Eliminating Recall &amp; Proxy Bias in Household Travel Surveys: Sample Correction Methodology under the Core-Satellite Fusion Paradigm

doi:10.21203/rs.3.rs-4832044/v1

Download PDF

Research Article

Eliminating Recall & Proxy Bias in Household Travel Surveys: Sample Correction Methodology under the Core-Satellite Fusion Paradigm

https://doi.org/10.21203/rs.3.rs-4832044/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Travel diary is one of the fundamental methods for collecting data critical for transportation planning, demand modelling, and analyses. Self-reported household travel surveys are known for recall and proxy biases. Both biases lead to underreporting of travel demand in the dataset. On the other hand, travel diaries collected through GPS-based methods are exempted from recall and proxy biases. Thus, this study will investigate the recall and proxy bias in self-reported travel surveys and propose correction procedures. The investigation and correction will be conducted under the core-satellite paradigm of urban passenger travel surveys under the core-satellite survey design paradigm. The study uses the Transportation Tomorrow survey (TTS) in the Greater Toronto and Hamilton area (GTHA) as the core dataset. The Google Timeline Travel Survey (GTTS) will be used as the satellite survey. The direct comparison between the core and satellite surveys shows that respondents who participated in both TTS and GTTS demonstrated bias-free diary reporting in the self-reported core survey. However, significant recall and proxy bias are identified for individuals who directly participate in the TTS but refuse to participate in GTTS and the proxy respondents in the TTS. Then, this study proposes a correction procedure based on respondents’ willingness-to-participate estimation for satellite surveys, dataset filtration, and re-weighting techniques. The effectiveness of the correction procedure is empirically demonstrated in this study.

Household travel survey

GPS-based survey

core-satellite paradigm

travel dairy

recall-bias

proxy-response bias

Travel diary is one of the most important and widespread methods to collect data critical for transportation modelling and demand analysis. Over time, travel survey methods have evolved dramatically. Initially, travel surveys were conducted physically through pen and paper and mail-back forms (Stopher, 1992). Then, as technology progressed, surveys could be administered through the computer-assisted telephone interviewing (CATI) method (Axhausen et al., 2002). After the internet became widely available, surveys shifted to the computer-assisted web interviewing (CAWI) method (Adler et al., 2002). All these methods require each respondent to recall their trips in the form of a self-reported travel diary. The latest trip diary reporting technology uses global positioning systems (GPS) and cellular-based methods. Using this technology, travel diaries are passively logged while each respondent conducts their daily activities (Murakami & Wagner, 1999; Chen et al., 2010; Kelly et al., 2013; Shen & Stopher, 2014; Li et al., 2023).

Household travel surveys using self-reported methods (e.g., CATI and CAWI) suffered from recall and proxy biases. Both biases lead to underreporting of travel demand in the dataset. Recall bias occurs when survey participants tend to forget the trips they have made (Stopher & Greaves, 2007). Proxy bias arises in household travel surveys that collect responses from a single household member on behalf of others. This method assumes that the respondent fully knows all household members’ travel activities, which could often be unrealistic. This assumption leads to proxy biases, resulting in the under-reporting of trips made by other household members (Morency, 2015; Weir et al., 2011). On the other hand, GPS-based diary collection methods are exempted from recall and proxy biases by passively collecting travel behaviour data (Stopher & Greaves, 2007)

Thus, this study will investigate the recall and proxy bias in self-reported travel surveys and propose correction procedures. The investigation and correction will be conducted under the core-satellite paradigm of urban passenger travel surveys (Goulias et al., 2013; Habib et al., 2018). The core-satellite paradigm considers large-scale household travel surveys as the core dataset and uses satellite surveys targeting specialized purposes to supplement the core. The study uses the Transportation Tomorrow survey (TTS), which is a regional household travel survey collected in the Greater Toronto and Hamilton area (GTHA), as the core dataset (Transportation Tomorrow survey, 2024). Meanwhile, the Google Timeline Travel Survey (GTTS) will be used as the satellite survey. The GTTS is a GPS-based travel survey in the GTHA (Li et al., 2023). Data collected from the self-reported TTS survey will be compared with GTTS. Then, a correction procedure will be developed based on identifying biases in trip rates in the core dataset using adjustment techniques founded on common variables in each dataset.

The remainder of the paper is outlined as follows. Section 2 presents literature reviews on survey biases. Section 3 reports descriptive statistics for the datasets used. Section 4 presents the results of comparing the core and satellite surveys. Section 5 presents the procedure to correct the core survey based on empirical analysis results. Finally, Section 6 concludes the study by stating key findings and plans for future work.

