You need to consider how the sparse sampling from the remote sensing data might affect the meaning of "3-day average".  You need to understand how the compositing/averaging is done in the product to see whether it is really representative.  I sometimes find it useful to sample the comparison data sets in the same way as the data set being tested to help ensure consistency.

The question is about a source of the systematic bias in averaged RS data. In further, it focuses on composite CATDS from SMOS that is taking probably the Soil Moisture (SM). Most of SM applications need long-term trends and thus it is natural that a user concern is put on a possible systematic bias. There are many reasons for creating composite CATDS products. The essential one is that data from particular half-orbits are taken in different times and a use of it for larger areas require several or many passes being only approximately around the local sun rise or the local sun set time for ascending and descending passes respectively. CATDS users can obtain products at http://www.catds.fr/sipad/login.do with temporal aggregations over 1, 3, 10 days or 1 month with a global coverage which are not only averaged over a particular interval but also under a proper retrieval of the composed data. Necessary documentation is available from http://www.catds.fr/Resources/Documentation especially for this subject question (SM) L3 L3 SM Algorithm Theoretical Based Document (ATBD) available at http://www.cesbio.ups-tlse.fr/SMOS_blog/?page_id=815. One can read in the documentation more essential and extended reasons for producing L3 CATDS SM products. One of them is keeping the control on consistency of the composed products with engaging retrievals and averaging. That is already done for a ready use of data. However, further responsibility for a use is put on the user. He/she obtains products in the time series generated with full available care about consistency on different levels, from the level L1C, L2 through L3. The document does not discuss effects of the bias, differently than few aspects of employing some auxiliary data which substantially introduce the bias due to snow and vegetation cover, but not to averaging. Because it was not necessary. Why? This is due to linearity of the radiometric system. The MIRAS system is sufficiently linear and calibrated for sufficiently clear RFI conditions. Only in RFI disturbed conditions it occurs unable to produce SM data. That happens in particular areas, in some only temporarily and then such "no-data" holes are filled (or not) by preceding or proceeding orbital passes if only the interval of integrating in a particular L3 product was long enough. All such incidental effects are documented in SMOS data by proper DQX measures. That is said only to ensure that no fear about "a bias due to averaging" should be red in this question.

However, some real fear about bias stay behind. It may be noticed by any user of SMOS SM data from L2 to L3, that the soil moisture values, time aggregated averaged and not averaged as well, can be ranged from 0 to 1, differently than it is expected on the ground and from proper hydrological models for a well defined and observed areas. It raises a fear that RS SM data values from SMOS could be "biased". Any other RS means also appear somehow biased and even more (but differently) than SMOS. RS data users live with that employing RS data not directly but with a use of hydrological models. A properly built, developed and anchored on the ground hydrological model is employed for managing with such "biases" for determining trends by quite advanced methods of matching RS and ground based data, within broad input from meteorological forcing, rain, and other data on the water drainage. But independently to instrumental error sources which surely exist and are controlled by RS systems (and in CATDS processing). The rest is left to the competency of user. What a rest? The need of separating eventual systematic biases from investigated trends in a properly long series of spatial and time aggregated data under proper choice of employed 1, 3, 10 days, 1 month or longer intervals. A user must decide a proper reference for his ROI, enabling the recognition between a suggested bias and a sought trend. Some necessary exercises with short data trains for managing effects are very useful for understanding effects. And for planning extended investigations.

SM data values from SMOS are proper to the Soil Moisture surface values (SM_surf) but are expressed by Dimension-Less (DL) measures being the same as SM_vol (volumetric or gravimetric) in [cm^3/cm^3]. The SM_surf values ranges 0…1(and even more accounting outliers). In large country scale areas, the average SM_sat may exceed 0.40, while from ground based measurements and hydrological models should be lower than 0.4. Under DRY conditions SM_surf usually fall lower than expected. Under WET conditions SM_surf happens larger than expected. The range expected on the ground is 0…0.40 or 0.50. The conclusion is that SM_surf response from SMOS is NON-LINEAR. It is natural that under averaging SM_surf over some interval - produce a suggested or suspected BIAS. That stays behind the question.

The reason is not in an imposed non-linearity of the radiometric system or the processing retrievals. The reason is in assigning the retrieved SM to the observed _surface, that is to the surface layer thickness equal or close to ZERO. However, the MIRAS radiometer integrates emission of the soil layer thick on the Penetration Depth (PD) in particular not known from SMUDP2 data. PD value is proper to the water content and other physical properties of soil but is not estimated by a particular model. The soil emission model CMEM enables modeling a stack of soil layers but it was not developed to an efficient tool and is not used under retrieving SM L2 from the emission determined on the base of L1C data in terms of TB. It is not implemented due to many reasons but one is in the lack of soil data determining the hydraulic state of water in soil allowing on distinguishing effects deserved by percolation and suction, on the base of electro-magnetic noise emission. Soil properties employed by CMEM and current version (v620) of SMUDP2 engage only the substantial sand, clay composition, while hydrological water states must be funded on many other soil components and soil properties. Therefore, SM retrieved values are constrained to the case SM_surface, though radiometric sensing performed with SMOS is in-depth with PD not defined. It is the reason for which effects of sensing SM_surf are not linear or remain considerably quasi-linear with respect to the water content under the surface. The problem is open not only for SMOS but also for other active radar sensing means. There was a justified expectation that advances in SMAP would have been able to proceed further. Unfortunately, the SMAP radar instrument failed and has not been recovered. Anyway, the suction and percolation of water in soil play very important role for applications focused not only on the surface soil moisture but on the water content over subsurface layers and its relations to the root zone questions. That what one can take for a bias in RS data for SM is in fact deserved by the lack of engaging the penetration depth proper to particular RS technique. One should not see it as deserved by RS instruments. However, the other side of this bias-like effects depend on averaging, but just the operation of averaging is also not a source of errors till the rest of an applied hydrological model fed by RS data, is set properly to goals.