	−0.41	(−2.53)	0.20	(1.96)
	Male	0.33	(3.5)			0.24	(2.25)
	Income 50–100 k$			0.27	(2.4)
	Income >100 k$			0.24	(1.98)
	${\pi }_1$					−1.59	(−17.15)
Religious	Intercept	−10.04	(−9.64)	0.90	(11.49)	−17.97	(−86.28)	0.55	(5.88)
	Age 18–34	−1.66	(−4.88)
	Age 35–54	−1.46	(−4.87)
	Age 55–74	−0.76	(−2.84)
	Age >74					1.13	(2.95)
	Male	0.26	(1.36)			0.51	(2.24)
	${\pi }_1$					−1.41	(−6.36)
Travel	Intercept	0.00	(fixed)	−7.56	(−7.48)	0.00	(fixed)	−14.17	(−128.4)
Other	Intercept	−10.54	(−10.32)	1.35	(6.47)	−17.18	(−120.67)	1.41	(7.21)
	Age 18–34	−0.30	(−1.54)
	Age 35–54	−0.28	(−1.62)
	Male	0.28	(1.8)
	Income >100 k$			0.59	(1.74)
Loglikelihood		−27,665.9				−51,957.0
Number of parameters		89				62

As Table 5 shows, coefficients signs and magnitudes are consistent across the aggregate and episode-based approach for the Leeds data. Travelling is the most popular activity (ceteris paribus) according to both the aggregate and episode-based model, followed by education and work. This reflects in these activities having the highest constants in their base utilities. In both models, participants are less likely to engage in work, education and other activities during the weekends, and instead are more likely to engage in shopping, private business, getting petrol, leisure and exercise activities during this period. Being at home is not significantly influenced by the weekend according to the aggregate approach, but it is instead positively influenced by it according to the episode-based approach. Older participants are less likely to engage in education activities according to both models, while there are other consistent effects of sex and income across both approaches. Concerning satiation, work exhibits the highest intercept in both models, meaning that -ceteris paribus- people spend more time working, compared to other activities. The overall effect, however, is also influenced by the base utility of the alternative, explaining why home is the activity consumed for the longest time (see Table 2). Satiation parameters are less influenced by the participants' characteristics, with just drop-off/pick-up, exercise, home and travel showing significant effects of covariates, all of which are consistent (or not significant) across the aggregate and episode-based approaches.

We included three penalty terms in the base utility of each alternative (i.e. $P_{\psi k}=3,\mathrm{\forall }k$), but progressively removed all non-significant parameters. We chose a third degree polynomial as a compromise between flexibility and parsimony. As Figure 3 shows, all remaining penalties have a net negative effect on the base utility of alternatives. As these only influence the base utility (ψ) from the second episode onwards (see Equation 5), we can conclude that the objective of making later episodes less likely to take place is achieved by our functional form. As expected, getting petrol is the activity whose penalty becomes negative most quickly, because the vast majority of participants perform at most one episode of this activity a day. Instead, at home grows much slower, to make multiple episodes of the activity more likely.

We initially also included three penalty terms in the satiation effect of each alternative, but few of them reached significance, leading us to only retain linear penalties in the model. We observe that work and exercise have negative penalty parameters, meaning that later episodes of these activities tend to be shorter. Leisure, on the other hand, has a positive penalty, meaning that later episodes tend to be longer than previous ones, implying that later episodes are usually performed during the evening, when individuals have more time to spend in recreational activities. These results are consistent with the average duration of episodes described in Figure 2.

According to the aggregate PSR model, home is the most likely activity for people engage to in, followed by shopping and personal business. The episode model also identifies home as the most likely activity to perform, but places work as the second. This difference is probably due to the different covariates retained in each model, based on their significance. Men are less likely than women to participate in work and more likely to participate in shopping, medical, personal business, pick-up/drop-off , leisure and religious activities according to both the aggregate and episode model. As expected, people above 75 years of age are less likely to participate in work according to both models. Yet only the episode model points to them being more likely to participate in medical and shopping activities. Income has no impact on the episode model, while it does significantly influence shopping (base utility), work, exercise, eating out and leisure (satiation) in the aggregate model. The constants of the satiation parameters exhibit the same sign for most activities in both the aggregate and in the episode-based model. Work has the highest positive value of the satiation constant followed by education, indicating people's propensity to spend longer in these activities when they perform them. Home and leisure have very similar magnitudes for their satiation parameter ${\theta }_k$.