A significant area of travel survey research focuses on identifying and mitigating biases. The four main types of biases are sampling bias, non-coverage bias, non-response bias, and measurement bias (Son et al., 2013). This study focuses explicitly on recall and proxy bias within the measurement bias spectrum. Recall bias occurs from respondents’ inability to accurately remember and report past travel activities. Rylander et al. (1995) assessed recall bias in a travel spending survey, finding that recall errors were related to trip complexity as much as to the passage of time. Bell et al. (2019) explored the use of smartphones to mitigate recall bias in household surveys. The study involved 500 individuals in rural Bangladesh who responded to survey components at different frequencies (weekly, monthly, or seasonally). Results showed that frequent data collection reduced recall decay.

Sfeir et al. (2023) examined recall bias in a household travel survey of public transit users. They matched individuals between the Danish National Travel Survey (TU) and the Danish Smart Card system (Rejsekort). They reported a median time reporting error of 11.34 minutes. They also found that males and individuals without fixed schedules were generally less accurate in reporting travel times. Some studies have attempted to address this issue. For instance, Spurr et al. (2014) compared travel demand data from the Montreal household travel survey with smart card data from the Montreal subway system to correct travel demand estimates. The analysis identified biases based on socioeconomic parameters, leading to a re-weighting of the demand data.

At the same time, proxy biases have been recognized as a significant concern in household travel surveys (Ahmadreza et al. 2018; Morency 2015; Wargelin & Kostyniuk 2006). The error arises when survey participants have imperfect knowledge of other household members’ travel behaviours. Badoe & Steuart (2002) found statistically significant proxy bias in the 1996 TTS, and the bias was higher for non-home-based trips. Morency et al. (2015) compared trip rates between proxy and primary respondents across five household travel surveys conducted in Montreal from 1987 to 2008. They found that proxy bias was lower for surveys with a smaller household size. This indicates smaller households might communicate better while reporting for household travel surveys. Maruyama et al. (2019) also modelled the likelihood of individuals participating in travel surveys as primary or proxy respondents. They found a statistically significant difference in reported trip rates between these groups. Wang et al. (2021) also found proxy bias in the 2016 TTS and proposed a hybrid data fusion methodology for correcting travel survey data.

Recently, a growing number of studies have examined the feasibility of using GPS-based surveys to supplement self-reported travel surveys. Wolf et al. (2003) provide an initial look into this procedure. Further studies on the usage of GPS tracking technology were undertaken by Bricka and Bhat (2006), Wolf (2006), Stopher et al. (2007), Bellemans et al. (2008), and Stopher and Greaves (2009). Several works focused on comparing GPS-based surveys to supplement self-reported travel surveys (Zhao et al., 2015; Patterson & Fitzsimmons, 2016; Stopher et al., 2018; and Hong et al., 2021). Their findings collectively confirmed that GPS-based surveys better minimized measurement errors and should be used to supplement and correct self-reported surveys.

This study utilizes two survey datasets: one core dataset from a self-reported survey and one satellite dataset from a GPS-based survey. The self-reported is the 2022 Transportation Tomorrow Survey (TTS), the regional household travel survey conducted in the Greater Toronto and Hamilton Area (GTHA) every five years since 1986. It is one of the most extensive household travel surveys worldwide, surveying around 5% of households in GTHA (Data Management Group., 2018). The dataset is taken from the latest Fall 2022 cycle of the TTS. The TTS records a self-reported travel diary from the primary and proxy respondents for a weekday using CAWI and CATI modes. Overall, the Fall 2022 cycle collected 32,138 primary and 41,261 proxy respondents.

The satellite dataset is the Google Timeline Travel Survey (GTTS) conducted in the GTHA (Li et al., 2023). The GTTS collected a whole week of travel dairy for each participant. The technical details of GTTS can be found in Li et al. (2023). The main cycle of GTTS was administered during Fall 2023. The main cycle was recruited from the email list of all participants in the 2022 TTS. After data cleaning, the GTTS collected 769 samples, directly participating in the 2022 TTS as primary respondents. This group provides the foundation for directly comparing the core and satellite datasets in this study.

Descriptive statistics of respondents in the 2022 TTS and 2023 GTTS can be found in Table 1. Comparing respondents’ socioeconomic characteristics, the TTS contained a significant portion of older respondents, especially those over 65. The GTTS has a much larger share of younger respondents under 40. This could be attributed to the complexity of responding to the GPS-based survey. The distributions between the two surveys were very similar regarding household size and dwelling types. The TTS observed a slightly larger share of individuals who were not employed individuals, while the GTTS had more full-time employees. This aligns with the discrepancy in age distribution. Lastly, driver’s license, transit pass ownership, and vehicle number are similar in the two datasets.

