Asymmetric cost pass-through and consumer search: empirical evidence from online platforms

Abstract

Prices often react stronger to rising than to falling costs. This asymmetric cost pass-through is still not fully understood, but recent theories suggest that asymmetric adjustments of consumers’ search efforts to rising and to falling prices may be one explanation for this pattern. I use novel panel data to investigate the interaction of consumer search intensity, pricing and cost pass-through of residential electricity tariffs on online price comparison sites. I find that consumers search slightly more when prices rise but drastically decrease search efforts when they fall. Moreover, I find direct evidence that cost pass-through heavily depends on consumers’ search efforts in that cost increases are passed-through less to the consumer when search intensity is high while cost decreases are passed-through more when search intensity is high. This finding may help upstream firms to better understand how their price changes will translate into retail price adjustments.

Appendices

Appendix A: Additional figures

Fig. 4

Screenshot of a price comparison website (https://www.Toptarif.de)

Fig. 5

Incumbent default tariffs (2012)

Fig. 6

Total costs (2012)

Fig. 7

Search intensity (2012)

Fig. 8

Range of prices (2012)

Fig. 9

Distribution of cost changes by year

Appendix B: First-stage results

Table 4 First-stage estimations of Table 2

Table 5 First-stage estimations of Table 3

Appendix C: Estimations of consumer search intensity without covariates

Table 6 IV estimation of consumer search – estimation without covariates

Appendix D: Estimations of consumer search intensity when search between November 15 to December 31 counts for the next year

Table 7 IV estimation of consumer search – search queries from November 15 to December 31 transferred to the next year

Appendix E: Estimations of cost pass-through with additional control variables

Table 8 Estimation of cost pass-through - additional control variables

Appendix F: Alternative consumption level

Table 9 Estimation of cost pass-through for a standard 5 person household (5,000 MWh/a)

Fig. 10

Marginal effects of Δμ on the cost pass-through rate for a 5 person household (5,000 MWh/a)

Appendix G: Alternative fixed effects

Table 10 Estimation of cost pass-through with incumbent FE

Fig. 11

Marginal effects of Δμ on the cost pass-through rate based on estimations with incumbent FE

Appendix H: Technical description of Lewbel’s (2012) IV method and results

1.1 H.1 Technical description

Consider the linear relationship Y = Xβ + Zγ + ε₁, where Z is potentially endogenous (the interactions of Δμ and the two cost change variables here) and γ is the parameter we wish to estimate. The equation that determines Z is Z = Xα + ε₂, where ε₁ and ε₂ may be correlated and no element of X can be used as an instrument, i.e. there is no outside instrument available. As usual, the requirement is that $E\left (X\varepsilon _{1}\right ) =0$, E(X_iε₂) = 0, and that $E\left (XX^{\prime }\right )$ is nonsingular. The additional assumptions for the identification in the absence of an outside instrument are that $Cov\left (X,\varepsilon _{1}\varepsilon _{2}\right ) =0$ and that there is some heteroskedasticity in the error of the first-stage, $Cov\left (X,{\varepsilon _{2}^{2}}\right ) \neq 0$. If these assumptions hold the variation in ε₂ can be used to identify the model parameters. γ (and β) can then be estimated consistently by using interactions of the mean-centered control variables and the residuals $(\left (X-\bar {X}\right ) \hat {\varepsilon }_{2})$ to instrument for Z.

The estimation procedure is then as follows:

• 1. Estimate $\hat {\alpha }$ by an OLS regression of Z on X to obtain $\hat {\varepsilon }_{2}=Z-X\hat \alpha$.

• 2. Use the interactions of the residuals $\hat {\varepsilon }_{2}$ and the mean-centered covariates $(X-\bar {X})$ as instruments for Z and estimate $Z=X\alpha + \gamma \left (X-\bar {X}\right ) \hat {\varepsilon }_{2} + \varepsilon _{3}$.

• 3. Obtain $\hat {\beta }$ and $\hat {\gamma }$ by estimating Y =$X \beta + \hat {Z} \gamma + \varepsilon _{4}$.

1.2 H.2 Results

As Lewbel (2012) shows, the model is identified if the errors from a regression of the endogenous variable on covariates from the main model are heteroskedastic and the variance of these errors is correlated with at least some of the covariates but not with the covariances of these errors and the second stage errors. I test the heteroskedasticity requirement based on the residuals of the first stage regression, using a modified Wald statistic for groupwise heteroskedasticity. The test rejects the null hypotheses of a constant variance as can be seen in Table 11.

The Kleibergen-Paap F-statistic suggests that the generated instruments are sufficiently strong to identify the endogenous variables in all estimations as the Stock and Yogo critical values are exceeded. Again, the results remain robust to this alternative IV.

Table 11 Lewbel (2012) IV estimation of cost pass-through

Fig. 12

Marginal effects of Δμ on the cost pass-through rate based on Lewbel (2012) IV estimations