As Table 6 shows, the significance of covariates varies strongly between the aggregate and episode-based approaches. By removing non-significant parameters, we end up with very different explanatory variables in the aggregate and episode-based approaches. Just by examining the model estimates, it is not possible to establish which of the two models reflects reality in a more truthful way. We can only judge them by the accuracy of their forecasts, which we measure in Section 4.2. Both approaches have similar precision at the aggregate level, but the episodes-based approach provides increased detail. This leads us to believe the episode-based approach to be a more reliable representation of individuals' behaviour.

As Figure 4 shows, all activities have negative values for their respective penalty terms except for home. The positive linear penalty for home indicates that this activity is more likely to be split into two episodes than in one episode as is the case for the other activity types (recall that the penalty term is applied starting from the second episode of an activity type). This is in line with the observed statistics and indicates the polynomial penalty terms were able to replicate the episode participation propensity of the individuals. Work and education have very negative penalties indicating these activities are more likely to be participated in one than in more episodes. On the other hand, the much lower magnitude of the negative penalty terms in the shopping and personal business activity indicate that many people are likely to participate in multiple episodes of these activities during the day, compared to participating in multiple episodes of work and education activities.

The log-likelihood of the aggregate and episode-based models is not comparable. While the aggregate approach in the Leeds sample has a final log-likelihood of −41,500, the episode-based approach peaks at −79,634. Similarly, for the PSR sample, the likelihood peaks at −27,665 for the aggregate model and at −51,957 for the episode model. The difference is due to the episode-based approach having to explain the allocation to more alternatives, meaning that the log-likelihood becomes more negative. McFadden's ${\rho }^2=1-\frac{{LL}^{\ast }}{{LL}_b}$ is not comparable across approaches either, as the base model ${LL}_b$ used for comparison is not well defined in the case of the MDCEV model. For logit models, ${LL}_b$ is usually the ‘null’ or ‘equiprobable’ model with all parameters set to zero, but setting all parameters to zero in the MDCEV model is not possible, as all satiation parameters must be positive. And while ${LL}_b$ could be defined as a model with constants only, it is not clear how many constants should be included in the episode-based formulation. Therefore, we assess the goodness of fit by comparing each model performance when forecasting out-of-sample, as described in the next section. This measure has the benefit of not being based on the model LL, not requiring the definition of a base model ${LL}_b$, and being a direct measure of a model's forecasting capabilities.

4.2. Forecast fit comparison

To measure the forecasting accuracy of both the aggregate and episode-based approaches, we estimated the model with 80% of the whole sample, and then used that model to forecast for the remaining 20% of the data i.e. the holdout sample. All fit measurements presented in this and the following sections are based on the holdout sample only. We measured the fit using the Root Mean Squared Error (RMSE) at the sample level, which we defined as follows:

$$RMSE=\sqrt{\frac{1}{K}\sum _k{(\sum _n\sum _t\sum _i{x}_{ntki}-\sum _n\sum _t\sum _i{\hat{x}}_{ntki})}^2}$$ (7)

where ${\hat{x}}_{ntki}$ is the forecasted time allocation to episode i of activity k for observation t from individual n, with the observed values given by $x_{ntki}$, and K is the total number of activity-types.

Table 7 presents forecast and fit indices for the Leeds sample. Under the ‘Time (hours)’ heading, we present the observed and predicted aggregated consumption of each activity. The forecasts are similar for the aggregate and episode-based approach, but with the second achieving a 15% smaller RMSE. This is an important improvement over the aggregate approach, achieved mostly due to better fit on the most popular activities (work, home, leisure). Such large improvement, however, is not observed in the PSR data set, so it may be dataset-specific. Under the ‘Activities (obs)’ heading, we present the observed and predicted number of participants during a day (observations) that engage in each activity, i.e. the number of observations that perform at least one episode of the corresponding activity. Once again, we see that the forecast is very similar between the aggregated and episode-based approaches, with the episode-based approach having a slightly (4%) smaller RMSE. Under the ‘Episodes (epi.)’ heading, we present the observed and predicted total number of episodes in the whole sample. As the aggregate approach cannot predict more than one episode, this heading does not apply to it. The episode-based approach achieves a RMSE of 128, with an average 17% error in its prediction.

Table 7.

Forecast fit comparison in the Leeds sample.