Table 1

Descriptive statistics of primary participants of TTS and GTTS
Attributes	2022 TTS	2023 GTTS
Gender
Male	57.2%	59.9%
Female	41.6%	38.2%
Other	1.2%	1.8%
Age
18–29	7.2%	12.5%
30–39	18.9%	30.0%
40–49	17.3%	22.8%
50–64	28.9%	23.1%
65+	27.3%	11.6%
Household Size
1	27.4%	22.1%
2	40.1%	39.5%
3	15.6%	16.0%
4	12.3%	15.2%
5+	4.6%	7.2%
Dwelling Type
House	54.6%	47.6%
Townhouse	10.7%	11.4%
Apartment	32.3%	40.6%
Other	2.5%	0.4%
Employment Type
Employed full-time ( > = 30 hours per week)	60.9%	71.4%
Employed part-time (< 30 hours per week)	7.3%	8.6%
Not employed	31.8%	20.0%
Workplace Arrangement
Work at a usual workplace regularly	74.6%	78.6%
Work exclusively from home	17.0%	18.5%
No fixed workplace address	8.3%	2.9%
Drivers License Ownership
Yes	93.5%	93.0%
No	6.4%	6.6%
Don’t know/prefer not to say	0.1%	0.4%
Transit Pass Ownership
Yes	8.7%	6.6%
No	91.2%	93.4%
Don’t know/prefer not to say	0.1%	0.0%
Number of vehicles
0	22.0%	15.4%
1	45.9%	45.6%
2	26.0%	30.4%
3	5.1%	6.3%
4	0.7%	1.7%
5+	0.4%	0.6%

Figure 1 summarizes the trip rates per person per day observed for primary and proxy respondents in the TTS and primary respondents in the GTTS. The TTS recorded lower trip rates across all modes compared to the GTTS. For TTS primary respondents, there is a 14.2% dairy trip difference between self-reported and GPS-based surveys. The difference increases to 42.9% for active modes such as walking and cycling. The discrepancy is more substantial for proxy respondents. In TTS, the global trip rate for proxy respondents is underreported by 38.7% compared to the GTTS.

Interestingly, the automobile trip rate for proxy respondents is unanimously lower than that reported for primary respondents. The difference is close to 1 trip between primary and proxy respondents. This strongly indicates that the difference is caused by imperfect knowledge of proxy respondents’ travel behaviours, and the trips were missed for proxy respondents in TTS.

This section will conduct individual-level comparisons of trip rates between the core and satellite surveys. Two groups of respondents will be analyzed. First, a direct comparison will be conducted for individuals who participated in the TTS and GTTS. Then, an indirect comparison will be performed for primary respondents in the TTS who refused to participate in the satellite survey and proxy respondents in the TTS.

4.1 Direct comparison

The most straightforward comparison is the difference in trip rates between respondents who participated in both TTS and GTTS surveys. Figure 2 presents the procedure used to compare the core and satellite datasets directly. Given the sampling strategy discussed in Section 3.0, all GTTS respondents in this study participated in the TTS. Thus, trip rates can be directly compared between individuals who participated in both surveys to reflect the core survey’s recall bias. Each GTTS respondent is matched with their corresponding observation from the TTS, and differences in trip rates for automobile, transit, and active modes and global trip rates are calculated.

The TTS collected travel diary for a typical weekday. However, the GTTS collected the travel diary for an entire week. Therefore, for direct comparison, the trip rate in TTS is compared with the weekday average and day-specific trip rate in the GTTS. By keeping the average day assumption of the TTS survey, the recorded trip rate for each TTS respondent may be considered their average rate for the whole week. The weekday average approach calculated the weekly average for each respondent in the GTTS and compared it with the same individual in the TTS.

On the other hand, the day-specific approach directly compares trip rates in the GTTS for the same day of the week that the individual is recorded in the TTS. This would lead to a weekday-specific trip rate difference instead.

Moreover, the trip rate difference distribution is also weighted to account for the variation in the composition across the weekdays in the TTS. The day-of-week composition in TTS is 0.139, 0.186, 0.188, 0.240, and 0.240 from Monday to Friday, respectively. The composition increases throughout the week and peaks on Thursday and Friday. However, the GTTS evenly represents day-of-week, given the nature of weeklong travel diaries. Therefore, weighting factors are calculated based on this distribution. Weekdays with fewer observations will be weighted higher than those with more observations. Using this approach, two types of trip rate differences can be evaluated: the naïve difference, which is the raw difference between the two datasets without any adjustments, and the weighted difference, which accounts for the weighting factors.

Figure 3 plots distributions of the naive and weighted direct differences between the core and satellite surveys. As evident from the analysis, direct trip rate differences for respondents between the two surveys appear to be randomly distributed and centred around a median of close to zero. The weekday difference for global trip rate is -0.092 and − 0.077 between TTS and GTTS for naïve and weighted approaches, respectively. This suggests a trivial level of under-reporting for the self-reported core survey. This points to the conclusion that travel diaries collected in the self-reported core by the subset of respondents who completed both surveys are reasonably accurate for all travel modes and free from recall bias.