	Time (hours)			Activities (obs)			Episodes (epi.)
		Forecast			Forecast			Forecast
	Obs	Agg.	Epi.	Obs	Agg.	Epi.	Obs	Epi.
Drop-off/Pick-up	372	250	231	210	189	277	350	309
Work	3304	3609	3263	488	450	548	814	733
Education	194	193	199	52	39	50	77	54
Shopping	642	481	485	398	311	396	656	459
Private Business	418	434	400	264	234	309	368	344
Petrol	3	11	12	31	26	26	31	26
Leisure	1357	1636	1542	424	393	501	722	618
Exercise	1016	603	554	249	148	191	335	202
Home	16808	15742	16247	1088	1051	1073	2377	2061
Travel	2205	3395	3421	996	1023	1045	996	1045
Other	105	69	69	46	38	37	46	37
TOTAL	26424	26424	26424	4246	3901	4452	6772	5887
RMSE (sample)		517	436		47	45		128

We observe a similar pattern in the PSR sample, as presented in Table 8. The aggregate and episode-based approach achieve very similar fit in terms of aggregated time consumption (column ‘Time’) and activity engagement (column ‘Activities’) with the exception that the episode-based approach is performing better while predicting the number of individuals engaging in different activities (unlike the Leeds data). The episode-based reaches an RMSE of 152 when predicting the number of episodes of each activity in the whole sample, with an average 28% error per activity.

Table 8.

Forecast fit comparison in the PSR sample.

	Time (h)			Activities (obs)			Episodes (epi.)
		Forecast			Forecast			Forecast
	Obs	Agg.	Epi.	Obs	Agg.	Epi.	Obs	Epi.
Home	11,361	10,489	10,481	714	705	693	1117	1323
Work	2928	2837	2836	356	305	353	492	398
Shopping	207	359	387	296	224	268	459	310
Education	145	114	117	27	17	19	34	19
Medical	105	145	141	83	55	55	88	55
Personal business	134	144	148	154	119	138	223	148
Drop-off/pick-up	42	65	70	83	71	94	129	100
Exercise	135	209	201	122	86	91	135	93
Eat Out	151	229	233	163	125	139	193	145
Leisure	523	433	444	139	95	108	167	112
Religious	92	100	106	36	23	27	45	27
Travel	1274	2019	1987	724	725	725	724	725
Other	257	210	201	40	34	32	40	32
TOTAL/ Budget	17,354	17,354	17,354	2937	2583	2742	4506	3487
RMSE (sample)		353	350		37	21		152

4.3. Episodes forecast analysis

In this section, we analyse the results from the episode-based approach in more detail, in particular its prediction of the number and duration of episodes. As the aggregate approach can only forecast a single episode per activity, we ignore it in this section. We begin by analysing the results from the Leeds dataset.

Tables 9 and 10 present, under the ‘Total time (hours) per episode’ column, the observed and predicted total time spent in each episode for each activity, from the first to the fifth episode. We observe that the total amount of time spent in each activity across the whole sample is decreasing with the order of the episodes, a phenomenon reproduced by our modelling.

Table 9.

Detailed episode forecasting in the Leeds sample.

	Time (hours) per episode										Observations per episode
	Observed					Forecasted					Observed					Forecasted
Episode:	1st	2nd	3rd	4th	5th	1st	2nd	3rd	4th	5th	1	2	3	4	5	1	2	3	4	5
Drop-off/Pick-up	248	54	40	5	26	142	56	22	8	3	120	60	17	6	7	246	29	1	0	0
Work	2476	538	192	57	42	2272	578	239	114	61	297	109	45	21	16	388	136	22	1	0
Education	142	37	5	9	0	144	39	12	3	1	35	12	2	3	0	46	4	0	0	0
Shopping	456	102	43	22	19	335	99	33	13	5	253	80	33	16	16	337	55	4	0	0
Private Business	308	65	29	7	10	274	81	28	11	5	201	37	15	7	4	277	31	1	0	0
Petrol	3	0	0	0	0	12	0	0	0	0	31	0	0	0	0	26	0	0	0	0
Leisure	863	303	113	37	41	964	361	134	56	27	247	106	38	16	17	395	94	10	1	0
Exercise	794	143	67	11	0	452	66	22	9	5	191	36	17	4	1	180	11	0	0	0
Home	9256	5218	1819	392	123	9069	5247	1450	383	97	226	536	245	61	20	323	531	202	17	0
Travel	2205					3421					996					1045
Other	105					69					46					37
RMSE (sample)						394	37	125	21	15						80	19	23	18	12

Table 10.

Detailed episode forecasting in the PSR sample.

	Time (hours) per episode										Observations per episode
	Observed					Forecasted					Observed					Forecasted
Episode:	1st	2nd	3rd	4th	5th	1st	2nd	3rd	4th	5th	1	2	3	4	5	1	2	3	4	5
Home	4485	4314	1912	471	179	5537	3601	1057	222	65	39	391	203	58	23	248	279	149	17	0
Work	2464	370	76	18	0	2148	497	132	42	16	257	71	20	7	1	309	42	1	0	0
Shopping	143	39	17	6	2	273	79	24	8	3	194	60	27	11	4	229	37	2	0	0
Education	124	18	3	0	0	107	9	1	0	0	21	5	1	0	0	19	0	0	0	0
Medical	99	4	1	0	0	132	9	1	0	0	79	3	1	0	0	54	0	0	0	0
Personal business	91	29	9	6	0	108	30	8	2	1	108	28	14	3	1	129	9	0	0	0
Drop-off/pick-up	27	12	2	1	0	39	23	7	1	1	48	27	5	3	0	87	6	0	0	0
Exercise	121	14	1	0	0	169	28	4	1	0	111	10	0	1	0	89	2	0	0	0
Eat out	122	24	2	2	0	194	33	5	1	0	140	18	3	2	0	133	6	0	0	0
Leisure	419	80	11	12	0	350	74	16	3	1	118	16	3	2	0	105	3	0	0	0
Religious	78	13	0	1	0	86	16	3	1	0	29	6	0	1	0	27	0	0	0	0
Travel	1274					1987					724	0	0	0	0	725
Other	257					201					40	0	0	0	0	32
RMSE (sample)						399	219	258	75	35						68	37	19	13	7

While the RMSE of the total time expenditure is higher for the first episode in both samples, this is only a scale effect. If we look at the RMSE as a percentage of the average duration of each episode across activities, we obtain 26, 5, 49, 35, and 51% for the first, second, third, fourth, and fifth episodes in the Leeds sample, and 53, 49, 140, 160, and 210% in the PSR sample. This points to larger mean errors for sparsely consumed episodes or, in other words, the model predicts less accurately for those activity-episodes that are less common in the sample.

The effect of the penalty is perhaps clearer when the predicted number of episodes is analysed. In the ‘Observations per episode’ columns in Tables 9 and 10, we present the observed and predicted number of individuals during a day (observations) performing one, two, three, four or five episodes for each activity. In this case, we did not consider the order in which the episodes were performed in the forecast, but only the total number of episodes. This is due to our forecasting algorithm not enforcing the order in which the episodes should be engaged with, as discussed in Section 2.4. To calculate these numbers, we register for each set of draws ${ϵ}_{ki}$ the number of episodes an individual performs. We then calculate the frequency of engaging in one, two, three, four or five episodes across all draws, which is our estimate for the probabilities of an individual engaging in each possible number of episodes. Finally, we obtain the expected number of individuals performing each number of episodes by summing these probabilities across individuals.

The (expected) number of individuals conducting each number of episodes confirms that the penalty parameterisation works as expected. In both the Leeds and PSR samples we observe that most individuals engage in two episodes of the home activity. On the other hand, no individuals engage in more than one episode of getting petrol in the Leeds sample, just as in the observed data. Similarly, education, medical and religious activities are only performed once a day in the PSR sample.

Once again, we observe that the RMSE of the ‘Observations per episode’ forecast decreases with the number of episodes, but again this is just a scale effect. If we calculate the ratio between these RMSE values and the average number of people engaging in each number of episodes, we obtain 33%, 17%, 50%, 120 % and 131% for the first, second, third, fourth and fifth episodes in the Leeds sample, and 46%, 64%, 77%, 163% and 263% for the PSR sample. In other words, the earlier episodes are predicted more accurately than the later ones. This is because our data contains many observations with a low number of episodes being performed, and limited observations with many episodes.

5. Discussion

In this paper, we propose a framework to enrich the MDCEV family of models with a new tool to model time use data. In particular, instead of modelling the total amount of time allocated to each activity across a whole day (or any other unit of time), we propose to model the duration of each instance or episode of the performed activities. In this framework, an episode is a continuous amount of time during which an individual engages in a given activity. There can be several episodes of the same activity within a single day, e.g. working in the morning, then performing another activity, then working again in the afternoon.

Describing and predicting time use at the episode level can provide valuable information on the number of trips performed during a day, as different activities need to be performed at different locations. This is a result that could not be inferred from activity-level time allocation alone. Furthermore, information about the episode duration is relevant as it informs the level of satiation for different activities. This could be important when planning the provision of services or understanding preferences for time use.