4.2 Indirect comparison

The next step should be a further in-depth comparison of trip rate differences in TTS respondents who did not participate in the GTTS. Figure 4 presents the indirect core-to-satellite survey respondent comparison procedure. A method is needed to match primary TTS respondents who did not complete the GTTS (termed the TTS-only) and proxy TTS respondents (termed the TTS-proxy) with GTTS records. The K-Nearest Neighbors (K-NN) algorithm is used for this analysis, employing the Nearest Neighbors module from the scikit-learn Python package (Pedregosa et al., 2018). In principle, K-NN classifies or predicts the value of a new data point based on the majority class or average value of its k-nearest neighbours in the training data.

In K-NN, all available data points, along with the labels for classification that correspond to each, are stored by the algorithm during the training phase. This study uses the sample of GTTS participants as the training set. The variables include all socioeconomic variables collected for each respondent. Then, during the prediction phase, K-NN calculates the distance between each TTS data point and every GTTS point based on socioeconomic parameters. Once distances are calculated, the algorithm identifies the k-nearest neighbours from the GTTS training dataset to each data point in the TTS-only and TTS-proxy datasets based on shortest distances. The parameter k is a user-specified value that denotes the number of neighbours for each data point. The list of k-neighbors represents the GTTS respondents’ most similar records in the TTS, enabling comparisons between them. In practice, the magnitude of parameter k must be balanced. A smaller k value means fewer neighbours, which is quicker to compute and may match better but is more affected by noise. Conversely, a larger k value reduces the noise effect but decreases the algorithm’s accuracy and increases computation time. This study determines that a k-value of five is appropriate.

Figure 5 shows the distributions of trip rate differences, revealing apparent under-reporting of trips across all modes for both TTS-only and TTS-proxy respondents. The average difference in global trip rates is -0.61 for TTS-only respondents and − 1.31 for TTS-proxy respondents. This indicates that measurement bias in the TTS arises from two leading causes. First, the bias stems from respondents’ reluctance to participate in travel surveys, as shown by the data quality differences between those willing to participate in a follow-up satellite survey and those who only participate in the TTS. Second, proxy responses in the TTS result in significant under-reporting. The empirical investigation indicates primary respondents do not have full knowledge of other household members’ travel behaviour, which contradicts the underlying assumption in proxy reporting.

Interestingly, the trip rate difference distribution for transit and active modes appears to have two distinct respondent behaviours. The distribution in the core survey simultaneously composites with recognizable underreporting and overreporting behaviour. In-depth analysis is then undertaken to identify which cohorts of respondents skew towards underreporting and which toward overreporting. Figure 6 compares key socioeconomic variables for underreporting and overreporting respondents in the core survey. The underreporting behaviour in the core survey for TTS-only and TTS-proxy respondents appears random. This is indicated by their dwelling type and age distribution closely matching the entire population in the core survey. However, significant socioeconomic attribute shifts can be identified for respondents who systematically overreport their transit and active trips in the core survey. These would be individuals who reside in apartments and individuals under the age of 39. It is evident that these socioeconomic classes overreport their transit and active trips. Further work may be pursued to investigate their motivation to overreport and find specific adjustment methods that are outside the scope of this study.

The above analysis demonstrates that respondents willing to participate in a satellite survey are exempted from measurement errors in the self-reported core survey. Therefore, this section will develop a general framework to identify and remove erroneous samples to improve the data quality in the core survey. Figure 7 presents the schematic of the proposed correction method. The procedure uses respondents’ willingness-to-participate estimation, dataset filtration, and re-weighting techniques. First, a sample selection model will be developed based on respondents’ willingness to participate in the follow-up survey to identify samples with higher reporting quality in the core survey. The estimated core and satellite participation likelihood (referred to hereafter as the willingness-to-participate values) can be extracted as the similarity between TTS and GTTS respondents. A larger willingness-to-participate value indicates a greater similarity to core-and-satellite participants. This value can shed insight into the correlation between willingness-to-participate and trip-reporting behaviours in the core survey. Afterward, erroneous samples will be removed from the core survey. The cut-off criterion is if the estimated willingness-to-participate is lower than the actual satellite survey participation market share. Finally, the filtered core survey will be re-weighted to restore the socioeconomic representation disturbed by the selection process.

5.1 Binary logit model sample selection model

The binary logit model generates core respondents’ likelihood to participate in the follow-up survey. Wang et al. (2022) developed a similar sample selection model using the binary probit model to analyze respondents’ willingness to participate in a follow-up survey after the Nation Household Travel Survey (NHTS) in the US. Thus, this study selects binary logit, which could produce results similar to binary probit but has parsimonious model formulation (Ben-Akiva & Lerman, 2018). The dependent variable is the probability that the individual in the TTS will complete the GTTS.

The model estimation results are presented in Table 2. Overall, the final model contains parameters corresponding to the most significant socioeconomic variables in determining willingness-to-participate for each respondent, including dwelling type, household size, participant age and gender, number of household vehicles, and transit pass ownership. The magnitudes and signs of each coefficient may be interpreted to determine the effect of each factor on the resulting willingness-to-participate estimate. Regarding dwelling type, respondents who live in an apartment are more likely to be similar to GTTS respondents, while respondents who live in other dwellings are less likely. Household size has a significant positive effect, meaning that respondents with larger households are more likely to participate in GTTS. Age parameters are coded as six different cohort-based dummy variables interacting with continuous ages within each cohort. Younger individuals are generally more likely to participate in GTTS. Individuals older than 70 are less likely to participate in the satellite survey. Lastly, having more vehicles and transit passes decreases the likelihood of participation in the satellite survey.

Table 2

Binary logit willingness-to-participate model
	Coefficient	z-stats
Constant
Intercept	-4.41	-21.72
Dwelling type
House	-	-
Townhouse	-0.0008	-0.01
Apartment	0.220	2.32
Other	-0.551	-1.52
Household size	0.106	3.08
Cohort-based age
18–29	0.049	5.20
30–39	0.037	5.35
40–49	0.024	4.21
50–59	0.016	3.46
60–69	0.008	1.89
70+	-0.0005	-0.18
Gender
Male	-	-
Female	-0.145	-1.99
Non-binary	0.211	0.57
Prefer not to answer	0.216	0.36
Number of household vehicles	-0.168	-3.17
Transit pass ownership
Yes	-0.268	-1.92
No	-	-
Number of observations	32138
Log-likelihood of full model	-3747.10
Log-likelihood of constant-only model	-3901.06
Rho-square value against constant-only model	0.0395

Then, the relationship between estimated willingness-to-participate and reported trip rates in the core survey is examined. Figure 8 presents the plots and the results of linear best-fit-line analysis. The positive coefficient empirically confirms the positive correlation between higher willingness-to-participate and higher trip rates in the core survey. Next, a cut-off point for willingness to participate can be established. TTS-only and TTS-proxy participants will be filtered to include only those individuals whose willingness-to-participate estimates are higher than this threshold. The cut-off point for willingness-to-participate is 2.62%, the portion of individuals in the original TTS that completed the GTTS satellite. Thus, a willingness-to-participate value of 0.0262 corresponds to around the 56th percentile of model estimates for the TTS-only respondents and the 57th percentile of estimates of the TTS-proxy set. Therefore, around 40% of the records in the core survey will be discarded due to data quality issues.

5.2 Re-weighting procedure

The filtered records in the core survey are not homogenous; thus, the distributions of socioeconomic attributes found in the filtered dataset no longer match those of the original data. Therefore, each individual must calculate weighting factors to restore the population representation in the core dataset. The iterative proportional fitting (IPF) procedure is used to compute weighting factors. The IPF is a reliable technique to adjust the cell values of a contingency table to match known marginal sums. This method is widely used in many fields. It is often chosen to ensure that data is consistent with external constraints or marginal distributions due to its simplicity and stability (Wang et al., 2021b). For this study, the algorithm will be utilized on two dimensions based on socioeconomic attributes to be balanced. In the case of the filtered TTS, the socioeconomic attributes to be balanced are selected based on the magnitude of differences between the filtered and original datasets. The most significant of these are found to be age and household size. Figure 9 reports the weighting factors, the intermediate distributions of the filtered datasets, and the final distributions of the weighted core survey. Weighting factors are normalized so that the lowest factor has a unit magnitude 1.00. The mean value for TTS-only and TTS-proxy is 5.02 and 2.84, respectively, which fall into a reasonable range.

5.3 Corrected trip rates in the core survey

Table 3 shows the trip rates in the core survey after being corrected by the procedure described in the previous section. Overall, the correction procedure mitigates the underreporting issue brought by the recall and proxy bias in the core dataset. Trip rates for TTS-only and TTS-proxy are comparable with the unbiased trip rates in the GTTS. The global difference has been reduced to approximately 0.08 trips for TTS-only respondents. The individual mode differences are also much closer than before.

However, the correction procedure appeared less effective for proxy respondents without the re-weighting procedure. The weighting procedure substantially increased the trip rate for TTS-proxy respondents from 2.39 to 4.15. However, caution should be exercised when interpreting the results. The weighting procedure might produce certain degrees of overreporting. Specific socioeconomic cohorts, such as those over 65, lack observation in the GTTS. Therefore, the willingness-to-participate sample selection model significantly removes this age cohort.

Consequently, the IFP algorithm has to assign large weighting factors for aged individuals. This issue must be fixed by increasing the observation of aged individuals in future GPS-based satellite surveys. Nonetheless, the correction procedure proposed in this study has substantially improved the travel demand in the unweighted core survey. The corrected core dataset can be used to produce unbiased results in travel demand modelling practices.