Our approach consists of creating multiple alternatives per activity, representing unique episodes. In terms of parameterisation, all alternatives belonging to the same activity share the same parameters measuring the impact of individual and activity characteristics on time use. At the same time, polynomial penalties are used to differentiate between the utilities of different episodes of the same activity type. When forecasting, the efficient algorithm proposed by Pinjari and Bhat (2011) can be applied, just as with the MDCEV model.

Our results indicate that the proposed episode-based approach to time use modelling is an improvement over current practice using the MDCEV model. While it does not improve the fit of the aggregate consumption as compared to a traditional MDCEV model, it does provide additional information in the form of the number of episodes each individual is likely to engage in. Such new insights do not impose additional burden in data collection, as most time use datasets are constructed from individuals' diaries recording their schedule. As a result, coding the information into aggregate time consumption per activity, or disaggregated time consumption across several episodes does not imply additional costs, other than additional data management. In other words, our approach provides new key information at marginally higher cost.

Nevertheless, we acknowledge two main limitations of the present framework. The first and most relevant one is that the current formulation does not enforce the orderly performance of episodes when forecasting. The episode-based formulation proposed in this paper only modifies the deterministic part of the base utility of each alternative, while keeping its stochastic part the same as in a traditional MDCEV model. This has the benefit of estimation and forecasting being the same as in the traditional MDCEV model, but it also assumes the error components ( ${\epsilon }_{ki}$) of each alternative to be independent, even across episodes of the same alternative. This can be problematic when simulating choices using an episode-based approach. When simulating, a single set of draws of the error terms is generated. It is possible that those draws lead to the base utility of later episodes being larger than that of earlier episodes of an alternative ( ${\psi }_{ki}<{\psi }_{k(i+1)}$), and thus to later episodes being consumed while earlier episodes are not (e.g. episode 2 is consumed while episode 1 is not). While this is mostly a theoretical issue (and a practical one in simulation), it is not a problem for common forecasting, because the forecast is obtained as the average over multiple sets of draws. When averaging, the deterministic penalty terms dominate over the stochastic error terms, effectively enforcing the ordered consumption of episodes. In cases where the model is used for simulation, the labelled order of the consumed episodes should be ignored, instead focusing only on the number of episodes consumed.

The episode-based approach does not consider individuals' overall schedule, instead looking at episodic consumption in a simultaneous way. It is reasonable to believe there might be scheduling effects across activities (see, e.g. Timmermans, Arentze, and Joh 2002; Wets et al. 2000; Allahviranloo, Regue, and Recker 2017). For example, if an individual has engaged in many episodes throughout the day, he or she might be more inclined to limit the number of episodes in the evening. Or, if a drop-off episode happened early in the day, another pick-up episode is likely to happen later in the day. However, including scheduling in the formulation of the problem would inevitably lead to an integer optimisation problem, and to a substantial increase in complexity. Instead, our approach seeks to be as efficient and as simple as possible. If scheduling is needed (e.g. for applications to activity based modelling), this can easily be achieved at a later stage with an additional algorithm.

While the present work represents an effort into improving the realism of our time use models, other elements could of course be incorporated to capture the full complexity of human behaviour. Activity engagement throughout the day is also known to be affected by social interactions. For example, it is likely that many activities are planned at the household level, and not independently by each individual (Timmermans, Arentze, and Joh 2002; Arentze and Timmermans 2009). These kinds of interactions are not included in the proposed modelling approach, though some of their effects could be captured by introducing correlations between the base utility of alternatives across individuals of the same household.

Further refinements to the episodes-based approach are possible. Especially in the presence of longer panels (such as the two-week Leeds data), a mixed MDCEV approach, i.e. one incorporating random heterogeneity across individuals, would be able to capture correlations across days for the same individual in terms of the frequency of conducting different activities.

In summary, the proposed episode-based modelling approach extends and furthers the methodology in time use research. This approach is capable of offering additional information at virtually no additional cost compared to the traditional time use modelling approach. This extra information can be key in understanding people's preferences and behaviour, and furthermore, it can more accurately predict the total number of trips during a day, in addition to the overall time expenditure. The approach can be applied to datasets with any number of activities and episodes, as its parametrisation does not lead to an explosion of parameters in situations with a high number of alternatives or episodes. Furthermore, the approach can be applied using any software capable of estimating MDCEV models, as it does not require any modification to the estimation or forecasting algorithm.

Funding Statement

The Leeds authors acknowledge the financial support by the European Research Council through the consolidator Grant 615596-DECISIONS.

Disclosure statement

No potential conflict of interest was reported by the author(s).