Table 3

Corrected trip rates in the core survey
	GTTS	TTS-only	TTS-proxy
Global
unweighted	3.23	3.15	2.39
weighted		3.88	4.15
Automobile
unweighted	1.84	1.72	1.45
weighted		1.43	2.47
Transit
unweighted	0.37	0.50	0.36
weighted		0.79	1.41
Active
unweighted	0.91	0.94	0.58
weighted		1.66	0.41

This study presents an in-depth analysis of recall and proxy bias in a self-reported household travel survey. The study uses the 2022 Transportation Tomorrow Survey (TTS) and the 2023 Google Timeline Travel Survey (GTTS). TTS is a self-reported survey using CAWI and CATI modes. The GTTS is a GPS-based survey, so the travel diaries are passively collected. Under the core-satellite paradigm, the TTS is the core dataset, and the GTTS is the satellite dataset. The GTTS was recruited directly from individuals who directly participated in the TTS. The comparison between the core and satellite shows that respondents who participated in both TTS and GTTS demonstrated bias-free diary reporting in the self-reported core survey. The direct comparison reveals that the trip rate differences are close to normally distributed with zero means between the core and satellite survey. However, significant recall and proxy bias are identified for individuals who directly participate in the TTS but refuse to participate in GTTS (TTS-only) and proxy respondents in the TTS (GTTS-proxy). The average difference in global trip rates is -0.61 for TTS-only respondents and − 1.31 for TTS-proxy respondents.

After identifying the sources of measurement bias, this study also proposes procedures to correct the core survey. The method uses respondents’ willingness-to-participate in the follow-up survey, dataset filtration, and re-weighting techniques. First, a sample selection model is estimated to predict respondents’ willingness to participate in the follow-up survey. The step will indicate the erroneous level for each record in the core survey. Afterward, erroneous samples will be removed from the core survey. Finally, the filtered core survey will be re-weighted to restore the socioeconomic representation disturbed by the selection process. The effectiveness of the correction procedure is demonstrated through empirical exercise in this study.

The results of this study demonstrate that, under specific conditions, self-reported travel surveys can generate data quality comparable to GPS-based surveys. Future self-reported household travel surveys should balance sample size and data quality (Habib et al., 2020). Empirical evidence clearly shows that individuals willing to participate in follow-up surveys provide substantially higher quality data. Therefore, large-scale web-based travel surveys should introduce a registration survey to identify willing respondents. The actual survey should be introduced as a pseudo-follow-up to the registration survey. This approach can help to screen a significant portion of low-quality respondents, thereby reducing measurement errors in the travel diary. Also, proxy respondents in household travel surveys report erroneous activity-travel behaviours. Proxy-respond travel diaries might be helpful in meta-level regional-level origin-destination (OD) travel analysis. However, they should not be used for modelling exercises such as activity scheduling and trip-chaining, as the data significantly suffered from underreport issues by almost one trip on a typical weekday.

There are several limitations that future research could address. First, the sample selection model in this study is parsimonious. Future research should explore the use of sophisticated modelling techniques, such as machine learning models, which have superior prediction performance (Wang et al., 2020; Liu et al., 2023). This could improve the overall performance of the correction procedure by capturing more heterogeneous survey participation behaviours. Additionally, as discussed earlier, the IPF weighting does not provide desirable results due to the lack of observations for elderly individuals in the satellite survey. Future GPS-based satellite surveys should specifically consider older people as a hard-to-reach group and aim to improve their representation. The correction procedure described in this study should be re-evaluated with a better-represented satellite dataset to assess performance improvements.

Author Contribution

Author Contribution StatementThe authors confirm their contribution to the paper as follows: Study conception and design: K. Wang, M. Li, K.M.N. Habib; Data collection: K. Wang, M. Li; Analysis and interpretation of results: K. Wang, M. Li; Draft manuscript preparation: M. Li, K. Wang, K.M.N. Habib; Overall project supervision: K.M.N. Habib. All authors reviewed the results and approved the final version of the manuscript.

Acknowledgment

An NSERC Discovery Grant funded the study. The authors are responsible for all results, interpretations, and comments the paper reports.

Adler, T., Rimmer, L., & Carpenter, D. (2002). Use of internet-based household travel diary survey instrument. Transportation Research Record: Journal of the Transportation Research Board, 1804(1), 134–143. https://doi.org/10.3141/1804-18
Axhausen, K. W., Zimmermann, A., Schönfelder, S., Rindsfüser, G., & Haupt, T. (2002). Observing the rhythms of daily life: A six-week travel diary. Transportation, 29(2), 95–124. https://doi.org/10.1023/a:1014247822322
Ahmadreza, F., Vaughan J., Miller EJ. (2018). Draft: Analysis of TTS 2016 Trip Rates. Retrieved July 2, 2020
Badoe, D. A., & Steuart, G. N. (2002). Impact of interviewing by proxy in travel survey conducted by telephone. Journal of Advanced Transportation, 36(1), 43–62. https://doi.org/10.1002/atr.5670360104
Bell, A., Ward, P., Tamal, M. E. H., & Killilea, M. (2019). Assessing recall bias and measurement error in high-frequency social data collection for human-environment research. Population and Environment, 40(3), 325–345. https://doi.org/10.1007/s11111-019-0314-1
Ben-Akiva, M., & Lerman, S. R. (2018). Discrete choice analysis: Theory and Application to Travel Demand. MIT Press.
Chen, C., Gong, H., Lawson, C., & Bialostozky, E. (2010). Evaluating the feasibility of a passive travel survey collection in a complex urban environment: Lessons learned from the new york city case study. Transportation Research Part A: Policy and Practice, 44(10), 830–840. https://doi.org/10.1016/j.tra.2010.08.004
Data Management Group. (2018). 2016 TTS: Data Guide (Issue February). http://dmg.utoronto.ca/pdf/tts/2016/2016TTS_DataGuide.pdf
Goulias, K. G., Pendyala, R. M., & Bhat, C. R. (2013). Keynote — Total Design Data Needs for the New Generation Large-Scale Activity Microsimulation Models. Transport Survey Methods: Best Practice for Decision Making, 21-46. doi:10.1108/9781781902882-002
Hong, S., Zhao, F., Livshits, V., Gershenfeld, S., Santos, J., & Ben-Akiva, M. (2021). Insights on data quality from a large-scale application of smartphone-based travel survey technology in the Phoenix Metropolitan Area, Arizona, USA. Transportation Research Part A: Policy and Practice, 154, 413–429. https://doi.org/10.1016/j.tra.2021.10.002
Habib, K.M.N., Miller, E.J., Srikukenthrian, S., Lee-Gosselin, M., Morency, C., Roorda, M., & Shalaby, A. (2018). TTS2.0: A research and development (R&D) project on passenger travel survey methods. Retrieved July 03, 2020, from https://www.sciencedirect.com/science/article/pii/S2352146518301558
Habib, K.M.N., El-Assi, W., & Lin, T. (2020). How Large is too Large? A Review of the Issues related to Sample Size Requirements of Regional Household Travel Surveys with a Case Study on the Greater Toronto and Hamilton Area (GTHA). arXiv.org. https://arxiv.org/abs/2005.00563
Kelly, P., Krenn, P., Titze, S., Stopher, P., & Foster, C. (2013). Quantifying the difference between self-reported and global positioning systems-measured journey durations: A systematic review. Transport Reviews, 33(4), 443–459. https://doi.org/10.1080/01441647.2013.815288
Liu, Y., Loa, P., Wang, K., & Habib, K.M.N. (2023). Theory-driven or data-driven? Modelling ride-sourcing mode choices using integrated choice and latent variable model and multi-task learning deep neural networks. Journal of Choice Modelling, 48, 100431. https://doi.org/10.1016/j.jocm.2023.100431
Li, M., Wang, K., Liu, Y., & Habib, K.M.N. (2023). Deriving Weeklong Activity-Travel Dairy from Google Location History: Survey Tool Development and A Field Test in Toronto. arXiv.org. https://arxiv.org/abs/2311.10210
Montini, L., Prost, S., Schrammel, J., Rieser-Schüssler, N., & Axhausen, K. W. (2015). Comparison of travel diaries generated from smartphone data and dedicated GPS devices. Transportation Research Procedia, 11, 227–241. https://doi.org/10.1016/j.trpro.2015.12.020
Murakami, E., & Wagner, D. P. (1999). Can using Global Positioning System (GPS) improve trip reporting? Transportation Research Part C: Emerging Technologies, 7(2–3), 149–165. https://doi.org/10.1016/s0968-090x(99)00017-0
Morency, C. (2015). What about Proxy Respondent bias Over time? - cirrelt.ca. Retrieved July 2, 2020, from https://www.cirrelt.ca/documentstravail/cirrelt-2015-55.pdf
Maruyama, T., Hosotani, K., & Kawano, T. (2019). Inferring proxy response in household travel surveys with unknown completer using a group-based choice model. Transportation, 0123456789. https://doi.org/10.1007/s11116-019-10053-3
Patterson, Z., & Fitzsimmons, K. (2016). DataMobile: Smartphone Travel Survey Experiment. Transportation Research Record: Journal of the Transportation Research Board, 2594(1), 35–43. https://doi.org/10.3141/2594-07
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Müller, A., Nothman, J., Louppe, G., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, É. (2018, June 5). Scikit-Learn: Machine learning in Python. arXiv.org. https://arxiv.org/abs/1201.0490
Rylander, R. G., Propst, D. B., & Mcmurtry, T. R. (1995). Non-response and recall biases in a survey of traveler spending. Journal of Travel Research, 33(4), 39–45. https://doi.org/10.1177/004728759503300406
Shen, L., & Stopher, P. R. (2014). Review of GPS travel survey and GPS data-processing methods. Transport Reviews, 34(3), 316–334. https://doi.org/10.1080/01441647.2014.903530
Spurr, T., Chapleau, R., & Piché, D. (2014). Use of subway smart card transactions for the discovery and partial correction of travel survey bias. Transportation Research Record: Journal of the Transportation Research Board, 2405(1), 57–67. https://doi.org/10.3141/2405-08
Stopher, P. R. (1992). Use of an activity-based diary to collect household travel data. Transportation, 19(2), 159–176. https://doi.org/10.1007/bf02132836
Stopher, P. R., Daigler, V., & Griffith, S. (2018). Smartphone app versus GPS Logger: A Comparative Study. Transportation Research Procedia, 32, 135–145. https://doi.org/10.1016/j.trpro.2018.10.026
Stopher, P. R., & Greaves, S. P. (2007). Household travel surveys: Where are we going? Transportation Research Part a Policy and Practice, 41(5), 367–381. https://doi.org/10.1016/j.tra.2006.09.005
Son, S., Khattak, A., Wang, X., Agnello, P., & Chen, J. (2013). Quantifying key errors in household travel surveys. Transportation Research Record Journal of the Transportation Research Board, 2354(1), 9–18. https://doi.org/10.3141/2354-02
Srikukenthiran, S., Loa, P., Hossain, S., Chung, B., Habib, K.N.M., & Miller, EJ. (2018). TTS 2.0 Research Report. Retrieved July 2, 2020, from https://uttri.utoronto.ca/files/2020/04/UTTRI-Report-TTS-2.0-Srikukenthiran-2018.pdf
Sfeir, G., Rodrigues, F., Zeid, M. A., & Pereira, F. C. (2023). Analyzing the reporting error of public transport trips in the Danish National Travel Survey using smart card data. arXiv.org. https://arxiv.org/abs/2308.01198
Transportation Tomorrow survey. (2024). Transportation Tomorrow Survey Shape tomorrow’s commute today. https://www.tts2023.ca/en/
Wang, K., Hossain, S., & Habib, K.M.N. (2021). A hybrid data fusion methodology for household travel surveys to reduce proxy biases and under-representation of specific sub-group of Population. Transportation, 49(6), 1801–1836. https://doi.org/10.1007/s11116-021-10228-x
Wang, K., Liu, Y., Mashrur, S. M., Loa, P., & Habib, K.M.N. (2021b). COVid-19 influenced households’ Interrupted Travel Schedules (COVHITS) survey: Lessons from the fall 2020 cycle. Transport Policy, 112, 43–62. https://doi.org/10.1016/j.tranpol.2021.08.009
Wang, S., Wang, Q., & Zhao, J. (2020). Deep neural networks for choice analysis: Extracting complete economic information for interpretation. Transportation Research Part C Emerging Technologies, 118, 102701. https://doi.org/10.1016/j.trc.2020.102701
Wang, X., Shaw, F. A., Mokhtarian, P. L., & Watkins, K. E. (2022). Response willingness in consecutive travel surveys: an investigation based on the National Household Travel Survey using a sample selection model. Transportation, 50(6), 2339–2373. https://doi.org/10.1007/s11116-022-10312-w
Weir, D., Faul, J., & Langa, K. (2011). Proxy interviews and bias in the distribution of cognitive abilities due to non-response in longitudinal studies: a comparison of HRS and ELSA. Longitudinal and life course studies, 2(2), 170–184. https://doi.org/10.14301/llcs.v2i2.116
Wolf, J., Loechl, M., Thompson, M., & Arce, C. (2003). Trip Rate Analysis in GPS-enhanced personal travel surveys. Transport Survey Quality and Innovation, 483–498. https://doi.org/10.1108/9781786359551-028
Zhao, F., Pereira, F. C., Ball, R., Kim, Y., Han, Y., Zegras, C., & Ben-Akiva, M. (2015). Exploratory analysis of a smartphone-based travel survey in Singapore. Transportation Research Record: Journal of the Transportation Research Board, 2494(1), 45–56. https://doi.org/10.3141/2494-06

No competing interests reported.

Download PDF

Editor assigned by journal
13 Aug, 2024
Submission checks completed at journal
03 Aug, 2024
First submitted to journal
30 Jul, 2024

You are reading this latest preprint version

Eliminating Recall & Proxy Bias in Household Travel Surveys: Sample Correction Methodology under the Core-Satellite Fusion Paradigm

Status:

Version 1

Abstract

Figures

1.0 Introduction

2.0 Literature Review

3.0 Descriptive statistics of core & satellite surveys used in this study

4.0 Individual-level comparison between core-satellite respondents

4.1 Direct comparison

4.2 Indirect comparison

5.0 Correction procedure for erroneous samples in the core survey

5.1 Binary logit model sample selection model

5.2 Re-weighting procedure

5.3 Corrected trip rates in the core survey

6.0 Conclusion & Discussion

Declarations

Author Contribution

Acknowledgment

References

Additional Declarations

Status:

Version 1